diff --git a/paper-releases/pbisc-diffusion/updates/20261006T033406Z-bulkonly-lora5-k1000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_20261006/analysis/wf/findings_ye2026_sepsis.json b/paper-releases/pbisc-diffusion/updates/20261006T033406Z-bulkonly-lora5-k1000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_20261006/analysis/wf/findings_ye2026_sepsis.json new file mode 100644 index 0000000000000000000000000000000000000000..e2d8bc4bca577ae66891ab1ef30baa4121aaf8be --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261006T033406Z-bulkonly-lora5-k1000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_20261006/analysis/wf/findings_ye2026_sepsis.json @@ -0,0 +1,4233 @@ +{ + "cohort": "ye2026_sepsis", + "date": "2026-10-06", + "report": "Returned as text in the synthesis agent's final response: the harness blocks subagents from writing report .md files, so /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/REPORT_ye2026_sepsis.md was not created by this agent.", + "grading_rule": "refuted if any verifier refuted; confirmed only if both confirmed; else qualified", + "counts": { + "confirmed": 0, + "qualified": 17, + "refuted": 1 + }, + "cross_cutting_conclusions": [ + "No finding is confirmed by both verifiers. 17 are qualified, 1 (SEP-CLIN-1) is refuted.", + "The generated cells reproduce, in direction, the known sepsis-vs-HC immune shifts in the 121 bulk-only adults: monocyte HLA class II loss, MS1/S100-high HLA-II-low monocytes, loss of MAIT/GZMK+ memory T and NK cells, activated CCL4+/NR4A2+ lymphocyte states, more plasma cells. All are in Ye 2026 or older work and all are visible in the bulk.", + "Magnitudes depend on the draw and follow the training donors (draw 0's MS1-rich donors; per-draw IEG/CCL effect sizes rank with the training donors' observed levels).", + "Nothing separates sepsis from non-septic ICU controls reliably: the trust layer has 5-6 ICU controls per draw, sepsis-vs-ICU calibration is the weakest contrast, and the generator pushes ICU controls toward sepsis for T-cell states while under-producing ICU MS1. ISG-high monocytes are the only lead.", + "The only content beyond the bulk is a modest cell-type-specific gene x type interaction (SEP-PROG-4).", + "No new population survives. Artefacts identified: chrY labeller inputs (sex contrasts), immediate-early/handling program, depth-dependent marker rules, generator chrY leakage, draw-specific and micro clusters, over-produced doublet-like clusters." + ], + "generated": { + "samples": [ + { + "set": "bulkonly", + "group": "adult_HC", + "n_samples": 36 + }, + { + "set": "bulkonly", + "group": "adult_ICU_ctrl", + "n_samples": 17 + }, + { + "set": "bulkonly", + "group": "adult_sepsis", + "n_samples": 68 + }, + { + "set": "bulkonly", + "group": "ped_ICU_ctrl", + "n_samples": 1 + }, + { + "set": "paired", + "group": "adult_HC", + "n_samples": 7 + }, + { + "set": "paired", + "group": "adult_ICU_ctrl", + "n_samples": 6 + }, + { + "set": "paired", + "group": "adult_sepsis", + "n_samples": 46 + }, + { + "set": "paired", + "group": "ped_HC", + "n_samples": 4 + }, + { + "set": "paired", + "group": "ped_ICU_ctrl", + "n_samples": 4 + }, + { + "set": "paired", + "group": "ped_sepsis", + "n_samples": 15 + } + ], + "cells": [ + { + "set": "observed", + "draw": -1, + "n_samples": 82, + "n_cells": 589478, + "cells_per_sample_min": 2107, + "cells_per_sample_median": 7153.5, + "cells_per_sample_max": 15576 + }, + { + "set": "paired_gen", + "draw": 0, + "n_samples": 77, + "n_cells": 115500, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "bulkonly", + "draw": 0, + "n_samples": 122, + "n_cells": 183000, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "paired_gen", + "draw": 1, + "n_samples": 77, + "n_cells": 115500, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "bulkonly", + "draw": 1, + "n_samples": 122, + "n_cells": 183000, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "paired_gen", + "draw": 2, + "n_samples": 77, + "n_cells": 115500, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "bulkonly", + "draw": 2, + "n_samples": 122, + "n_cells": 183000, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "paired_gen", + "draw": 3, + "n_samples": 77, + "n_cells": 115500, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "bulkonly", + "draw": 3, + "n_samples": 122, + "n_cells": 183000, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "paired_gen", + "draw": 4, + "n_samples": 77, + "n_cells": 115500, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + }, + { + "set": "bulkonly", + "draw": 4, + "n_samples": 122, + "n_cells": 183000, + "cells_per_sample_min": 1500, + "cells_per_sample_median": 1500.0, + "cells_per_sample_max": 1500 + } + ], + "training_donors": [ + { + "draw": 0, + "train_donors": "HC07;Abd-S66;Res-S150;Skin-S74;Res-S77", + "train_groups": "adult_HC;adult_sepsis;adult_sepsis;adult_sepsis;adult_sepsis" + }, + { + "draw": 1, + "train_donors": "HC12;Uri-S134;Uri-S147;Res-S164;Res-S166", + "train_groups": "adult_HC;adult_sepsis;adult_sepsis;adult_sepsis;adult_sepsis" + }, + { + "draw": 2, + "train_donors": "PIC01;Res-S177;Skin-S74;Res-S136;Res-S77", + "train_groups": "ped_ICU_ctrl;adult_sepsis;adult_sepsis;adult_sepsis;adult_sepsis" + }, + { + "draw": 3, + "train_donors": "PHC05;Res-S139;Res-S174;Skin-S142;Abd-S57", + "train_groups": "ped_HC;adult_sepsis;adult_sepsis;adult_sepsis;adult_sepsis" + }, + { + "draw": 4, + "train_donors": "IC23;Skin-S142;Res-PS22;Abd-S102;Skin-S46", + "train_groups": "adult_ICU_ctrl;adult_sepsis;ped_sepsis;adult_sepsis;adult_sepsis" + } + ], + "sources": [ + "synthesis_ye2026_sepsis/y00_samples_by_set_group.csv", + "synthesis_ye2026_sepsis/y00_cells_by_set_draw.csv", + "synthesis_ye2026_sepsis/y00_training_donors.csv" + ] + }, + "calibration_summary": { + "gates": "level: 0.5<=ratio<=2 and Spearman>=0.3; PASS = level + same sign + gen/obs contrast in [0.5,2]; DIR = 1/3<=ratio<=3, Spearman>=0.2 and same sign; else FAIL", + "per_draw": [ + { + "draw": 0, + "n_features": 50, + "level_strict": 21, + "level_lenient": 31, + "spearman_ge_0p3": 27, + "ratio_0p5_2": 31, + "S_vs_HC_sign": 42, + "S_vs_HC_PASS": 13, + "S_vs_HC_DIR": 12, + "S_vs_HC_FAIL": 25, + "S_vs_ICU_sign": 30, + "S_vs_ICU_PASS": 4, + "S_vs_ICU_DIR": 16, + "S_vs_ICU_FAIL": 30, + "S_vs_allctrl_sign": 33, + "S_vs_allctrl_PASS": 9, + "S_vs_allctrl_DIR": 12, + "S_vs_allctrl_FAIL": 29, + "S_vs_HC_PASS_composition": 11, + "n_composition": 35, + "S_vs_HC_PASS_state": 2, + "n_state": 15 + }, + { + "draw": 1, + "n_features": 50, + "level_strict": 18, + "level_lenient": 26, + "spearman_ge_0p3": 29, + "ratio_0p5_2": 27, + "S_vs_HC_sign": 38, + "S_vs_HC_PASS": 5, + "S_vs_HC_DIR": 14, + "S_vs_HC_FAIL": 31, + "S_vs_ICU_sign": 37, + "S_vs_ICU_PASS": 4, + "S_vs_ICU_DIR": 14, + "S_vs_ICU_FAIL": 32, + "S_vs_allctrl_sign": 32, + "S_vs_allctrl_PASS": 4, + "S_vs_allctrl_DIR": 12, + "S_vs_allctrl_FAIL": 34, + "S_vs_HC_PASS_composition": 4, + "n_composition": 35, + "S_vs_HC_PASS_state": 1, + "n_state": 15 + }, + { + "draw": 2, + "n_features": 50, + "level_strict": 16, + "level_lenient": 29, + "spearman_ge_0p3": 25, + "ratio_0p5_2": 27, + "S_vs_HC_sign": 39, + "S_vs_HC_PASS": 6, + "S_vs_HC_DIR": 18, + "S_vs_HC_FAIL": 26, + "S_vs_ICU_sign": 32, + "S_vs_ICU_PASS": 8, + "S_vs_ICU_DIR": 14, + "S_vs_ICU_FAIL": 28, + "S_vs_allctrl_sign": 35, + "S_vs_allctrl_PASS": 7, + "S_vs_allctrl_DIR": 14, + "S_vs_allctrl_FAIL": 29, + "S_vs_HC_PASS_composition": 4, + "n_composition": 35, + "S_vs_HC_PASS_state": 2, + "n_state": 15 + }, + { + "draw": 3, + "n_features": 50, + "level_strict": 22, + "level_lenient": 30, + "spearman_ge_0p3": 28, + "ratio_0p5_2": 37, + "S_vs_HC_sign": 38, + "S_vs_HC_PASS": 6, + "S_vs_HC_DIR": 19, + "S_vs_HC_FAIL": 25, + "S_vs_ICU_sign": 33, + "S_vs_ICU_PASS": 6, + "S_vs_ICU_DIR": 17, + "S_vs_ICU_FAIL": 27, + "S_vs_allctrl_sign": 34, + "S_vs_allctrl_PASS": 8, + "S_vs_allctrl_DIR": 14, + "S_vs_allctrl_FAIL": 28, + "S_vs_HC_PASS_composition": 4, + "n_composition": 35, + "S_vs_HC_PASS_state": 2, + "n_state": 15 + }, + { + "draw": 4, + "n_features": 50, + "level_strict": 23, + "level_lenient": 36, + "spearman_ge_0p3": 31, + "ratio_0p5_2": 34, + "S_vs_HC_sign": 36, + "S_vs_HC_PASS": 11, + "S_vs_HC_DIR": 14, + "S_vs_HC_FAIL": 25, + "S_vs_ICU_sign": 38, + "S_vs_ICU_PASS": 10, + "S_vs_ICU_DIR": 19, + "S_vs_ICU_FAIL": 21, + "S_vs_allctrl_sign": 31, + "S_vs_allctrl_PASS": 8, + "S_vs_allctrl_DIR": 17, + "S_vs_allctrl_FAIL": 25, + "S_vs_HC_PASS_composition": 10, + "n_composition": 35, + "S_vs_HC_PASS_state": 1, + "n_state": 15 + } + ], + "pooled": [ + { + "kind": "composition", + "feature": "l1 | B cell", + "gen_mean": 0.150298, + "obs_mean": 0.175599, + "ratio": 0.85938, + "spearman": 0.717931, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "composition", + "feature": "l1 | DC", + "gen_mean": 0.00507013, + "obs_mean": 0.0055057, + "ratio": 0.929965, + "spearman": 0.409779, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l1 | Monocyte", + "gen_mean": 0.178388, + "obs_mean": 0.206889, + "ratio": 0.868074, + "spearman": 0.492623, + "adult sepsis vs adult HC: draws same sign": "3/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "composition", + "feature": "l1 | NK cell", + "gen_mean": 0.0469126, + "obs_mean": 0.0800551, + "ratio": 0.588939, + "spearman": 0.47805, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "1/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "composition", + "feature": "l1 | Other", + "gen_mean": 0.0392693, + "obs_mean": 0.0488678, + "ratio": 0.804254, + "spearman": 0.3149, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "2/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l1 | T cell", + "gen_mean": 0.580062, + "obs_mean": 0.483084, + "ratio": 1.20159, + "spearman": 0.450056, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "composition", + "feature": "l2 | ASDC", + "gen_mean": 6.23377e-05, + "obs_mean": 0.000151251, + "ratio": 0.420336, + "spearman": 0.0197944, + "adult sepsis vs adult HC: draws same sign": "0/5", + "adult sepsis vs adult ICU control: draws same sign": "1/5", + "sepsis vs all controls: draws same sign": "1/5" + }, + { + "kind": "composition", + "feature": "l2 | CD14 monocyte", + "gen_mean": 0.108426, + "obs_mean": 0.153453, + "ratio": 0.708289, + "spearman": 0.448764, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "composition", + "feature": "l2 | CD16 monocyte", + "gen_mean": 0.0299706, + "obs_mean": 0.0365749, + "ratio": 0.836578, + "spearman": 0.362139, + "adult sepsis vs adult HC: draws same sign": "2/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "2/5" + }, + { + "kind": "composition", + "feature": "l2 | CD56bright NK cell", + "gen_mean": 0.00598615, + "obs_mean": 0.00699504, + "ratio": 0.858594, + "spearman": 0.113721, + "adult sepsis vs adult HC: draws same sign": "3/5", + "adult sepsis vs adult ICU control: draws same sign": "1/5", + "sepsis vs all controls: draws same sign": "2/5" + }, + { + "kind": "composition", + "feature": "l2 | CD56dim NK cell", + "gen_mean": 0.032942, + "obs_mean": 0.0585976, + "ratio": 0.565946, + "spearman": 0.496084, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "0/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "composition", + "feature": "l2 | CD8aa", + "gen_mean": 0.0211498, + "obs_mean": 0.0283451, + "ratio": 0.751028, + "spearman": 0.0218263, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l2 | DN T cell", + "gen_mean": 0.0352398, + "obs_mean": 0.0183763, + "ratio": 1.92172, + "spearman": 0.221698, + "adult sepsis vs adult HC: draws same sign": "2/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "0/5" + }, + { + "kind": "composition", + "feature": "l2 | Effector B cell", + "gen_mean": 0.0189922, + "obs_mean": 0.0164096, + "ratio": 1.1628, + "spearman": 0.369317, + "adult sepsis vs adult HC: draws same sign": "1/5", + "adult sepsis vs adult ICU control: draws same sign": "1/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "composition", + "feature": "l2 | Erythrocyte", + "gen_mean": 0.0105281, + "obs_mean": 0.0180918, + "ratio": 0.586754, + "spearman": 0.186637, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "3/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l2 | ILC", + "gen_mean": 0.00350649, + "obs_mean": 0.00379754, + "ratio": 0.937479, + "spearman": -0.111151, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "3/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l2 | Intermediate monocyte", + "gen_mean": 0.0399913, + "obs_mean": 0.0168611, + "ratio": 2.45971, + "spearman": 0.400325, + "adult sepsis vs adult HC: draws same sign": "3/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "composition", + "feature": "l2 | MAIT", + "gen_mean": 0.00610216, + "obs_mean": 0.0112594, + "ratio": 0.541799, + "spearman": 0.418919, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l2 | Memory B cell", + "gen_mean": 0.0317853, + "obs_mean": 0.0348292, + "ratio": 0.920796, + "spearman": 0.41146, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "composition", + "feature": "l2 | Memory CD4 T cell", + "gen_mean": 0.131917, + "obs_mean": 0.120175, + "ratio": 1.0995, + "spearman": 0.431728, + "adult sepsis vs adult HC: draws same sign": "2/5", + "adult sepsis vs adult ICU control: draws same sign": "2/5", + "sepsis vs all controls: draws same sign": "2/5" + }, + { + "kind": "composition", + "feature": "l2 | Memory CD8 T cell", + "gen_mean": 0.11813, + "obs_mean": 0.0829867, + "ratio": 1.42429, + "spearman": 0.642392, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "2/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "composition", + "feature": "l2 | Naive B cell", + "gen_mean": 0.0852727, + "obs_mean": 0.0626012, + "ratio": 1.37338, + "spearman": 0.582123, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "composition", + "feature": "l2 | Naive CD4 T cell", + "gen_mean": 0.103262, + "obs_mean": 0.0839498, + "ratio": 1.23146, + "spearman": 0.631546, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l2 | Naive CD8 T cell", + "gen_mean": 0.0704468, + "obs_mean": 0.056122, + "ratio": 1.26779, + "spearman": 0.630498, + "adult sepsis vs adult HC: draws same sign": "2/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "composition", + "feature": "l2 | Plasma cell", + "gen_mean": 0.00507532, + "obs_mean": 0.015238, + "ratio": 0.334373, + "spearman": 0.283359, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l2 | Platelet", + "gen_mean": 0.0221905, + "obs_mean": 0.0256883, + "ratio": 0.863247, + "spearman": 0.321513, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "1/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "composition", + "feature": "l2 | Progenitor cell", + "gen_mean": 0.00304416, + "obs_mean": 0.00129011, + "ratio": 2.38216, + "spearman": 0.114896, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "1/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l2 | Proliferating NK cell", + "gen_mean": 0.00798442, + "obs_mean": 0.0144624, + "ratio": 0.553192, + "spearman": 0.353844, + "adult sepsis vs adult HC: draws same sign": "3/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "composition", + "feature": "l2 | Proliferating T cell", + "gen_mean": 0.00780606, + "obs_mean": 0.00921837, + "ratio": 0.850739, + "spearman": 0.239214, + "adult sepsis vs adult HC: draws same sign": "3/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "0/5" + }, + { + "kind": "composition", + "feature": "l2 | Transitional B cell", + "gen_mean": 0.00917229, + "obs_mean": 0.0465207, + "ratio": 0.198269, + "spearman": 0.448315, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "2/5" + }, + { + "kind": "composition", + "feature": "l2 | Treg", + "gen_mean": 0.0241264, + "obs_mean": 0.0204165, + "ratio": 1.18346, + "spearman": 0.220493, + "adult sepsis vs adult HC: draws same sign": "1/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "1/5" + }, + { + "kind": "composition", + "feature": "l2 | cDC1", + "gen_mean": 0.00035671, + "obs_mean": 0.000288272, + "ratio": 1.2427, + "spearman": 0.0208758, + "adult sepsis vs adult HC: draws same sign": "0/5", + "adult sepsis vs adult ICU control: draws same sign": "2/5", + "sepsis vs all controls: draws same sign": "1/5" + }, + { + "kind": "composition", + "feature": "l2 | cDC2", + "gen_mean": 0.00382338, + "obs_mean": 0.00401928, + "ratio": 0.959136, + "spearman": 0.419878, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "composition", + "feature": "l2 | gdT", + "gen_mean": 0.0618823, + "obs_mean": 0.0522348, + "ratio": 1.18874, + "spearman": 0.556223, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "2/5", + "sepsis vs all controls: draws same sign": "2/5" + }, + { + "kind": "composition", + "feature": "l2 | pDC", + "gen_mean": 0.000827706, + "obs_mean": 0.00104691, + "ratio": 0.809386, + "spearman": 0.166715, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "3/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "state", + "feature": "C1Q-high monocyte (reference)", + "gen_mean": 0.0115385, + "obs_mean": 0.0637201, + "ratio": 0.181992, + "spearman": 0.185724, + "adult sepsis vs adult HC: draws same sign": "3/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "state", + "feature": "Cycling T/NK", + "gen_mean": 0.00433106, + "obs_mean": 0.0107062, + "ratio": 0.405054, + "spearman": 0.223907, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "0/5" + }, + { + "kind": "state", + "feature": "Cycling activated T (CD38/HLA-DR)", + "gen_mean": 0.00517207, + "obs_mean": 0.0197585, + "ratio": 0.261391, + "spearman": 0.267919, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "0/5" + }, + { + "kind": "state", + "feature": "Cycling monocyte (Ye 2026)", + "gen_mean": 0.00258637, + "obs_mean": 0.0333968, + "ratio": 0.0763468, + "spearman": 0.240666, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "state", + "feature": "Cytotoxic CD4 T", + "gen_mean": 0.0259646, + "obs_mean": 0.017299, + "ratio": 1.49254, + "spearman": 0.571523, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "2/5", + "sepsis vs all controls: draws same sign": "2/5" + }, + { + "kind": "state", + "feature": "DN2 / ABC B", + "gen_mean": 0.0232406, + "obs_mean": 0.0546338, + "ratio": 0.437401, + "spearman": 0.16455, + "adult sepsis vs adult HC: draws same sign": "3/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "state", + "feature": "Granulocyte-like (LDG)", + "gen_mean": 0.00111861, + "obs_mean": 0.0256158, + "ratio": 0.0447189, + "spearman": 0.340554, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "3/5" + }, + { + "kind": "state", + "feature": "HSPC", + "gen_mean": 0.00191515, + "obs_mean": 0.00120085, + "ratio": 1.61801, + "spearman": 0.0373594, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "0/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "state", + "feature": "ISG-high T", + "gen_mean": 0.0187968, + "obs_mean": 0.036554, + "ratio": 0.519188, + "spearman": 0.524691, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "state", + "feature": "ISG-high monocyte (positive control)", + "gen_mean": 0.0929607, + "obs_mean": 0.149538, + "ratio": 0.636239, + "spearman": 0.564337, + "adult sepsis vs adult HC: draws same sign": "4/5", + "adult sepsis vs adult ICU control: draws same sign": "5/5", + "sepsis vs all controls: draws same sign": "4/5" + }, + { + "kind": "state", + "feature": "MS1 monocyte (IL1R2+)", + "gen_mean": 0.0357019, + "obs_mean": 0.153137, + "ratio": 0.251236, + "spearman": 0.418951, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "0/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "state", + "feature": "Plasmablast (label)", + "gen_mean": 0.00507532, + "obs_mean": 0.015238, + "ratio": 0.334373, + "spearman": 0.283359, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "state", + "feature": "Plasmablast (markers)", + "gen_mean": 0.00681558, + "obs_mean": 0.0418537, + "ratio": 0.167236, + "spearman": 0.262342, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "2/5" + }, + { + "kind": "state", + "feature": "Platelet signal in monocytes (negative control)", + "gen_mean": 0.0925024, + "obs_mean": 0.135679, + "ratio": 0.676094, + "spearman": 0.186832, + "adult sepsis vs adult HC: draws same sign": "5/5", + "adult sepsis vs adult ICU control: draws same sign": "3/5", + "sepsis vs all controls: draws same sign": "5/5" + }, + { + "kind": "state", + "feature": "Proliferating T (label)", + "gen_mean": 0.0145379, + "obs_mean": 0.0180398, + "ratio": 0.805395, + "spearman": 0.285628, + "adult sepsis vs adult HC: draws same sign": "2/5", + "adult sepsis vs adult ICU control: draws same sign": "4/5", + "sepsis vs all controls: draws same sign": "0/5" + } + ], + "sources": [ + "synthesis_ye2026_sepsis/y01_draw_summary.csv", + "synthesis_ye2026_sepsis/y01_pooled.csv", + "synthesis_ye2026_sepsis/y01_codes_S_vs_HC.csv", + "synthesis_ye2026_sepsis/y01_codes_S_vs_ICU.csv", + "synthesis_ye2026_sepsis/y01_codes_S_vs_allctrl.csv" + ] + }, + "findings": [ + { + "id": "SEP-PROG-1", + "lens": "sep_programs", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 1, + "theme": "monocyte HLA class II", + "synthesis_strength": "moderate for direction vs HC; weak beyond bulk; none for sepsis specificity", + "reportable_statement": "In bulk-only adults, generated CD14 and intermediate monocytes of sepsis patients express less HLA class II than those of healthy controls in all five draws (CD14: -0.93/-1.06/-0.71/-1.37/-1.13 log2). The direction is calibrated in 5/5 draws but the size is attenuated (0.38-0.77x of observed), so these are lower bounds. Non-septic ICU controls carry 55-75% of the same drop; the sepsis-vs-ICU part is not calibrated.", + "carrying_draws": { + "CD14 / intermediate monocytes, sepsis vs HC": "0,1,2,3,4 (draw 2 weakest)", + "CD16 monocytes": "1,3,4", + "monocyte-specific residual beyond bulk": "1,4 only", + "sepsis vs ICU": "none" + }, + "honesty": { + "calibration": "Program not in calibration.csv; paired-test check passes 5/5 for CD14 and intermediate monocytes (ratio 0.48/0.55/0.38/0.77/0.59, bootstrap CI excludes 1 in 4/5; Spearman 0.51-0.64, 0.30-0.46 within adult sepsis).", + "replication": "5/5 draws same sign, all q<=4.8e-7; survives age, sex, site, sex_flag, cell number, depth, label confidence, compositional renormalisation.", + "observed": "Observed paired CD14 -1.56 log2 (43 vs 7, p 4.5e-12); paediatric replication -1.38 (13 vs 4). Observed ICU-HC -0.97 (p 0.019).", + "beyond_bulk": "Direction and per-sample ranking: no (bulk HLA-II -0.76; bulk ranks observed CD14 HLA-II at Spearman 0.68-0.78 vs 0.51-0.64 generated). Localisation inside monocytes is reproduced at 0.33-0.51x of the observed bulk-adjusted residual, significant only in draws 1 and 4; cross-type tracking over all 21-22 types is r 0.14-0.45." + }, + "verdict_reasons": [ + "not sepsis-specific (ICU controls carry 55-75% of the drop)", + "sepsis-vs-ICU calibration fails (generated ~0 in draws 2-4)", + "cross-type 'pattern tracks' carried only by the 3 monocyte types", + "magnitude attenuated; draws 0-1 whole-sample HLA-DRA moves opposite to bulk", + "bulk already shows the fall and ranks patients better" + ], + "new_vs_cohort_papers": "Known: Ye 2026 Mono_c01-CD14-HLA-DRA reduced in sepsis (PDF p.6, p.14); Reyes 2020; Winkler 2017.", + "next_steps": [ + "Per-patient monocyte HLA-DR (flow) for bulk-only patients, if Ye's flow cohort overlaps them; test whether generated CD14 HLA-II adds to bulk HLA-DRA.", + "Add the HLA-II program, ICU-vs-HC and magnitude ratios with CIs to the trust layer.", + "More paired ICU controls (only 5-6 per draw) before any sepsis-specific claim." + ], + "key_sources": [ + "verify/SEP-PROG-1_recompute/v01_bulkonly_effect.csv", + "verify/SEP-PROG-1_recompute/v02_calibration_program.csv", + "verify/SEP-PROG-1_recompute/v07_icu.csv", + "verify/SEP-PROG-1_calib/v05_bulk_adjusted.csv", + "verify/SEP-PROG-1_calib/v02_type_specificity_corr.csv", + "sep_programs/programs_calib.csv" + ] + }, + "finding": { + "id": "SEP-PROG-1", + "cohort": "ye2026_sepsis", + "claim": "In bulk-only adult sepsis, the generated CD14 monocytes express less HLA class II than in adult HC (CD14 HLA-II score −1.04 log2 on average, −0.71 to −1.37 per draw), and so do the intermediate monocytes. This matches the observed paired monocytes in direction (−1.56) and every draw calibrates. The bulk itself already shows HLA-II falling and ranks samples better, so the only new content is that the loss sits in the monocytes.", + "kind": "program", + "feature": "HLA-II score = mean edgeR log2 CPM (prior.count 2) of HLA-DRA, HLA-DRB1, HLA-DPA1, HLA-DPB1, HLA-DMA, HLA-DMB, CD74 in the per-sample AIFI-L2 CD14 monocyte pseudobulk (>=10 cells); also intermediate and CD16 monocytes", + "comparison": "adult sepsis vs adult HC, bulk-only (50–68 vs 33–36 samples with >=10 CD14 cells), tested within each draw", + "effect_bulkonly": "CD14: −0.93, −1.06, −0.71, −1.37, −1.13 log2 (draws 0–4; mean −1.04, Cohen's d −2.1). Intermediate: −0.30, −0.86, −0.31, −0.90, −0.74. CD16: −0.02, −0.42, −0.19, −0.20, −0.65. In CD14, all 7 genes fall in 5/5 draws (gene means −0.80 to −1.73). Absolute levels are not compared across draws.", + "stats": "Per-draw Welch t on sample scores, CD14: p = 1.4e-19, 2.1e-27, 5.2e-8, 3.7e-17, 3.9e-21. BH q over all program × type × contrast tests within a draw <= 4.8e-7. Mann–Whitney p <= 7e-7. OLS adjusted for age and sex: −0.92, −1.06, −0.71, −1.40, −1.12. Sepsis restricted to the HC hospital and sex-flag samples dropped: −0.93, −1.07, −0.64, −1.40, −1.20 (p < 1e-4). With a 20-cell minimum: mean −1.02.", + "draws_same_sign": "5/5 (q<0.05 in 5/5)", + "paired_calibration": "Lens program calibration on each draw's paired TEST samples (programs_calib.csv), CD14: PASS in 5/5. Generated/observed sepsis−HC = −0.71/−1.47, −0.89/−1.62, −0.58/−1.53, −1.20/−1.56, −0.90/−1.52, so the ratio is 0.38–0.77 (attenuated). Per-sample Spearman generated vs observed 0.51–0.64. Intermediate monocytes PASS 5/5; CD16 PASS 4/5 (draw 2 sign only). Type-level logFC calibration (calib_types.csv), CD14: PASS in draws 0,1,3,4 (r 0.37–0.47, top-100 sign agreement 0.90–0.99), WEAK in draw 2 (r 0.28). calibration.csv row 'l2 | CD14 monocyte' (abundance, not this program): sepsis−HC sign kept in draws 0–3 (ratio 0.51–1.08, Spearman 0.33–0.53), reversed in draw 4.", + "paired_observed_concordance": "Observed paired cells (43 adult sepsis vs 7 HC): CD14 −1.56 log2 (d −1.86, p 4.5e-12), intermediate −0.79, CD16 −0.64. All 7 genes fall (−0.92 to −2.32). Across 12 L2 types, the bulk-only generated HLA-II change pattern tracks the observed one (r 0.52–0.78 in all draws). Observed memory CD4/Treg also fall (−0.60/−0.54); the generated cells reproduce this inconsistently.", + "beyond_bulk": "Direction: no. Whole-PBMC bulk HLA-II genes fall in the bulk-only samples (−0.76 log2; −1.71 relative to monocyte markers). Per sample: no. On paired test samples the bulk score ranks observed CD14 HLA-II better (Spearman 0.68–0.78 raw, 0.59–0.67 monocyte-normalised) than the generated CD14 score (0.51–0.64); the partial Spearman of generated given bulk is 0.25–0.34. What is added is the location: the loss is placed within CD14/intermediate monocytes. Observed cells show the bulk HLA-DRA drop is a mixture (DC loss −518 CPM, CD14+intermediate within-cell loss, B-cell gain +457). The generated whole-sample HLA-DRA change has the wrong sign in draws 0–1 (+0.64/+0.59 vs bulk −0.69), so the generated cells do not decompose the bulk drop reliably; only the within-monocyte direction is trustworthy.", + "novelty": "Known. Loss of HLA-DR on monocytes is the classic marker of sepsis immunosuppression. Ye 2026 (Nat Immunol 27:150; reference-papers/paired-cohorts PDF p.6) reports Mono_c01-CD14-HLA-DRA reduced in most sepsis, and on p.14 decreased monocyte HLA-DR. Reyes 2020 (rare-disease-states PDF pp.2–4) describes HLA-DRlo IL1R2hi MS1 cells; Winkler 2017 describes HLA-DRA mRNA. New here only as a calibrated replication in the 122 bulk-only samples from bulk input.", + "strength": "strong", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/lens_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s03_programs.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_draws.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_summary_min20.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s05_program_specificity.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_type_specificity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_types_matrix.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s06_bulk_baselines.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/bulk_baselines.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s11_site_sensitivity.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/site_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s01_calibration_types.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/calib_types.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s04_attribution.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_lineage_fine.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_genes_fine.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a15fcd8d8f65159b7", + "verdict": "qualified", + "recomputed": "The effect reproduces exactly from my own code. Bulk-only CD14 HLA-II score: mean edgeR log2 CPM (prior 2) of 7 genes, library = measured panel, >=10 cells. Adult sepsis minus adult HC per draw 0-4: -0.931, -1.059, -0.706, -1.372, -1.132 (mean -1.040; mean Cohen d -2.10). Welch p 1.4e-19, 2.1e-27, 5.2e-8, 3.7e-17, 3.9e-21; Mann-Whitney max p 7.2e-7. n = 64/68/50/57/58 sepsis vs 34/36/33/36/35 HC. Intermediate: -0.30, -0.86, -0.31, -0.90, -0.74. CD16: -0.02 (p 0.70), -0.42, -0.19, -0.20, -0.65. All 7 genes are negative in 5/5 draws (gene means -0.80 for CD74 to -1.73 for HLA-DMB). Same result with log2(CPM+1), summed-CPM and all-6k-library scores. Pooled L1 monocytes: -0.94, -0.73, -0.55, -1.32, -0.94.\n\nConfounder controls, CD14 (negative in 5/5 draws, all p <= 2e-4 unless noted):\n- OLS age+sex: -0.92/-1.06/-0.71/-1.40/-1.12. Adding log n cells and log UMI per cell: -0.80/-0.80/-0.60/-1.35/-0.94.\n- HC hospital only, sex_flag dropped: -0.93/-1.07/-0.64/-1.40/-1.20. 1:1 age/sex matched: -0.87/-1.04/-0.78/-1.44/-1.06.\n- Cell-number checks: >=20 cells, mean -1.01. >=30 cells: -0.98/-1.08/-0.56/-1.38/-1.08. Exactly 30 cells per sample (HLA-DRA+DRB1, per-cell markers.npz): -0.54/-0.98/-0.37/-1.28/-1.10.\n- Library checks: binomially thinned to a common library: -0.84/-0.98/-0.48/-0.95/-0.98. Per-cell depth quartiles: 20/20 negative.\n- Label confidence >=0.97: -0.52/-1.07/-0.57/-1.37/-1.19.\n- Composition checks: library without the top-20 genes (S100A8/9 etc.): -0.66/-0.89/-0.65/-1.10/-1.01. Median-of-ratios normalisation: -0.70/-0.82/-0.65/-1.18/-0.95. HLA-II set falls below all 2,000 expression-matched random 7-gene sets in every draw.\n- Removing LDG-like, platelet-complex or MS1-like cells from the CD14 label changes nothing.\n- All four infection sources are negative in 5/5 draws.\n- Donors: 122 unique bulk-only donors; none paired, none a training donor; test sets contain no training donors.\n\nPaired-test calibration (my re-implementation):\n- CD14 generated/observed sepsis-HC: -0.71/-1.47, -0.89/-1.62, -0.58/-1.53, -1.20/-1.56, -0.89/-1.52.\n- Ratio 0.48, 0.55, 0.38, 0.77, 0.59; bootstrap 95% CI 0.32-0.68, 0.41-0.74, 0.21-0.59, 0.52-1.06, 0.40-0.81.\n- Spearman 0.64, 0.58, 0.51, 0.55, 0.61; 0.30-0.46 within adult sepsis only.\n- Gate: CD14 PASS 5/5, intermediate PASS 5/5, CD16 PASS 4/5 (draw 2 sign only).\n- Generated CD14 level is within -0.24 to +0.30 log2 of observed. Each draw's bulk-only effect divided by its own ratio: -1.93, -1.93, -1.86, -1.78, -1.93.\n- Genome-wide CD14 logFC calibration: r 0.37, 0.43, 0.28, 0.46, 0.44 (top-100 sign agreement 0.90-0.99); whole-PBMC bulk broadcast r 0.67-0.74.\n\ncalibration.csv rows:\n- 'l2 | CD14 monocyte' (abundance): sepsis-HC sign kept in draws 0-3, reversed in draw 4; contrast ratio 0.94, 1.94, 0.28, 0.10, -0.23.\n- 'l1 | Monocyte': reversed in draws 3-4.\n- 'MS1 monocyte (IL1R2+)': sign kept 5/5; level ratio 0.98 in draw 0, 0.03-0.14 in draws 1-4.\n\nObserved paired cells (43 sepsis vs 7 HC):\n- CD14 -1.56 (d -1.86, p 4.5e-12); -1.64 with sepsis from the HC hospital only; -1.56 adjusted for age and sex.\n- Intermediate -0.79, CD16 -0.64; genes -0.92 to -2.32.\n\nICU controls (critical illness vs sepsis):\n- Bulk-only generated, CD14, ICU control minus HC: -0.70, -0.59, -0.41, -0.91, -0.62 (p <= 0.01), which is 55-75% of the sepsis-HC drop.\n- Sepsis minus ICU control: -0.23, -0.47, -0.30, -0.46, -0.51 (p 0.13, 8e-4, 0.08, 0.09, 0.003).\n- Paired test, sepsis minus ICU: generated -0.34, -0.18, +0.03, -0.03, +0.02 vs observed -0.59, -0.65, -0.56, -0.59, -0.49.\n- Paired test, ICU minus HC: generated -0.37, -0.71, -0.61, -1.18, -0.92 vs observed -0.88 to -1.04.\n- Genome-wide CD14 sepsis-vs-ICU logFC r is -0.00 to 0.14 in all draws.\n- Observed (all adults): ICU minus HC -0.97 (p 0.019); sepsis minus ICU -0.59 (p 0.11).\n\nLocation:\n- The 12-type r (0.77, 0.74, 0.52, 0.78, 0.75) reproduces exactly.\n- Without the 3 monocyte types: 0.59, -0.22, -0.11, -0.37, -0.17. Over all 21-22 types: 0.14-0.45.\n- Generated whole-sample HLA-DRA sepsis-HC: +0.64, +0.65, -0.12, -0.91, -0.24, vs bulk -0.69 and observed single-cell aggregate -0.38.\n\nBulk:\n- Bulk-only whole-PBMC HLA-II -0.76 (p 1.3e-12); -1.71 relative to monocyte markers.\n- Paired test: bulk ranks observed CD14 HLA-II with Spearman 0.69-0.78, vs 0.51-0.64 for generated; partial Spearman of generated given bulk 0.25-0.34.\n\nBulk-proxy adjustment and set offset:\n- Adding all five bulk composition/QC proxies (MT, ribosomal, Hb, neutrophil and platelet shares): -0.43, -0.66, -0.10 (p 0.60), -0.86, -0.82. The same adjustment cuts the observed paired-test effect from about -1.5 to -0.49 to -0.97. No single proxy removes the effect.\n- Set offset: generated CD14 HLA-II is lower in bulk-only than in paired-test samples of the same group (sepsis -0.24 to -0.58; HC -0.04 to -0.47).", + "issues": [ + "The core result is confirmed. Every quoted number reproduces from my own code, and the CD14 and intermediate sepsis-vs-HC drop survives age, sex, site, sex_flag, matching, cell number, library size, depth, label confidence, compositional re-normalisation and contaminating-state exclusion in 5/5 draws. Draws carrying the result: CD14 and intermediate in all 5 draws (draw 2 weakest); CD16 in draws 1, 3 and 4 only.", + "The drop is not specific to sepsis. Adult ICU controls show 55-75% of the generated CD14 drop vs HC (-0.41 to -0.91, all p<=0.01); the observed data agree (ICU-HC -0.97). The sepsis-vs-ICU part is small (-0.23 to -0.51; p<0.05 in only 2/5 draws) and uncalibrated. On paired test samples the generated sepsis-ICU difference is about 0 in draws 2-4 (+0.03/-0.03/+0.02 vs observed -0.49 to -0.59). Genome-wide CD14 sepsis-vs-ICU calibration fails in 5/5 draws. The generated cells therefore support 'critically ill vs healthy' HLA-II loss, not a sepsis-specific immunosuppression marker, and the novelty framing should say so.", + "The statement 'pattern tracks the observed one across 12 L2 types (r 0.52-0.78)' is reproduced but carried entirely by the 3 monocyte types. Without them r = 0.59 (draw 0) and -0.37 to -0.11 (draws 1-4); over all 21-22 types r = 0.14-0.45. The observed loss in memory CD4, Treg and cDC2 is not reproduced in draws 1-4, and draw 1 shows spurious HLA-II rises in NK and CD8 types. The 'location' content is right only for the main site (CD14 > intermediate).", + "Magnitude is attenuated: the generated paired-test contrast is 38-77% of observed, and the bootstrap CI excludes 1 in 4/5 draws. Bulk-only effect sizes are lower bounds; rescaled by each draw's own ratio they are about -1.8 to -1.9.", + "Within sepsis, the generated per-sample HLA-II is tied more strongly to whole-bulk features than the observed is: Spearman with bulk Hb share -0.45 vs -0.14 observed, platelet -0.55 vs -0.18, ribosomal +0.57 vs +0.30. Adjusting for all five bulk proxies removes the draw-2 effect (-0.10, p 0.60). The same adjustment also cuts the observed effect by 40-70%, so this is over-adjustment for disease biology, not a confounder. It still means part of the per-sample signal is a bulk-wide sepsis axis, which agrees with 'per sample: no' beyond bulk.", + "Draw heterogeneity: draw 2 is weakest in every layer (ratio 0.38; genome-wide r 0.28, WEAK; equal-cell HLA-DR -0.37; abundance contrast ratio 0.28). Draw 4 reverses CD14 abundance in calibration.csv. Draws 0-1 over-generate monocytes in sepsis (median CD14 cells per sample 271 vs 59 and 287 vs 76), which makes the generated whole-sample HLA-DRA rise (+0.64/+0.65) while the bulk falls (-0.69).", + "Set offset: bulk-only generated CD14 HLA-II is lower than paired-test generated in both groups (sepsis -0.24 to -0.58, HC -0.04 to -0.47; whole-PBMC bulk HLA-II is lower too, not significant). There is no batch variable, so a wave imbalance between bulk-only HC and sepsis cannot be excluded beyond the site restriction, which passes.", + "Beyond bulk: whole-PBMC bulk HLA-II relative to monocyte markers (-1.71) is already closer to the observed within-CD14 loss (-1.56) than the generated estimate (-1.04). The bulk ranks samples better, and the genome-wide CD14 bulk broadcast correlates better with the observed logFC (r 0.67-0.74 vs 0.28-0.46).", + "Minor numeric differences: >=20-cell mean is -1.01 (claim -1.02); draw 1 generated aggregate HLA-DRA is +0.65 (claim +0.59); observed memory CD4 is -0.69 with all 7 genes (claim -0.60, expressed-gene filter).", + "No archive delta from this verifier: only analysis files in the verify dir were written. The coordinator owns any PBISC-Diffusion task-end archive update." + ], + "reasoning": "I could not refute the claim.\n\nFrom the raw pb_l2/pb_l1, markers.npz, subsample.npz, cells.parquet and bulk npz files I rebuilt the score, contrasts, Welch/Mann-Whitney tests, covariate models, the program calibration gate, observed concordance and the bulk baselines. Every headline number matches to the second decimal, and the CD14/intermediate drop vs HC holds in 5/5 draws under every reasonable confounder control.\n\nIt is real and robust. A Hb/platelet/neutrophil/ribosomal/MT proxy model removes draw 2's effect, but it also halves the observed truth. These proxies are disease mediators, so this is over-adjustment.\n\nCalibration: the HLA-II program passes my re-implemented gate in 5/5 draws for CD14 and intermediate monocytes. The generated levels match the observed levels, but contrasts are attenuated to 38-77%. calibration.csv has no HLA-II feature; its CD14 abundance row is weak in draws 2-3 and reversed in draw 4, exactly as the claim states.\n\nI still grade it qualified rather than confirmed, for three reasons:\n1. The natural control for hospital sampling and critical illness, adult ICU controls, carries most of the generated drop. The sepsis-specific remainder is not calibrated: generated is about 0 on paired test in draws 2-4, and genome-wide sepsis-vs-ICU calibration fails in 5/5. The generated cells say nothing calibrated about sepsis-specific HLA-II loss.\n2. The cross-type 'pattern tracks' statistic that supports the claim's only new content (location) comes entirely from the monocyte types. Outside monocytes it is negative in draws 1-4.\n3. Magnitudes are lower bounds, and draws 0-1 do not reproduce the bulk aggregate.\n\nSuggested wording: 'In the bulk-only samples, generated CD14 and intermediate monocytes of critically ill adults (sepsis and, to 55-75% of the extent, non-septic ICU controls) show lower HLA class II than healthy controls. The sepsis-vs-HC part is calibrated in direction in 5/5 draws but attenuated to 38-77%; the sepsis-vs-ICU part is not calibrated.' Strength: strong for direction vs HC, weak for added value over bulk, none for sepsis specificity.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/vc.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v01_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v01_scores.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v01_bulkonly_effect.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v01_bulkonly_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v02_calib_observed_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v02_calibration_program.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v02_calibration_persample.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v02_observed_contrasts.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v02_bulkonly_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v02_paired_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v02_calibration_csv_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v03_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v03_confounders.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v03_set_wave_check.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v03_donor_check.json", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v04_cells.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v04_cells.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v04_confidence.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v05_location.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v05_types.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v05_type_corr.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v05_aggregate.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v05b_type_corr_main12.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v05b_type_corr_main12.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v06_extra.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v06_checks.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v06_bulk_qc_proxies.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v06_ols_bulk_proxies.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v06_levels_rescaled.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v06b_proxy_adjust.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v06b_proxy_adjust.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v06b_within_sepsis_proxy_rho.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v07_icu.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v07_icu.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v08_cd14_logfc_calib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v08_cd14_logfc_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v09_exclusions.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v09_exclusions.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_recompute/v09_state_shares.csv" + ] + }, + { + "verifier": "calib", + "agent_id": "a7602ad6f57b59758", + "verdict": "qualified", + "recomputed": "I recomputed everything with my own code, without using the candidate's lens_common. All numbers are edgeR log2 CPM (prior 2) with at least 10 cells, adult sepsis minus adult HC, analysed within each draw.\n\n**Bulk-only CD14 HLA-II, draws 0–4**\n- Main result: −0.93, −1.06, −0.71, −1.37, −1.13. Welch p from 1.4e-19 to 5.2e-8; Mann–Whitney p ≤ 7.2e-7.\n- Adjusted for age and sex: −0.92, −1.06, −0.71, −1.40, −1.12. The age ranges overlap: sepsis median 69, HC median 64.5.\n- HC hospital only, sex-flag samples dropped: −0.93, −1.07, −0.64, −1.40, −1.20.\n- Compositional-robust library (top-100 genes removed): −0.78, −0.97, −0.87, −1.35, −0.95.\n- Median-of-ratios normalisation: −0.72, −0.83, −0.79, −1.36, −1.04.\n- Per-cell, label confidence ≥ 0.9: −0.48, −0.47, −0.30, −0.64, −0.55.\n- All 7 genes fall in all 5 draws; gene means range from −0.80 (CD74) to −1.73 (HLA-DMB).\n- Intermediate monocytes: −0.30, −0.86, −0.31, −0.90, −0.74.\n\n**Paired TEST calibration (my equivalent check; this feature is not in calibration.csv)**\n- Generated / observed difference: −0.71/−1.47, −0.89/−1.62, −0.58/−1.53, −1.20/−1.56, −0.89/−1.52.\n- Sign agrees in 5/5 draws. Ratio 0.48, 0.55, 0.38, 0.77, 0.59; under median-of-ratios 0.44–0.77.\n- Spearman over all test samples 0.51–0.64; within adult sepsis only 0.30–0.46.\n- Pooled over draws: per-draw centred scores give Spearman 0.63 (all) and 0.46 (within sepsis); mean generated diff −0.85 vs observed −1.54.\n- Leaving out each observed HC in turn keeps the sign in 33/33 cases.\n- Every draw is tested against the same 6–7 observed adult HC (HC01/05/07/09/12/13/16).\n- Per-cell, confidence ≥ 0.9: ratio 0.35, 0.45, 0.12, 0.29, 0.38; in draw 2 the generated contrast is not significant (p = 0.58).\n- Intermediate monocytes: sign agrees 5/5, ratio 0.29–1.09, pooled Spearman 0.72.\n\n**Observed paired cells**\n- Adult CD14: −1.56 (43 vs 7, p = 4.5e-12); −1.36 to −1.56 across normalisations; −0.92 per cell.\n- Independent replication in paediatric sepsis vs paediatric HC: −1.38 (13 vs 4, p = 1e-4).\n- Intermediate −0.79, CD16 −0.64.\n\n**Bulk alone**\n- Whole-PBMC bulk HLA-II score in bulk-only samples: −0.76; −1.71 after subtracting monocyte markers.\n- Within adult sepsis on paired test samples, bulk ranks observed CD14 HLA-II better than the generated score: Spearman 0.45–0.63 vs 0.30–0.46.\n- Partial Spearman of generated given bulk: −0.06 to 0.17. Partial of bulk given generated: 0.30 to 0.50.\n\n**Is the monocyte location more than the bulk level?**\n- Observed CD14 effect after adjusting for bulk HLA-II: −1.02 (p = 0.005).\n- Generated, same adjustment, draws 0–4: −0.26, −0.53, −0.43, −0.52, −0.40 (p = 0.30, 0.0005, 0.16, 0.11, 0.012). That is 0.33–0.51 of the observed size, significant only in draws 1 and 4.\n\n**Whole-sample aggregate**\n- The generated whole-sample HLA-II score rises in sepsis in draws 0–1 (+0.36, +0.24) and falls in draws 2–4 (−0.28, −0.77, −0.31). The bulk falls (−0.76) and so do the observed cells (−0.42).\n- In observed cells, the whole-sample HLA-DRA change is −374 CPM: DC −518, CD14 + intermediate −232, B +467.\n\n**Pattern across cell types**\n- Generated vs observed HLA-II change, Pearson r over all 21–22 L2 types with cells: 0.44, 0.15, 0.30, 0.37, 0.35.\n- The candidate's 0.52–0.78 is over its 12 main types, and three of those are the monocyte types.\n\n**Compositional share**\n- S100A8/A9/A12 make up 6% of the observed CD14 library in HC and 19% in sepsis. That alone dilutes every other gene by about −0.21 log2.", + "issues": [ + "Calibration layer: this HLA-II program is not in calibration.csv. The row the claim cites, 'l2 | CD14 monocyte', is cell abundance, a different feature, and its sign reverses in draw 4. The program calibration was computed by the lens; I reproduced it independently and it holds.", + "'Every draw calibrates' is 5 looks at one small control set, not 5 independent confirmations. All five draws are calibrated against the same 6–7 observed adult HC donors (HC01, 05, 07, 09, 12, 13, 16). Leaving out each HC in turn keeps the sign (33/33).", + "The generated effect is attenuated: ratio 0.38–0.77 on pseudobulk and 0.12–0.45 per cell with label confidence ≥ 0.9. Draw 2 is the weakest draw: per-cell generated contrast p = 0.58, type-level logFC WEAK (r 0.28), bulk-only effect −0.71.", + "Per-patient information within sepsis comes from the bulk. Within adult sepsis, the partial Spearman of generated given bulk is −0.06 to 0.17, while bulk given generated is 0.30 to 0.50.", + "The claimed new content (the loss sits in monocytes) is only partly supported. Observed CD14 HLA-II falls by −1.02 beyond what the bulk HLA-II level predicts (p = 0.005). The generated cells reproduce that residual with the right sign in 5/5 draws, but at 0.33–0.51 of its size and significantly only in draws 1 and 4.", + "The generated populations do not add up to the bulk. In draws 0–1 the generated whole-sample HLA-II score rises in sepsis (+0.36, +0.24), while the bulk (−0.76) and the observed cells (−0.42) fall. The DC loss that drives most of the observed whole-sample HLA-DRA drop (−518 CPM) is not reproduced, because generated DCs are too rare to measure in most draws.", + "The cross-type tracking figure (r 0.52–0.78) depends on which types are included. Over all 21–22 L2 types with cells it is r 0.15–0.44. Generated CD8aa cells gain HLA-II in all 5 draws, and in draw 1 every T-cell subset gains it (+0.47 to +0.78). The observed cells show losses in memory CD4 (−0.68), Treg (−0.54) and DN T (−0.92).", + "The CD16 monocyte part is fragile. The bulk-only effect is about zero in draw 0 (−0.02) and changes sign in draw 3 under a compositional-robust library (+0.02). It is robust only in draws 1 and 4.", + "Part of the magnitude is compositional. The S100A8/A9/A12 share of the CD14 library rises from 6% to 19% in observed cells, which dilutes every other gene by about −0.21 log2. The effect survives robust and median-of-ratios normalisation (−0.72 to −1.36), so it is not an artifact.", + "Novelty: this is already reported for this cohort. Ye 2026 shows the HLA-DRA-high CD14 subset shrinking in most/all sepsis groups (PDF p.6 = journal p.155; PDF p.14 = p.163; whole blood on PDF p.12 = p.161). Low monocyte HLA-DR is not specific to sepsis (Reyes 2020 PDF p.4: uncomplicated infection shows it too), and whole-blood HLA-DRA mRNA already captures it (Winkler 2017 PDF p.1). The sepsis vs ICU-control contrast is weak, and its calibration is uninformative in 4/5 draws (only 6 observed ICU controls)." + ], + "reasoning": "I could not refute the core claim. My independent recomputation reproduces every number the candidate quotes. The direction holds under four normalisations, per-cell scoring with high-confidence labels only, age/sex adjustment, restriction to the HC hospital, and a 20- or 50-cell minimum.\n\n**1. Calibration.** The feature is not in calibration.csv, so I ran the equivalent paired-test check myself. CD14 HLA-II sepsis vs HC keeps its sign in all five draws, with ratios 0.38–0.77 and Spearman 0.51–0.64 (pooled 0.63). The pass rests on one small control set: the same 6–7 observed HC donors are reused in every draw, though leaving out any single HC keeps the sign every time. Draw 2 is the weakest.\n\n**2. Observed cells.** These support the claim strongly: −1.56 in adults, and the paediatric sepsis vs paediatric HC comparison replicates it (−1.38). The observed loss also exceeds what the bulk HLA-II level predicts.\n\n**3. Biology.** This is known, and no source contradicts it. The cohort paper itself reports the HLA-DRA-high CD14 subset shrinking in sepsis (Ye 2026 PDF p.6 and p.14). Reyes 2020, Winkler 2017, COMBAT 2022 and Yao 2023 agree. Reyes 2020 adds that low monocyte HLA-DR is not specific to sepsis.\n\n**4. Beyond the bulk.** The bulk already shows the fall, ranks patients better, and leaves almost nothing for the generated score within sepsis. The only new content the claim offers is the location of the loss, and that is weaker than stated:\n- the bulk-adjusted monocyte residual is reproduced at only 0.33–0.51 of its observed size, and is significant in 2 of 5 draws;\n- in draws 0–1 the generated whole-sample HLA-II moves the opposite way to the bulk;\n- the DC part of the observed drop is not reproduced;\n- the cross-type tracking of r 0.52–0.78 comes from a 12-type subset; over all 21–22 types it is r 0.15–0.44.\n\n**Draws that carry the result.** CD14 holds in draws 0–4, with draw 2 weakest. Intermediate monocytes hold in draws 0–4, weak in 0 and 2. CD16 holds robustly only in draws 1 and 4. The location beyond the bulk level is significant only in draws 1 and 4.\n\n**Verdict.** This is a calibrated replication of a known effect in the 122 bulk-only samples. It is strong for the within-monocyte direction and weak as information beyond the bulk.\n\n**Process.** All work was CPU-only under the memory-capped scope. Nothing was written outside my verify directory. This is a verifier-only analysis, so there is no paper-archive change from me; the coordinating agent owns the POLICY.md archive update.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v01_hla_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v01_bulkonly_perdraw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v01_calib_perdraw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v01_calib_pooled.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v01_calib_loo_hc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v01_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v01_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v02_specificity_aggregate_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v02_type_specificity_wide.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v02_type_specificity_corr.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v02_aggregate_whole_sample.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v02_hladra_decomposition_lineage.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v02_bulk_baselines_paired_test.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v02_bulkonly_bulk_hla.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v02_s100_share_cd14.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v03_percell_confidence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v03_percell_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v03_percell_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v03_label_confidence.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v04_bulk_tracking_age_train.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v04_bulk_tracking.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v04_age.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v04_train_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v05_bulk_adjusted_group_effect.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v05_bulk_adjusted.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v06_pooled_and_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v06_pooled_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v06_verdict_numbers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v00_literature_quotes.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/v00_pdf_extract.sh", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-1_calib/pdf_text" + ] + } + ] + }, + { + "id": "SEP-POP-2", + "lens": "sep_populations", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 2, + "theme": "monocyte HLA class II", + "synthesis_strength": "moderate for the cluster-free HLA-DR-high share; weak for cluster c4", + "reportable_statement": "The share of HLA-DR-high monocytes (cluster-free, depth-matched rule) is lower in bulk-only sepsis than HC in all five draws (-0.154/-0.178/-0.082/-0.199/-0.221) and passes the paired gate in 5/5 draws. The Leiden cluster c4 itself reproduces the observed contrast size only in draw 1. Same biology as SEP-PROG-1.", + "carrying_draws": { + "cluster c4": "1 (3 partly: 0.45x)", + "cluster-free HLA-DR-high share": "0,1,2,3,4 (0.35-0.89x; draw 2 weakest)" + }, + "honesty": { + "calibration": "c4: paired contrast 0.29/1.10/0.04/0.45/0.22x of observed; draw 2 Spearman 0.13. Cluster-free share: Spearman 0.40-0.63, contrast 0.35-0.89x, 5/5 gate.", + "replication": "5/5 draws negative under both definitions; survives age, sex, site, sex_flag, matching, label confidence.", + "observed": "Observed HLA-DR-high share -0.268 (43 vs 7, p 6.8e-5); c4 -0.243 (p 0.0038).", + "beyond_bulk": "None: bulk HLA index AUC 0.088; raw bulk HLA-DRA tracks the observed per-sample share better than generated in 5/5 draws (rho 0.71-0.74 vs 0.40-0.62)." + }, + "verdict_reasons": [ + "c4 mixes HLA-II with an uncalibrated AP-1 axis", + "sample sizes misreported (>=10-monocyte filter)", + "only draw 1 reproduces c4 size; d0 HC arm has 3 samples", + "part of the d0 effect is compositional (c4 share of all cells unchanged)", + "beyond-bulk overstated" + ], + "new_vs_cohort_papers": "Known: Ye 2026 Fig. 4d, Fig. 7e, ED Fig. 9a (deconvolution of the same in-house bulk), ED Fig. 9e (flow).", + "next_steps": [ + "Report the cluster-free HLA-DR-high share, not c4.", + "Same per-patient flow HLA-DR check as SEP-PROG-1." + ], + "key_sources": [ + "verify/SEP-POP-2_calib/tables/v06_per_draw_summary.csv", + "verify/SEP-POP-2_recompute/tables/v09_summary.csv", + "verify/SEP-POP-2_calib/tables/v04_tracking.csv" + ] + }, + "finding": { + "id": "SEP-POP-2", + "cohort": "ye2026_sepsis", + "claim": "In bulk-only generated cells, the HLA class II-high CD14 monocyte state (HLA-DRA, HLA-DRB1, HLA-DPB1, CD74, LGALS2, MS4A6A; MS2 / Mono_c01-CD14-HLA-DRA-like) is a smaller share of monocytes in adult sepsis than in adult HC in 5/5 draws. Draws 0, 1, 3 and 4 pass their own calibration. The observed paired cells show the same loss, but generated contrasts are weaker than observed in draws 0, 2 and 4.", + "kind": "state", + "feature": "Monocyte|r1.0|c4 (HLA-II-high CD14 monocytes; its r2.0 sub-states c7 [HLA-II-high, AP-1-low] and c13 [LGALS2/VCAN/HLA-II] change in the same direction). Generated top markers: FOSB, HLA-DRB1, LGALS2, HLA-DRA, JUN, FOS, MS4A6A, CD74, HLA-DPB1. HLA-II signature lfc +0.79 generated / +0.70 observed. Share of monocytes: bulk-only 11.5%, paired_gen 16.3%, observed 19.4%.", + "comparison": "Adult sepsis (68) vs adult HC (36), bulk-only, per draw; state fraction = cluster cells / monocytes. Also vs all controls and vs ICU.", + "effect_bulkonly": "Sepsis vs HC: d0 0.097 vs 0.305 (diff -0.207, AUC 0.08); d1 0.021 vs 0.182 (-0.161, AUC 0.09); d2 0.132 vs 0.258 (-0.125, AUC 0.16); d3 0.092 vs 0.261 (-0.169, AUC 0.15); d4 0.119 vs 0.218 (-0.099, AUC 0.19). Vs all controls -0.169/-0.112/-0.102/-0.143/-0.097 (q<=0.001). Vs ICU -0.013 to -0.090, q 0.07-0.43 (n.s.). Sub-state r2.0 c7: -0.094/-0.059/-0.027/-0.060/-0.070 (q<0.05 in d0, d1, d3, d4).", + "stats": "MWU per draw; BH over 143 cluster features per draw x contrast x measure. Sepsis vs HC q = 2.0e-8, 8.3e-11, 7.9e-4, 1.0e-5, 1.4e-5. sex_flag samples dropped: q <= 7.2e-5.", + "draws_same_sign": "5/5 (q<0.05 in 5/5)", + "paired_calibration": "Own rows (s05_calibration.csv). Ratio 1.04/0.66/0.86/0.97/1.06; Spearman 0.44/0.72/0.13/0.52/0.42. Sepsis-HC diff, generated vs observed: -0.069/-0.238, -0.224/-0.204, -0.010/-0.255, -0.105/-0.235, -0.054/-0.241. Sign agrees 5/5; gate passes draws 0, 1, 3, 4 (draw 2 fails, Spearman 0.13). The generated contrast is only 0.04-0.45x observed in draws 0, 2, 4. Related calibration.csv 'l2 | CD14 monocyte' (parent share): the HC contrast agrees in d0-d3 and fails in d4.", + "paired_observed_concordance": "Yes. Observed (co-clustered) adult sepsis vs adult HC diff -0.243 (p=0.0038; 30 vs 7); kNN-mapped -0.253 (p=0.0023); vs all controls -0.139 (p=0.011); vs ICU -0.118 (p=0.16; 30 vs 6). Counterpart: 642 observed cells; marker concordance Spearman 0.905. Observed cells of this coarse cluster also carry IL1B/CCL3/CXCL8. The AP-1-low HLA-II-high sub-state r2.0 c7 shows observed -0.091 (p=0.047).", + "beyond_bulk": "Partly. The 20-marker bulk score of this cluster is not associated with sepsis in bulk-only samples (AUC 0.47, p=0.60). However, a simple monocyte-normalised bulk HLA-II index (mean log CPM of HLA-DRA/DRB1/DPA1/DPB1/CD74 minus CD14/LYZ/FCN1/CSF1R/VCAN) already separates sepsis from HC (AUC 0.088, p<1e-4). Tracking the observed fraction on the same test samples, the generated fraction beats that index in 4/5 draws: generated 0.44/0.72/0.13/0.52/0.42 vs index 0.40/0.28/0.26/0.18/0.30. Rank-OLS: generated adds beyond the 20-marker score in d1 (p=0.001) and d3 (p=0.041). Net gain: a per-monocyte estimate that tracks observed cells somewhat better than a bulk ratio.", + "novelty": "Known. Ye 2026: Mono_c01-CD14-HLA-DRA reduced in most sepsis (PDF p.6 = printed p.155, Fig. 4d; whole-blood validation PDF p.12 = printed p.161, Fig. 7e). Lower monocyte HLA-DR is the classic sepsis immunosuppression marker; Winkler 2017 PLoS ONE shows that HLA-DRA mRNA tracks the protein. Not new.", + "strength": "moderate", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s04_profiles_r1.0.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s07_candidate_detail.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s09_bulk_derivability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s10_hla_bulk_index.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s12_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/s10_within_group_and_ratios.py" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "afdea1de81ea34dec", + "verdict": "qualified", + "recomputed": "Exact recompute from the discovery labels (v02, own code). Adult sepsis vs adult HC, c4 share of monocytes per draw: -0.207/-0.161/-0.125/-0.169/-0.099, AUC 0.077/0.086/0.161/0.145/0.195, p 1.8e-9/3.5e-12/1.6e-4/1.4e-6/2.9e-6. These match the claim exactly, including n. Vs all controls: -0.169/-0.112/-0.102/-0.143/-0.097. Vs ICU: -0.090/-0.013/-0.041/-0.059/-0.080 (p 0.006-0.11). Sub-state r2.0 c7: -0.094/-0.059/-0.027/-0.060/-0.070. The n_mono>=10 rule drops samples unevenly between groups: draw 2 drops 42/68 sepsis and 18/36 HC; draw 3 drops 35/68 sepsis and 6/36 HC; draw 0 drops 12/36 HC. A binomial GLM that keeps every sample with at least one monocyte gives log-OR -1.34/-2.85/-0.65/-1.24/-0.64 (all p<1e-4). Abundance (c4 cells / all cells): -0.002 (p=0.16, n.s.) / -0.015 / -0.002 (p=0.017) / -0.015 / -0.017.\nIndependent, cluster-free HLA-DR-high share (v03): uses all 1,500 cells per sample, HLA-DRA+DRB1 CP10k above the draw's q75. Diff -0.154/-0.178/-0.082/-0.199/-0.221, AUC 0.08/0.05/0.22/0.10/0.03, all p<=2e-5. Within high-confidence CD14 monocytes: -0.075/-0.135/-0.027/-0.133/-0.152 (p<=0.002). Mean log HLA-DR per monocyte: -0.87/-0.47/-0.62/-0.99/-0.87.\nMy own per-draw Leiden (v08/v08b): clusters matched to c4 by cell overlap fall in all 5 draws: d0 r2 -0.194, d1 r2 -0.141, d2 r1 -0.053, d3 r1 -0.314, d4 r1 -0.126.\nCalibration of the claimed state on the paired test samples, generated vs observed (v02):\n- Spearman 0.44/0.72/0.13/0.52/0.42; ratio 1.04/0.66/0.86/0.97/1.06.\n- Sepsis-HC diff -0.069/-0.238, -0.224/-0.204, -0.010/-0.255, -0.105/-0.235, -0.054/-0.241, i.e. generated/observed = 0.29/1.10/0.04/0.45/0.22.\n- HC n on the generated side is 3/6/6/7/7; on the observed side 6/6/7/7/7.\nCalibration of my HLA-DR-high share (observed cells thinned to generated depth): Spearman 0.55/0.45/0.41/0.43/0.63; ratio 1.56/1.43/1.06/1.18/1.15; generated/observed contrast 0.71/0.64/0.35/0.77/0.89, same sign in 5/5.\nObserved paired cells: c4 -0.243 (p=0.0038, 30 vs 7); First Affiliated only -0.262 (p=0.0019); HLA-DR-high share -0.268 (p=1e-4, 43 vs 7).\nBulk (v06):\n- HLA index AUC 0.088.\n- The claim's 6 named markers as a plain bulk score: AUC 0.18 (p<1e-4).\n- Tracking the observed c4 share on the same test samples, the cluster's 20-marker bulk score scores 0.58/0.67/0.46/0.56/0.66, against 0.44/0.72/0.13/0.52/0.42 for the generated share. The bulk score is better in 4/5 draws.", + "issues": [ + "Reproduces. My own code on the discovery labels gives identical diffs, AUCs, p and n. An independent cluster-free HLA-DR-high share (all cells, from markers.npz) and c4-matched clusters in my own per-draw Leiden both fall in sepsis in 5/5 draws. The direction is not an artefact of the embedding.", + "Age, sex, site, sex_flag and infection source do not explain it. Rank-OLS with age + sex + EMC site gives sepsis p<=2.7e-5 in every draw. The effect also holds with only First Affiliated Hospital samples, without sex_flag samples, with ages 50-85, and in age/sex-matched pairs (Wilcoxon p<=0.008 in every draw), and every infection source with n>5 falls. The observed contrast also survives site restriction and age/sex adjustment. The Imperial repeated-measures and trained-donor rules do not apply here: 0 test or bulk-only samples come from a training donor, and there is one sample per donor.", + "The reported n hide uneven missingness. The >=10-monocyte rule on the 200-cell subsample drops 42/68 sepsis samples in draw 2 and 35/68 in draw 3, so those per-draw tests use 26 and 33 sepsis samples. The all-sample GLM and the all-cells measure confirm the direction.", + "Technical and compositional covariates weaken the effect in some draws:\n- Adding log monocyte count, monocyte library size, label confidence and CD14 share makes the c4 effect borderline in d1 (p=0.052) and d4 (p=0.034).\n- In d0 and d2, the drop in the HLA-DR-high share is absorbed by the shift from intermediate- to CD14-labelled monocytes (p 0.63 and 0.34).\n- These covariates are partly mediators, so this does not refute the claim. Mean HLA-DR within CD14 monocytes stays lower under every adjustment.", + "The cluster is not purely CD14 monocytes: it is 35% intermediate-labelled in generated cells (CD14 62%). The drop is seen in both label parts. In d0 the c4 cells as a share of all cells do not change (p=0.16): the share of monocytes falls because the monocyte compartment expands. The abundance calibration has the wrong sign in d0 and d2.", + "Calibration of the claimed state is weak:\n- Generated/observed contrast is 0.29/1.10/0.04/0.45/0.22, so only d1 reproduces the size. The claim lists the weak draws as 0, 2 and 4 and omits d3 (0.45x).\n- The paired HC side is only 6-7 samples, and in d0 only 3 generated HC samples meet the >=10-monocyte rule.\n- d2 fails (Spearman 0.13).\n- d1 carries the best c4 calibration, but its monocyte compartment fails the trust layer: l1 Monocyte ratio 1.70 with a 4.8x sepsis-HC difference; l2 Intermediate monocyte ratio 7.55, Spearman 0.06, wrong-sign contrast.\n- l1 Monocyte sepsis-HC fails in d3 and d4, and l2 CD14 monocyte fails in d4.", + "The 'beyond bulk' claim is overstated by selective comparison:\n- The bulk already gives the contrast: HLA index AUC 0.088, and the claim's own 6-marker bulk score AUC 0.18, against generated AUC 0.08-0.19.\n- On the same test samples, the cluster's 20-marker bulk score tracks the observed c4 share better than the generated share in 4/5 draws (all except d1).\n- The claim compared only against the HLA index, the weakest bulk tracker.\n- For the HLA-DR-high share and mean HLA-DR, the bulk HLA index beats the generated cells in 5/5 draws.\n- The generated share adds beyond the bulk score only in d1 and d3.", + "'HLA-II-high' is a relative label. In my clustering, the most HLA-II-high generated monocyte clusters are ISG-high and CD16-like. A fixed rule (HLA-II z>=0.5, not CD16-like) at r=1.0 gives mixed signs: d0 +0.058 (p=5e-4), d4 -0.062. Only clusters matched to c4 reproduce the drop: LGALS2/MS4A6A/HLA-II/AP-1 classical monocytes, contrasted with S100A/MS1-like sepsis monocytes.", + "Which draws carry it. All 5 draws show the bulk-only drop. For the claimed c4 state, d1 reproduces the observed contrast fully and d3 partly (0.45x); d0 and d4 reproduce the sign only, and d2 fails. For the cluster-free HLA-DR-high share, all 5 draws pass the gate, and d0, d1, d3 and d4 reproduce at least 0.64x of the observed contrast (d2 0.35x)." + ], + "reasoning": "I could not refute the core direction:\n- In bulk-only generated cells, adult sepsis has a smaller share of HLA-II-high (LGALS2/MS4A6A/HLA-II) classical monocytes than adult HC in 5/5 draws.\n- The discovery labels give identical numbers with my own code.\n- The drop holds with my own definitions: a cluster-free per-cell HLA-DR-high share on all cells, a within-CD14 version with high label confidence, and c4-matched clusters from an independent per-draw Leiden clustering.\n- It survives age, sex, site, sex_flag, infection source, age/sex matching and label-confidence restrictions.\n- Observed paired cells show the same loss, robust to site and age/sex, but with only 7 adult HC.\n- This agrees with known sepsis biology.\n\nThe finding is qualified, not confirmed, for four reasons:\n1. The trust layer for this state is weak. Only d1 reproduces the observed contrast magnitude: d0, d2, d3 and d4 reach 0.04-0.45x, and the claim omits d3. The HC side has 3-7 samples. d1's monocyte compartment fails calibration.csv (monocyte share 1.7x, intermediate monocytes 7.6x with the wrong sign).\n2. Part of the effect is compositional. In d0 the c4 cells as a share of all cells do not change: the share of monocytes falls because monocytes expand, and that abundance calibration has the wrong sign in d0 and d2. The cluster is 35% intermediate-labelled, and the d0/d2 HLA-DR-high share drop is absorbed by the CD14/intermediate label shift.\n3. The claimed per-draw tests rest on uneven sample loss from the >=10-monocyte subsample rule. The all-sample GLM still confirms the direction.\n4. The beyond-bulk value is overstated. Simple bulk HLA-II scores give the same sepsis-HC separation, and bulk scores track the observed per-sample fractions as well as or better than the generated cells in most draws. The claim reported only the comparison against the weakest bulk index.\n\nNet: a real but known effect, which the bulk largely already shows. It is supported in the generated cells mainly by d1 (and partly d3) for the claimed cluster, and by d0/d1/d3/d4 for the cluster-free HLA-DR-high share. CPU only; no GPU work was started.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/vcommon.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v00_meta_overview.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v01_inspect_files.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v02_recompute_from_labels.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v03_hladr_allcells.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v04_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v04b_technical_covariates.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v05_calibration_rows.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v06_bulk_derivability.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v07_cluster_identity.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v08_own_leiden.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v08b_cluster_characterisation.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/v09_observed_checks_and_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v09_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v02_compare_discovery.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v02_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v02_missingness.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v02_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v03_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v03_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v04_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v04_restrictions.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v04_matched.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v04b_single_cov.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v05_calibration_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v06_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v06_paired_tracking.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v07_label_split.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v07_cluster_lfc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v08_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v08b_c4_matched.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v08b_hla_vs_effect.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/tables/v09_observed_confounders.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_recompute/logs" + ] + }, + { + "verifier": "calib", + "agent_id": "ae7acf1338481f4fc", + "verdict": "qualified", + "recomputed": "(1) CALIBRATION, the claimed cluster Monocyte|r1.0|c4 (v01). I re-ran it on matched paired-TEST samples: a sample counts only if both sets have at least 10 monocytes in the 200-cell subsample. Sepsis minus HC, generated vs observed:\n- d0: -0.072 vs -0.427 (17 vs 3 samples; generated p=0.31; bootstrap CI [-0.147, 0.006])\n- d1: -0.198 vs -0.204 (26 vs 6; p=0.003)\n- d2: +0.002 vs -0.217 (17 vs 6; p=1.0)\n- d3: -0.076 vs -0.194 (20 vs 7; p=0.18)\n- d4: -0.056 vs -0.228 (26 vs 7; p=0.24)\nSpearman reproduces the candidate: 0.44/0.72/0.13/0.52/0.42. Under a strict gate (same sign, generated p<0.05, at least 0.33x of observed, rho>=0.3) only draw 1 passes.\n\nCLUSTER-FREE ALL-CELL CHECK (v02/v03/v05). Measure: share of monocytes with HLA-DRA+HLA-DRB1 CP10k at or above the observed 75th percentile, with observed cells thinned to each draw's depth (hi_abs75).\n- Paired test, sepsis minus HC, generated vs observed: d0 -0.144 vs -0.239 (p=0.010); d1 -0.158 vs -0.276 (p=0.0008); d2 -0.098 vs -0.261 (p=0.005); d3 -0.193 vs -0.254 (p=0.0005); d4 -0.202 vs -0.250 (p=1e-4).\n- Spearman 0.52/0.40/0.48/0.41/0.62. The gate passes in 5/5 draws at 0.38-0.81x of observed.\n- Monocyte pseudobulk HLA-II (9 genes, log2), generated vs observed: -0.95/-1.28, -0.74/-1.45, -0.61/-1.31, -1.18/-1.33, -0.89/-1.31. Generated p<=0.001 in every draw.\n- The CD14-only parent gives the same result.\n- AP-1 genes (the cluster's top markers): observed sepsis minus HC is -0.40 log2 (p=0.58), and the generated-vs-observed Spearman is 0.05-0.23. AP-1 is uncalibrated and not linked to sepsis.\n\nBULK-ONLY, hi_abs75 sepsis minus HC: -0.185 (AUC 0.08), -0.195 (0.04), -0.085 (0.20), -0.200 (0.10), -0.229 (0.02). All q<=1.3e-5. Robust to dropping sex_flag samples and the 10 Emergency-Center sepsis samples.\n- The candidate's cluster numbers reproduce: -0.207/-0.161/-0.125/-0.169/-0.099.\n- But they were computed on 59/66/26/33/50 sepsis vs 24/31/18/30/33 HC, not 68 vs 36.\n- Vs ICU: observed paired shows no difference (hi_abs75 -0.026, p=0.60). In bulk-only, generated d1 (q=0.001) and d4 (q=0.03) show nominal decreases that the observed cells do not support.\n\n(2) OBSERVED, all cells of the 82 paired samples. HLA-DR-high share 0.189 vs 0.457 (-0.268, AUC 0.076, p=6.8e-5; 43 vs 7). CD14-only -0.372 (p=1e-6). Pseudobulk HLA-II -1.37 log2 (p=1e-5).\n\n(4) BEYOND BULK (v04).\n- Bulk HLA-DRA log CPM tracks the observed all-cell HLA-DR-high share at rho 0.74/0.74/0.74/0.71/0.73. The generated cells reach only 0.52/0.40/0.48/0.41/0.62, so the bulk wins in 5/5 draws.\n- Partial rho(generated, observed | bulk HLA-DRA) is 0.16/0.12/0.28/0.16/0.36.\n- For the observed cluster target, bulk HLA-DRA gives 0.55/0.48/0.40/0.36/0.52 vs generated 0.44/0.72/0.13/0.52/0.42; the generated cells win only in d1 and d3. The candidate's own 20-marker bulk score (s09: 0.58/0.67/0.46/0.56/0.66) beats the generated cells in 4/5.\n- The bulk index alone gives AUC 0.088 for sepsis vs HC.", + "issues": [ + "The candidate's calibration for the claimed cluster is overstated. Its sign agreement in 5/5 draws compares generated and observed differences computed on different sample subsets. On matched samples only draw 1 reproduces the sepsis-HC contrast (-0.198 vs -0.204, p=0.003). Draws 0, 3 and 4 keep the sign, but the generated contrast is 0.17-0.39x of observed and not significant (p=0.31/0.18/0.24; draw 0 has 3 HC). Draw 2 flips to +0.002.", + "The sample sizes are misreported. The bulk-only state tests did not compare 68 vs 36. After the filter requiring at least 10 monocytes among the 200 subsampled cells they compared 59/66/26/33/50 sepsis vs 24/31/18/30/33 HC; draw 2 kept only 26 of 68 sepsis samples. The observed contrast used 30 of 46 adult sepsis samples.", + "The cluster is not a clean HLA-II-high state. 8 of its top 15 generated markers are AP-1 / immediate-early genes (FOSB, JUN, FOS, IER2, JUND, JUNB, DUSP1, ZFP36), and its AP-1 signature (+0.91) is higher than its HLA-II signature (+0.79). The observed counterpart also carries IL1B/CCL3/CXCL8. That resembles Ye's Mono_c02-CD14-IL1B, which Ye reports is increased in sepsis.", + "The AP-1 part of the cluster plays no role in sepsis. In observed monocytes, AP-1 does not differ between sepsis and HC (p=0.58), and generated AP-1 tracks observed poorly (Spearman 0.05-0.23). The sepsis signal comes entirely from HLA-II. The all-cell HLA-DR-high share or the monocyte pseudobulk HLA-II is the calibrated feature, and it passes in 5/5 draws.", + "The 'net gain beyond bulk' claim is not supported. Raw bulk HLA-DRA tracks the observed per-sample HLA-DR-high share better than the generated cells in 5/5 draws (rho about 0.71-0.74 vs 0.40-0.62). For the cluster target, the candidate's own 20-marker bulk score (s09) beats the generated fraction in 4/5 draws. The 'beats the index in 4/5' result holds only against the weakest bulk comparator (the monocyte-normalised index).", + "The cohort paper already reported this from the same bulk data. Ye 2026 deconvolved its in-house PBMC bulk RNA-seq (Extended Data Fig. 9a, PDF p.39; methods PDF p.19-20) and found Mono-c01-CD14-HLA-DRA highest in HC and lower in ICU controls and sepsis. So the bulk-only result is neither new nor beyond what the bulk already shows.", + "Not novel. Ye 2026 PDF p.6 / printed p.155 and Fig. 4d (PDF p.9 / printed p.158) show Mono_c01 reduced vs HC in IC*, Abd***, Res**, Uri***, CNS*, with no difference between ICU controls and sepsis. The same appears on PDF p.12 / printed p.161 (Fig. 7e, whole blood) and in Extended Data Fig. 9e flow cytometry: HC about 47% vs sepsis 16-19% (p=0.0044-0.0244) and ICU controls about 23%. Reyes 2020 (PDF p.2, p.4) defines MS2 as the class II-high state and notes that low HLA-DR is not sepsis-specific. Winkler 2017 (PDF p.7) shows whole-blood HLA-DRA mRNA 70% lower in sepsis, tracking surface HLA-DR at rho = 0.9.", + "There is a ratio error in the claim. It says the generated contrast is '0.04-0.45x in draws 0, 2, 4', but 0.45x is draw 3. Draws 0/2/4 are 0.29/0.04/0.22x unmatched and 0.17/-0.01/0.24x matched.", + "Vs ICU, the all-cell bulk-only measure shows nominal sepsis-below-ICU differences in draw 1 (q=0.001) and draw 4 (q=0.03). The observed paired cells (p=0.60) and Ye Fig. 4d do not show this. The ICU contrast is uncalibrated and should be reported as no difference.", + "In draw 0 the drop in share of monocytes comes mostly from the denominator: the cluster's abundance among all cells does not change (diff -0.002, q=0.21). In draw 4, generated sepsis monocytes are deeper (AUC 0.74), and depth correlates with the threshold-based HLA-DR share (rho about -0.6). The pseudobulk ratio measure is not affected by this." + ], + "reasoning": "I tried to refute the claim but could not. The direction holds: monocyte HLA class II and the HLA-DR-high share are lower in adult sepsis than in adult HC. The observed paired cells support it strongly, an independent cluster-free calibration supports it in all five draws, and the biology is well established. The claim as written still needs qualification on four points.\n\n(1) Calibration. This Leiden cluster is not in calibration.csv, so I checked it myself.\n- Draw by draw on matched paired-test samples, the cluster's sepsis-HC contrast is reproduced with similar size and significance only in draw 1. That is the draw whose training donors include adult HC12.\n- Draws 0, 3 and 4 keep the sign only, at 0.17-0.39x of observed and not significant, with 3-7 HC per draw. Draw 2 fails.\n- The cluster mixes an uncalibrated AP-1 axis, which has no sepsis association in observed cells, with HLA-II.\n- A cluster-free measure on all cells passes the gate in draws 0-4, with 0.38-0.81x attenuation and draw 2 weakest. The same holds for CD14 monocytes alone and for HLA-II pseudobulk. That measure is the HLA-DRA+HLA-DRB1-high share of monocytes, with observed cells thinned to each draw's depth. So the HLA-II loss is reportable in all five draws, while the specific cluster is carried robustly only by draw 1.\n- Sepsis vs ICU has no observed signal to calibrate against.\n\n(2) Observed support. The association is clearly present in the observed paired cells: -0.268 HLA-DR-high share, p=6.8e-5, and HLA-II pseudobulk -1.37 log2. The candidate's co-clustered estimate (-0.243, p=0.004) agrees.\n\n(3) Biology and novelty. This is known and reported for this exact cohort. Ye 2026 shows it in scRNA-seq (Fig. 4d), whole blood (Fig. 7e), flow cytometry (Extended Data Fig. 9e) and CIBERSORTx deconvolution of the same in-house bulk RNA-seq (Extended Data Fig. 9a). It is also the classic monocyte HLA-DR immunosuppression marker (Reyes 2020 MS2; Winkler 2017).\n\n(4) Beyond bulk. The generated cells add nothing measurable. The bulk alone separates sepsis from HC (AUC 0.088), and raw bulk HLA-DRA tracks the observed per-sample HLA-DR-high share better than the generated cells in every draw.\n\nNet: real and calibrated in direction (cluster-free in 5/5 draws; the claimed cluster robustly only in draw 1), confirmed by observed cells, not new, and not beyond bulk. The verdict is qualified, not confirmed.\n\nI started no GPU work; everything ran CPU-only under systemd-run with MemoryMax=6G. This verification changed no paper assets, so there is no archive delta; the coordinating agent owns the archive update.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/vcommon.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/v01_recalib_clusters.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/v02_hladr_allcells.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/v03_pb_programs.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/v04_beyond_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/v05_sensitivity_depth.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/v06_compile.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v06_per_draw_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v01_cluster_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v01_monocyte_counts_by_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v02_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v02_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v02_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v02_thresholds.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v02_persample.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v03_pb_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v03_pb_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v03_pb_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v04_tracking.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v04_bulkonly_vs_index.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v04_bulk_index.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v05_parent_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/tables/v05_depth_by_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/lit/edf9a_thisstudy.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/lit/edf9e_mono_c01.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/lit/fig4d.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/lit/ye_p6.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/lit/ye_p12.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/lit/ye_p14.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/lit/reyes_p4.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/lit/winkler_p7.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-2_calib/logs/v06_compile.log" + ] + } + ] + }, + { + "id": "SEP-PROG-2", + "lens": "sep_programs", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 3, + "theme": "MS1 / S100 monocyte program", + "synthesis_strength": "moderate for direction; weak for magnitude (draws 0-1 only); none beyond bulk", + "reportable_statement": "Generated CD14 and intermediate monocytes of bulk-only sepsis patients express more MS1 genes (+2.25/+1.31/+0.40/+1.09/+0.60 log2 in CD14) and S100A8/A9/A12 than HC in all five draws. Only draws 0 and 1 reproduce the observed size; draws 2-4 give the direction only. The CD16 part and several named genes do not hold.", + "carrying_draws": { + "MS1 score in CD14 (magnitude)": "0,1 (3 borderline)", + "MS1 direction": "0,1,2,3,4", + "S100 in CD14": "0,1,3", + "CD16 part": "none at gene level" + }, + "honesty": { + "calibration": "MS1 in CD14 paired ratio 1.06/0.58/0.19/0.45/0.26, Spearman 0.62-0.73; calibration.csv MS1 state ratio 0.98 (d0) vs 0.03-0.14 (d1-d4); sepsis-vs-ICU state sign 0/5.", + "replication": "5/5 draws same sign, q<1e-5; robust to age, sex, site, sex_flag, depth, label confidence, granulocyte leakage.", + "observed": "Observed CD14 MS1 +1.93 (p 6e-15), S100 +1.55 (p 2e-11); RETN and CYP1B1 do not rise in observed CD14.", + "beyond_bulk": "None: bulk MS1 +2.17 and S100 +2.59 (t ~14); bulk ranks observed CD14 MS1 better (0.83-0.88 vs 0.62-0.73); in draws 2 and 4 the generated contrast is fully explained by bulk." + }, + "verdict_reasons": [ + "lenient PASS window (0.25-4) hid attenuation", + "S100 not monocyte-specific in generated cells; CD16>CD14 reversed", + "CLU partly a platelet confound", + "PLAC8/MCEMP1/ALOX5AP/S100A12 flat or reversed in some draws", + "not sepsis-specific" + ], + "new_vs_cohort_papers": "Known: Reyes 2020 MS1; Ye 2026 Mono_c03-CD14-RETN (PDF p.6, 8, 14); Yao 2023.", + "next_steps": [ + "Draws whose training donors carry MS1 cells (only draw 0 does now) or more draws, to separate imprint from bulk signal.", + "Use IL1R2/CD163/S100A8 (genes that hold in all draws) and depth-matched scores; drop CD16.", + "Benchmark against bulk MS1 deconvolution (Reyes 2020; Leite 2024) on paired samples." + ], + "key_sources": [ + "verify/SEP-PROG-2_recompute/v99_headline.csv", + "verify/SEP-PROG-2_calib/v05_per_draw_verdict.csv", + "verify/SEP-PROG-2_calib/v02_gene_table.csv", + "verify/SEP-PROG-2_calib/v03_gene_pattern_concordance.csv" + ] + }, + "finding": { + "id": "SEP-PROG-2", + "cohort": "ye2026_sepsis", + "claim": "In bulk-only sepsis vs HC, the generated CD14, intermediate and CD16 monocytes express more MS1 genes (IL1R2, CLU, MCEMP1, CD163, RETN, ALOX5AP, PLAC8) and more S100A8/A9/A12: CD14 MS1 +1.13 log2, S100 +0.71. The observed paired monocytes show the same (+1.93, +1.55). Only draw 0 reaches the observed MS1 magnitude; the other draws are attenuated. The bulk alone already shows these genes rising.", + "kind": "program", + "feature": "MS1 score (IL1R2, RETN, ALOX5AP, CYP1B1, MCEMP1, PLAC8, CLU, CD163) and S100 score (S100A8, S100A9, S100A12) in AIFI-L2 monocyte pseudobulks", + "comparison": "adult sepsis vs adult HC, bulk-only (50–68 vs 33–36 per type), within draw", + "effect_bulkonly": "MS1 in CD14: +2.25, +1.31, +0.40, +1.09, +0.60 log2 (mean +1.13, d 2.3); intermediate +1.45/+1.53/+0.37/+1.05/+0.48; CD16 +0.81/+1.66/+0.27/+0.42/+0.43. S100 in CD14: +1.37, +1.00, +0.14, +0.84, +0.19; CD16 +1.69, +2.70, +0.87, +1.42, +0.80. IL1R2 within CD14: +5.9, +1.6, +1.1, +1.8, +2.0. CYP1B1 is inconsistent (3/5 draws).", + "stats": "Welch per draw. MS1 in CD14: p 9e-35, 1e-19, 1e-7, 5e-18, 1e-16. S100 in CD14: p 1e-27, 2e-19, 0.004, 3e-14, 0.0015. S100 in CD16: p <= 3e-7. Within-draw BH q < 0.01 for all. Mann–Whitney p <= 0.003. Robust to age+sex adjustment, to restricting sepsis to the same hospital, to dropping sex-flag samples and to a 20-cell minimum.", + "draws_same_sign": "5/5 (MS1 and S100 in CD14, intermediate and CD16; q<0.05 in 5/5)", + "paired_calibration": "programs_calib.csv. MS1 in CD14: PASS in draws 0,1,3,4 (paired-test generated/observed 1.99/1.88, 1.14/1.98, 0.86/1.90, 0.52/1.99; Spearman 0.62–0.73); draw 2 SIGN-ONLY (0.36/1.91, ratio 0.19). S100 in CD14: PASS in 0,1,3; SIGN-ONLY in 2 and 4 (ratio 0.13, 0.11). S100 in CD16: PASS 5/5. MS1 in intermediate: PASS 0,1,3,4. calibration.csv row 'MS1 monocyte (IL1R2+)': sign kept 5/5. Draw 0 ratio 0.98 (sepsis−HC 0.167 vs 0.151); draws 1–4 ratio 0.03–0.14 (Spearman 0.29–0.52). The DATA_GUIDE notes draw 0's training donors are MS1-rich sepsis donors.", + "paired_observed_concordance": "Observed paired cells (43 vs 7). MS1: CD14 +1.93 (p 6e-15), intermediate +1.13, CD16 +1.04. S100: CD14 +1.55 (p 2e-11), intermediate +1.60, CD16 +0.89. Individual genes in CD14: IL1R2 +4.6, CLU +4.3, MCEMP1 +1.9 log2. The MS1 cross-type pattern, generated vs observed, has r 0.71–0.97 (confined to monocytes). The S100 cross-type r is about 0, because observed S100 also rises in T/B pseudobulks.", + "beyond_bulk": "Little. In the bulk-only samples the bulk MS1 genes rise +2.17 log2 and S100 +2.59 (Welch t 14.6 and 14.0). An HC reference atlas already places S100A8/A9/A12, MCEMP1, CD163 and RETN in CD14 monocytes. Per sample, the bulk ranks observed CD14 MS1 better than the generated cells (Spearman 0.83–0.88 vs 0.62–0.73). What is added: localisation to monocyte subsets, and for IL1R2 (healthy carrier = DC) within-CD14 induction in 5/5 draws. Against this, for S100A8/A9 the generated decomposition calls monocyte expansion in 4/5 draws where observed cells show within-cell induction. The generated rise is also larger in CD16 than in CD14 monocytes, the reverse of observed.", + "novelty": "Known. Reyes 2020 Nat Med (rare-disease-states PDF pp.2–4): MS1 = RETN-, ALOX5AP- and IL1R2-high CD14+ monocytes. Ye 2026 PDF p.6: Mono_c03-CD14-RETN (S100A8/S100A12-high, HLA-DR-low, MS1-like) increased in most sepsis. Yao 2023: HLA-DRlow S100Ahigh monocytes.", + "strength": "strong", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s03_programs.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_draws.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_type_specificity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/bulk_baselines.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/site_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_genes_fine.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a0620fb5dc7c7b14b", + "verdict": "qualified", + "recomputed": "My own code worked from pb_l2/pb_l1, cells.parquet + markers.npz, subsample.npz, the bulk npz files and calibration.csv. Cell-level MS1 sums reproduce pb_l2 exactly. Score = edgeR log-CPM (prior.count 2), cells ≥10.\n\nBULK-ONLY, adult sepsis vs adult HC, per draw d0/d1/d2/d3/d4:\n- MS1, CD14: +2.25/+1.31/+0.40/+1.09/+0.60 (p 9e-35, 1e-19, 1e-7, 5e-18, 1e-16; mean +1.13).\n- MS1, intermediate: +1.45/+1.53/+0.37/+1.05/+0.48.\n- MS1, CD16: +0.81/+1.66/+0.27/+0.42/+0.43.\n- S100, CD14: +1.37/+1.00/+0.14/+0.84/+0.19 (p 1e-27, 2e-19, 0.004, 3e-14, 0.0015; mean +0.71).\n- S100, CD16: +1.69/+2.70/+0.87/+1.42/+0.80.\n- Same sign in 5/5 draws; within-draw BH q<0.01 (6 tests per draw).\n- IL1R2 in CD14: +5.9/+1.6/+1.1/+1.8/+2.0.\nThese match the discovery agent's numbers to 3 decimals.\n\nCONFOUNDER CONTROLS, MS1 in CD14:\n- Age+sex OLS: 2.22/1.30/0.40/1.10/0.58.\n- Sepsis restricted to the HC hospital: 2.24/1.29/0.34/1.04/0.62.\n- sex_flag samples dropped: 2.30/1.32/0.41/1.13/0.63.\n- 20-cell minimum: 2.28/1.32/0.44/1.14/0.60.\n- Adjusted for log cell number and log library: 1.91/0.67/0.25/0.99/0.49.\n- Per-cell depth reweighting: 2.26/1.33/0.35/1.13/0.69.\n- Label confidence ≥0.99: 2.20/1.28/0.41/1.14/0.60.\n- Depth-only thinning: 2.29/1.29/0.35/1.18/0.63.\n- Strict combined control (one hospital + no flags + ≥20 cells + thinned with raised prior + age/sex): 1.63/0.96/0.19/0.80/0.45, all p≤1.4e-4.\n\nOBSERVED PAIRED ADULTS (43 vs 7):\n- MS1: CD14 +1.93 (p 6e-15), intermediate +1.13, CD16 +1.04.\n- S100: CD14 +1.55 (p 2e-11), intermediate +1.60, CD16 +0.89.\n\nPAIRED CALIBRATION, MS1 in CD14 (test samples):\n- Generated/observed difference: 2.00/1.88, 1.15/1.98, 0.36/1.91, 0.86/1.90, 0.52/1.99.\n- Ratio: 1.06/0.58/0.19/0.45/0.26 (depth-matched 1.14/0.59/0.19/0.49/0.27).\n- Spearman: 0.73/0.68/0.71/0.71/0.62.\n- Magnitude ratio, bulk-only vs observed at equal depth: 1.18/0.64/0.17/0.66/0.29.\n\nPAIRED CALIBRATION, OTHER CELLS:\n- S100 in CD14 ratio: 0.80/0.59/0.13/0.54/0.11 (Spearman 0.52/0.60/0.36/0.59/0.22).\n- S100 in CD16 ratio: 1.31/2.83/0.52/1.08/0.62.\n\nCALIBRATION.CSV, 'MS1 monocyte (IL1R2+)' row, sepsis vs HC:\n- Sign kept in 5/5 draws.\n- Ratio of the differences: 1.11/0.04/0.06/0.13/0.08.\n- Level ratio: 0.98/0.033/0.048/0.138/0.061.\n- Spearman: 0.52/0.46/0.29/0.37/0.47.\n- Sepsis vs ICU control: sign kept in 0/5 draws.\n- There is no S100 row.\n\nBULK ALONE:\n- Bulk-only cohort: MS1 +2.17 (t 14.6), S100 +2.59 (t 14.0).\n- Paired cohort: MS1 +1.67, S100 +2.27.\n- Spearman with observed CD14 MS1: bulk 0.83–0.87 vs generated 0.62–0.73. Partial Spearman of generated vs observed given the bulk: −0.24 to 0.28.\n- Sepsis effect left after adjusting for the bulk MS1 level: observed +0.85 (p 0.004); bulk-only generated d0 +1.00, d1 +0.78, d2 −0.04 (p 0.70), d3 +0.55, d4 +0.11 (p 0.18).\n\nOTHER CHECKS:\n- Granulocyte genes in generated monocytes (FCGR3B, MMP9, LCN2, CAMP): about 0 CPM and no change, so the effect is not granulocyte RNA leaking into monocytes.\n- No donor appears twice; no training donor is among the bulk-only samples; the highest bulk correlation with a training donor is 0.991–0.995, against 0.997 between bulk-only samples themselves.", + "issues": [ + "Calibration is weaker than 'PASS in 0,1,3,4'. The discovery gate accepted any ratio from 0.25 to 4. With a 0.5–2 window, MS1 in CD14 is calibrated in magnitude only in draws 0 (1.06) and 1 (0.58). Draws 3 (0.45) and 4 (0.26) are attenuated and draw 2 (0.19) has the right sign only. On the trust layer's MS1-state row, draws 1–4 generate the state at 0.03–0.14x observed and only draw 0 passes (0.98).", + "S100 in CD14 is only carried by draws 0, 1 and 3. Draws 2 and 4 have the right sign only (ratio 0.13, 0.11; Spearman 0.36, 0.22). Their bulk-only effect is fragile: at a 3-cell minimum, draw 2 gives −0.01 (p 0.89) and draw 4 +0.10 (p 0.17). Pooled L1 monocytes in draw 4 give +0.075 (p 0.5), and S100A12 in draw 4 is 0.00 (p 1). calibration.csv has no S100 row, so S100 rests only on the lens's own program calibration.", + "The generated S100 rise is not specific to monocytes. Draw 1 raises S100 in T cells by +2.3 to +3.6 log2, more than in CD14 (+1.0). Draw 4 lowers it in T/B cells (−0.5 to −0.95), the opposite of observed (+0.6 to +1.4). Across all draws, CD16 rises more than CD14, the reverse of observed. 'S100 in CD16 PASS 5/5' also hides draw 1's ratio of 2.83 (inflated). The observed CD16 test contrast is weak (p 0.008–0.04).", + "Little beyond the bulk. The bulk-only bulk contrast (MS1 +2.17, S100 +2.59; t about 14) is stronger than the paired bulk contrast. The bulk predicts observed CD14 MS1 better than the generated cells do. In draws 2 and 4 the generated sepsis−HC difference is fully explained by each sample's bulk MS1 level (sepsis coefficient −0.04 and +0.11, not significant). Observed cells keep an extra sepsis effect beyond the bulk (+0.85, p 0.004); generated paired-test cells reproduce it only in draws 0 and 1.", + "Smaller misstatements in the claim. (1) Draws 1–4 Spearman in calibration.csv is 0.29–0.47, not 0.29–0.52 (0.52 belongs to draw 0). (2) The Mann–Whitney p for S100 in draw 2 is 0.0034–0.0037, above the stated 0.003. (3) PLAC8 does not consistently rise: it falls in draw 4 (−0.23, p 0.01) and is flat in draw 2; MCEMP1 is flat in draw 2 and ALOX5AP flat in draw 4. (4) The cross-type MS1 correlation of 0.71–0.97 is driven by 3 monocyte points; the Spearman is 0.26–0.71.", + "The size of the effect depends on the score definition; its sign mostly does not. MS1 in CD14, draw 0, ranges from +1.54 to +2.61 log2 across score variants, and draw 4 from +0.21 to +0.75. One sign flip: the summed-CPM score for MS1 in CD16, draw 4, is −0.20 (p 0.02).", + "Selection by the cell minimum. With 10 cells required, CD16 and intermediate analyses in draws 2 and 3 keep only 23–32 of 68 sepsis samples, so they rest on a selected subset. Draw 2 MS1 in intermediate monocytes is fragile: +0.15 (p 0.18) at a 3-cell minimum and +0.10 (p 0.4) at label confidence ≥0.99. MS1 in CD14 and in pooled L1 monocytes stays positive in 5/5 draws.", + "The signal is not sepsis-specific. Observed paired sepsis vs ICU control is not significant (MS1 in CD14 +0.39, p 0.34, 6 controls). The calibration.csv MS1-state row disagrees in sign with observed for sepsis vs ICU in 0/5 draws. Composition calibration of the monocyte subsets is also poor in some draws: the CD16 sepsis−HC sign is wrong in draws 2–4, and draw 1 makes 7.6x the observed share of intermediate monocytes (Spearman 0.06). This weakens the 'localisation to subsets' that the claim gives as its added value." + ], + "reasoning": "I recomputed the effect from the raw compact files with independent code. The bulk-only per-draw effects, p-values, Mann–Whitney tests, observed paired effects and program calibration numbers all reproduce the claim, mostly to three decimals.\n\nNo reasonable confounder control removes the MS1 effect in CD14 in any draw:\n- age and sex;\n- restricting sepsis to the HC hospital (the only site variable in this cohort);\n- dropping sex_flag samples;\n- repeated donors and training donors (none exist here, and there are no near-duplicate bulk profiles);\n- the cell minimum (3/10/20/50), adjusting for log cell number and library, thinning the pseudobulk to a common library, reweighting cells to a common depth;\n- label confidence ≥0.9 or ≥0.99;\n- granulocyte transcripts leaking into monocytes.\n\nSo the core direction stands: 5/5 draws, every draw significant. The observed paired monocytes agree (+1.93 MS1, +1.55 S100 in CD14).\n\nThe verdict is qualified rather than confirmed for three reasons:\n1. **Calibration is weak outside draws 0–1.** Using each draw's own row, MS1 in CD14 is calibrated in magnitude only in draws 0 and 1. Draws 3 and 4 are attenuated (0.45, 0.26) and draw 2 has the right sign only. The trust-layer MS1-state row in calibration.csv passes only draw 0; draws 1–4 sit at 0.03–0.14x.\n2. **The S100 part is weaker still.** It is carried by draws 0, 1 and 3 only; draws 2 and 4 have the right sign only and lose significance with a looser cell minimum or pooled monocytes. calibration.csv has no S100 row. The generated S100 rise is not monocyte-specific (draw 1 broadcasts it into T cells; draw 4 inverts it), and the CD16 > CD14 ordering is the reverse of observed.\n3. **It adds little beyond the bulk.** The bulk shows the same genes rising more strongly. The bulk predicts observed CD14 MS1 better than the generated cells do. In draws 2 and 4 the generated contrast is fully explained by the per-sample bulk MS1 level.\n\nReportable carriers:\n- MS1 in CD14 and intermediate: draws 0 and 1 at calibrated magnitude (draw 0's training donors are MS1-rich: mean MS1-state fraction 0.29 vs a sepsis median of 0.087). Draws 3 and 4 give the direction only, attenuated.\n- S100 in CD14: draws 0, 1 and 3.\n\nThe biology is already known (Reyes 2020, MS1). Overall: a reproduced, confounder-robust recovery of a known signal that the bulk already shows. It is not 'strong' novel evidence.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v99_headline.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/vlib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v00_extract.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v00_checks.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v01_bulkonly.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v01_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v01_bulkonly_primary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v01_gene_level.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v01_selection.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v02_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v02_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v02_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v02_obs_vs_bulkonly_thin.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v03_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v03_bulk_effects.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v03_gen_vs_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v03_selection_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v03_by_source.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v04_confidence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v04_ms1_conf.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v04_depth_conf_by_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v04_subsample_conf.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v04b_cell_depth.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v04b_cell_depth.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v05_crosstype.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v05_type_effects.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v05_crosstype_corr.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v06_calibcsv_states.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v06_calibration_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v06_ms1_state_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v07_extra.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v07a_neutrophil_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v07b_residual_S.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v07c_bulk_dupes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v07d_gene_carriers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v07e_depth_only_thin.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v07f_S_vs_ICU.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v08_train_donors.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/v08_train_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_recompute/pb_long.parquet" + ] + }, + { + "verifier": "calib", + "agent_id": "acdedb71cc3fe249e", + "verdict": "qualified", + "recomputed": "The analysis was rebuilt from scratch (v00_common.py; it does not import lens_common) and reproduces the candidate's numbers exactly. Unless noted, values use edgeR-style logCPM with prior 2, at least 10 cells per pseudobulk, and adult sepsis vs adult HC within each draw.\n\nBULK-ONLY, MS1 score in CD14:\n- Differences by draw: +2.251, +1.307, +0.395, +1.092, +0.602.\n- Largest within-draw BH q is 2.3e-6.\n- With log2(CPM+1) instead: +2.61, +1.48, +0.53, +1.30, +0.75.\n- The sign holds at minimums of 20 and 50 cells, with age and sex in an OLS model, without sex_flag samples, and with sepsis restricted to the HC hospital.\n\nBULK-ONLY, S100 score in CD14:\n- Differences by draw: +1.367, +1.002, +0.140, +0.841, +0.191.\n- Largest q is 0.016.\n- In CD16: +1.69, +2.70, +0.87, +1.42, +0.80.\n\nPAIRED CALIBRATION, MS1 in CD14 (draw test samples, generated vs observed):\n- Contrast ratio with edgeR-style norm: 1.06, 0.58, 0.19, 0.45, 0.26.\n- Contrast ratio with log2(CPM+1): 1.37, 0.65, 0.28, 0.51, 0.35.\n- Spearman, generated vs observed: 0.73, 0.68, 0.70, 0.71, 0.62.\n- Spearman, bulk vs observed: 0.87, 0.83, 0.86, 0.83, 0.87.\n- Partial Spearman, generated vs observed given bulk: 0.09, 0.26, 0.28, 0.19, -0.23.\n\nPAIRED CALIBRATION, S100 in CD14:\n- Contrast ratio: 0.80, 0.59, 0.13, 0.54, 0.11.\n- Spearman: 0.52, 0.60, 0.36, 0.59, 0.22.\n\nMS1 AS A DISCRETE CELL STATE (paired_gen vs observed_thin):\n- Contrast ratio by draw: 1.12, 0.04, 0.07, 0.17, 0.07.\n- Bulk-only sepsis minus HC state fraction: 0.268 in draw 0, but only 0.012 to 0.044 in draws 1 to 4.\n- Observed: 0.199 in sepsis vs 0.007 in HC.\n\nSTRICTER PER-DRAW GATE (same sign, Spearman at least 0.3, ratio 0.5 to 2):\n- MS1 in CD14 is carried by draws 0 and 1. Draw 3 is borderline (0.45 with edgeR-style norm, 0.51 with log2(CPM+1)).\n- S100 in CD14 is carried by draws 0, 1 and 3.\n- The MS1 cell state is carried by draw 0 only.\n- Draws 2 and 4 keep only the sign.\n\nOBSERVED ADULTS (43 sepsis vs 7 HC):\n- MS1 in CD14 +1.93 (p 6e-15); S100 in CD14 +1.55 (p 2e-11).\n- Restricted to the HC hospital: IL1R2 +4.74, CLU +4.55, MCEMP1 +1.97, S100A8 +1.90.\n- RETN in CD14 +0.99 (p 0.13, not significant); CYP1B1 -0.38 (not significant).\n\nPER GENE IN CD14, BULK-ONLY (draws up at q<0.05, out of 5):\n- 5 of 5: IL1R2, CLU, CD163, S100A8, RETN. RETN is only 0.39 to 0.43 in draws 0 to 2.\n- MCEMP1 4 of 5 (draw 2: 0.12).\n- ALOX5AP 4 of 5 (draw 4: 0.10).\n- S100A9 4 of 5; S100A12 4 of 5 (draw 4: 0.00).\n- PLAC8 3 of 5: draw 4 goes down (-0.23, q<0.05) and draw 2 is 0.08.\n- CYP1B1 up in 1 of 5, down in 1 of 5.\n\nPLATELET CONFOUND IN CD14:\n- Platelet-gene score, sepsis minus HC: +0.71 to +1.58 in generated test pseudobulks and +0.32 to +1.48 in bulk-only, vs +0.35 (p 0.37) observed.\n- Spearman of CLU with the platelet score: 0.34 to 0.73 in generated pseudobulks vs 0.02 observed.\n- Adjusting for platelet changes the CLU difference from 3.32 to 2.32 (draw 1) and from 1.37 to 0.81 (draw 4).\n- The platelet-in-monocyte state overshoots: ratio 2.96 in draw 1 and 2.55 in draw 4.\n\nGENE-PATTERN CONCORDANCE (Pearson r over the 11 MS1 and S100 genes, generated vs observed on paired test samples):\n- CD14: 0.81, 0.73, 0.61, 0.78, 0.74, vs 0.49 to 0.55 for the bulk.\n- CD16: 0.42, -0.14, -0.13, -0.33, 0.09, vs 0.52 to 0.62 for the bulk.\n- CD16, S100A12: generated +1.5 to +3.3 vs observed about 0.\n- CD16, ALOX5AP: same sign in 0 of 5 draws.\n- CD16, CLU: overshoot of 2.6 to 7.5 times.\n\nBULK ALONE:\n- In bulk-only samples: MS1 +2.17 (t 14.6), S100 +2.59 (t 14.0). After subtracting monocyte lineage genes: +1.22 and +1.64.\n- Error in predicting the observed within-CD14 contrast, MS1: generated 0.11, 0.84, 1.55, 1.04, 1.47 vs raw bulk 0.04 to 0.26.\n- Same error, S100: generated 0.30 to 1.42 vs lineage-corrected bulk 0.08 to 0.16.\n\nHEALTHY CARRIERS (observed HC cells):\n- IL1R2: cDC2 holds 0.67 of the UMIs.\n- CLU: platelets hold 0.70.\n- ALOX5AP and PLAC8: mostly NK and T cells.\n- RETN, CYP1B1, MCEMP1, CD163, S100A8, S100A9, S100A12: CD14 monocytes.\n\nTRAINING DONORS: mean observed MS1-state fraction is 0.29 for draw 0's donors vs 0.06 to 0.12 for draws 1 to 4.", + "issues": [ + "Magnitude is carried by few draws. The closest calibration.csv feature, the MS1 cell state, reproduces the contrast only in draw 0 (ratio 1.12; draws 1 to 4: 0.04 to 0.17). The candidate's PASS window (ratio 0.25 to 4) is lenient: draw 4's ratio of 0.26 counts as PASS. With a 0.5 to 2 window, MS1 in CD14 is carried by draws 0 and 1 (draw 3 borderline at 0.45/0.51) and S100 in CD14 by draws 0, 1 and 3. Draws 2 and 4 keep only the sign.", + "Draws 1 to 4 shift MS1 genes up across the whole pseudobulk but produce almost no MS1 cells (generated state fraction 0.008 to 0.024 vs 0.14 to 0.20 in thinned observed cells). Only draw 0 reproduces MS1 as a cell state. Its five training donors are the most MS1-rich (mean observed MS1 fraction 0.29 vs 0.06 to 0.12), so draw 0's magnitude is probably learned from those donors.", + "The gene list in the claim is overstated. PLAC8 goes down in draw 4 (q<0.05) and is flat in draw 2. MCEMP1 is flat in draw 2, ALOX5AP is flat in draw 4 and S100A12 is flat in draw 4. RETN is not significant in observed CD14 cells (+0.99, p 0.13) and is only about +0.4 in draws 0 to 2. CD163 has the wrong calibration sign in draw 3. Only IL1R2, CLU, CD163 and S100A8 rise at q<0.05 in all 5 draws.", + "CLU is partly a platelet confound. Generated CD14 pseudobulks gain platelet genes in sepsis that observed cells do not show (+0.7 to +1.6 vs +0.35, not significant). Generated CLU tracks the platelet score (Spearman 0.34 to 0.73 vs 0.02 observed), and adjusting for platelets cuts the generated CLU rise by up to about 40%. The platelet-in-monocyte state overshoots about 2.5 to 3 times in draws 1 and 4.", + "The CD16 part fails at the gene level. Across the 11 genes, the generated CD16 pattern correlates with observed at r 0.42, -0.14, -0.13, -0.33, 0.09, while the bulk alone reaches 0.52 to 0.62. Generated S100A12 rises +1.0 to +3.8 in CD16 but observed CD16 shows about 0, so 'S100 in CD16 PASS 5/5' partly rests on this artefact. ALOX5AP has the opposite sign in all 5 draws and CLU overshoots 2.6 to 7.5 times. The genes that truly rise in observed CD16 (IL1R2, MCEMP1, CD163) are attenuated to 0.6 times or less in draws 1 to 4. In intermediate monocytes, the bulk beats the generated gene pattern in 4 of 5 draws.", + "Little is added beyond the bulk. The bulk ranks samples' observed CD14 MS1 better than the generated cells in every draw, and the partial Spearman given bulk is -0.23 to 0.28. Raw bulk log fold change predicts the observed within-CD14 MS1 contrast better than every draw; lineage-corrected bulk does the same for S100. MS1 abundance can already be deconvolved from bulk: Reyes 2020 PDF pp.3-4 (11 cohorts, AUC 0.90) and Leite 2024 p.1 (CIBERSORTx). The only real addition is CD14 gene-level localisation (r 0.61 to 0.81 vs 0.49 to 0.55 for bulk), including IL1R2 (healthy carrier cDC2) and CLU, ALOX5AP and PLAC8 (healthy carriers are platelets, NK or T cells).", + "The finding is not new for this cohort. Ye 2026 PDF p.6 (journal p.155) reports Mono_c03-CD14-RETN: S100A8/S100A12-high, HLA-DR-low, MS1-like, and increased in most individuals with sepsis. PDF p.8 (journal p.157) reports that the MS1 signature (RETN-high, ALOX5AP-high) dominates in Mono_c03. PDF p.14 (journal p.163) repeats this in the Discussion. Reyes 2020 p.2 defines MS1 as CD14+ cells high in RETN, ALOX5AP and IL1R2; p.3 shows it expands in sepsis, with PLAC8 and CLU separating sepsis within MS1 cells; p.4 adds HLA-DR-low IL1R2-high flow cytometry. Yao 2023 p.7 describes HLA-DR-low S100A-high (S100A8, S100A12) and HLA-DR-low CD163-high monocytes.", + "The result is limited to sepsis vs healthy. For adult sepsis vs adult ICU control, the MS1-state calibration has the wrong sign in 5 of 5 draws (generated positive, observed negative). The generated MS1 rise therefore marks critical illness vs health, not sepsis specificity, which fits Reyes 2020 p.3 (sepsis vs ICU without sepsis, FDR 0.27) and Leite 2024 p.2.", + "Minor accuracy points. The claimed 'Mann-Whitney p<=0.003' is slightly off: the maximum is 0.0034 (S100 in CD14, draw 2) and 0.0037 (S100 in intermediate, draw 2). 'BH q<0.01 for all' depends on the test family; a wider family gives q 0.016 for S100 in CD14, draw 2. The edgeR-style prior shrinks the shallower generated pseudobulks more. With log2(CPM+1), ratios rise modestly but the conclusions do not change." + ], + "reasoning": "Verdict: qualified. The direction holds across the calibration, the observed cells and the literature. The magnitude, several of the named genes and the CD16 part do not hold, and the generated cells add little beyond the bulk.\n\nI recomputed everything independently (scripts v00 to v05, CPU only, under systemd-run with MemoryMax=6G and swap off). My numbers match sep_programs exactly.\n\n(1) Calibration\n- Sign: the paired test contrast keeps the sign in 5 of 5 draws for MS1 and S100 in all three monocyte types. The calibration.csv MS1-state row keeps it too (S vs HC 5 of 5).\n- Magnitude: the trust-layer row (MS1 state) gives contrast ratios of 1.12 for draw 0 and 0.04 to 0.17 for draws 1 to 4. Under DATA_GUIDE rule 2, only draw 0 carries MS1 as a cell state.\n- At the pseudobulk program level, magnitude is carried by draws 0 and 1 (draw 3 borderline) for MS1 in CD14, and by draws 0, 1 and 3 for S100 in CD14. Draws 2 and 4 keep only the sign. Draws 1 to 4 shift MS1 genes across the whole pseudobulk without making MS1 cells.\n\n(2) Observed paired cells\n- Strong support in CD14 (MS1 +1.93, S100 +1.55, 43 vs 7), and it survives restricting sepsis to the HC hospital.\n- Gene by gene, RETN and CYP1B1 do not rise in observed CD14 cells.\n- In CD16 the generated gene composition contradicts the observed one: S100A12, ALOX5AP and CLU rise in generated cells but not in observed ones, while IL1R2, MCEMP1 and CD163 are under-generated. That part of the claim fails.\n- Generated CLU is partly confounded by a platelet signal in generated monocytes that observed cells do not have.\n\n(3) Biology and novelty\n- Already known, including for this cohort's own paper: Ye 2026 PDF pp.6, 8 and 14 (journal pp.155, 157, 163); Reyes 2020 pp.2 to 4; Yao 2023 p.7. Nothing here is new or contradicted.\n- The effect is not specific to sepsis: the S vs ICU state calibration has the wrong sign in all draws.\n\n(4) Beyond the bulk\n- The bulk shows MS1 +2.17 and S100 +2.59 (t about 14).\n- The bulk ranks samples better than the generated cells, and a simple bulk log fold change predicts the observed within-CD14 contrast at least as well as any draw.\n- MS1 can already be deconvolved from bulk (Reyes 2020, Leite 2024).\n- The only gain is localising the gene pattern to CD14 monocytes in draws 0, 1, 3 and 4.\n\nNo GPU was used. All writes are confined to my verify directory; no paper-archive delta was published by this verifier.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v00_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v01_programs.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v01_programs.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v01_bulkonly_draws.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v01_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v01_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v01_gene_sets.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v02_genes_platelet.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v02_genes_platelet.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v02_gene_table.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v02_clu_platelet.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v03_beyond_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v03_beyond_bulk.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v03_bulk_alone_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v03_bulk_vs_generated_pairedtest.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v03_gene_pattern_concordance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v03_carriers_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v03_monocyte_composition.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v04_states.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v04_states.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v04_states.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v05_training_donors_and_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v05_training_donors_and_summary.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v05_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-2_calib/v05_per_draw_verdict.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/features/ye2026_sepsis/calibration.csv", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Reyes2020_Immune_cell_signature_bacterial_sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Yao2023_HLADRlow_S100Ahigh_monocytes_sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Leite2024_MS1_deconvolution_critically_ill_sepsis.pdf" + ] + } + ] + }, + { + "id": "SEP-POP-1", + "lens": "sep_populations", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 4, + "theme": "MS1 / S100 monocyte program", + "synthesis_strength": "moderate (draws 2-4 clean, draw 0 overshoots); partial MS1 identity", + "reportable_statement": "An MS1-like, HLA-II-low CD14 monocyte state (S100A12, MCEMP1, RETN, ALOX5AP, PLBD1, CLU) is a larger share of monocytes in bulk-only sepsis than HC in draws 0, 2, 3 and 4; draw 0 overshoots 2-3x and draw 1 does not carry it. Outside draw 0 the generated state lacks IL1R2. Not sepsis-specific and nothing beyond bulk.", + "carrying_draws": { + "sepsis vs HC": "2,3,4 cleanly (0.74-1.44x); 0 direction only (2.0-3.3x); 1 none (0.15x)", + "sepsis vs ICU": "none (wrong sign in draws 0-3)" + }, + "honesty": { + "calibration": "Own rows: ratio 2.00/0.46/1.75/1.05/1.16, Spearman 0.37/0.36/0.72/0.55/0.45; matched contrast 3.3/0.15/1.44/0.74/0.95x. Trust-layer MS1 (IL1R2+) row passes only in d0.", + "replication": "Bulk-only q<=3e-4 in 5/5, but draw 1 effect ~1% of monocytes, pooled Fisher p 0.17 and null in independent clustering.", + "observed": "Observed +0.070 (30 vs 7, p 0.0014); ICU controls >= sepsis in observed (0.107 vs 0.070); generated puts ICU below sepsis.", + "beyond_bulk": "None: bulk MS1 genes AUC 0.96-0.97; bulk tracks the observed fraction better in 5/5 draws." + }, + "verdict_reasons": [ + "draw 1 does not carry the state", + "IL1R2 detected in 2-6% of generated c8 cells (d1-d4) vs 42-45% observed", + "observed counterpart holds only ~16% of observed MS1-rule cells", + "draw 0 overshoots; HC arm 3 samples", + "ICU-control direction wrong" + ], + "new_vs_cohort_papers": "Known: Reyes 2020 MS1 (PDF p.2-4); Ye 2026 Mono_c03-CD14-RETN (printed p.155, Fig. 4d incl. ICU controls); Leite 2024.", + "next_steps": [ + "Define the state by an embedding-free, depth-matched rule (HLA-DR-zero S100A12+, IL1R2-free) and calibrate per draw.", + "Fix IL1R2 under-generation (training donors with MS1 cells).", + "More paired ICU controls." + ], + "key_sources": [ + "verify/SEP-POP-1_calib/tables/v07_per_draw_summary.csv", + "verify/SEP-POP-1_recompute/tables/v10_summary.csv", + "verify/SEP-POP-1_calib/tables/v05_obs_ms1_by_cluster.csv", + "verify/SEP-POP-1_calib/tables/v04_bulk_auc.csv" + ] + }, + "finding": { + "id": "SEP-POP-1", + "cohort": "ye2026_sepsis", + "claim": "In bulk-only generated cells, an MS1-like, HLA class II-low CD14 monocyte state (S100A12, MCEMP1, RETN, ALOX5AP, PLAC8, CXCL8, IL1R2) makes up a larger share of monocytes in adult sepsis than in adult healthy controls in 5/5 draws, all of which pass their own calibration (draw 1 only marginally). It has an observed counterpart with the same markers and the same sepsis > HC shift. However, a bulk marker score already shows it, and it does not separate sepsis from ICU controls.", + "kind": "state", + "feature": "Monocyte|r2.0|c8: Leiden r=2.0 inside six-type 'Monocyte' (joint embedding of observed + paired_gen + bulkonly, per-domain standardised). Generated top markers: S100A12, MCEMP1, CTSD, RETN, ALOX5AP, VCAN, CXCL8, PLBD1, PLAC8, SELL, IL1R2. HLA-II signature lfc -0.92 generated / -0.77 observed. Share of monocytes: bulk-only 8.9%, paired_gen 7.0%, observed 6.8%. Subset of r1.0 c3 (98.7% nested).", + "comparison": "Adult sepsis (68) vs adult HC (36), bulk-only, per draw. State fraction = cluster cells / monocytes per sample (200-cell subsample; samples with >=10 monocytes). Also vs all controls (54) and vs adult ICU controls (17).", + "effect_bulkonly": "State fraction, sepsis vs HC: d0 0.228 vs 0.010 (diff +0.217, AUC 0.96); d1 0.0094 vs 0.0017 (+0.008, AUC 0.71); d2 0.175 vs 0.019 (+0.157, AUC 0.96); d3 0.195 vs 0.013 (+0.182, AUC 0.94); d4 0.134 vs 0.019 (+0.115, AUC 0.93). Abundance (cluster / all cells) diff +0.053/+0.003/+0.014/+0.023/+0.029. Vs all controls diff +0.181/+0.006/+0.137/+0.163/+0.101. Vs ICU diff +0.002 to +0.115, q 0.07-0.63 (n.s.).", + "stats": "Two-sided Mann-Whitney U per draw on per-sample fractions. BH across the 143 cluster features (r=1.0 and r=2.0 clusters with >=150 cells, 4 parent types) within each draw x contrast x measure. Sepsis vs HC q = 1.0e-9, 3.1e-4, 1.5e-6, 1.2e-8, 6.0e-10. Dropping the sex_flag samples (HC23, HC30, Res-S34): q <= 4.3e-4 in all draws.", + "draws_same_sign": "5/5 (q<0.05 in 5/5; draw 1 effect ~10x smaller than observed)", + "paired_calibration": "Own calibration rows (new feature, not in calibration.csv; s05_calibration.csv, each draw's 77 paired TEST samples, observed by co-clustering). Ratio 2.00/0.46/1.75/1.05/1.16; Spearman 0.37/0.36/0.72/0.55/0.45. Sepsis-HC diff, generated vs observed: +0.155/+0.071, +0.008/+0.076, +0.085/+0.063, +0.075/+0.070, +0.066/+0.069. Sign agrees 5/5; gate (same sign, Spearman>=0.2, ratio 1/3-3) passes 5/5. The ICU contrast sign fails in draws 0-3. Related calibration.csv row 'MS1 monocyte (IL1R2+)': d0 ratio 0.98 (HC +0.167 generated vs +0.151 observed); d1-d4 ratio 0.03-0.14, sign kept. The IL1R2-gated rule under-calls generated MS1 because IL1R2 is detected in 20-25% of generated MS1-like cells vs 42% of observed cells thinned to generated depth (s08_ms1_detection.csv).", + "paired_observed_concordance": "Yes, for sepsis vs HC. Observed cells (82 samples, co-clustered): adult sepsis vs adult HC diff +0.070 (MWU p=0.0014; 30 vs 7 samples with >=10 monocytes); kNN-mapped view +0.095 (p=0.0006); vs all controls +0.032 (p=0.003). Not for ICU: sepsis vs ICU -0.038 (p=0.93, 30 vs 6), so ICU controls carry as much. Observed counterpart: 227 co-clustered observed cells with markers S100A12, MCEMP1, ALOX5AP, RETN, IL1R2, CLU, PLBD1, RNASE2, HP, CXCL8, CD63. Marker-profile concordance Spearman 0.90; 30/30 top generated markers also up in observed.", + "beyond_bulk": "Little. A bulk score of its 20 markers gives AUC 0.956 for sepsis vs HC in bulk-only samples, equal to the generated fraction (AUC 0.71-0.96). In paired samples the bulk score tracks the observed fraction (Spearman 0.51-0.77) as well as or better than the generated fraction (0.36-0.72). The generated fraction adds information beyond the bulk score only in draws 1-2 (rank-OLS p=0.002, <0.001), not in draws 0, 3, 4. Within adult sepsis: bulk 0.55 vs generated -0.00 to 0.51. Bulk alone already says this (as in Reyes 2020's bulk deconvolution).", + "novelty": "Known. Reyes 2020 Nat Med: MS1 = CD14+ with high RETN, ALOX5AP, IL1R2 (PDF p.2); HLA-DRlo IL1R2hi CD14+ (PDF p.4). Ye 2026 Nat Immunol: Mono_c03-CD14-RETN (S100A8/S100A12, low HLA-DR, MS1-like) increased in sepsis (PDF p.6 = printed p.155, Fig. 4d; external validation PDF p.12 = printed p.161, Fig. 7d). Yao 2023 Mil Med Res: HLA-DRlow S100Ahigh monocytes (PDF p.7). New only as a calibration note: the transcriptome-wide state is reproduced at about the right frequency in draws 2-4 even though the IL1R2-gated rule fails there.", + "strength": "moderate", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s04_profiles_r2.0.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s04_markers_r2.0.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s05_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s05_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s05_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s07_candidate_detail.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s08_ms1_detection.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s09_bulk_derivability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s10_within_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s12_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/s03_parent_embed.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/s05_disease_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/figures/s14_forest.png" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a0e921b37ca3276af", + "verdict": "qualified", + "recomputed": "The numbers reproduce exactly. They were rebuilt from the raw compact files, using the discovery's c8 membership; the parent monocyte set was confirmed identical to mine.\n\n**c8 state fraction, adult sepsis vs adult HC (bulk-only)**\n\n| Draw | Sepsis | HC | AUC | p |\n|---|---|---|---|---|\n| d0 | 0.228 | 0.010 | 0.959 | 4.3e-11 |\n| d1 | 0.0094 | 0.0017 | 0.713 | 5.8e-5 |\n| d2 | 0.175 | 0.018 | 0.959 | 2.1e-7 |\n| d3 | 0.195 | 0.013 | 0.942 | 5.8e-10 |\n| d4 | 0.134 | 0.019 | 0.929 | 3.8e-11 |\n\n- BH over the 143 features gives the same q (q = 1.0e-9, 3.1e-4, 1.5e-6, 1.2e-8, 6.0e-10; largest difference 2e-15).\n- Without the sex_flag samples: largest q is 4.3e-4.\n- vs ICU: state difference 0.115 / 0.002 / 0.106 / 0.108 / 0.069, all n.s.\n\n**Confounder controls (c8)**\n- Rank-INT OLS with HC3 SEs, adjusted for age, sex, site, 4 bulk-QC covariates, monocyte share, monocyte library size, label confidence and log monocyte count: sepsis p = 0.034 / 0.0031 / 0.014 / 8e-7 / 6e-4.\n- Main site only, plus no sex_flag, plus 1:1 age matching: AUC 0.97 / 0.72 / 1.00 / 0.92 / 0.90, all p ≤ 1.8e-4.\n- Males only, females only, age ≥ 50, bulk library above the 10th percentile, and ≥ 20 monocytes per sample: p < 0.05 in all draws (weakest: draw 1 females, p = 0.035).\n- Monocytes with label confidence ≥ 0.9 only: AUC 0.95 / 0.72 / 0.99 / 0.96 / 0.96.\n- Minimum-monocyte rule set to ≥ 1, 5, 10 or 20: the effect holds. One exception: draw 1 with pooled counts, Fisher p = 0.17.\n\n**Independent definitions**\n- Mean per-cell MS1-minus-HLA-II index over all 1,500 cells: AUC 0.96 / 0.98 / 0.91 / 0.94 / 0.97.\n- Index computed after downsampling every cell to 1,000 counts: p ≤ 3.5e-4 in all draws.\n- Monocyte pseudobulk log2(MS1/HLA-II): AUC 0.97 / 0.96 / 0.80 / 0.92 / 0.87.\n- My own Leiden clusters (thinned observed, HVGs from generated cells, kNN-mapped observed), MS1/HLA-low union: AUC 0.95 / 0.48 / 0.88 / 0.88 / 0.82. Draw 1 p = 0.80.\n\n**Paired calibration (c8)**\n- Ratio 2.00 / 0.46 / 1.75 / 1.05 / 1.16.\n- Spearman 0.37 / 0.36 / 0.72 / 0.55 / 0.45.\n- Sepsis−HC difference, generated vs observed: 0.155 vs 0.071; 0.008 vs 0.076; 0.085 vs 0.063; 0.075 vs 0.070; 0.066 vs 0.069.\n- Generated paired contrast MWU p = 0.040 / 0.64 / 0.016 / 0.005 / 0.002.\n- Bootstrap sign agreement 1.00 / 0.68 / 1.00 / 1.00 / 1.00.\n- Generated HC arm: 3 / 6 / 6 / 7 / 7 samples with ≥ 10 monocytes.\n- ICU sign fails in draws 0–3.\n- The continuous indices pass the gate in 5/5 draws: Spearman 0.35–0.62, generated paired contrast p ≤ 0.036.\n- calibration.csv row MS1 (IL1R2+): ratio 0.98 / 0.033 / 0.048 / 0.138 / 0.061.\n\n**IL1R2 detection in generated c8 cells (bulk-only)**\n- 61% / 5% / 2% / 5% / 4% by draw.\n- Observed c8 cells thinned to each draw's depth: 44–45%.\n\n**Observed counterpart (82 paired samples)**\n- Sepsis vs HC 0.070 vs 0 (30 vs 7 samples, p = 0.0014).\n- Removing the 30/227 granulocyte-like cells: p = 0.0021.\n- vs ICU: −0.038, p = 0.93.\n\n**Bulk derivability**\n- Bulk-only AUC: 20-marker score 0.956; MS1 genes alone 0.971.\n- Paired Spearman with the observed fraction: bulk 0.67 / 0.54 / 0.79 / 0.63 / 0.72 vs generated 0.37 / 0.36 / 0.72 / 0.55 / 0.45.", + "issues": [ + "Draw 1 does not carry the discrete-state result. Its effect is about 1% of monocytes. There are 47 c8 cells across 68 sepsis samples and 2 across 36 HC samples. Pooled-count Fisher p = 0.17. Its paired generated sepsis-HC contrast is not significant (p = 0.64), with bootstrap sign agreement 0.68 and 10% of the observed magnitude. My independent Leiden clusterings show no effect (AUC 0.48, p = 0.80). Draw-1 c8 cells carry FOS/JUN and lack IL1R2. Draw 1 passes the claim's gate only nominally; it shows only a graded MS1-up/HLA-II-down program shift (continuous index AUC 0.98, calibrated). So '5/5 draws, all passing calibration' should read 4/5 (draws 0, 2, 3, 4).", + "In draws 1-4 the generated state is only partly MS1-like. IL1R2 is detected in 2-5% of generated c8 cells vs 42-45% of observed c8 cells at matched depth. CXCL8 is weak in draws 3-4 (lfc 0.11-0.33 vs 0.67 observed). S100A12, MCEMP1, RETN, ALOX5AP, PLBD1, CLU up and HLA-II down are reproduced. The claim's pooled IL1R2 figure (20-25%) is dominated by draw 0 (61%). The trust-layer row 'MS1 monocyte (IL1R2+)' passes only in draw 0 (ratio 0.98); draws 1-4 have ratio 0.03-0.14.", + "Draw 0 overshoots: calibration ratio 2.0 and generated contrast 0.155 vs 0.071 observed, consistent with its MS1-rich sepsis training donors. Its generated paired HC arm has only 3 samples with at least 10 monocytes.", + "Every paired calibration and observed contrast rests on 6-7 adult HC and 5-6 adult ICU controls. All 7 observed HC have zero c8 cells, so the observed difference is simply the sepsis mean.", + "The ≥10-monocyte rule drops samples unevenly. For example, draw 2 keeps 26 of 68 sepsis and 18 of 36 HC samples. This does not change the conclusion: the effect holds with all-cell measures (98-104 samples per draw) and with thresholds of 1, 5 or 20.", + "The effect depends on the definition, though only for draw 1. My cluster closest to c8 at r2.0 overlaps it by only 27-47%. The most extreme IL1R2-high cluster is almost entirely draw 0 (917 of 1,003 bulk-only cells).", + "Bulk alone already shows it. Bulk-only AUC is 0.956 for the 20-marker score and 0.971 for MS1 genes alone. In paired samples the bulk score tracks the observed c8 fraction better than the generated fraction in 5/5 draws. With my bulk score, the generated fraction adds beyond bulk at p<0.05 in draws 0-2, not in draws 3-4 (the discovery reported draws 1-2 only). Logistic 'generated beyond bulk' tests are unstable because the bulk score nearly separates the groups.", + "There is no separation from ICU controls, and the generated ICU-contrast sign is wrong vs observed in draws 0-3. ICU controls are also 16 years younger (48.9 vs 65.2), so age confounds that contrast.", + "Bulk technical metrics differ between sepsis and HC: assigned reads AUC 0.33, no-feature fraction 0.26, multimapping 0.71, MT fraction 0.79. Adjusting for them does not remove the effect. Collection wave or sequencing batch cannot be tested: GSM numbers follow sample names, and sample-ID numbering is separate by group.", + "Confounders that do not apply or show nothing: repeated measures (donor_id is unique, one sample per donor), training-donor overlap (no bulk-only donor trained any draw, and test sets exclude training donors), and gating leakage (at least 90% of MS1-high myeloid cells fall inside the Monocyte gate in every group and draw)." + ], + "reasoning": "I recomputed the claim with my own code from the raw compact files. The sepsis > HC shift in the c8 state reproduces exactly, including the BH q-values and the calibration rows, and it survives every reasonable confounder control I tried: age, sex, site, sex_flag, bulk library size and quality, generated monocyte share, library size and label confidence, monocyte count, depth downsampling, a label-confidence filter, gating leakage, and granulocyte-like cells in the observed counterpart. It is therefore not refuted.\n\nIt is not fully confirmed either.\n- **Draw 1:** the discrete state is not supported. The effect is tiny, the paired generated contrast is not significant, and independent clusterings show nothing. Draw 1 shows only a continuous MS1/HLA-II program shift.\n- **IL1R2:** in draws 1–4 the generated state lacks IL1R2 (2–5% detection vs about 44% observed at matched depth). The official IL1R2-gated calibration row fails there, so 'MS1-like' is only partial outside draw 0.\n- **Draw 0:** it overshoots about 2x and its generated HC calibration arm has 3 samples.\n- **Calibration size:** every paired calibration contrast rests on 6–7 HC.\n\n**Which draws carry the result:** draws 2, 3 and 4 cleanly (ratio 1.05–1.75, Spearman 0.45–0.72, generated paired contrast p 0.002–0.016). Draw 0 also carries it, with overshoot. Draw 1 does not carry the discrete state.\n\n**The four honesty criteria:**\n1. Calibration: adequate in draws 0 and 2–4 for c8; the IL1R2 rule is calibrated in draw 0 only.\n2. Replication across draws: 4/5 robust.\n3. Observed agreement: yes vs HC (n = 7 HC); no vs ICU.\n4. Beyond bulk: little. Bulk AUC is 0.956–0.971, and bulk tracks the observed fraction better than the generated fraction.\n\nNo GPU work was started. I wrote only under my lens dir. There is no paper-archive change from this verifier; the coordinator owns that step.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/vcommon.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v01_meta_arith.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v02_own_embedding.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v03_measures.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v04_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v05_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v06_bulk_derivability.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v07_identity_gating.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v08_extra_checks.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v09_per_draw_profile.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/v10_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v10_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v01_arith_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v01_meta_checks.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v01_covariate_balance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v02_choice.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v02_cluster_profiles.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v03_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v03_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v03_per_sample.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v04_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v04_subsets.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v04_covariate_balance_by_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v04_within_group_cov_corr.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v05_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v05_calibration_csv_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v06_auc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v06_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v06_paired.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v07_identity_detection.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v07_obs_c8_granulocyte_like.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v07_gating_leak.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v08_minparent.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v08_draw1_anatomy.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v08_bh_recheck.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v08_c8_marker_detection.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v08_id_order.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v09_claimed_marker_lfc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/tables/v09_top_up_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/figures/v10_summary.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_recompute/logs" + ] + }, + { + "verifier": "calib", + "agent_id": "ac0c7d8a4f4f01e55", + "verdict": "qualified", + "recomputed": "Matched paired-test calibration of Monocyte|r2.0|c8 (samples with >=10 monocytes in both generated and observed, co-clustered view), draws 0/1/2/3/4:\n- Level ratio: 2.00 / 0.46 / 1.75 / 1.05 / 1.16.\n- Spearman over test samples: 0.37 / 0.36 / 0.72 / 0.55 / 0.45. Within adult sepsis only: 0.14 / 0.39 / 0.51 / 0.18 / -0.00.\n- Adult sepsis - adult HC, generated vs observed: +0.172 vs +0.052 (3.3x); +0.012 vs +0.076 (0.15x; generated 95% CI -0.027 to +0.045); +0.086 vs +0.060 (1.44x); +0.056 vs +0.076 (0.74x); +0.065 vs +0.068 (0.95x).\n- Same contrast ratios with the kNN-mapped observed view: 2.04 / 0.11 / 1.11 / 0.60 / 0.69.\n\nObserved, all 82 paired samples:\n- Sepsis vs HC: +0.070 (95% CI 0.046 to 0.096), p=0.0014. All 7 HC samples have zero c8 cells; 23 of 30 sepsis samples have some.\n- ICU vs HC: +0.107, p=0.0056.\n- Sepsis vs ICU: -0.038, p=0.93.\n\nBulk-only, generated:\n- Sepsis vs HC holds whatever the minimum monocyte count per sample (1, 5, 10 or 20) and under a pooled binomial model, in all draws. The draw 1 effect is tiny (+0.004 to +0.010).\n- Sepsis - ICU: +0.115 / +0.002 / +0.106 / +0.108 / +0.069 (AUC 0.74 / 0.56 / 0.80 / 0.75 / 0.76; nominal p 0.008 / 0.39 / 0.009 / 0.020 / 0.007).\n- The r1.0 parent cluster c3 reverses in draw 1 (-0.031, AUC 0.23).\n- Paired-generated sepsis means are 0.155 / 0.034 / 0.110 / 0.094 / 0.084, against observed 0.071 / 0.076 / 0.063 / 0.070 / 0.069 on the same samples.\n\nIL1R2 detection inside c8:\n- Generated bulk-only: 0.61 / 0.05 / 0.02 / 0.05 / 0.04.\n- Generated paired-test: 0.69 / 0.04 / 0.02 / 0.04 / 0.06.\n- Observed c8, thinned to each draw's depth: 0.42 to 0.44.\n- IL1R2 log-fold difference (c8 vs other monocytes), bulk-only: 1.25 / 0.11 / 0.03 / 0.06 / 0.05; observed 0.90. HP: 0.03 / 0.17 / 0.06 / 0.37 / 0.41; observed 0.74.\n- Per-draw marker concordance with the observed c8 (Spearman): 0.85 / 0.89 / 0.92 / 0.91 / 0.90.\n\nWhere the observed MS1 cells sit:\n- Only 15.6% of observed MS1-rule cells co-cluster in c8. An equal share is in the low-library cluster c0; 29% are in the r1.0 parent c3.\n- Per sample, observed c8 fraction vs MS1-rule fraction: Spearman 0.46 (n=60). Res-S150: 84% MS1-rule monocytes, 2% c8.\n\nEmbedding-free calibration (all cells, observed thinned), sepsis - HC contrast ratio per draw:\n- IL1R2-free rule (>=3 of RETN, ALOX5AP, CYP1B1, MCEMP1, PLAC8, CLU, CD163, with HLA-DRA undetected): 0.47 / 0.25 / 0.30 / 0.70 / 0.40. Level ratio 0.38 to 0.91.\n- Union of c0 and c8: 1.25 / 0.07 / 0.70 / 0.62 / 0.52.\n- Trust-layer IL1R2 rule: 1.09 / 0.04 / 0.06 / 0.14 / 0.09.\n\nBulk alone:\n- Bulk-only, sepsis vs HC AUC: the claim's 20 markers 0.956; core MS1 genes 0.958; MS1 minus HLA-II 0.938.\n- Bulk-only, sepsis vs ICU AUC: 0.64 / 0.66 / 0.71.\n- Paired samples, sepsis vs HC AUC 0.91 / 0.93 / 0.99; sepsis vs ICU 0.54 to 0.64.\n- Partial Spearman, generated vs observed given the bulk score, within adult sepsis: 0.00 / 0.46 / 0.09 / 0.19 / -0.14. Only draw 1 is significant (p=0.017).\n- Bulk-only sepsis samples have higher bulk MS1 scores than paired sepsis samples (core MS1 genes +0.26 z, p=0.038).", + "issues": [ + "Draw 1 does not reproduce the paired contrast. On matched test samples the generated sepsis - HC difference is +0.012 vs +0.076 observed (0.15x), and the generated 95% CI (-0.027 to +0.045) includes 0. In draw 1 bulk-only data the r1.0 parent cluster c3 even reverses (-0.031, AUC 0.23). The claim's gate checks sign, Spearman and level ratio, but not the size of the contrast, so '5/5 pass' is too generous. Under the DATA_GUIDE rule the draws that carry the result are 2, 3 and 4 (contrast 0.74x to 1.44x). Draw 0 counts for direction only: it over-calls the contrast 2.0x (kNN-mapped) to 3.3x (co-clustered) and the level 2x.", + "The IL1R2 explanation hides a per-draw split. The claim's figure (IL1R2 in 20-25% of generated MS1-like cells vs 42% observed) is pooled over draws and comes almost entirely from draw 0, where 61-69% of c8 cells have IL1R2. In draws 1-4 only 2-6% do, against 42-44% of observed c8 cells thinned to the same depth. HP is also low. So the marker list (which includes IL1R2) and 'same markers as the observed counterpart' hold only for draw 0. In draws 1-4 the state is an IL1R2-negative, S100A12/MCEMP1/RETN/ALOX5AP-high, HLA-II-low CD14 monocyte. Its other markers still agree with the observed cells (Spearman 0.85-0.92).", + "The observed counterpart is only part of the observed MS1 population. Just 15.6% of observed MS1-rule cells co-cluster in c8; as many sit in the low-library cluster c0, and per sample c8 tracks the MS1-rule fraction at Spearman 0.46 (Res-S150: 84% MS1-rule monocytes, 2% c8). So 'reproduced at about the right frequency in draws 2-4' holds only against this part. Without the joint embedding, an IL1R2-free rule (>=3 MS1 genes, HLA-DRA undetected) keeps the sign in all 5 draws but reproduces only 0.25-0.70x of the observed sepsis - HC contrast and 0.38-0.91x of the level.", + "ICU controls go the wrong way. In the observed cells, adult ICU controls carry at least as much of the state as sepsis (0.107 vs 0.070; ICU vs HC +0.107, p=0.0056). Ye 2026 Fig. 4d shows the same (Mono_c03 up in ICU controls vs HC, P<0.01, no sepsis-vs-ICU difference). The generated bulk-only cells instead put ICU controls below sepsis in draws 0, 2, 3 and 4 (+0.07 to +0.11, AUC 0.74-0.80, nominal p 0.007-0.02). On matched paired samples the sepsis vs ICU sign is wrong in draws 0-2, because generated sepsis levels are over-called 1.2-2.2x. 'Does not separate sepsis from ICU controls' understates this: the generated fraction partly tracks how sepsis-like the bulk looks. It must not be used to separate sepsis from non-septic critical illness.", + "The bulk-only effect sizes (+0.115 to +0.217 in the draws that carry the result) are 1.6-3x the observed paired contrast (+0.070), so the absolute differences should not be quoted as estimates of the true expansion. Part of the gap comes from the bulk itself: bulk-only sepsis samples have higher bulk MS1 scores than paired sepsis samples (+0.26 z, p=0.038).", + "Nothing beyond bulk for this contrast. Bulk MS1 scores separate sepsis from HC at least as well (AUC 0.956-0.958) as the generated fraction (0.71-0.96). Within adult sepsis, the generated fraction tracks the observed fraction beyond the bulk score only in draw 1 (partial rho 0.46, p=0.017), and draw 1 is the one that fails calibration.", + "Not new. The cohort paper itself reports it: Ye 2026 PDF p.6 (printed p.155) and Fig. 4d (PDF p.9, printed p.158), including the rise in ICU controls. Ye 2026 also read this state from its own bulk RNA-seq (n=207, which includes the bulk-only samples): Fig. 4e survival by bulk signature (P<0.0001); methods PDF p.19 (bulk mRNA-seq of 174 adults) and p.19-20 (CIBERSORTx on the in-house PBMC bulk). Two critical-illness references the claim did not cite: Reyes 2020 PDF p.3-4 (MS1 only slightly higher in sepsis than in ICU-NoSEP, FDR 0.27; 'the MS1 fraction alone cannot be used as a sepsis classifier in this context') and Leite 2024 PDF p.1-2 (MS1 from bulk 'highly similar' in non-infectious ICU patients).", + "Minor, outside this claim: in sep_populations/tables/s11_training_donor_imprint.csv the train_groups column is not in the same order as train_donors (for example, in draw 3 PHC05, a paediatric HC, is listed as adult_sepsis)." + ], + "reasoning": "Verdict: qualified. Not refuted, because the observed paired cells and the literature support a sepsis > HC rise of an MS1-like, HLA-II-low CD14 monocyte state, and at least 3 draws reproduce that contrast in their own paired test. Not confirmed, because several specific statements in the claim fail.\n\n1. Calibration. The feature is not in calibration.csv, so I redid the paired check independently on matched test samples (v01).\n- Draws that reproduce the sepsis - HC contrast: 2, 3 and 4 (1.44x, 0.74x, 0.95x observed; kNN view 1.11x, 0.60x, 0.69x). Level ratios 1.05-1.75; Spearman 0.45-0.72.\n- Draw 0 reproduces the direction but over-calls the contrast 2.0-3.3x and the level 2x. Its training donors are MS1-rich: Res-S150 is 82% MS1-rule cells, Abd-S66 38%, Res-S77 24% (v04).\n- Draw 1 fails: 0.15x observed, CI includes zero, and its r1.0 parent cluster reverses in bulk-only cells.\n- The trust-layer row 'MS1 monocyte (IL1R2+)' fails in draws 1-4 because those draws' MS1-like cells almost never express IL1R2 (2-6% vs 42-44% observed), not because of a moderate under-call. Only draw 0 makes IL1R2+ MS1 cells (v02, v02b).\n- Without the joint embedding, IL1R2-free MS1-gene rules keep the sign in all draws but give 0.25-0.70x of the observed contrast (v03).\n- The observed 'counterpart' holds only about 16% of observed MS1-rule cells. So the claimed level match is against that part of the observed MS1 cells, not all of them (v05, v06).\n\n2. Observed support.\n- Yes for sepsis vs HC: +0.070, p=0.0014 co-clustered; +0.095 kNN-mapped. All IL1R2-free marker-rule definitions agree.\n- But observed adult ICU controls are as high or higher than sepsis (ICU vs HC p=0.0056). The generated cells place ICU controls below sepsis because they over-call sepsis (v04 group levels).\n\n3. Biology and novelty.\n- Identity is right. The observed c8 cells express CD14, LYZ, FCN1 and VCAN and lack FCGR3B, CXCR2 and MME, so they are monocytes, not low-density granulocytes. Some ELANE, MPO and S100P are present, which fits MS1's origin in emergency myelopoiesis.\n- Already reported. Ye 2026 PDF p.6 (printed p.155) describes Mono_c03-CD14-RETN (S100A8/S100A12-high, HLA-DR-low, 'resembling MS1-like/HLA-DRloS100Ahi') as increased in sepsis. Fig. 4d on PDF p.9 (printed p.158) shows it also increased in ICU controls vs HC, with no difference between sepsis and ICU controls. The external replication is on PDF p.12 (printed p.161), Fig. 7d.\n- Reyes 2020: MS1 defined as CD14+ cells high in RETN, ALOX5AP and IL1R2 (PDF p.2); sepsis vs ICU-NoSEP FDR 0.27 (p.3); bulk deconvolution, and the HLA-DRlo IL1R2hi CD14+ flow gate (p.4).\n- Leite 2024 PDF p.1-2: MS1 estimated from bulk is similar in non-infectious ICU patients. Yao 2023 PDF p.7: HLA-DRlow S100Ahigh monocytes.\n\n4. Beyond bulk: nothing for this contrast.\n- Bulk scores give AUC about 0.96 for sepsis vs HC. Ye 2026 already used this cohort's bulk (which includes the bulk-only samples) to read out the state, by bulk signature for survival (Fig. 4e) and by CIBERSORTx deconvolution.\n- The generated fraction adds information within sepsis only in draw 1, which is uncalibrated.\n\nWhat the evidence supports: in draws 2-4, with draw 0 for direction only, the generated cells reproduce the known sepsis > HC rise of an MS1-like, HLA-II-low CD14 monocyte state. That state lacks IL1R2 except in draw 0. It adds nothing to the bulk, and it should not be used for sepsis vs critical illness, where it gets the ICU-control direction wrong.\n\nAll work was CPU-only, inside systemd-run scopes with MemoryMax=6G and swap off. MemAvailable was checked to be >=48 GiB before loading the subsample counts. I wrote only to my verify directory; there is no archive delta from this verifier, and archiving is left to the coordinator.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/v01_recompute_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/v02_identity_markers.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/v02b_per_draw_concordance.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/v03_rule_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/v04_bulk_and_groups.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/v05_obs_ms1_in_clusters.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/v06_ms1_compartment.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/v07_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v07_per_draw_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v01_calibration_matched.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v01_observed_contrasts.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v01_bulkonly_robust.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v02_identity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v02_rule_hits.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v02b_per_draw_concordance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v02b_lfc_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v03_rule_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v03_rule_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v03_rule_specificity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v04_bulk_auc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v04_bulk_joint_levels.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v04_bulk_joint_set_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v04_partial.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v04_group_levels.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v04_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v05_obs_ms1_by_cluster.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v05_obs_sample_c8_vs_rule.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/tables/v06_ms1_compartment.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/figures/v07_calibration.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/lit/ye2026_fig4d_crop.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/lit/ye2026_raw.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/lit/reyes2020_raw.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/lit/leite2024_raw.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/lit/yao2023_raw.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-1_calib/lit/v00_pdf_text.sh" + ] + } + ] + }, + { + "id": "SEP-POP-6", + "lens": "sep_populations", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 5, + "theme": "MS1 / S100 monocyte program", + "synthesis_strength": "weak as a separate population (artefact-suspect); its disease signal is the SEP-POP-1 state", + "reportable_statement": "The low-library S100A12+ HLA-DRA-low monocyte cluster is not granulocyte-like: at matched depth its observed members lack neutrophil genes (FCGR3B 0.6% vs 9.1% in other monocytes) and lean MS1. Its sepsis > HC signal survives library adjustment and is the HLA-II-low S100A-high monocyte shift (draws 0, 2, 3, 4 on an embedding-free rule). As a population it is defined partly by low library size and is not recovered by independent clustering.", + "carrying_draws": { + "c0 cluster": "0,3 (significant); gate passes 0,2,3", + "HLA-DR-zero S100A12+ rule / all-cell proxy": "0,2,3,4" + }, + "honesty": { + "calibration": "c0: ratio 0.64/0.50/0.81/1.05/0.72, Spearman 0.11-0.33, contrast 0.14-0.71x; HLA-DR-zero rule Spearman 0.54-0.59 in d0,d2-d4, -0.01 in d1.", + "replication": "5/5 positive; q<0.05 in 2/5 (subsample) and 5/5 on all cells.", + "observed": "Observed +0.117 (30 vs 7, p 1.4e-4; all 7 HC zero); vs ICU +0.044 (p 0.24).", + "beyond_bulk": "None: bulk S100A12-minus-HLA-DRA index AUC 0.98; generated share adds nothing once bulk is in the model (p>=0.23)." + }, + "verdict_reasons": [ + "granulocyte/LDG reading contradicted at matched depth", + "library size defines membership, not the disease contrast", + "CSF3R is lower in c0, not a marker", + "concordance 0.39 was metric-dependent (0.70-0.84 on a broad depth-matched set)", + "cluster boundary unstable (Jaccard <=0.35)" + ], + "new_vs_cohort_papers": "Known state: Ye 2026 Mono_c03-CD14-RETN; Reyes 2020 MS1; Schulte-Schrepping 2020 HLA-DRlo S100Ahi.", + "next_steps": [ + "Drop the separate population; report the HLA-DR-zero S100A12+ share with SEP-POP-1." + ], + "key_sources": [ + "verify/SEP-POP-6_recompute/tables/v09_key_numbers.csv", + "verify/SEP-POP-6_calib/tables/v08_key_numbers.csv", + "verify/SEP-POP-6_calib/tables/v02_gene_detection.csv", + "verify/SEP-POP-6_calib/tables/v05_hla0_calibration.csv" + ] + }, + "finding": { + "id": "SEP-POP-6", + "cohort": "ye2026_sepsis", + "claim": "A low-library S100A12+ HLA-DRA-low monocyte cluster is higher in sepsis in bulk-only draws 0 and 3 and in observed cells. It is library-size driven in both domains, and its observed members carry granulocyte genes (IL1R2, HP, S100P, CSF3R) that the generated members lack, so the generated and observed cells are not the same population. Artefact-suspect.", + "kind": "population", + "feature": "Monocyte|r2.0|c0. Generated median 827 counts / 185 genes vs 2,223 / 351 for the rest of monocytes (AUC of log counts 0.18); observed 1,717 vs 3,073 (AUC 0.29). Share of monocytes: bulk-only 7.2%, paired_gen 6.9%, observed 9.4%. Generated markers: only S100A12; HLA-DRA lfc -1.47.", + "comparison": "Adult sepsis vs adult HC, bulk-only, per draw; state fraction of monocytes.", + "effect_bulkonly": "Sepsis vs HC: d0 0.092 vs 0.038 (+0.054); d1 +0.001; d2 +0.024; d3 0.168 vs 0.091 (+0.077); d4 +0.006.", + "stats": "MWU per draw; BH over 143 features. q = 4e-4, 0.53, 0.056, 0.004, 0.83.", + "draws_same_sign": "5/5 positive (q<0.05 in 2/5)", + "paired_calibration": "Own rows. Ratio 0.64/0.50/0.81/1.05/0.72; Spearman 0.33/0.11/0.29/0.32/0.12. Sepsis-HC generated vs observed +0.085/+0.120, +0.018/+0.123, +0.020/+0.122, +0.071/+0.120, +0.024/+0.120. Gate passes d0, d2, d3; generated contrast is 0.15-0.7x observed.", + "paired_observed_concordance": "Observed adult sepsis vs HC +0.117 (p=0.0001); vs all controls +0.072 (p=0.0012); vs ICU +0.044 (p=0.24). Marker concordance is only 0.39, the lowest of all populated clusters. Observed members detect IL1R2 35% vs 5% generated, HP 29% vs 7%, S100P 16% vs 1%, CSF3R 48% vs 12-15%. The observed side looks like low-RNA granulocyte-like cells, which the generator does not make.", + "beyond_bulk": "Not evaluated as a population, because the cluster is library-size driven and its two domains differ in composition.", + "novelty": "The observed side resembles IL1R2+ immature neutrophils / low-density granulocytes (Kwok 2023 Nat Immunol; Schulte-Schrepping 2020 Cell), which this generator cannot produce (calibration.csv 'Granulocyte-like (LDG)' ratio 0.01-0.09). Not checked further.", + "strength": "artifact-suspect", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s04_profiles_r2.0.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s07_candidate_detail.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s13_marker_checks.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s14_artefact_clusters.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a1d94868f01573734", + "verdict": "qualified", + "recomputed": "I rebuilt every number from the raw subsample.npz, cells.parquet and meta.csv files with my own code. The discovery Leiden labels (s03_cells_Monocyte.parquet) were used only to say which cells are in the cluster; all 40,541 monocytes matched the raw files. The BH correction ran over the 143 features.\n\nREPRODUCTION (exact)\n- Bulk-only, adult sepsis vs adult HC, state fraction (c0 cells / monocytes):\n - d0: 0.0921 vs 0.0380, +0.054, p 1.4e-4, q 4.3e-4, n 59 vs 24\n - d1: +0.001, q 0.53\n - d2: +0.024, q 0.056, n 26 vs 18\n - d3: 0.1676 vs 0.0909, +0.077, p 1.0e-3, q 0.0041, n 33 vs 30\n - d4: +0.006, q 0.83\n - 5/5 positive; q < 0.05 in d0 and d3.\n- Own calibration rows, d0/d1/d2/d3/d4:\n - ratio 0.64/0.50/0.81/1.05/0.72\n - Spearman 0.33/0.11/0.29/0.32/0.12\n - sepsis − HC, generated vs observed: +0.085/+0.120, +0.018/+0.123, +0.020/+0.122, +0.071/+0.120, +0.024/+0.120\n - gate passes d0, d2, d3; the generated contrast is 0.14–0.71× observed.\n- Observed cells:\n - adult sepsis vs adult HC: 0.117 vs 0.000, p 1.4e-4. All 7 HC have zero c0 cells among 168 monocytes.\n - vs ICU controls: +0.044, p 0.24. Vs all controls: +0.072, p 0.0012.\n- Depth:\n - generated: 826 vs 2,223 counts and 185 vs 351 genes (members vs rest); AUC of log counts 0.185\n - observed: 1,717 vs 3,073 counts; AUC 0.290\n - share of monocytes: bulk-only 7.2%, paired_gen 6.9%, observed 9.4%\n- Raw marker detection, observed members vs bulk-only generated members:\n - IL1R2 0.350 vs 0.048\n - HP 0.293 vs 0.074\n - S100P 0.161 vs 0.014\n - CSF3R 0.479 vs 0.124\n\nCONFOUNDERS (the d0 and d3 effects survive all of them)\n- Age and sex are balanced: p 0.32 and 0.61.\n- Site, First Affiliated Hospital only (all HC were collected there):\n - d0 +0.049, p 4e-4\n - d3 +0.068, p 0.0015\n - observed +0.127, p 2e-4\n- sex_flag samples dropped: q 9e-4 (d0) and 0.0097 (d3).\n- Minimum-monocyte threshold set to 1, 5, 10, 20 or 30: d0, d3 and observed are significant at every threshold (d0 p 7e-5 to 0.019; d3 p 1e-4 to 0.018).\n- Quasi-binomial GLM adjusted for age, sex and log median library:\n - d0 logOR 0.92, p 3e-5\n - d3 logOR 0.99, p 1e-10\n - d4 logOR 0.37, p 0.005\n - d2 logOR 0.53, p 0.049\n- Label confidence ≥ 0.9 / ≥ 0.99: d0 p 1.5e-4 / 0.0067; d3 p 0.0071 / 0.012. Members have fewer low-confidence labels than non-members.\n\nLIBRARY SIZE\n- Monocyte library size is the same in sepsis and HC: AUC 0.51 (d0), 0.55 (d3), 0.53 (observed).\n- Cell-level GEE with a library spline: d0 logOR 0.80 (p 1.8e-4), d3 1.04 (p 8e-6), d4 0.51 (p 0.0013).\n- The share of the contrast predicted by library size alone:\n - d0: +0.022 of +0.054\n - d3: −0.006 of +0.077\n - observed: −0.015 of +0.117\n- Inside a library-matched window (sepsis vs HC library AUC 0.41–0.53), the c0 rate is:\n - d0: 14.4% vs 5.9%\n - d3: 21.3% vs 8.9%\n - observed: 13.7% vs 0%\n- Low library alone carries no sepsis signal: observed −0.053 (p 0.29), d3 +0.002, d4 −0.054. Its calibration fails in every draw (Spearman ≤ 0).\n- HLA-II-negative (HLA-DRA + HLA-DRB1 = 0) S100A12+ monocytes at any depth:\n - bulk-only +0.134/+0.058/+0.128/+0.243/+0.140 (d0–d4), all p ≤ 2e-4; still p < 1e-5 after library adjustment\n - observed +0.151, p 2e-4\n - own calibration gate passes d0, d2, d3 and d4 (Spearman 0.44/0.16/0.34/0.33/0.21).\n\nMARKERS, DEPTH-MATCHED (observed cells binomially thinned to generated depth)\n- IL1R2: observed members 0.275 vs bulk-only pooled 0.048. But draw 0 makes it: bulk-only d0 0.204; on the same samples, paired_gen d0 0.277 vs observed 0.267.\n- HP: 0.214 vs 0.074. Draws 3 and 4 make 0.117 and 0.148; same samples d3 0.127 vs 0.215.\n- S100P: 0.098 vs 0.014.\n- CSF3R: 0.330 vs 0.124. It is lower in members than in the rest in both domains (observed 0.33 vs 0.59; generated 0.12 vs 0.36).\n- Neutrophil-specific genes in observed members:\n - FCGR3B 0.006 (rest 0.091), CXCR2 0.006, MMP9 0.006, CAMP 0, LCN2 0\n - ≥ 2 of 14 neutrophil genes: 6.1% of members vs 15.2% of the rest (thinned: 2.6% vs 11.7%)\n- MS1 rule: 32.8% of members vs 18.2% of the rest.\n- Marker concordance measured on a fixed gene set and at matched depth: c0 scores 0.70, ranked 9 of 20 clusters (range 0.49–0.90).\n\nINDEPENDENT RE-DERIVATION\n- My own embedding and Leiden clustering:\n - the best-matching clusters overlap c0 with Jaccard 0.21–0.35\n - their observed sepsis − HC difference ranges from +0.023 (p 0.88) to +0.133 (p 0.016)\n - bulk-only is positive in d0, d2 and d3 at every resolution.\n\nBULK ALONE\n- In the bulk-only samples, bulk HLA-DRA has AUC 0.18, S100A12 0.98, and the S100A12 − HLA-DRA index 0.98 for sepsis vs HC.\n- Within sepsis, generated c0 does not follow the bulk index in d0 (rho −0.05); it does in d3 (rho 0.37).", + "issues": [ + "The library-size explanation is half right. Cells in c0 are low-library (AUC 0.185 generated, 0.290 observed), but the sepsis vs HC difference is not caused by library size. Monocyte library is the same in sepsis and HC (AUC 0.51–0.55). The difference stays after library-spline GEE (d0 logOR 0.80, d3 1.04) and inside a library-matched window. Library size alone predicts about 0 of the contrast in d3 (−0.006 of +0.077) and in observed cells (−0.015 of +0.117). Low library alone has no sepsis association in observed cells (−0.053) or in d3/d4.", + "The sepsis signal comes from the HLA-II-low S100A12+ phenotype, not from low depth. HLA-II-negative S100A12+ monocytes at any depth are higher in sepsis in all 5 draws and in observed cells (+0.151, p 2e-4), survive library adjustment, and pass their own calibration gate in d0, d2, d3 and d4. This is the known HLA-DR-low monocyte shift in sepsis.", + "The granulocyte / LDG reading is not supported. Observed c0 members essentially lack neutrophil-specific genes (FCGR3B 0.6% vs 9.1% in the rest; CXCR2, MMP9, CAMP, LCN2, CYP4F3 and MME at 0–0.6%). Only 6.1% of members have ≥ 2 of 14 neutrophil genes, vs 15.2% of the rest. Members are instead enriched for the MS1 rule (32.8% vs 18.2%) and for hemoglobin (HBB). The novelty statement linking them to IL1R2+ immature neutrophils / LDG (Kwok 2023, Schulte-Schrepping 2020) is therefore not supported.", + "The claim compared marker detection without thinning, against the DATA_GUIDE depth rule. After thinning observed cells to generated depth, the gaps shrink: IL1R2 35%→27.5%, HP 29%→21%, S100P 16%→10%, CSF3R 48%→33%.", + "'Generated members lack IL1R2' is false for draw 0, one of the two draws that carry the result. On the same samples, paired_gen d0 members detect IL1R2 at 0.277 vs 0.267 observed. The IL1R2 gap follows each draw's overall MS1 calibration (official MS1 ratio 0.98 in d0 vs 0.03–0.14 in d1–d4). HP is partly made by d3 and d4.", + "CSF3R is not a granulocyte marker of c0. It is lower in members than in the rest in both domains, and the generator under-produces it in all monocytes.", + "The 'concordance 0.39, lowest of all clusters' result depends on the metric. On a fixed gene set at matched depth, c0 scores 0.70 and ranks 9 of 20. The two domains share the core phenotype: HLA-II down 1.0–1.6, S100A8/9/12 up, FCN1/CD14/FCGR3A/NAMPT down, and low library. So 'not the same population' is overstated; the generator mainly misses the MS1 and hemoglobin components.", + "The c0 boundary depends on the embedding. Independent clusterings recover it only with Jaccard ≤ 0.35, and the observed effect ranges from +0.02 (p 0.88) to +0.13 depending on the cut. The artefact-suspect label is fair for c0 as a separate population, but not for the disease signal it carries.", + "Calibration is weak. Spearman is ≤ 0.33. In d0 the paired_gen HC group has only 3 samples with ≥ 10 monocytes. The d2 gate pass depends on the estimator (paired_gen d2 GLM logOR −0.20). The generated contrast is 0.14–0.71× observed. Only d0 and d3 carry a reportable result (d2 q 0.056). d4 becomes positive after library adjustment, but it fails the gate.", + "Selection by the ≥ 10 monocyte rule is heavy: it drops 35/68 sepsis samples in d3 and 42/68 in d2. Thresholds from 1 to 30 do not change the d0, d3 or observed conclusions.", + "The signal separates critical illness from health, not sepsis from ICU controls. Observed sepsis vs ICU is +0.044 (p 0.24). Bulk-only sepsis vs ICU is not significant in any draw, and the paired sepsis-vs-ICU contrast has the wrong sign in d3.", + "The observed contrast rests on 7 HC that all have zero c0 cells; the binomial GLM separates completely there.", + "Beyond-bulk was not evaluated in the claim. The bulk already separates sepsis from HC almost perfectly (HLA-DRA AUC 0.18; S100A12 AUC 0.98; index AUC 0.98), so the c0 contrast adds nothing at group level.", + "Draw imprinting: d3 and d4 both trained on Skin-S142 (observed c0 fraction 16.8%), but only d3 carries the effect. Draw 0's training donors have little c0. Imprinting therefore does not explain which draws carry the result." + ], + "reasoning": "Every number in SEP-POP-6 reproduces exactly from the raw compact files with my own code, including the per-draw effects, q-values over 143 features, calibration rows, observed contrasts, depth statistics and raw detection rates. The bulk-only sepsis > HC effect in draws 0 and 3, and the observed effect, survive every confounder I tested: age, sex, site restriction, sex_flag removal, the monocyte-count threshold, label confidence, a sample-level library covariate, a cell-level library spline and a library-matched window. So the disease contrast is not refuted. The sepsis cohort has no repeated measures, and its bulk-only donors are disjoint from the training donors, so those checks do not apply.\n\nWhat does not hold up is the mechanism the claim gives for calling it an artefact:\n- Library size defines who is in the cluster but does not create the sepsis − HC difference. Library sizes are the same in both groups, low library alone carries no disease signal, and the HLA-II-negative S100A12+ phenotype at any depth carries the signal in all 5 draws and in observed cells, with better calibration.\n- Once the comparison is depth-matched, the observed members are not granulocyte-like: neutrophil-specific genes are absent and lower than in the rest. They look like HLA-II-low, MS1-leaning monocytes with some hemoglobin signal.\n- Draw 0 reproduces IL1R2 at the observed rate, and CSF3R is not a c0 marker.\n- With a cluster-agnostic, depth-matched metric, marker concordance is typical (0.70), not the lowest.\n\nThe artefact-suspect label is still defensible for c0 as a separate population: independent clusterings recover it poorly (Jaccard ≤ 0.35), calibration is weak (Spearman ≤ 0.33, generated contrast 0.14–0.71× observed, d0 HC n = 3), and the observed effect depends on where the cluster is cut. The signal also says nothing beyond the bulk, which already separates sepsis from HC on HLA-DRA and S100A12 (AUC 0.98).\n\nVerdict: qualified. The reportable bulk-only result is carried by draws 0 and 3, and the claim's library-size and granulocyte explanations are wrong.\n\nAll work was CPU-only under systemd-run with a 6G cap and no swap. Outputs are only under my verify directory. This verifier made no archive delta; the pbisc-diffusion archive update is for the coordinator to decide.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/vcommon.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v00_inspect.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v01_build_mono_table.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v02_recompute_c0.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v03_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v04_markers_thinned.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v05a_own_embedding.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v05b_own_population_tests.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v06_bulk_check.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v07_concordance_ambient.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v08_phenotype_library_control.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/v09_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v09_key_numbers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v09_official_calibration_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v02_c0_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v02_c0_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v02_c0_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v02_c0_depth.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v02_bulkonly_tests_all.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v03_A_threshold.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v03_B_glm.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v03_C_library.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v03_D_confidence.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v03_E_covariates.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v04_detection_pooled.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v04_detection_same_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v04_within_domain_enrichment.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v04_scores.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v04_obs_member_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v05_own_overlap.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v05_population_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v05_population_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v06_bulk_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v06_bulk_vs_c0.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v07_concordance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v07_ambient.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/tables/v08_phenotype_library_gee.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_recompute/logs/v09_summary.log" + ] + }, + { + "verifier": "calib", + "agent_id": "ac2f73db70a8409a8", + "verdict": "qualified", + "recomputed": "(1) Reproduction from the agent's own labels (subsample, Monocyte|r2.0|c0, v01): every quoted number reproduces.\n- Bulk-only, adult sepsis minus adult HC: d0 +0.054 (q 4e-4, 59 vs 24), d1 +0.001 (q 0.53), d2 +0.024 (q 0.056, 26 vs 18), d3 +0.077 (q 0.004, 33 vs 30), d4 +0.006 (q 0.83). BH is over the agent's 143 features.\n- Sample drop-out: samples with fewer than 10 subsampled monocytes are dropped. That removes 42 of 68 adult sepsis samples in d2, 35 of 68 in d3, and 16 of 46 in observed.\n- Paired calibration: ratio 0.64/0.50/0.81/1.05/0.72; Spearman 0.33/0.11/0.29/0.32/0.12. The generated sepsis-HC contrast is 0.71/0.14/0.16/0.59/0.20 of the observed one. The agent's gate passes d0, d2, d3.\n- Observed: 0.117 vs 0.000 (30 vs 7, p=1e-4). None of the 7 adult HC has a single c0 cell.\n\n(2) All-cells re-measurement (v04). A per-domain classifier trained on the subsample c0 labels uses lib_measured plus 14 marker genes from markers.npz. CV AUC is 0.84 observed and 0.88 generated. Its per-sample agreement with the subsample share is Spearman 0.09 to 0.79 (weakest in d4).\n- Bulk-only sepsis minus HC: +0.062/+0.013/+0.041/+0.055/+0.020 (AUC 0.92/0.68/0.87/0.89/0.82). q ≤ 0.007 in all 5 draws, and the result holds after dropping sex_flag samples.\n- Bulk-only sepsis vs ICU: +0.005 to +0.016, not significant (d4 q 0.058).\n- Calibration: ratio 1.00/0.63/0.89/1.13/0.88; Spearman 0.39/0.01/0.38/0.40/0.32; gen/obs contrast 0.56/0.25/0.52/0.51/0.19. The gate passes d0, d2, d3, d4 and fails d1.\n- Observed: 0.101 vs 0.025 (43 vs 7, AUC 0.99, p<1e-4). Against ICU controls it is 0.101 vs 0.072 (p=0.17).\n- Depth adjustment: adjusting for each sample's median monocyte library leaves the sepsis coefficient unchanged (+0.077, p<1e-4). Sepsis samples have higher monocyte libraries than HC (3,375 vs 3,025), not lower.\n- The pure low-library monocyte fraction is not sepsis-associated in observed cells (p=0.70). It is sepsis-associated in generated d0 and d2 (+0.13/+0.11), but it is uncalibrated (Spearman -0.08 to -0.20).\n\n(3) Depth-matched HLA-DR-zero monocytes (v05; observed cells thinned to each draw's generated depth):\n- Calibration Spearman 0.56/-0.01/0.55/0.54/0.59; ratio 0.84/0.62/1.25/1.23/1.06; gen/obs contrast 0.68/0.24/0.49/0.73/0.55.\n- Bulk-only sepsis minus HC: +0.18/+0.04/+0.16/+0.20/+0.14 (AUC 0.94/0.75/0.90/0.89/0.91).\n\n(4) Identity, depth-matched (v02, v03, v06, v07):\n- Neutrophil genes: in observed c0, the share of cells with ≥2 of 13 strict neutrophil genes is 3.9%, vs 10.3% in other observed monocytes. FCGR3B is 0.6% vs 9.1%, CXCR2 0.6% vs 2.7%. CD24, ARG1, LCN2 and CTSG are all 0%. MPO and ELANE are at the level of the rest.\n- Monocyte genes: CD14 79%, LYZ 99%, FCN1 88%. scANVI labels 98% of these cells CD14 monocyte.\n- CSF3R is lower in c0 than in the rest in every domain (observed -0.49, generated -0.37).\n- MS1 genes are up in observed c0: IL1R2 +0.29, HP +0.23, S100P +0.16, RETN +0.82.\n- At equal depth (within depth deciles), c0 still has 1.0 to 1.25 fewer HLA-II genes detected and +1.3 to +1.6 log units of S100A12, in both domains.\n- Predicting membership: library alone gives AUC 0.71 (observed) / 0.82 (generated); library plus HLA-II plus S100A12 gives 0.84 to 0.89.\n- Broad depth-matched concordance (1,221 genes): Pearson 0.84 with paired_gen and 0.81 with bulk-only, 0.71 to 0.86 per draw. That ranks 7th lowest of 20 monocyte clusters (range 0.75 to 0.96), not the lowest.\n- IL1R2 is present in draw-0 generated c0 (27.7% vs 26.7% in observed, thinned) and absent in d1-d4 (≤2.5%). HP is enriched in d3/d4 generated c0 (+0.16/+0.18).\n- Per observed sample, the c0 share tracks the MS1 (IL1R2+) state (Spearman 0.51, p<1e-4) and not the LDG state (-0.12).\n\n(5) Beyond bulk (v05):\n- The bulk log2 S100A12 minus log2 HLA-DRA index alone separates sepsis from HC with AUC 0.98. The generated proxy reaches 0.68 to 0.92.\n- In paired test samples the bulk index tracks the observed share (Spearman 0.56 to 0.63) better than the generated share does (0.01 to 0.40).\n- In a rank OLS, the generated share adds nothing beyond the bulk index: p 0.25/0.89/0.23/0.23/0.35.\n\n(6) Training donors: draw 0 has three c0-rich sepsis training donors (observed proxy 0.11 to 0.15). Draw 1's training donors are mostly c0-poor (mean 0.067).", + "issues": [ + "Contradicted by observed data: the observed c0 members are not granulocyte-like. At matched depth, strict neutrophil genes are depleted relative to other observed monocytes (FCGR3B 0.6% vs 9.1%; CXCR2 0.6% vs 2.7%; ≥2 neutrophil genes 3.9% vs 10.3%). The immature-neutrophil and LDN genes named by Kwok 2023 (CD24, ARG1, MPO/ELANE/CTSG) and Schulte-Schrepping 2020 (FCGR3B, LCN2, CD24) are absent or at background. The cells are CD14+ LYZ+ FCN1+ monocytes (scANVI 98% CD14 monocyte). The per-sample c0 share correlates with the MS1 (IL1R2+) state (0.51), not with the LDG state (-0.12).", + "CSF3R is wrongly listed as a gene that observed c0 carries. It is lower in c0 than in other monocytes in every domain (observed -0.49, generated -0.37). The '48% vs 12-15%' gap is a domain-wide and depth gap: the rest of the monocytes show 65% vs 36%.", + "The detection comparisons (IL1R2 35% vs 5% and so on) are not depth-matched, which DATA_GUIDE requires. After thinning, observed c0 detects IL1R2 at 25%, HP 17%, S100P 11%, CSF3R 29%. 'Generated members lack' these genes holds mainly for draws 1-2. IL1R2 is present in draw 0, matching observed (27.7% vs 26.7%), and HP is enriched in draws 3-4. This follows the known MS1 limitation in calibration.csv (MS1 level is reproduced only by draw 0).", + "'Marker concordance 0.39, the lowest' comes from a 51-gene set with a single up gene (S100A12). With a broad, depth-matched concordance over 1,221 genes, observed vs generated c0 correlate at Pearson 0.84 (paired_gen) and 0.81 (bulk-only), mid-ranked among 20 monocyte clusters. Both domains share the core axis: HLA-DRA -1.33 vs -1.63, CD74 -1.37 vs -1.31, S100A12 +1.21 vs +1.20, S100A8 +1.45 vs +1.23, and RETN up. 'Not the same population' is not supported; the generator renders the same HLA-DR-low/S100A-high state without part of its MS1 program.", + "The replication count is understated by noisy 200-cell subsample fractions with heavy drop-out (up to 42 of 68 sepsis samples). Re-measured on all cells, sepsis > HC holds in 5/5 draws (q ≤ 0.007). The calibration gate passes d0, d2, d3 and d4; d1 fails under every definition (Spearman about 0), and its training donors are c0-poor.", + "'Library-size driven in both domains' is only half right. Cluster formation is partly depth (library-only membership AUC 0.71 observed, 0.82 generated). But the observed sepsis association is not a depth effect: the pure low-library fraction is not associated (p=0.70), and adjusting for sample depth leaves the effect unchanged. The real artefact component is generated-only: low-library monocytes rise with sepsis in d0 and d2 but are uncalibrated (Spearman < 0).", + "Not novel for this cohort. Ye 2026 reports this state as Mono_c03-CD14-RETN (S100A8/S100A12-high, HLA-DR-low, MS1-like), increased in most sepsis patients. Ye 2026 also states that neutrophils were absent from their PBMC data.", + "'Beyond bulk' was not evaluated in the claim. Tested here, the population adds nothing beyond the bulk S100A12/HLA-DRA ratio: AUC 0.98, and generated-share p ≥ 0.23 once bulk is in the model.", + "Not sepsis-specific. Observed ICU controls are also raised (0.072 vs HC 0.025; sepsis vs ICU p=0.17). Generated sepsis vs ICU is about 0.", + "Caveat on this verifier's own work: the all-cells measure is a classifier reconstruction of c0 (CV AUC 0.84/0.88; agreement with the subsample share 0.09-0.79). The v04 'hla_low' and 'lowlib_hlalow' rows are invalid for generated sets (ties at zero); v05's HLA-DR-zero rule replaces them.", + "This verifier wrote only under its own directory, ran CPU-only and started no GPU work. No PBISC-Diffusion archive delta was prepared; the coordinator owns the POLICY.md task-end update." + ], + "reasoning": "The claim mixes reproducible numbers with an identity argument that the observed data contradict. The verdict is qualified, not refuted, because its practical conclusion still stands: this is not a reportable new finding. That conclusion holds for different reasons than the claim gives.\n\nCalibration. Monocyte|r2.0|c0 is not in calibration.csv, so I recomputed the paired check myself on each draw's test samples.\n- On the agent's subsample labels the direction agrees in 5/5 draws for sepsis vs HC. Spearman is weak (0.11-0.33) and the generated contrast is 0.14-0.71× observed.\n- On all cells (embedding-free proxy), the gate passes d0, d2, d3, d4 (Spearman 0.32-0.40) and fails d1 (0.01). The generated contrast is 0.19-0.56× observed.\n- The depth-matched HLA-DR-zero rule gives the same draw pattern (Spearman 0.54-0.59 in d0, d2, d3, d4; -0.01 in d1).\n- Draws that carry a calibrated sepsis > HC effect: d0, d2, d3, d4. d1 does not.\n- Related rows in calibration.csv: the MS1 (IL1R2+) level is reproduced only by draw 0 (ratio 0.98 vs 0.03-0.14). That explains why generated c0 lacks IL1R2/HP/S100P in some draws. The LDG row (ratio 0.01-0.09) is irrelevant, because c0 is not LDG.\n\nReplication across draws. All 5 draws are positive. On subsample fractions q<0.05 in 2/5; on all cells q ≤ 0.007 in 5/5.\n\nObserved agreement. Strong:\n- Observed c0 share is 0.101 vs 0.025 (AUC 0.99), or 0.117 vs 0 on the subsample.\n- It is not explained by sample depth.\n- It is not sepsis-specific vs ICU controls (p=0.17).\n\nBiology. Observed c0 are HLA-DR-low, S100A8/9/12-high CD14 monocytes with lower CD74, CD68, MAFB, CSF1R, NAMPT, CXCL8 and IL1B, and higher RETN, IL1R2, HP and S100P. That is the MS1 / HLA-DRlo S100Ahi state, not neutrophils:\n- Ye 2026 Nat Immunol 27, p.155: \"Mono_c03-CD14-RETN was enriched in S100A genes (S100A8 and S100A12) and had low HLA-DR expression, resembling previously reported MS1-like/HLA-DRloS100Ahi monocytes\"; increased in most sepsis patients.\n- Ye 2026, p.157: the MS1 signature (RETNhi, ALOX5APhi) predominated in Mono_c03-CD14-RETN.\n- Ye 2026, p.163: this subset was \"consistently elevated across various infection sources\"; \"neutrophils ... were absent from our dataset\". QC kept cells with >200 genes (PDF p.18).\n- Reyes 2020 Nat Med, author manuscript p.2: \"MS1, CD14+ cells characterized by high expression of RETN, ALOX5AP, and IL1R2\". p.4: \"low HLA-DR and high IL1R2 expression can be used to quantify the fraction of MS1 cells\".\n- Schulte-Schrepping 2020 Cell 182, p.1427: HLA-DRlo S100Ahi monocytes (S100A12) in severe COVID-19.\n- Yao 2023 Mil Med Res 10:27, p.7 of 20: HLA-DRlow S100Ahigh monocytes in sepsis PBMC, with PLAC8, RETN and CLU.\n- The claim's comparators do not match. Kwok 2023, accepted manuscript p.1 and p.25: IL1R2+ immature neutrophils come from whole blood and carry CD24, ARG1, MPO, ELANE and CTSG. Schulte-Schrepping 2020, p.1423: PBMC LDNs express FCGR3B, CXCL8, LCN2 or CD24, PGLYRP1, DEFA3/4. None of these is enriched in observed c0.\n- Novelty: already reported for this exact cohort.\n\nBeyond bulk. Nothing. The bulk S100A12-HLA-DRA ratio already separates sepsis from HC (AUC 0.98). It tracks the observed share better than the generated cells do, and the generated share adds no information once bulk is in the model in any draw.\n\nNet. Keep SEP-POP-6 as a non-finding, but rewrite it:\n- the cluster is the known HLA-DR-low S100A-high (MS1-like, Ye's Mono_c03-CD14-RETN) monocyte state in both domains;\n- it is reproduced at reduced size in d0, d2, d3, d4, and d1 is uncalibrated;\n- the generated low-library component is an uncalibrated artefact;\n- it is bulk-derivable and not novel;\n- drop the granulocyte/LDG interpretation and the 'not the same population' conclusion.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v01_reproduce_c0.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v02_identity_depthmatched.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v03_depth_vs_phenotype.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v03b_print_ms1_by_draw.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v04_fullcell_proxy_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v05_bulk_thin_imprint.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v06_extra_identity.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v07_ms1genes_per_draw.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/v08_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v08_key_numbers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v01_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v01_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v01_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v01_nan_dropout.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v02_gene_detection.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v02_cell_scores.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v03_concordance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v03_depth_stratified.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v03_membership_auc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v04_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v04_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v04_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v04_observed_ols.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v04_sample_features.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v04_classifier_cv.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v05_hla0_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v05_hla0_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v05_bulk_derivability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v05_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v06_extra_detection.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v06_obs_state_correlation.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v07_ms1genes_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/tables/v08_proxy_vs_subsample.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/logs/v08_summary.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/lit/ye2026_paged.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/lit/Kwok2023_Neutrophils_emergency_granulopoiesis_sepsis_SRS1_paged.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/lit/Reyes2020_Immune_cell_signature_bacterial_sepsis_paged.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/lit/SchulteSchrepping2020_Severe_COVID19_dysregulated_myeloid_paged.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-6_calib/lit/Yao2023_HLADRlow_S100Ahigh_monocytes_sepsis_paged.txt" + ] + } + ] + }, + { + "id": "SEP-PROG-3", + "lens": "sep_programs", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 6, + "theme": "where the bulk change sits", + "synthesis_strength": "weak: descriptive agreement only; no gain over a monocyte prior", + "reportable_statement": "Draws 0, 2, 3 and 4 place most top bulk-up sepsis genes in CD14 monocytes, as the observed paired cells do (agreement 72-92%). A constant 'CD14 monocyte' prior (80-84%) and bulk deconvolution do as well; only draw 0 beats a healthy-reference lookup in paired calibration. No added value beyond the bulk is shown.", + "carrying_draws": { + "descriptive carrier agreement": "0,2,3,4", + "gain over HC reference (paired calibration)": "0 only", + "bulk-decreased genes": "none" + }, + "honesty": { + "calibration": "Paired top-30 carrier agreement 0.87/0.37/0.80/0.77/0.73 vs reference 0.67-0.70, vs constant-CD14 0.80-0.86; d2-d4 worse than the constant at top 100 (q 0.005-0.016).", + "replication": "4/5 draws (draw 1 fails: intermediate-monocyte calibration ratio 7.55).", + "observed": "Observed: 24/30 top-up genes carried by CD14 monocytes, 27/30 expression-driven.", + "beyond_bulk": "Not shown: constant CD14 prior, NNLS deconvolution and a lookup in each draw's own training donors match or beat every draw." + }, + "verdict_reasons": [ + "weak reference baseline", + "library-size confounding removes d2-d4 gains", + "'mostly within-cell induction' false for d0 and d2", + "IL1R2 '96% within-cell' is ~51%", + "down-gene carriers 1-11/30 vs 29/30 for a constant NK call" + ], + "new_vs_cohort_papers": "Monocyte origin known (Reyes 2020/2021; Ye 2026 Mono_c03); VSIG4/ADAMTS2/TPST1/FAM20A as monocyte glucocorticoid genes in Molendijk & Koenderman 2026; Hofer 2008.", + "next_steps": [ + "Score against constant-prior and deconvolution baselines on held-out paired samples with a library-equalised decomposition.", + "Check glucocorticoid exposure (FKBP5/PER1 up; no steroid column in meta.csv)." + ], + "key_sources": [ + "verify/SEP-PROG-3_recompute/v08_verdict_table.csv", + "verify/SEP-PROG-3_recompute/v03_baselines.csv", + "verify/SEP-PROG-3_calib/v02_summary.csv", + "verify/SEP-PROG-3_recompute/v07_calib_baselines.csv" + ] + }, + "finding": { + "id": "SEP-PROG-3", + "cohort": "ye2026_sepsis", + "claim": "For the 30–100 bulk genes most increased in bulk-only sepsis vs HC, draws 0, 2, 3 and 4 place the increase in CD14 monocytes, mostly as within-cell induction. Their carrier matches the observed paired cells' carrier for 72–92% of genes, against 64–70% for a healthy-reference atlas lookup. This includes genes whose healthy carrier is another type: IL1R2, TPST1, GRB10, CD63, FAM20A, VSIG4. Draw 1 fails, and no draw beats the reference on the bulk-decreased genes.", + "kind": "other", + "feature": "Cell-type carrier of the bulk sepsis−HC change. The change in whole-sample pseudobulk CPM is decomposed exactly into AIFI-L2 types and summed to 13 lineages; the carrier is the lineage with the largest change in the direction of the total. An approximate split separates composition from within-cell expression.", + "comparison": "adult sepsis vs adult HC. Bulk-only generated cells (68 vs 36) vs observed paired cells (46 vs 7). Gene lists: top 30/50/100 bulk DE genes up (and down) in bulkonly_raw.npz, on the 6k panel, q<0.05, ranked by Welch t.", + "effect_bulkonly": "Top-30 up genes: carrier is CD14 mono for 25, 7, 20, 24, 22 genes (draws 0–4). It equals the observed carrier for 27, 7, 23, 27, 26 of 30; the HC reference atlas gets 21 of 30. The reference is wrong for 9 genes (IL1R2→DC, TPST1→B, GRB10→CD4 T, CD63→NK, FAM20A/MERTK→CD16 mono, VSIG4→Int mono, HIST1H1C, CEBPB); the generated cells get 8, 2, 4, 7, 7 of these right. Carrier and driver both equal observed for 16, 4, 14, 20, 23 of 30. Top 50: 0.92, 0.22, 0.82, 0.84, 0.84 vs reference 0.66. Top 100: 0.89, 0.17, 0.76, 0.72, 0.81 vs 0.64.", + "stats": "Exact McNemar test, generated vs reference, each scored against the observed carrier. Top 50 up: p = 0.002, <0.001 (draw 1 worse), 0.039, 0.012, 0.049. Top 100: <0.001, <0.001 (worse), 0.029, 0.17, 0.008. Top 30: 0.11, 0.001 (worse), 0.69, 0.07, 0.18. Not corrected across draws or lists. The generated whole-sample change has the bulk's sign for 30, 30, 28, 29, 28 of the top-30 up genes.", + "draws_same_sign": "4/5 draws (0, 2, 3, 4) beat the reference baseline; draw 1 is worse", + "paired_calibration": "Same decomposition on each draw's paired TEST samples, generated vs observed carrier (attribution_calib.csv). Top-30 up agreement: 0.87, 0.37, 0.80, 0.77, 0.73, vs reference 0.70, 0.70, 0.70, 0.70, 0.67. Driver agreement given carrier agreement: 0.58, 0.45, 0.67, 0.91, 0.95. Top 50: 0.88, 0.34, 0.80, 0.74, 0.66. calibration.csv row 'l2 | Intermediate monocyte' in draw 1: ratio 7.55, Spearman 0.06, wrong sepsis−HC sign (+0.117 vs −0.012). That explains draw 1's 'Int mono' carriers. Row 'l2 | CD14 monocyte': sepsis−HC sign kept in draws 0–3, not in draw 4.", + "paired_observed_concordance": "Observed paired cells: 24 of the 30 top-up genes are carried by CD14 monocytes, and 27 of 30 rise through within-cell expression rather than monocyte expansion. Example: IL1R2 rises 5.26 log2 in total, 96% of it from CD14 within-cell induction. Draws 3–4 reproduce the expression-driven reading (21 and 24 of 30 expression-driven); draws 0–2 call more genes composition-driven.", + "beyond_bulk": "Yes, modestly. The bulk gives the change but not its carrier. A healthy-reference atlas lookup, which is what a bulk analyst can add, gets 64–70% of carriers right. The generated cells of draws 0, 2, 3 and 4 get 72–92% across top 30–100, with significant gains at top 50 in 4 draws. This is group-level only. The paired observed cells give the same group attribution, so the added value is for cohorts without paired cells.", + "novelty": "The monocyte origin of IL1R2, MCEMP1, S100A12, CD163 and RETN is known (Reyes 2020 MS1; Ye 2026 PDF p.6, Mono_c03-CD14-RETN). Ye 2026 does not report gene-by-gene carrier calls for ADAMTS2, FAM20A, TPST1, GRB10, VSIG4, MERTK or CD63 in this cohort; other literature was not checked.", + "strength": "moderate", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s02_bulkonly_de.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/de_bulk.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s04_attribution.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_genes_fine.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_genes_coarse.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_lineage_fine.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_types_long.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_beyond_ref.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s04b_attribution_topN.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_topN_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s04c_attribution_mcnemar.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_mcnemar.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s04d_attribution_counts.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_counts_fine.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "aa59c29ee07abca50", + "verdict": "qualified", + "recomputed": "I wrote independent CPU code (vlib.py, v01–v08) that reads the raw compact files directly.\n\nREPRODUCTION\n- My bulk DE (68 vs 36, log2(CPM+1) Welch, BH over genes with mean CPM ≥ 1) gives top 30/50/100 lists identical to the discovery's (max |Δt| 5e-5).\n\nTop-30 up genes, draws d0–d4:\n- Generated carrier is CD14 mono for 25/7/20/24/22.\n- Generated carrier equals the observed paired carrier (46 vs 7) for 27/7/23/27/26 of 30. The 7-HC reference gets 21.\n- The reference is wrong for 9 genes; the generated cells get 8/2/4/7/7 of them right.\n- Carrier and driver both equal observed: 16/4/14/20/23.\n- Generated total has the bulk's sign: 30/30/28/29/28.\n\nTop 50 and top 100 up:\n- Top 50: 46/11/41/42/42 vs reference 33. McNemar p = 0.002 / <0.001 (worse) / 0.039 / 0.012 / 0.049.\n- Top 100: 89/18/76/73/81 vs reference 64. The claim's 17 (d1) and 72 (d3) differ by one gene each, from near-tied lineages (SOCS3/ASPH/SIAE in d1, SRGN/CD55 in d3). d3 p = 0.12 (claim 0.17).\n\nPaired calibration, top-30 up:\n- Carrier agreement 0.87/0.37/0.80/0.77/0.73 vs reference 0.70–0.67.\n- Driver agreement given carrier agreement 0.58/0.45/0.67/0.91/0.95.\n\nCalibration rows (calibration.csv, l2 CD14 monocyte, sepsis−HC, generated vs observed on test samples):\n- d0: +0.067 vs +0.072\n- d1: +0.122 vs +0.063 (twice observed)\n- d2: +0.019 vs +0.068 (0.28×)\n- d3: +0.006 vs +0.060 (0.10×)\n- d4: −0.015 vs +0.065 (wrong sign)\n- Int mono in d1: ratio 7.55, Spearman 0.06, +0.117 vs −0.012. MS1 (IL1R2+) level: 0.98× in d0, 0.03–0.14× in d1–d4.\n\nSTRONGER BASELINES (counts correct for top-30/50/100 up)\n- Constant \"CD14 mono\": 24/42/83.\n - Only d0 beats it: bootstrap difference +0.10 [0.03, 0.13], +0.08 [0.04, 0.14], +0.06 [−0.01, 0.10].\n - d2–d4 differences run from −0.10 to +0.10, and every CI includes 0.\n- Bulk NNLS deconvolution with HC lineage signatures (no disease single cells): 23/40/82.\n - Generated vs this baseline: d2–d4 p ≥ 0.08; d0 p = 0.125 / 0.031 / 0.118.\n- Lookup in the observed cells of each draw's 4 training sepsis donors, scored on truth without those donors:\n - Lookup: 28,27,26,28,26 / 47,46,42,45,45 / 89,84,80,85,83.\n - Generated: 27,7,23,27,27 / 46,11,41,42,43 / 88,18,76,73,82.\n - The lookup is equal or better in every draw (d3 top 100 p = 0.008).\n- Reference enlarged to 11 HC (adds 4 paediatric HC): 22/35/66. Top-50 p for d2–d4 becomes 0.11 / 0.092 / 0.14.\n- HC jackknife (leave one of 7 HC out), top-50 p ranges: d2 0.021–0.146, d3 0.002–0.065, d4 0.035–0.143.\n- BH over the 15 up-list tests: q < 0.05 only for d0 (top 50, top 100), d3 (top 50) and d4 (top 100). d2 never.\n- Sample-level bootstrap, generated minus reference:\n - d0: +0.20 [0.07, 0.33], +0.26 [0.16, 0.32], +0.25 [0.12, 0.32]\n - d2: CI includes 0 at every N\n - d3: [0.00, 0.30], [−0.02, 0.24], [−0.08, 0.18]\n - d4: [0.03, 0.27], [0.02, 0.28], [−0.03, 0.25]\n- Paired-test calibration against the same baselines:\n - Constant CD14 scores 0.80–0.86; the training-donor lookup scores 0.80–0.94.\n - Generated d2–d4 are not better than the HC reference (p 0.09–1.0).\n - At top 100, generated d2–d4 (0.68 / 0.65 / 0.61) are worse than the constant (p = 0.008 / 0.003 / 0.001).\n\nCONFOUNDERS (generated vs reference, McNemar p)\n- Truth without the draw's training donors: unchanged (top 50: 46/11/41/42/43 vs 33/33/33/33/32).\n- Site: all HC are from FAH. Restricting to FAH (bulk-only 58 vs 36, truth 36 vs 7) holds: top-50 p = 0.0003 / 0.11 / 0.021 / 0.035 for d0, d2, d3, d4.\n- No bulk-only donor trained any draw. There are no repeated sepsis donors.\n- Dropping the sex_flag samples: holds.\n- Age + sex OLS-adjusted gene list and decomposition: holds at top 100 (p <0.001 / 0.007 / 0.024 / 0.006), but at top 50 d2–d4 p = 0.51 / 0.15 / 0.15.\n- Label confidence ≥ 0.9 (subsample): holds (p ≤ 0.03).\n- Library-equalised (cell-weighted) decomposition: d0 holds (p 0.012 / 0.0003 / 0.00006); d2–d4 lose it (top 50 p 0.057 / 0.30 / 0.057; top 100 p 0.36 / 1.0 / 0.22).\n- Generated CD14-mono library size, sepsis vs HC: d0 +12% (p 0.002), d2 −17% (p 0.001), d4 +25% (p < 1e-4). Observed: +9% (p 0.37).\n\nDRIVER AND MAGNITUDE\n- Among CD14-carried top-30 genes, expression-driven calls: 11/25, 3/7, 9/20, 16/24, 18/22. Observed: 22/24.\n- A constant \"expression\" driver scores as well as or better than the generated driver calibration at top 30 (0.92/1.0/0.92/0.96/0.95 vs 0.58/0.45/0.67/0.91/0.95).\n- Median generated/observed log2FC for the top 30: 0.94 / 1.40 / 0.27 / 0.65 / 0.36.\n- IL1R2: observed +5.26 log2; CD14 carries 95.5% of the change. By the approximate split, the CD14 within-cell part is 52% of the total, the composition part 12%, and the rest is non-separable.\n- Generated IL1R2: +6.72 / +2.33 / +0.72 / +1.87 / +1.73 log2.\n\nDOWN LISTS\n- Generated: 7/3/3/11/1 of 30. Reference: 11. Constant \"NK\": 29. Bulk deconvolution: 20.", + "issues": [ + "Relative to the claim's 7-HC reference, the generated gain is real, but stronger bulk-side baselines erase it. Draws 2–4 never beat a constant 'CD14 mono' rule; d0 beats it only modestly (+0.06 to +0.10). A bulk NNLS deconvolution baseline with HC signatures ties d2–d4, and d0 beats it only at top 50 (p = 0.031, uncorrected).", + "Every draw is matched or beaten by a lookup in the observed cells of that draw's own 4 training sepsis donors (e.g. top 50: 47/46/42/45/45 vs generated 46/11/41/42/43). LoRA-5 needs those paired donors anyway, so the stated added value for cohorts without paired cells is not shown.", + "The paired-test calibration (the trust layer) supports a gain over the HC reference only in d0. For d2–d4 the test-sample agreement is not significantly above the reference (p 0.09–1.0). At top 100, d2–d4 are significantly worse than a constant CD14 guess (p = 0.008 / 0.003 / 0.001).", + "The d2–d4 significance is fragile. Under BH over 15 tests, d2 is never significant. With an 11-HC reference, the top-50 p for d2–d4 becomes 0.11 / 0.09 / 0.14. Leaving out a single HC moves those p-values between 0.002 and 0.15. Bootstrap CIs for d2 always include 0, and d3/d4 are borderline.", + "Library size is a confounder for d2–d4. Generated CD14-mono library sizes differ by group (d2 −17%, d4 +25%; observed +9%, n.s.). With the library-equalised (cell-weighted) decomposition, the d2–d4 gains disappear (top 100 p = 0.36 / 1.0 / 0.22). d0 survives.", + "'Mostly as within-cell induction' is wrong for d0 and d2: only 11/25 and 9/20 of their CD14-carried genes are expression-driven. d3–d4 reproduce the expression-driven reading, but their CD14 composition calibration is attenuated (0.10×) or has the wrong sign. A constant 'expression' driver call matches their driver calibration (0.96 / 0.95 vs 0.91 / 0.95).", + "Magnitudes are badly attenuated in d2–d4: the median generated/observed log2FC is 0.27 / 0.65 / 0.36. Several 'reference-wrong' genes rise only slightly in d2–d4 (VSIG4 0.06–0.48 vs 3.55 log2; GRB10 0.10–0.22 vs 1.34; TPST1 0.15–0.41 vs 1.94). The carriers are being called on near-null generated changes. This is consistent with MS1 being generated at only 0.03–0.14× in d1–d4.", + "The IL1R2 example overstates the split. CD14 carries 95.5% of the change, but by the claim's own approximate split the CD14 within-cell part is only 52% of the total (composition 12%, about a third non-separable). '96% from CD14 within-cell induction' is not correct.", + "The reference is built from only 7 HC and several of its 'wrong' calls are unstable under the HC jackknife: GRB10, VSIG4, CEBPB and HIST1H1C change call. The 11-HC reference gets CEBPB right.", + "Down genes: the claim says no draw beats the reference. The bigger problem is that the generated carriers (7/3/3/11/1 of 30) are far worse than a constant NK call (29) and bulk deconvolution (20), with paired calibration of 0.0–0.33.", + "Minor: the top-100 counts for d1/d3 (17/72) are off by one gene from my 18/73 because of near-tied lineages, and the d3 top-100 p is 0.12, not 0.17. Draw 1's failure is the CD14-vs-Int-mono split (Int mono calibration ratio 7.55), as stated.", + "d0, the only draw that is robust everywhere, is the draw whose training donors are rich in MS1 cells (MS1 calibration ratio 0.98). Its result tracks that training data, not general bulk-to-cell inference." + ], + "reasoning": "My independent code reproduces the claim's descriptive numbers almost exactly: same gene lists, same carrier counts, same McNemar p-values at top 30 and 50, and one-gene near-tie differences at top 100. The same holds for the paired calibration and the calibration.csv rows. The reproduced finding is: for most bulk-up sepsis genes, the generated cells of d0, d2, d3 and d4 call CD14 monocytes as the carrier, agreeing with the observed paired cells 72–92% of the time, against 64–70% for a 7-HC share lookup.\n\nThe comparison with the HC reference survives most requested confounder controls:\n- truth excluding each draw's training donors (and no bulk-only donor trained any draw)\n- restriction to the one site that holds all HC\n- removal of the sex_flag samples\n- age + sex OLS adjustment, at top 100\n- label confidence ≥ 0.9\n\nIt does not fully survive library-size equalisation. Generated CD14 libraries differ by group in a draw-specific way (d2 −17%, d4 +25%), and under the equalised decomposition only d0 keeps its gain.\n\nThe interpretation fails against fair baselines:\n- A constant \"CD14 mono\" rule matches d2–d4.\n- A bulk deconvolution baseline with HC signatures ties them.\n- A lookup in each draw's own 4 training sepsis donors matches or beats every draw.\n\nThe calibration layer agrees. On paired test samples, d2–d4 are not significantly better than the HC reference and are worse than a constant at top 100. d2–d4 significance also depends on uncorrected testing, the particular 7-HC reference and individual HC donors. The \"mostly within-cell induction\" statement is wrong for d0 and d2. Where d3/d4 agree with the observed driver, it coincides with their failed monocyte-composition calibration, and a constant \"expression\" call matches it. Generated magnitudes in d2–d4 are a third to two-thirds of observed.\n\nVerdict: qualified.\n- Holds: the carrier attribution itself, as a description, and d0's gain over the HC reference.\n- Fragile: d2–d4's gain over the reference.\n- Not supported: the \"beyond bulk\" value, since generated carriers add nothing over simple bulk-side or small sepsis-atlas baselines.\n- Overstated: the driver (\"within-cell induction\") and IL1R2 statements.\n\nRecommended restatement: descriptive agreement only, with strength downgraded to weak for any beyond-bulk claim.\n\nNo GPU was used. All runs were CPU, under systemd-run with MemoryMax=6G and no swap. All writes are in my verify directory. This verifier makes no archive delta; the coordinator owns any paper-archive update.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/vlib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v00_inspect.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v00_inspect.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01_genes_top30_up.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01_genes_top30_down.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01_genes_top50_up.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01_genes_top100_up.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01_bulk_de.csv.gz", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01_confusion_top30_up.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01b_list_diff.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01c_ties_driver.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01c_near_ties.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v01c_driver_const.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v02_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v02_confounders.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v03_baselines.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v03_baselines.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v03_baselines_genes_top30_up.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v03_jackknife.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v03_jackknife_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v03_deconv_F.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v05_library_composition.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v05_group_qc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v05_calibration_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v05_driver_by_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v06_confidence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v06_confidence.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v07_bootstrap_calib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v07_bootstrap.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v07_calib_baselines.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v07_mcnemar_bh.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v07_magnitude.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v08_consolidate.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v08_verdict_table.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_recompute/v08_il1r2.csv" + ] + }, + { + "verifier": "calib", + "agent_id": "a0c829a6e956f8771", + "verdict": "qualified", + "recomputed": "I re-implemented the whole analysis myself (v01–v06) instead of importing the candidate's code.\n\n**Reproduction: exact.**\n- Bulk DE recomputed from bulkonly_raw.npz gives identical top 30/50/100 up and down lists (2,684 eligible genes, 760 up). Dropping the 3 sex_flag samples keeps 29/30, 49/50 and 95/100 of the up lists.\n- All candidate numbers reproduce:\n - Bulk-only carrier equals the observed carrier for 27, 7, 23, 27, 26 of the top-30 up genes (draws 0–4), against 21 for the reference.\n - Top 50: 0.92, 0.22, 0.82, 0.84, 0.84 vs 0.66. Top 100: 0.89, 0.17, 0.76, 0.72, 0.81 vs 0.64.\n - Paired-test calibration, top 30: 0.87, 0.37, 0.80, 0.77, 0.73 vs reference 0.70 (0.67 in draw 4).\n\n**New baselines (bulk-only vs observed, top 30/50/100 up).**\n- An \"always CD14 mono\" call scores 0.80/0.84/0.83.\n- A crude NNLS bulk deconvolution, composition only, scores 0.73/0.78/0.69.\n- Generated vs always-CD14 (exact McNemar): no draw is significantly better.\n - Draw 0: 27 vs 24, 46 vs 42, 89 vs 83 (p = 0.25, 0.22, 0.21).\n - Draws 2–4 tie or lose; draw 3 at top 100 is 72 vs 83 (p = 0.035).\n\n**Paired-test calibration (the trust layer).**\n- Generated vs the candidate's reference, top 50 / top 100:\n - draw 0: 44 vs 33 (BH q = 0.024); 83 vs 65 (q = 0.005)\n - draw 2: 40 vs 33 (p = 0.09); 67 vs 64 (p = 0.71)\n - draw 3: 37 vs 33 (p = 0.45); 65 vs 64 (p = 1.0)\n - draw 4: 33 vs 32 (p = 1.0); 61 vs 63 (p = 0.87)\n- Generated vs always-CD14, top 100:\n - draw 0: 83 vs 82\n - draw 2: 67 vs 83 (q = 0.016)\n - draw 3: 65 vs 83 (q = 0.010)\n - draw 4: 61 vs 84 (q = 0.005)\n - No draw beats it at any list size.\n- Kappa (agreement corrected for chance), top 50: 0.59, 0.49, 0.36, 0.33 for draws 0, 2, 3, 4, against reference 0.32/0.27 and always-CD14 0.\n\n**The 9 genes the reference gets wrong in the top 30.** Generated draws get 8, 2, 4, 7, 7 right; always-CD14 gets 8 of 9.\n\n**Stability (bootstrap, 300 resamples).**\n- Observed carriers are stable (mean 0.95).\n- Reference calls are unstable for GRB10 (0.60), VSIG4 (0.43) and CEBPB (0.49), because only 7 HC are available.\n\n**Driver (composition vs within-cell expression).**\n- Bulk-only, top 30, expression-driven among CD14-carried genes: 11/25, 3/7, 9/20, 16/24, 18/22. Over all 30 genes: 15, 8, 17, 21, 24. Observed: 27/30.\n- Paired calibration, driver agreement where the carrier agrees: 0.58, 0.45, 0.67, 0.91, 0.95. An \"always expression\" call gets 0.92, 1.00, 0.92, 0.96, 0.95.\n\n**Within-CD14 induction, paired test, top 30.**\n- Direction is right for 87–100% of genes; Spearman 0.20–0.63.\n- Median generated/observed magnitude: 0.75 (draw 0), 0.57, 0.30, 0.44, 0.32.\n\n**calibration.csv rows.**\n- CD14 monocyte fraction, sepsis − HC (generated vs observed):\n - draw 0: 0.067 vs 0.072\n - draw 1: 0.122 vs 0.063\n - draw 2: 0.019 vs 0.068\n - draw 3: 0.006 vs 0.060\n - draw 4: −0.015 vs 0.065\n- MS1 ratio: 0.98 in draw 0, 0.03–0.14 in draws 1–4. LDG ratio: 0.01–0.09.\n\n**IL1R2 (observed).** CD14 carries 95.5% of the change. By the candidate's own split, CD14 within-cell expression is 51% of the total and CD14 expansion 12%; the rest is a covariance term.\n\n**Bulk vs single cells on the same paired samples.**\n- The single-cell change is a median 72% of the bulk log2 change (Spearman 0.91).\n- Exceptions: CD63 is 0.24 vs 0.91 log2 (26%); CEBPB is 0.12 vs 0.85 (14%).\n\n**Bulk-decreased genes.** A per-cell HC lookup gets 0.86–0.87 of observed carriers; the generated draws get 0.03–0.40.", + "issues": [ + "Weak baseline: the healthy-reference lookup ignores the direction of change. A constant 'CD14 mono' call scores 80–84% (top 30/50/100) and no draw significantly beats it. In the paired-test calibration, draws 2–4 are significantly worse than it at top 100 (BH q = 0.016, 0.010, 0.005). The 'beyond bulk: yes, modestly' claim is not supported.", + "The paired calibration supports the gain over the reference only in draw 0 (top 50: 44 vs 33, q = 0.024; top 100: 83 vs 65, q = 0.005). The bulk-only top-50 wins of draws 2–4 (p = 0.039, 0.012, 0.049) do not reproduce on their own paired test samples: draw 2 40 vs 33 (p = 0.09), draw 3 37 vs 33 (p = 0.45), draw 4 33 vs 32 (p = 1.0). At top 100 they are 67 vs 64, 65 vs 64 and 61 vs 63. The candidate reported the calibration-vs-reference comparison only at top 30.", + "'Mostly as within-cell induction' is false for draws 0 and 2: 11/25 and 9/20 of their CD14-carried genes are expression-driven. Draws 3–4 call expression (16/24, 18/22), but they fail the CD14 monocyte composition calibration (0.10x; reversed sign in draw 4), so they cannot generate an expansion and the 'expression' call is partly an artifact. In calibration, an 'always expression' driver call matches or beats every draw at top 30.", + "The showcased genes add nothing over a simple prior: 'always CD14 mono' gets 8 of the 9 reference-wrong genes, as many as the best draw. With only 7 HC, the reference calls for GRB10, VSIG4 and CEBPB are unstable (bootstrap 0.60, 0.43, 0.49). CD63 and CEBPB barely change in the observed cells (0.24 and 0.12 log2, i.e. 26% and 14% of the paired bulk change), so their carrier explains little of the bulk signal.", + "The IL1R2 example is misquoted. 96% is CD14's share of the change, not the within-cell part. By the candidate's own split, CD14 within-cell expression is about 51% of the total and CD14 expansion about 12%; the rest is a covariance term.", + "Statistics are optimistic. After BH over the 30 McNemar tests, the bulk-only top-50 wins survive only for draws 0 and 3. Genes come from co-regulated programs (MS1, glucocorticoid) and the top-30/50/100 lists are nested, so the gene-level McNemar p-values are anti-conservative.", + "Novelty is low. Reyes 2020 shows IL1R2/RETN/ALOX5AP on MS1 CD14 monocytes and lists CD63 among the top-6 MS1 markers (PDF p.2, p.4, p.15), with bulk deconvolution of MS1 in 11 cohorts (p.3). The Reyes 2021 MS1 program includes CTSD, MCEMP1, TIMP1, ACSL1, NAMPT and CD163 (PDF p.2/4/6). Ye 2026 reports expansion of the MS1-like Mono_c03-CD14-RETN subset (PDF p.6, p.8, p.14 = journal p.155/157/163) and deconvolved its own bulk data (PDF p.12, journal p.161). Molendijk & Koenderman 2026 (J Leukoc Biol 118:qiag068) report VSIG4, ADAMTS2, FLT3 and CD163 as monocyte-biased glucocorticoid-driven genes in septic shock, cover IL1R2 and TPST1, and show FAM20A induced in sorted monocytes and neutrophils. Hofer 2008 (J Mol Med 86:323) shows ADAMTS2 induced 440-fold in CD14++ monocytes by glucocorticoids. GRB10 and MERTK were not found in the sources checked.", + "Possible glucocorticoid confounding: FKBP5 (+1.27 log2) and PER1 are also up in the bulk, which fits steroid or cortisol exposure. meta.csv has no steroid-treatment column, so this cannot be tested.", + "Scope limits. Neutrophils and LDGs are absent from the single-cell data (Ye 2026 PDF p.14), and LDG generation fails calibration (ratio 0.01–0.09). The carrier is defined only within the captured PBMC lineages, and IL1R2 is also an immature-neutrophil gene in sepsis (Kwok 2023, PDF p.1/4/5). The observed truth has only 7 HC, and the 'all observed paired cells' set includes the training donors of draws 0 and 1 (HC07, HC12); the paired calibration excludes them.", + "Bulk-decreased genes: generated carriers are poor (0.03–0.40), while a per-cell HC lookup reaches 0.86–0.87. The candidate's 'no draw beats the reference' understates this." + ], + "reasoning": "**Calibration.** The descriptive core of the claim holds. My independent re-implementation reproduces every number exactly. Draws 0, 2, 3 and 4 place most top bulk-up genes in CD14 monocytes (25, 20, 24, 22 of 30). Their paired-test carrier agreement (0.73–0.87 at top 30) passes calibration; draw 1 fails.\n\nThe value-added part does not hold. The candidate's reference picks the lineage with the largest HC share of the gene, ignoring the direction of change. A trivial \"sepsis-up genes are CD14-monocyte genes\" prior already scores 0.80–0.84, a crude bulk deconvolution scores 0.69–0.80, and no draw beats the prior in any test. In the paired calibration (the trust layer), only draw 0 beats even the weak reference. Draws 3–4's bulk-only wins do not reproduce on their own test samples, and draws 2–4 are significantly worse than the prior at top 100. The gene-level driver calls do not beat an \"always expression\" call in calibration. \"Mostly within-cell induction\" in the generated cells is false for draws 0 and 2. In draws 3–4 it coincides with a monocyte-composition contrast that is 0.1x or reversed.\n\nSo only draw 0 carries the \"better than the reference atlas\" result. That fits its MS1-rich training donors: it is the only draw that reproduces MS1 (ratio 0.98) and the monocyte expansion.\n\n**Observed support.** The observed paired cells confirm the biology: 24/30 CD14 carriers and 27/30 expression-driven, stable under bootstrap. Two corrections apply:\n- The IL1R2 \"96% within-cell\" is really about 51% by the candidate's own split.\n- CD63 and CEBPB barely change in the single cells, so their carrier explains little of the bulk change.\n\n**Biology and novelty.** The CD14-monocyte origin of the sepsis-up program is established: MS1 (Reyes 2020/2021; Ye 2026 Mono_c03-CD14-RETN). The headline \"reference-wrong\" genes are also documented as monocyte-induced glucocorticoid or STAT3 genes in septic shock (Molendijk & Koenderman 2026; Hofer 2008). Novelty for this cohort is limited to a gene-level restatement that the paired observed cells already provide.\n\n**Beyond bulk.** Not shown. A bulk analyst with the monocyte prior or a deconvolution matches draws 2–4 and is not significantly beaten by draw 0.\n\n**Verdict.** Qualified, because support is partial. A defensible restatement: \"Generated cells of draws 0, 2, 3 and 4 reproduce the observed CD14-monocyte carrier of bulk sepsis-up genes about as often as a 'sepsis-up = CD14 monocyte' prior. Only draw 0 beats a healthy-reference lookup in paired calibration. No added value beyond the bulk and a monocyte prior is demonstrated.\"\n\n**Conduct.** CPU only, no GPU used. All writes stayed in my lens dir, so this verifier adds no paper-archive delta.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v01_bulk_de.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v01_bulk_lists.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v01_check.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v02_attribution.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v02_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v02_mcnemar.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v02_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v02_deconv_fractions.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v02.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v03_discrimination_bootstrap.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v03_kappa.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v03_bootstrap_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v03_bootstrap_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v03_named_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v04_bulk_vs_sc_and_cd14_calib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v04_bulk_vs_sc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v04_cd14_induction_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v04_driver_cd14.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v04_calib_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v05_gc_program_check.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v05_gc_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v06_key_numbers.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/v06_key_numbers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-3_calib/lit_txt/", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Reyes2020_Immune_cell_signature_bacterial_sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Reyes2021_Sepsis_COVID_plasma_induces_suppressive_myelopoiesis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Kwok2023_Neutrophils_emergency_granulopoiesis_sepsis_SRS1.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Leite2024_MS1_deconvolution_critically_ill_sepsis.pdf" + ] + } + ] + }, + { + "id": "SEP-CLIN-5", + "lens": "sep_clinical", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 7, + "theme": "MS1 in ICU controls", + "synthesis_strength": "weak: draw 0 only; bulk-derived; already reported", + "reportable_statement": "In draw 0 (the only MS1-calibrated draw), bulk-only ICU controls get more MS1 monocytes than HC (12.7% vs 0.33% of monocytes). This restates bulk IL1R2 (+3.89 log2) and is reported in Ye 2026. The generated sepsis > ICU ordering is a model bias.", + "carrying_draws": { + "ICU vs HC": "0 only (d1-d4 at 0.02-0.04x; median 0-1 MS1 cells per ICU sample)" + }, + "honesty": { + "calibration": "ICU-vs-HC not in calibration.csv; paired check sign 5/5, magnitude 0.32x (d0) and 0.02-0.04x (d1-d4); sepsis-vs-ICU sign 0/5; ICU under-generated at fixed bulk IL1R2 in 5/5 draws.", + "replication": "Count/rank tests positive 5/5; logit transform 4/5 (draw 1 zero-count artefact).", + "observed": "6 vs 7: 24.4% vs 0.67% (exact permutation p 0.017, MWU 0.051, HC3 p 0.20); carried by IC16 (87%) and IC19 (53%).", + "beyond_bulk": "None: bulk IL1R2 explains R2 0.71-0.78 of observed MS1 on paired tests; generated adds 0.01-0.03." + }, + "verdict_reasons": [ + "only draw 0 calibrated", + "observed truth rests on two ICU samples", + "fragile 'ICU at or above sepsis' reference", + "factual fixes (Skin-S74 not MS1-rich; Reyes page; '8-fold' not reproducible)" + ], + "new_vs_cohort_papers": "Reported: Ye 2026 Fig. 4d (Mono_c03 IC>HC), ED Fig. 9a (bulk deconvolution), ED Fig. 9e (flow); Reyes 2020; Leite 2024.", + "next_steps": [ + "Draws trained on MS1-rich ICU controls; more paired ICU controls; compare with bulk IL1R2 baseline per sample." + ], + "key_sources": [ + "verify/SEP-CLIN-5_recompute/v08_key_numbers.csv", + "verify/SEP-CLIN-5_calib/v01_paired_calib_contrasts.csv", + "verify/SEP-CLIN-5_calib/v06_mapping.csv", + "verify/SEP-CLIN-5_calib/v04_bulk_vs_gen_paired.csv" + ] + }, + "finding": { + "id": "SEP-CLIN-5", + "cohort": "ye2026_sepsis", + "claim": "Generated MS1 (IL1R2+) monocytes are higher in adult ICU controls than in HC (3.2% vs 0.1% of monocytes). Only draw 0 reproduces MS1 levels, and no draw reproduces the observed ordering where ICU controls are at or above sepsis. This is a bulk-driven, only partly calibrated result.", + "kind": "state", + "feature": "state | MS1 monocyte (IL1R2+)", + "comparison": "Adult ICU control (17) vs adult HC (36); adult sepsis (68) vs ICU control", + "effect_bulkonly": "Five-draw mean of monocytes: ICU 3.2%, HC 0.12%, sepsis 7.0%. Adjusted logit, ICU vs HC: +1.33 (SE 0.33). By draw: d0 +2.73 (the only MS1-calibrated draw), d1 -0.10, d2 +0.51, d3 +1.43, d4 +1.32. Sepsis vs ICU: +0.38 (p=0.26).", + "stats": "Model: OLS with HC3 errors, mean of draw logits ~ ICU + age + sex, n=53: p=4.6e-5, BH q=1.8e-4; rank-ANCOVA q=6e-6. Robust to the age window (p=0.003) and to matching (p=2e-4).", + "draws_same_sign": "4/5, with 3/5 at p<0.05 (d0, d3, d4). Only draw 0 carries the result. Its training donors (Abd-S66, Res-S150, Skin-S74, Res-S77) are MS1-rich sepsis samples.", + "paired_calibration": "calibration.csv, MS1: ratio 0.98/0.03/0.05/0.14/0.06, Spearman 0.52/0.46/0.29/0.37/0.47. 'Adult sepsis vs adult ICU control': same sign 0/5 (gen +0.100/+0.002/+0.008/+0.016/+0.002 vs obs -0.060/-0.035/-0.035/-0.036/-0.089). Sepsis vs HC: 5/5. ICU vs HC, own paired check: the sign agrees 5/5, but the raw difference is gen 0.067 vs obs 0.210 in d0 and 0.004-0.010 vs 0.20-0.25 in d1-d4.", + "paired_observed_concordance": "Observed paired: ICU 24.4% vs HC 0.7% (6 vs 7; logit +2.48, p=0.20; MWU p=0.051). Sepsis 19.9% vs ICU 24.4% (+0.33, p=0.82). The ICU > HC direction agrees. Relative to sepsis, generated ICU MS1 is about 8-fold too low.", + "beyond_bulk": "No. Bulk MS1 score, ICU vs HC: t=7.7, against t=4.1 for the cells. IL1R2 +3.89 log2 (q<1e-4), RETN +2.30, CD163 +2.22. The cell effect after adjusting for the bulk MS1 score: p=0.05; after adjusting for IL1R2: p=0.27.", + "novelty": "Reported. Reyes 2020 (rare-disease-states PDF, p.3) finds MS1 only slightly lower in non-septic ICU patients than in sepsis (FDR 0.27), and expanded in sterile ICU inflammation (p.4).", + "strength": "weak", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s02_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s03_paired_calib_long.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s04_observed_assoc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s06_feature_vs_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s08_bulk_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s10_evidence.txt" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a6eda0f39c607a762", + "verdict": "qualified", + "recomputed": "State recompute: my own code reads cells.parquet and markers.npz and applies the states.json rule (parent = AIFI L1/L2 Monocyte; IL1R2 detected and at least 2 of 7 genes detected). It reproduces states.parquet exactly: parent and MS1 counts are equal in all 122 bulk-only x 5 draws, 77 paired-gen x 5 draws and 82 observed samples. Thinning with my own seed matches observed_thin_d (r >= 0.996).\n\nFive-draw mean MS1 share of monocytes: ICU 3.19% (n 17), HC 0.12% (n 36), sepsis 7.01% (n 68).\n\nClaim's model (per-draw logit((n+0.5)/(N+1)), N >= 10 filter, nanmean over draws, OLS HC3 ~ ICU + age + sex, n 53):\n- +1.332 (SE 0.327), p 4.6e-5.\n- BH q 1.76e-4 (rank 13 of 50 features); rank-ANCOVA q 6.4e-6. Both exactly as claimed.\n- Other estimators:\n - complete cases: +0.86, p 1.1e-4\n - no N filter: +1.35, p 2.1e-4\n - stacked logit with draw fixed effects (clustered by sample): +1.35, p 5.4e-5\n - stacked binomial with draw fixed effects (clustered): +3.31, p 7.9e-19; heterogeneity p 0.0023, all draw effects positive\n\nPer draw, ICU vs HC (group means; claim transform; quasi-binomial with corrected scale):\n- d0: 12.7% vs 0.33%; +2.73 (p 5e-13); +3.68 (p 1.1e-5)\n- d1: 0.44% vs 0.05%; -0.10 (p 0.77); +2.21 (p 0.0066)\n- d2: 0.48% vs 0.05%; +0.51 (p 0.09); +2.25 (p 0.001)\n- d3: 1.04% vs 0.16%; +1.43 (p 2e-7); +2.52 (p 7e-10)\n- d4: 1.32% vs 0.01%; +1.32 (p 0.010); +3.60 (p 9e-5)\n- Mann-Whitney p <= 0.0016 and Fisher p <= 0.003 in every draw, so the sign is 5/5 with count or rank models.\n\nSepsis vs ICU: the claim's +0.38 (p 0.26) reproduces. The stacked binomial gives +0.89 (p 0.0066), with all 5 draws positive.\n\nConfounder controls on d0: every one of the following keeps the effect, at p <= 1e-4 by at least one model:\n- age squared; ages 40-80 (10 vs 33, quasi-binomial p 1e-6); age-overlap window; 1:1 age matching; 5-year caliper matching (11 vs 11, Mann-Whitney p 4e-4)\n- dropping sex_flag samples\n- log parent cell number; log monocyte library size; bulk library size and QC\n- monocyte L2 confidence >= 0.90 or >= 0.99\n- k = 3 rule; IL1R2-only rule\n- per-cell depth equalisation (Mann-Whitney p 1.7e-8)\n- leave-one-ICU-out (worst p 2.5e-5)\n- bulk neutrophil score (p 1.3e-6)\n\nOther structure checks:\n- Age works against the claim: MS1 rises with age within HC (d0 p 0.002), and ICU controls are younger (48.9 vs 63.8 years).\n- Site: ICU and HC come from the same hospital.\n- GSM accession order is fully collinear with group, so it cannot serve as a batch proxy.\n- Bulk depth and QC do not differ between groups (p >= 0.49).\n- Bulk-only donors are unique; none is paired and none trained a draw.\n\nCalibration: the calibration.csv MS1 rows reproduce exactly.\n- ratio d0-d4: 0.98 / 0.033 / 0.048 / 0.14 / 0.061\n- Spearman d0-d4: 0.52 / 0.46 / 0.29 / 0.37 / 0.47\n- sepsis minus ICU: same sign in 0/5 draws\n- ICU minus HC (own paired check, 6 vs 6-7 samples): generated 0.067 / 0.005 / 0.004 / 0.007 / 0.010 vs observed 0.210 / 0.229 / 0.226 / 0.204 / 0.255. Sign agrees 5/5. Draw 0 fold: 50x generated vs 52x observed.\n- Spearman generated vs observed within adult controls: 0.40 / 0.23 / 0.37 / 0.14 / 0.65.\n\nObserved paired cells (native depth):\n- ICU mean 24.4% (median 2.5%) vs HC 0.67%. Mann-Whitney p 0.051; quasi-binomial p 0.38; logit OLS with age and sex +2.48, p 0.20.\n- The ICU mean comes from 2 of 6 samples: IC16 87% and IC19 53%. The other four are 0.4-3.7%.\n- Sepsis vs ICU: means 19.6% vs 24.4%, medians 8.2% vs 2.5%; Mann-Whitney AUC 0.55, p 0.70.\n- Every draw generates IC16 as the top ICU control (34% in d0); every draw misses IC19 (0-1%).\n\nBulk:\n- Bulk-only, ICU vs HC: MS1 score t 7.74; IL1R2 +3.89 log2 (t 5.2); RETN +2.30; CD163 +2.22.\n- The ICU effect adjusted for bulk: for the MS1 score, mean-logit p 0.051, stacked binomial p 0.027, d0 quasi-binomial p 0.018. For IL1R2, mean-logit p 0.27, stacked binomial p 1.4e-5, d0 quasi-binomial p 5e-4.\n- Paired test samples: bulk IL1R2 alone has Spearman 0.82-0.87 with observed MS1 (R2 0.71-0.78); the generated cells have 0.46-0.64. Adding the generated value raises R2 by only 0.01-0.03 (partial rho 0.22-0.30).\n- Bulk-only controls: d0 generated MS1 follows bulk IL1R2 (rho 0.63; 0.70 within ICU), and 12 of 17 ICU samples have IL1R2 above the HC maximum.", + "issues": [ + "Calibration is weak, and only draw 0 carries the result. Draws 1-4 make MS1 at 0.03-0.14x of observed, so they fail the guide's level gate. The ICU-vs-HC contrast is not in calibration.csv. The paired check has 6 ICU vs 6-7 HC per draw, and the observed ICU-vs-HC difference is itself not significant (Mann-Whitney p 0.051; quasi-binomial p 0.38). In draw 0 the generated absolute difference is one third of observed (0.067 vs 0.210), although the fold matches (50x vs 52x).", + "The observed ICU MS1 level is bimodal. The 24.4% mean comes from IC16 (87%) and IC19 (53%); the other four ICU controls are 0.4-3.7% (median 2.5%). Every draw generates IC16 as the top ICU control, but every draw misses IC19 (0-1%). Within adult controls, generated-vs-observed Spearman is -0.33 to 0.43 by draw, against 0.64-0.76 for bulk IL1R2 alone.", + "The 'observed ordering where ICU controls are at or above sepsis' is fragile. It holds only for means driven by the same two samples: medians are 2.5% (ICU) vs 8.2% (sepsis), Mann-Whitney AUC 0.55 (p 0.70), quasi-binomial log-odds ratio 0.07 (p 0.89). Reyes 2020 reports MS1 slightly higher in sepsis than in ICU-NoSEP (FDR 0.27) and a bulk meta-analysis effect of 0.32 for sepsis over ICU controls. The defensible statement is narrower: the generator makes sepsis > ICU (stacked binomial +0.89, p 0.007; 5/5 draws) while the paired observed cells show no difference.", + "The d1 sign (-0.10, giving '4/5 same sign') is an artefact of the logit((n+0.5)/(N+1)) transform. In d1, ICU samples have 3.5x more generated monocytes (median 585 vs 175), and 32/36 HC and 7/17 ICU samples have zero MS1 cells, so the zero-count logits are driven by N. Count and rank tests are all positive: quasi-binomial +2.21 (p 0.0066), beta-binomial +1.63 (p 0.0076), Mann-Whitney p 4e-4, Fisher 10/17 vs 4/36 (p 5e-4). The sign is 5/5, not 4/5. This does not change the calibration gate.", + "The nanmean of draw logits after the N >= 10 filter mixes draws with very different levels whenever a draw is missing for a sample. ICU samples miss 10 sample-draw cells and HC samples miss 5. The complete-case estimate is +0.86 (p 1.1e-4) against the claimed +1.33, so the magnitude depends on the estimator; significance does not.", + "'Relative to sepsis, generated ICU MS1 is about 8-fold too low' does not reproduce. The ICU/sepsis ratio is 0.33-0.82 generated vs 1.18-1.52 observed per draw on paired test samples (d0: 0.40 vs 1.38, i.e. 3.4x). About 7.6x appears only when the five-draw mean ICU level (3.2%) is set against the observed paired ICU level (24.4%), which mixes uncalibrated draws and different sample sets.", + "'Not beyond bulk' is supported, and more strongly by the paired information test than by the adjusted p. The claim's 'after adjusting for IL1R2: p=0.27' holds only for the mean-logit model; the stacked binomial gives p 1.4e-5 and the d0 quasi-binomial p 5e-4. The stronger argument: in paired test samples, bulk IL1R2 alone explains R2 0.71-0.78 of observed MS1, and adding the generated cells adds 0.01-0.03.", + "The bulk-only ICU-vs-HC contrast is not specific to MS1. Many generated features shift. The top hit is the platelet-in-monocyte negative-control state (p 6e-12), and MS1 ranks 13 of 50. The generated MS1 shift is one readout of a global bulk shift; in bulk, IL1R2 is above the HC maximum in 12 of 17 ICU samples.", + "Minor: of draw 0's sepsis training donors, Skin-S74 has only 8 observed monocytes and 0 MS1 cells. Only Abd-S66 (38%), Res-S150 (82%) and Res-S77 (24%) are MS1-rich; the HC training donor HC07 is 0.8%.", + "Minor: in Reyes 2020 the 'expanded in ICU patients with sterile inflammation' sentence is on p.5 (Discussion), not p.4. Page 4 has the bulk meta-analysis (effect 0.32, FDR 0.08). The FDR 0.27 sentence is correctly on p.3.", + "Method caution for other lenses: in statsmodels 0.14.6, GLM Binomial with a 2-column (successes, failures) endog and fit(scale='X2') ignores n_trials in the Pearson scale. It gave scale 0.06 instead of 10.4, i.e. SEs about 13x too small. Use proportions with var_weights=N instead (v02b_qbin_check.log)." + ], + "reasoning": "The effect reproduces exactly, so it is not refuted, but the calibration is weak, so the verdict is qualified.\n\n**Reproduction.** I wrote my own code from cells.parquet and markers.npz. It reproduced the shared state table exactly, then every number in the claim: the +1.332 (SE 0.327) mean-logit effect, p 4.6e-5, BH q 1.8e-4, the per-draw coefficients, the five-draw means (3.2% / 0.12% / 7.0%), the calibration.csv rows and the observed paired numbers.\n\n**Confounders.** The draw 0 ICU > HC effect survived every control I tried:\n- age covariate, age squared, ages 40-80, age-overlap window, 1:1 matching and 5-year caliper matching\n- dropping sex-flag samples\n- parent cell number and monocyte library size\n- per-cell depth equalisation\n- bulk library size and QC\n- L2 label confidence, and stricter or simpler marker rules\n- leave-one-out\n- a bulk granulocyte score\n\nAge biases against the claim (MS1 rises with age in HC, and ICU controls are younger). ICU and HC come from the same site. Bulk-only donors are unique and never trained a draw. Count models make the direction 5/5 rather than 4/5.\n\n**Why qualified, not confirmed.**\n1. Only draw 0 is level-calibrated. Draws 1-4 make MS1 at 0.03-0.14x, so under the guide only draw 0 carries the result.\n2. The ICU-vs-HC paired check is underpowered (6 vs 6-7) and rests on one sample. Observed ICU MS1 is bimodal (IC16 and IC19). The generator recovers IC16 but misses IC19, and the observed ICU-vs-HC difference is not significant.\n3. The generated cells add nothing useful beyond the bulk. Bulk IL1R2 alone tracks observed MS1 better (Spearman about 0.85 vs about 0.55), and adding the generated value raises R2 by only 0.01-0.03.\n4. Several sub-statements need correction: the d1 sign artefact, the '8-fold' figure, the fragile 'ICU at or above sepsis' reference, Skin-S74, and the Reyes page number.\n\nThe claim's own framing ('weak, bulk-driven, only partly calibrated') fits this. A defensible report is: in draw 0, the only MS1-calibrated draw, bulk-only ICU controls get more MS1 cells than HC. The effect is robust to technical and demographic confounders, but it is a cell-level restatement of bulk IL1R2 and is not independently validated beyond one paired ICU sample.\n\nI ran CPU only and wrote only to my verify directory. This verification changed no paper evidence asset, so any archive update under the paper maintenance policy is left to the coordinating agent.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v01_recompute_ms1.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v01_ms1_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v01_check_vs_states.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v00_bulk_qc.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v02_bulkonly_models.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v02_perdraw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v02_stacked.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v02_groupmeans.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v02_loo.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v02b_qbin_check.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v03_paired_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v03_calibration_mine.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v03_calib_compare.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v03_controls_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v03_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v04_control_samples.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v04_paired_controls_gen_obs.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v05_bulk_derivability.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v05_bulk_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v05_adjusted.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v05_paired_beyond_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v06_reyes_pages.sh", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v07_bh_age_batch.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v07_bh_features.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v07_age.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v07_batch.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v08_key_numbers.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v08_key_numbers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v09_age_balance.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v09_age_balance.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v10_bulkonly_gen_vs_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_recompute/v10_bulkonly_gen_vs_bulk.csv" + ] + }, + { + "verifier": "calib", + "agent_id": "a6a2908fe798a81ad", + "verdict": "qualified", + "recomputed": "Independent recomputation, CPU only, in the lens dir (v01-v08).\n\nBULK-ONLY ADULTS (17 ICU / 36 HC / 68 sepsis), MS1 as fraction of monocytes\n- Five-draw mean: ICU 3.19%, HC 0.12%, sepsis 7.01%. Medians: 2.15%, 0.04%, 5.65%.\n- ICU vs HC logit (+age +sex, HC3), per draw: d0 +2.73 (p 5e-13), d1 -0.10 (0.77), d2 +0.51 (0.09), d3 +1.43 (2e-7), d4 +1.32 (0.010). This matches the claim exactly.\n- By rank, all 5 draws show ICU > HC (MWU p <= 4e-4, AUC 0.74-0.89).\n- Share of samples with at least one MS1 cell is higher in ICU in all 5 draws (Fisher p <= 0.003).\n- Draw as a fixed effect: ICU +1.20 (SE 0.27, p 7.7e-6). ICU x draw heterogeneity p 6e-21.\n- Group means, ICU vs HC: d0 12.7% vs 0.35%; d1-d4 0.44-1.4% vs 0.01-0.16%.\n- Median generated MS1 cells per ICU sample: d0 10; d1-d4 0-1.\n- Dropping the sex_flag samples (HC23, HC30) changes nothing.\n\nPAIRED CALIBRATION, ICU vs HC (computed here: gen vs observed_thin_d, 5-6 ICU vs 6-7 HC test adults)\n- Sign agrees 5/5.\n- Mean difference, gen vs obs: d0 0.067 vs 0.210 (0.32x); d1-d4 0.004-0.010 vs 0.20-0.25 (0.02-0.04x).\n- Share of the observed logit effect reproduced: d0 0.63; d1-d4 0.08-0.24.\n- ICU-group level ratio, d0: 0.32 by mean, 1.04 by median (1.65% vs 1.58%). d1-d4: 0.02-0.04.\n- Sepsis-group level ratio, d0: 1.15. So draw 0's all-sample 0.98 comes from the sepsis samples.\n- Sepsis vs ICU: the mean-difference sign agrees 0/5 (my code reproduces calibration.csv). On the logit scale it agrees 4/5.\n- At fixed bulk IL1R2, the generator gives ICU less MS1 than sepsis in 5/5 draws (offset -0.58 to -1.10 logit). Observed data show no such offset (+0.18, p 0.68; R2 0.82).\n\nOBSERVED PAIRED ADULTS\n- ICU (6) vs HC (7): mean 24.4% vs 0.67%; median 2.5% vs 0.54%; AUC 0.83; exact MWU p 0.051; exact permutation p 0.017 (mean) and 0.020 (logit); HC3-adjusted p 0.20.\n- Two samples carry the mean: IC16 87% and IC19 53%. The other four ICU are 0.4-3.7%. Without IC16, p 0.11.\n- Sepsis (46) vs ICU: mean 19.6% vs 24.4%, but median 8.2% vs 2.5%; AUC 0.55; logit +0.41 (p 0.77).\n\nBULK\n- Bulk-only ICU vs HC, 8-gene MS1 score: t 7.74 (p 1e-14, AUC 0.92). IL1R2 +3.89 log2 (AUC 0.96); RETN +2.30; CD163 +2.22.\n- Paired test samples: bulk IL1R2 alone ranks observed MS1 at rho 0.83-0.86, against 0.38-0.63 for generated MS1.\n- Generated MS1 adds nothing once the bulk is known (p 0.28-0.68). The bulk still adds once the generated values are known (p <= 0.023).\n- Paired bulk score, ICU vs HC: AUC 0.95 (exact p 0.005), stronger than the observed cells (AUC 0.83).\n- Bulk IL1R2 tracks observed MS1 (rho 0.83), not the LDG state (rho -0.04).\n\nDRAW 0 TRAINING DONORS (observed MS1)\n- HC07 0.8%, Res-S150 82%, Abd-S66 38%, Res-S77 24%.\n- Skin-S74: 0 MS1 cells out of only 8 monocytes, so uninformative.", + "issues": [ + "Calibration supports the direction only. The ICU>HC sign agrees on paired test samples in 5/5 draws, but the size of the difference is reproduced at 0.32x in d0 and 0.02-0.04x in d1-d4 (mean scale). Draws 1-4 fail the level gate (calibration.csv ratio 0.03-0.14). Only draw 0 can carry the result, as the claim says.", + "Draw 0's calibration.csv ratio of 0.98 is driven by sepsis samples (gen/obs 1.15). In ICU controls its mean ratio is 0.32, though the median matches (1.65% vs 1.58%). It misses the MS1-high ICU tail: IC19 observed 53% vs generated 1.0%; IC16 87% vs 34%.", + "Observed support is suggestive, with small n, and driven by outliers. 6 ICU vs 7 HC: exact permutation p 0.017, MWU p 0.051, HC3 p 0.20. IC16 and IC19 carry the 24.4% ICU mean; the ICU median is 2.5% vs 0.54% in HC. Without IC16, p 0.11.", + "The sepsis-vs-ICU sentence needs restating. In the paired set, 'ICU at or above sepsis' holds only for means. By median (2.5% vs 8.2%), rank (AUC 0.55) and logit (+0.41 for sepsis, p 0.77) the observed contrast is null or reversed, and gen and obs logit signs agree 4/5. However, Ye 2026's independent flow cohort (Extended Data Fig. 9e, PDF p.39) does put ICU controls above sepsis: Mono_c03-CD14-RETN about 78% vs 42-60% of CD14+ cells; IC>Abd p=0.0378, IC>Uri p=0.0067. The generator under-calls ICU MS1 at fixed bulk IL1R2 in 5/5 draws. So the generated sepsis>ICU ordering is a model bias and should not be reported.", + "Not novel; it was already reported in this cohort. Ye 2026 Fig. 4d (journal p.158, PDF p.9): the MS1-like Mono_c03-CD14-RETN is higher in IC vs HC (** P<0.01, Kruskal-Wallis + Dunn), with no significant IC vs sepsis-source differences. That analysis used the same scRNA-seq adults (8 ICU / 7 HC; Fig. 1b, p.152), so it is not independent of our observed check. Also in Ye 2026: Extended Data Fig. 9a (PDF p.39), deconvolution of the cohort's own bulk RNA-seq, shows HC < IC < Abd/Res; Extended Data Fig. 9e (PDF p.39), flow cytometry, IC vs HC p<0.0001. Elsewhere: Leite 2024 (p.1 abstract; p.9), MS1 from bulk deconvolution, raised in non-infected ICU patients vs HC and no different from sepsis.", + "Citation fix. The Reyes 2020 sentence 'expanded in ICU patients with sterile inflammation' is on PDF p.5 (Discussion), not p.4. Page 4 has the 7-dataset bulk meta-analysis: sepsis vs non-infectious SIRS ICU controls, effect size 0.32, FDR 0.08. The p.3 statement (FDR 0.27) is correct.", + "Fact fix. Skin-S74 is not MS1-rich: it has 0 MS1 cells among only 8 observed monocytes. Draw 0 has three MS1-rich sepsis training donors (Res-S150 82%, Abd-S66 38%, Res-S77 24%) plus HC07 (0.8%).", + "Cannot reproduce 'generated ICU MS1 is about 8-fold too low relative to sepsis'. No saved computation in sep_clinical produces it. The closest figures: generated ICU five-draw mean 3.2% vs observed paired ICU mean 24.4% (7.6x, an absolute gap across different samples); ICU/sepsis ratio by mean 0.46 generated vs 1.24 observed (2.7x); by median 0.38 vs 0.30.", + "The d1 logit null (-0.10) is an artefact: many zero counts combined with unequal monocyte counts (ICU median 585 vs HC 175). By rank and by the any-MS1-cell test, all 5 draws show ICU>HC. But in d1-d4 the effect rests on a median of 0-1 MS1 cells per ICU sample, i.e. noise level.", + "Adds nothing beyond bulk. Bulk alone shows ICU>HC more strongly (t 7.7, AUC 0.92-0.96). On paired test samples bulk IL1R2 alone predicts observed MS1 better than any draw (rho 0.83-0.86 vs 0.38-0.63), and generated MS1 adds nothing. The residual ICU effect beyond the bulk score in d0/d3 (p about 0.005) is not supported by the paired data.", + "Paper archive: no archive delta from this verifier. It only wrote verification tables and scripts under the workflow verify dir; any archive update belongs to the coordinating agent." + ], + "reasoning": "Verdict: qualified, not refuted. Calibration, observed data and biology all point the same way on the main claim (ICU > HC), but each only partly. The secondary statements need corrections.\n\n(1) Calibration. calibration.csv has no ICU-vs-HC contrast, so I computed it per draw: paired_gen vs observed_thin_d, on each draw's adult test samples. The ICU>HC sign is reproduced in 5/5 draws. Magnitude is reproduced only in part: d0 0.32x on the mean difference (0.63x on the logit effect); d1-d4 0.02-0.04x. The feature itself is level-calibrated only in d0 (ratio 0.98, Spearman 0.52). By the DATA_GUIDE trust rule, d1-d4 are not usable, so only draw 0 carries the result, as the claim says. Even d0 under-generates the MS1-high ICU controls. The sepsis-vs-ICU calibration failure (0/5 on the mean difference) is real. On paired samples, at fixed bulk IL1R2, the generator under-produces ICU MS1 relative to sepsis in all draws, while the observed data show no such offset. The generated sepsis > ICU ordering is therefore a model bias.\n\n(2) Observed paired cells. ICU > HC: exact permutation p 0.017, MWU p 0.051. This is suggestive, but n is 6 vs 7 and the effect is carried by 2 ICU samples. Ye 2026's independent flow cohort (Extended Data Fig. 9e) gives strong ICU > HC support (p<0.0001). It also places ICU at or above sepsis, which is where the generated ordering fails.\n\n(3) Biology and novelty. The pattern is plausible and already reported. The cohort paper itself shows the MS1-like Mono_c03-CD14-RETN higher in ICU controls than HC: Fig. 4d (p.158), in its own bulk deconvolution (Extended Data Fig. 9a) and by flow (Extended Data Fig. 9e). Reyes 2020 (pp.3-5) and Leite 2024 (p.1, p.9) show MS1 expansion in non-infected critically ill patients, and Leite 2024 recovers it from bulk deconvolution. Nothing here is new for this cohort.\n\n(4) Beyond bulk. Nothing. Bulk IL1R2 alone is a better MS1 readout than the generated cells on held-out paired samples in every draw.\n\nThe claim's own framing ('weak', 'bulk-driven, only partly calibrated', 'Reported', beyond_bulk 'No') is fair, and its per-draw and paired numbers reproduce. Remaining fixes: the Skin-S74 statement, the Reyes page number, the '8-fold' figure, a mean-only observed ordering in the paired set, and a missing citation to Ye 2026's own figures.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v01_paired_calib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v01_paired_calib.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v01_paired_calib_contrasts.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v01_paired_calib_group_levels.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v02_observed_paired.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v02_observed_paired.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v02_observed_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v02_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v03_bulkonly_per_draw.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v03_bulkonly_per_draw.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v03_bulkonly_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v03_bulkonly_gating.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v04_bulk_derivability.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v04_bulk_derivability.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v04_bulk_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v04_bulk_vs_gen_paired.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v04_bulk_scores.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v05_sensitivity.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v05_sensitivity.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v05_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v06_bulk_to_cell_mapping.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v06_bulk_to_cell_mapping.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v06_mapping.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v06_icu_residuals.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v07_training_donors.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v07_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v08_quoted_numbers.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v08_quoted_numbers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_pdf_text.sh", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_text_ye2026.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_text_reyes2020.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_text_leite2024.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_ye2026_p9_fig4d_header_crop.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_ye2026_p9_fig4d_crop.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_ye2026_p3_fig1b_table.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_ye2026_p39_EDF9a_thisstudy.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-5_calib/v00_ye2026_p39_EDF9e_RETN_zoom.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/features/ye2026_sepsis/calibration.csv", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Reyes2020_Immune_cell_signature_bacterial_sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Leite2024_MS1_deconvolution_critically_ill_sepsis.pdf" + ] + } + ] + }, + { + "id": "SEP-CLIN-2", + "lens": "sep_clinical", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 8, + "theme": "interferon", + "synthesis_strength": "weak lead: the only sepsis-vs-ICU signal; observed truth not significant", + "reportable_statement": "Generated ISG-high monocytes are more frequent in bulk-only sepsis than in non-septic ICU controls in all five draws (mean of draw logits +0.83 to +0.90, p 5e-5 to 3e-4); this is the only one of 50 features with q<0.05 for sepsis vs ICU. The direction is calibrated in 5/5 draws, but the observed paired contrast (6 ICU controls) is not significant and comes from 6 of 46 very-high-ISG patients, while the generator shifts the typical patient. 'Sepsis equals HC' is a generator artefact (observed sepsis < HC).", + "carrying_draws": { + "sepsis vs ICU direction": "0,1,2,3,4", + "sepsis vs ICU magnitude": "1,2,3,4 (0.38-0.96x); 0 at 0.09x", + "ICU vs HC": "3 (4 by the lens gate), at <=0.34x", + "sepsis vs HC": "none (wrong sign 5/5 after adjustment)" + }, + "honesty": { + "calibration": "calibration.csv ratio 0.43-0.93, Spearman 0.48-0.71; sepsis-vs-ICU sign 5/5; adjusted sepsis-vs-HC sign wrong 5/5; generator offset sepsis-HC +0.70 logit at equal truth (p 3.5e-6).", + "replication": "Sign 5/5; p<0.05 in 4-5/5 depending on the monocyte-count rule; beta-binomial 4/5; quasi-binomial 0/5.", + "observed": "Sepsis vs ICU +0.71 to +0.94 logit (p 0.20-0.35), equal medians (4.8% vs 4.4%); ICU vs HC -1.72 to -1.88 (p 0.003-0.053).", + "beyond_bulk": "Partly: bulk IFN programs do not separate sepsis from ICU (Hallmark IFN-alpha p 0.33); truth keeps +0.38 after bulk IFN adjustment, generated +0.68; but real paired cells separate them no better than bulk IFI27." + }, + "verdict_reasons": [ + "observed truth for sepsis vs ICU not significant", + "model-dependent (quasi-binomial null)", + "draw 0 reproduces 9% of the paired difference", + "generator over-produces IFI27 in monocytes", + "ICUICU not significant in Ye's abundance data; Fig. 4i per-cell IFN-alpha score has ICU controls lowest. Not established as new.", + "next_steps": [ + "More paired ICU controls (decisive gap: 5-6 per draw).", + "Median/quantile tests separating a high-ISG tail from a typical-patient shift.", + "Confirm with an external cohort with bulk + scRNA of ICU controls." + ], + "key_sources": [ + "verify/SEP-CLIN-2_recompute/v02_summary.txt", + "verify/SEP-CLIN-2_recompute/v11_gated.txt", + "verify/SEP-CLIN-2_calib/v10_summary.csv", + "verify/SEP-CLIN-2_calib/v03_group_quantiles.csv", + "verify/SEP-CLIN-2_recompute/v09_conditional.txt" + ] + }, + "finding": { + "id": "SEP-CLIN-2", + "cohort": "ye2026_sepsis", + "claim": "In bulk-only generated cells, ISG-high monocytes are lowest in non-septic adult ICU controls. They are higher in adult sepsis than in ICU controls (8.2% vs 2.4% of monocytes), which makes this the only feature separating sepsis from ICU controls at BH q<0.05. They are lower in ICU controls than in healthy controls (2.4% vs 6.5%). Sepsis and HC do not differ.", + "kind": "state", + "feature": "state | ISG-high monocyte (positive control): at least 4 of 10 ISGs detected in a monocyte", + "comparison": "Adult sepsis (68) vs adult ICU control (17); adult ICU control (17) vs adult HC (36)", + "effect_bulkonly": "Five-draw mean fraction of monocytes: sepsis 8.2%, ICU 2.4%, HC 6.5%. Adjusted logit coefficients: sepsis vs ICU +0.90 (SE 0.22); ICU vs HC -0.68 (SE 0.24); sepsis vs HC -0.05 (n.s.). The depth-normalised within-monocyte pseudobulk ISG CPM, sepsis vs ICU, log2 by draw: +0.68/+0.50/+0.67/+0.94/+0.36 (p<0.003 in d0-d3, p=0.077 in d4). The within-T ISG CPM is also higher (+0.35 to +0.42, all p<0.013). Median monocyte library size shows no consistent group difference (only d3 has p=0.03), so this is not a depth artifact.", + "stats": "Model: OLS with HC3 errors on the mean of the draw logits ~ group + age + sex. Sepsis vs ICU, n=85: p=5.0e-5, BH q=0.0025 over 50 features; asin q=0.004; rank-ANCOVA q=0.008; MWU AUC 0.73, q=0.037. ICU vs HC, n=53: p=0.0052, q=0.014; rank-ANCOVA q=0.011. Sensitivity for sepsis vs ICU: without the sex-flag samples q=0.0024; without the Emergency-Center sepsis samples q=0.0043; age 40-80 p=6.5e-4; 1:1 age- and sex-matched sepsis p=0.017. Gated on draws 0, 3 and 4 only: coefficient 0.91, p=2e-4.", + "draws_same_sign": "Sepsis vs ICU: 5/5, all 5 with p<0.05 (d0 0.97, d1 0.90, d2 0.91, d3 1.34, d4 0.57); per-draw BH q<0.05 in d1, d2 and d3. ICU vs HC: 5/5, 3/5 with p<0.05. Calibration-gated carrying draws: sepsis vs ICU 0,3,4; ICU vs HC 3,4. Draws 1 and 2 fail only on ratio (<0.5). Draw 4 was trained on an adult ICU control (IC23) and shows the weakest within-monocyte ISG shift.", + "paired_calibration": "calibration.csv, ISG-high monocyte, draws 0-4: ratio 0.75/0.47/0.43/0.93/0.60, Spearman 0.52/0.60/0.48/0.52/0.71. 'Adult sepsis vs adult ICU control': same sign 5/5 (gen +0.008/+0.050/+0.051/+0.087/+0.049 vs obs +0.090/+0.102/+0.105/+0.090/+0.128). The same-model paired coefficient also agrees 5/5 (gen 0.15-1.36 vs obs 0.59-1.14). ICU vs HC, from my own paired check: 4/5 same sign, generated at 0.15-0.45x of observed, d0 wrong sign. The third contrast fails: 'adult sepsis vs adult HC' agrees in raw sign 4/5, but the age- and sex-adjusted paired coefficient has the wrong sign in 5/5 draws (gen +0.02 to +0.87 vs obs -0.84 to -1.09).", + "paired_observed_concordance": "Observed paired adults. Sepsis vs ICU (46 vs 6): 14.9% vs 6.4% (logit +0.71, p=0.35; thinned +0.73, p=0.34). ICU vs HC (6 vs 7): 6.4% vs 17.9% (-1.72, p=0.053; rank-ANCOVA p=0.011; thinned p=0.048). Within-monocyte ISG CPM, sepsis 2393 vs ICU 677 (log2 +0.50, p=0.35). The ordering, with ICU lowest, is the same, but the observed contrasts are underpowered. After adjustment, observed sepsis is below HC (-0.85, p=0.031); the generated cells do not reproduce this.", + "beyond_bulk": "Partly beyond the bulk. For sepsis vs ICU, the 10-gene bulk ISG score differs only weakly (+0.38 z, p=0.12, q=0.33). Single bulk genes: IFI27 +1.19 log2 (p=0.0045), IFI6 +0.60 (p=0.016), IFIT3 +0.68 (p=0.04); the other 7 are n.s. The cell-level effect survives adjustment for the bulk ISG score (+0.63, p=1.8e-4). So the cells separate sepsis from ICU controls more cleanly than the crude bulk score, and resolve the shift to monocytes and T cells, but the bulk does carry the IFI27/IFI6 differences. For ICU vs HC, the bulk alone shows it: ISG score p=0.015; IFI44L -1.62, RSAD2 -1.03, IFIT1 -1.14 log2 (q<=0.0014).", + "novelty": "Not stated as such. Ye 2026 reports that interferon-inducible Mono_c05-CD14-IFITM1 monocytes are reduced in most sepsis vs HC (PDF p.6). It also reports an interferon-responsiveness MOFA program attenuated in both sepsis and ICU controls (p.14). The sepsis > ICU-control ISG-monocyte contrast, and ICU < HC with 17 vs 36 samples, are not in the text.", + "strength": "moderate", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s02_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s03_paired_calib_long.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s04_observed_assoc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s06_feature_vs_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s07_artifact_checks.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s07_mono_isg_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s08_bulk_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s08_age_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s10_evidence.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s12_gated.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "ace14208435dd1d2d", + "verdict": "qualified", + "recomputed": "I recomputed everything from the raw bulkonly, paired_gen and observed cells.parquet and markers.npz files with my own code. My ISG-high monocyte calls (≥4 of 10 ISGs in AIFI-L1 monocytes) match states.parquet sample for sample: all 122 bulk-only and 77 paired-gen samples in every draw.\n\n**Group levels.** Five-draw mean fraction: sepsis 8.20%, ICU 2.41%, HC 6.46%. Per draw (S/I/H): d0 10.4/4.1/7.0; d1 6.9/2.2/6.7; d2 4.9/0.6/2.1; d3 11.3/2.3/8.5; d4 6.9/2.8/8.0.\n\n**Headline model.** The discovery agent's rule sets draw fractions with fewer than 10 monocytes to NaN and then averages over the remaining draws. With that rule I reproduce the claim exactly: S vs I +0.898 (SE 0.221, p=5.0e-5), I vs H -0.682 (p=0.0052), S vs H -0.055 (p=0.76).\n- Without that rule (only zero-monocyte samples dropped): S vs I +0.834 (p=3.3e-4), I vs H -0.538 (p=0.024).\n- Stacked model with draw as fixed effect and sample-clustered errors: S vs I +0.821 (p=1.3e-4). Heterogeneity across draws: p=0.085 for S vs I, 0.0028 for I vs H, 1e-11 for S vs H.\n\n**S vs I per draw.**\n- Logit OLS, all samples: 0.94/0.90/0.52/1.08/0.66; 4/5 with p<0.05 (d2 p=0.16).\n- Logit OLS, with the <10-monocyte rule: 0.97/0.90/0.91/1.34/0.57; 5/5 with p<0.05.\n- Beta-binomial: 0.76/0.53/0.78/0.87/0.46 (p 0.019/0.044/0.046/0.021/0.11).\n- Quasi-binomial (count-weighted, phi about 90): p=0.07–0.15 in every draw.\n- Cell-level logistic with a spline in each cell's library size, sample-clustered: 0.99/1.10/1.01/1.18/0.60, all p<0.025.\n\n**I vs H per draw.**\n- Logit OLS, all samples: -0.67/-0.87/+0.21/-0.65/-0.71.\n- With the <10-monocyte rule: -0.63/-0.87/-0.43/-0.81/-0.74.\n- Beta-binomial: negative in 5/5.\n\n**Within-monocyte pseudobulk ISG CPM (≥10 monocytes), log2.** S vs I: +0.68/+0.50/+0.67/+0.94/+0.36, which reproduces the claim. I vs H: +0.10/-0.04/-0.26/-0.32/-0.66; only d4 has p<0.05.\n\n**Confounders, S vs I** (five-draw mean logit; all-samples rule / <10-monocyte rule):\n- Age spline: 0.66 (p=0.003) / 0.72 (p=2e-4).\n- Age 40–80: 0.83 / 0.88 (p≤0.002).\n- Nearest-age same-sex matched, 15 ICU vs 37 sepsis: 0.85 / 0.88 (p≤7e-4).\n- Without sex-flag samples; without Emergency-Center sepsis: both p≤4e-4.\n- Plus log monocyte count; plus log monocyte library: 0.85–0.90.\n- High-confidence monocytes only 0.82 / 0.92; CD14 only 0.80 / 0.97; plus subtype shares 0.76 / 0.81.\n- Plus bulk QC 0.70 / 0.77; plus bulk 10-ISG 0.58 / 0.63; plus bulk Hallmark IFN-alpha 0.67 / 0.73.\n- All covariates together: 0.38 (p=0.033) / 0.43 (p=0.0096).\n- Single-thinning depth-standardised call: 0.50 (p=0.20) / 0.68 (p=0.033).\n- Leave-one-ICU-out: max p 6.8e-4. Leave-one-draw-out: max p 8.5e-4.\n- By source vs ICU: respiratory 0.95 (p=0.004), urinary 0.85 (p=0.02), abdominal 0.43 (p=0.13), non-respiratory 0.68 (p=0.004).\n- Monocyte library does not differ S vs I in any draw (all p>0.05).\n\n**Confounders, I vs H:**\n- Age-matched 9 vs 9: -0.26 (p=0.35); 10 vs 19: -0.52 (p=0.054).\n- Plus bulk IFN-alpha: -0.20 (p=0.31). Plus bulk monocyte/granulocyte proxies: -0.22 (p=0.51). Plus subtype shares: -0.43 (p=0.13).\n- Leave-one-ICU-out: max p 0.064.\n\n**50-feature screen.** ISG-high monocyte is the only feature with S vs I q<0.05 (q=0.017 all-samples rule; q=0.0025 with the <10-monocyte rule). I vs H moves dozens of features, including the negative control \"platelet signal in monocytes\" (p=1e-9).\n\n**Calibration recheck** (matches calibration.csv):\n- Level ratio 0.76/0.47/0.43/0.92/0.59; Spearman 0.51/0.61/0.50/0.52/0.68.\n- Paired S−I raw difference, generated/observed: 0.09/0.49/0.48/0.96/0.38; same sign 5/5.\n- Paired I−H: raw sign agrees in 3/5 draws (d0 wrong, d2 zero); magnitude at most 0.34× of observed; adjusted sign 4/5.\n- Observed paired truth (46 S / 6 I / 7 H): S vs I +0.94 (p=0.20; rank-ANCOVA p=0.44; MWU p=0.47; medians 4.8% vs 4.4%). I vs H -1.72 (p=0.053). S vs H -0.65 (p=0.099; rank-ANCOVA p=4e-4).\n\n**Generator offset at equal observed truth** (paired test adults, stacked, clustered): S−H +0.70 (SE 0.15, p=3.5e-6); S−I +0.34 (SE 0.30, p=0.26); I−H +0.37 (p=0.26).\n\n**Bulk.**\n- Bulk IFN programs do not separate S from I: Hallmark IFN-alpha +0.15 z (p=0.33), IFN-gamma p=0.24, rule-10 score p=0.13.\n- They do separate I from H (IFN-alpha p=0.006), and S is below H (IFN-alpha -0.39 z, p=4e-5).\n- In paired adults, bulk IFN scores predict the observed ISG-high monocyte fraction with rho 0.78–0.83.\n- Truth conditional on bulk IFN-alpha: S−I +0.38 (p=0.026), S−H -0.05. Generated bulk-only, same model: S−I +0.68 (p=2e-5), S−H +0.38 (p=0.012).\n\n**Gated summaries:**\n- S vs I: claim gate {0,3,4} +0.90 (p=1.8e-4); contrast-reproduction gate {1,2,3,4} +0.81–0.91 (p≤8.6e-4); strict gate {3} +1.08–1.34 (p≤0.001).\n- I vs H: no draw passes the strict gate; contrast gate {3} -0.65 to -0.81 (p≈0.03–0.04).", + "issues": [ + "The headline p-values depend on the discovery agent's MIN_PARENT=10 + nanmean rule. That rule averages each sample over a different subset of draws and drops 16/68 sepsis and 6/17 ICU samples in draw 2. Without it: S vs I p=3.3e-4 (still the only q<0.05 feature), I vs H p=0.024, draw-2 S vs I loses significance (p=0.16) and its I vs H sign flips to +0.21. The claim 'all 5 draws p<0.05' holds only under that rule.", + "The result depends on the count model. A quasi-binomial (count-weighted) S vs I is n.s. in every draw (p=0.07-0.15). A beta-binomial supports it in 4/5 draws (d4 p=0.11). A cell-level logistic with a per-cell library spline supports it in 5/5.", + "The calibration gate is level-based, not contrast-based. Draw 0 is listed as a carrying draw but reproduces only 9% of the paired sepsis-ICU difference (0.008 vs 0.090). Draws 1 and 2, which are excluded, reproduce 48-49%. The effect survives either gate (claim gate p=1.8e-4; contrast gate {1,2,3,4} p<=8.6e-4; draw 3 alone p<=0.001).", + "The observed paired truth for sepsis vs ICU is itself not significant: 6 ICU controls, adjusted +0.94 p=0.20, rank-ANCOVA p=0.44, MWU p=0.47, equal medians (4.8% vs 4.4%). The observed mean gap comes from a few very high-ISG sepsis donors.", + "Generator offset. At equal observed truth (paired test adults) the generator puts more ISG-high monocytes into sepsis than HC by +0.70 logit (p=3.5e-6). The sepsis-vs-ICU offset is +0.34 (SE 0.30, 95% CI about -0.24 to 0.92). With 6 paired ICU controls this cannot exclude that a large part (point estimate ~40%) of the bulk-only gap of 0.83-0.90 is model offset rather than measurement.", + "Beyond-bulk is real but smaller than generated. No a priori bulk IFN program separates sepsis from ICU (Hallmark IFN-alpha p=0.33, IFN-gamma p=0.24, rule-10 p=0.13), and bulk IFN predicts the observed state with rho 0.78-0.83. The truth itself has a sepsis>ICU increment conditional on bulk IFN (+0.38, p=0.026), which supports a genuine beyond-bulk component. The generated conditional increment (+0.68) exceeds it.", + "ICU < HC is bulk-derived and poorly calibrated. The bulk shows it directly (IFN-alpha p=0.006; IFI44L, RSAD2, IFIT1 are the genes that drive it in the generated monocytes too). No draw reproduces >=0.5x of the paired gap and d0 has the wrong sign.", + "ICU < HC is fragile. Age-matched: p=0.054-0.35. It vanishes after bulk IFN-alpha or monocyte/granulocyte-proxy adjustment (p=0.16-0.56). Leave-one-ICU-out gives max p=0.064. Within-monocyte ISG CPM and ISG UMIs per 10k show it only in draw 4, which was trained on ICU control IC23. ICU vs HC also shifts dozens of features, including the platelet-in-monocyte negative control (p=1e-9).", + "'Sepsis and HC do not differ' is a generator artifact and should be dropped. The bulk shows sepsis < HC (IFN-alpha -0.39 z, p=4e-5). The paired truth shows sepsis < HC (adjusted -0.65 to -0.84, rank-ANCOVA p=4e-4). The generator inflates sepsis vs HC at equal truth (+0.70).", + "Age. ICU controls are 16 years younger (mean 49 vs 65). HC show a positive within-group age slope (+0.22 logit per 10 y, p=0.04). Sepsis vs ICU survives age-matching (15 vs 37, p=7e-4), but the gap shrinks at older ages (at age 70: +0.52, p=0.12; interaction p=0.12).", + "Depth. Monocyte library size does not differ between sepsis and ICU, and cell-level library adjustment keeps the effect. A strict depth-standardised call attenuates it (all-samples rule p=0.20, <10-monocyte rule p=0.033) but keeps a positive sign in 5/5 draws.", + "Not applicable to this cohort: repeated measures, and donor-trained samples (every bulk-only sepsis donor is unique and none trained a draw). Draw-specific control training donors (IC23 in d4; PIC01 in d2, observed ISG-high 0.73) plausibly shape those draws' ICU contrasts.", + "Minor: statsmodels 0.14.6 mis-estimates scale='X2' for a two-column binomial endog, so I set the Pearson-chi2/df scale explicitly. My first all-samples run scored zero-monocyte samples as logit 0; this was corrected before every table quoted here." + ], + "reasoning": "**Part by part:**\n- **Sepsis > ICU reproduces and survives confounder controls.** The numbers reproduce exactly under the discovery agent's rule and keep their direction under every rule I tried. It survives age (spline, restriction, nearest-age same-sex matching), sex, sex-flag removal, collection site, monocyte count, library size (including per-cell adjustment), label confidence (high-confidence and CD14-only monocytes), monocyte subtype mix, bulk sequencing QC, bulk IFN scores, bulk composition proxies, leave-one-out, and infection source. It is the only one of 50 features with q<0.05. So it is not refuted.\n- **Sepsis > ICU: the calibration is weaker than presented.** Draw 0, one of the claimed carrying draws, reproduces only 9% of the paired sepsis-ICU gap. The paired observed truth is itself not significant (6 ICU controls, equal medians). At equal observed truth the generator over-produces ISG-high monocytes in sepsis relative to HC (+0.70, p=3.5e-6) and possibly relative to ICU (+0.34, CI up to about the full effect). The magnitude of the bulk-only gap is therefore uncertain, though the truth does show a beyond-bulk-IFN sepsis>ICU increment (+0.38, p=0.026).\n- **ICU < HC is weak and bulk-derived.** It is fragile under age matching and adjustment, and the generator reproduces it at most 0.34× in paired tests.\n- **'Sepsis = HC' is contradicted.** Both the bulk and the paired truth show sepsis < HC; the generated equality comes from the model's sepsis inflation.\n\n**Draws carrying the sepsis-vs-ICU result:**\n- All five agree in sign under every rule and in the beta-binomial model.\n- The claim's level-based gate gives {0,3,4}; a contrast-reproduction gate gives {1,2,3,4}; only draw 3 passes a strict gate. All three gated summaries are significant.\n- For ICU vs HC only draw 3 (and, under the claim gate, draw 4) carries it, at 0.22–0.34× reproduction.\n\n**Suggested strength:** sepsis vs ICU moderate-to-weak (calibrated sign, uncertain magnitude); ICU vs HC weak and already in the bulk; the sepsis-vs-HC statement not supported.\n\n**Scope:** CPU only and no GPU work; I wrote only into the SEP-CLIN-2_recompute lens dir. There is no PBISC-Diffusion archive delta from this verifier: it changed no paper asset, so the task-end archive update stays with the coordinator.\n\n**Run order:** v01 → v02/v02b → v04 → v03 → v05 → v07 → v06 → v08 → v09 → v10 → v11. v03 needs v04's covariates, v09 needs v08's scores, v11 needs v06.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v01_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v01_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v01_crosscheck.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v02_assoc.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v02_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v02_summary.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v02b_min_parent.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v02b_min_parent.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v03_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v03_summary.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v03_variants.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v03_betabinom.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v03_matched.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v03_age.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v03_influence.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v04_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v04_bulk.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v04_bulk_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v04_gene_screen.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v05_depth.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v05_depth.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v06_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v06_calibration.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v06_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v06_observed_assoc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v06_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v07_genes.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v07_genes.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v08_bulk_ifn.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v08_bulk_ifn.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v09_conditional.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v09_conditional.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v10_screen.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v10_screen.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v11_gated.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_recompute/v11_gated.txt" + ] + }, + { + "verifier": "calib", + "agent_id": "ac3c50eb97adc6179", + "verdict": "qualified", + "recomputed": "All numbers below were recomputed from states.parquet, meta.csv, the per-cell work files and the bulk npz files. The sep_clinical intermediates were not used.\n\nBULK-ONLY RESULT (reproduced)\n- Sepsis vs ICU, logit coefficient by draw: d0 0.97 (p=0.004), d1 0.90 (p=1e-4), d2 0.91 (p=8e-4), d3 1.34 (p=2e-4), d4 0.57 (p=0.021).\n- Sepsis vs ICU, mean logit across draws: 0.90 (SE 0.22, p=5e-5, AUC 0.74). Dropping any one of the 17 ICU controls leaves every draw at p<0.04.\n- ICU vs HC, logit coefficient by draw: -0.63, -0.87, -0.43, -0.81, -0.74 (3 of 5 draws p<0.05). Mean across draws -0.68 (p=0.005).\n- Sepsis vs HC: -0.05 (p=0.76).\n- Gating on the draws whose own paired test reproduces the contrast:\n - sepsis vs ICU, draws 1-4: 0.91 (p=2e-5);\n - ICU vs HC, draws 2-4: -0.57 (p=0.010);\n - the lens's choice of draws 0, 3, 4 gives 0.91 (p=2e-4).\n- With every monocyte downsampled to the same depth (1,000 UMI) and draws stacked:\n - sepsis vs ICU 0.66 (p=0.016), but only draws 0 and 3 have p<0.05 on their own;\n - ICU vs HC -0.61 (p=0.018).\n\nPAIRED CALIBRATION (adults; thinned observed cells; adjusted logit)\n- Sepsis vs ICU:\n - generated 0.15 / 0.69 / 1.36 / 1.24 / 0.64 vs observed 0.77 / 0.59 / 0.86 / 0.71 / 1.14;\n - stacked: generated 0.81 (p=0.019) vs observed 0.80 (p=0.21).\n - Draw 0 does not reproduce it: AUC 0.49, and its paired ICU median (9.2%) is above its sepsis median (6.7%).\n- ICU vs HC:\n - generated +0.32 / -0.29 / -0.48 / -0.84 / -0.62 vs observed -1.71 / -1.95 / -1.89 / -1.87 / -2.09;\n - stacked: generated -0.38 (p=0.31) vs observed -1.88 (p=0.003).\n- Sepsis vs HC:\n - generated +0.47 / +0.36 / +0.87 / +0.35 / +0.02 vs observed -0.84 to -1.09, the wrong sign in all 5 draws;\n - stacked: generated +0.41 vs observed -0.94 (p=0.007);\n - per sample, the generated value sits +0.79 to +1.73 logit above observed for sepsis relative to HC (p≤0.03 in every draw).\n\nOBSERVED PAIRED CELLS (native depth)\n- ISG-high monocyte fraction, median (mean):\n - HC 16.7% (17.9%)\n - ICU 4.4% (6.4%)\n - sepsis 4.7% (14.9%)\n- Without the 6 sepsis patients who have very high interferon-stimulated-gene (ISG) levels, sepsis is 4.5% median (6.1% mean), the same as ICU.\n- Median regression: sepsis vs ICU -0.002 (p=0.96); sepsis vs HC -0.12 (p=0.0005).\n- At equal per-cell depth:\n - sepsis vs ICU +0.49 (p=0.62, AUC 0.47);\n - ICU vs HC -2.13 (permutation p=0.04);\n - sepsis vs HC -1.38 (p=0.004).\n\nBULK\n- A leave-one-out ridge on the 10 bulk ISGs plus a monocyte score predicts the generated feature with R²=0.66 and uses no group labels.\n- That bulk-only predictor already gives sepsis vs ICU +0.59 (p=0.012, AUC 0.68).\n- After adjusting for the 10 ISGs and the monocyte score, the generated feature gives:\n - sepsis vs ICU +0.37 (p=0.010);\n - ICU vs HC -0.36 (p=0.27).", + "issues": [ + "Calibration and observed data both contradict 'Sepsis and HC do not differ'. The adjusted paired coefficient has the wrong sign in 5 of 5 draws (stacked: generated +0.41 vs observed -0.94, p=0.007). The generator places sepsis +0.79 to +1.73 logit too high relative to HC in every draw (p≤0.03). In the observed cells sepsis is below HC: median 4.7% vs 16.7%, median regression p=0.0005, p=0.004 at equal depth. The bulk itself shows sepsis below HC for the classic ISGs: IFI44L -1.39, RSAD2 -0.84 and IFIT1 -0.88 log2, q≤0.001. Ye 2026 says the interferon-inducible Mono_c05-CD14-IFITM1 subset is reduced in most sepsis (PDF p.6, journal p.155; Fig. 4d, PDF p.9). This sentence has to be dropped.", + "Observed support for sepsis > ICU exists only at the level of the mean, and it comes from 6 of 46 sepsis patients with very high ISG levels: Res-S139 98%, Res-S150 78%, Abd-S79 78%, Res-S166 74%, Uri-S157 48%, Skin-S142 38%. Without them, observed sepsis equals ICU: mean 6.1% vs 6.4%, median 4.5% vs 4.4%. Median regression gives -0.002 (p=0.96), and the observed contrast is not significant in any draw (p=0.09-0.46). The bulk-only generated effect is different in kind: it shifts the typical patient (AUC 0.74) and survives removal of each draw's top 10% (coefficients 0.45-0.99, p≤0.06). The same artifact shows in the paired test. After removing the same 6 patients from both sets, draws 1-3 still give a generated coefficient of 0.53-1.05 (AUC 0.62-0.67), while the observed coefficient is 0.07-0.21 (AUC 0.50).", + "Observed depth confound. Observed sepsis monocytes are deeper than ICU monocytes (per-sample median library 3,372 vs 2,566 UMI). The single global thinning factor per draw does not remove this, so the detection-based state call inflates the observed sepsis-ICU difference. With every monocyte downsampled to 1,000 UMI, observed sepsis vs ICU becomes +0.49 (p=0.62, AUC 0.47). The bulk-only effect also weakens at equal depth: stacked 0.66 (p=0.016), with only 2 of 5 draws at p<0.05.", + "The lens gated the wrong draws for sepsis vs ICU. Draw 0 passes on the level ratio, but it does not reproduce the contrast in its own paired test. calibration.csv gives a generated difference of 0.008 vs 0.090 observed. The adjusted coefficient is 0.15 vs 0.77 (p=0.76), AUC 0.49, and its paired ICU median is above its sepsis median. Draws 1-4 reproduce the contrast (0.64-1.36 vs 0.59-1.14). The bulk-only conclusion does not change: draws 1-4 give 0.91, p=2e-5.", + "ICU < HC is weakly calibrated. In draws 1-4 the generator reproduces 0.15-0.45x of the observed coefficient, and none is significant on its own; draw 0 has the wrong sign. Stacked: generated -0.38 (p=0.31) vs observed -1.88 (p=0.003). Only draws 3-4 (and draw 2 at about 0.25x) carry it. The observed data support it (AUC 0.12; permutation p=0.04 at equal depth), but the bulk alone already shows it: bulk ISG score p=0.015, IFI44L/RSAD2/IFIT1 q≤0.003. After adjusting for bulk ISGs and the monocyte score, the generated effect is -0.24 to -0.36 (p=0.27-0.43). The cells add nothing beyond the bulk here.", + "The novelty statement is inaccurate. Ye 2026 Fig. 4d (PDF p.9, journal p.158) shows Mono_c05-CD14-IFITM1 reduced in ICU controls vs HC (Dunn's test, **). The same heat map shows no significant difference between ICU controls and any sepsis source. MOFA factor 3 ('interferon responsiveness') is 'reduced in both sepsis and ICU controls' (PDF p.14, journal p.163). Only Fig. 4i, a per-cell test, has ICU controls lowest and respiratory sepsis highest on the 'interferon alpha response' score. So ICU < HC is already published. Sepsis > ICU is not significant in the paper's own single-cell data.", + "The 'beyond bulk' value is not supported by the paired data. A bulk-only readout explains 65% of the generated feature. In the paired data the real cells separate sepsis from ICU no better than the bulk does: observed AUC 0.59, paired bulk ISG score AUC 0.52, bulk IFI27 AUC 0.84. Two signs that the extra separation comes from the generator: (1) generated monocytes follow bulk IFI27 more than real ones do (rho 0.62-0.64 vs 0.34 in draws 1-2); (2) in draws 0-3, generated sepsis monocytes carry a median of 8-100 CPM of IFI27, against about 1 CPM in real monocytes of every group. Within-monocyte pseudobulk tells the same story. For the other 9 ISGs, the generator gives sepsis vs ICU +0.64 to +1.07 log2 (AUC 0.72-0.89), against +0.46 observed (p=0.38). It gives sepsis vs HC +0.49 to +0.89 in draws 0-3, against -0.47 observed (AUC 0.25). Dropping IFI27 from the state rule changes nothing.", + "The paired truth for every ICU contrast rests on the same 6 ICU controls (5 in draw 4). 'Same sign in 5 of 5 draws' means five models scored on the same 6 samples, not five independent replications. These 6 controls match the 17 bulk-only ICU controls on bulk ISGs (all p>0.1). The training donors explain the differences in level between draws: draws 0, 1, 3 and 4 each trained on a sepsis donor with very high ISG levels (Res-S150, Res-S166, Res-S139, Res-PS22), draw 2 on a high-ISG paediatric ICU control (PIC01), and draw 4 on IC23, the only adult ICU control with high ISG levels." + ], + "reasoning": "I recomputed everything from the raw feature and cell files.\n\n**What holds up**\n- The bulk-only replication is solid: sepsis > ICU in 5 of 5 draws. It is robust to dropping any one ICU control, to regating on the draws that reproduce the contrast, and (in the stacked model) to equal per-cell depth.\n- The calibration trust layer passes the sepsis vs ICU direction in all 5 draws. The stacked paired magnitude matches (0.81 generated vs 0.80 observed).\n- The direction is consistent with the observed means, with Ye 2026 Fig. 4i, and with the published reduction in ICU controls.\n\n**What weakens it**\n- The claimed pattern (ICU uniquely low, sepsis at HC level) is not what the paired data show.\n - The observed cells give HC highest, with typical sepsis equal to ICU.\n - The whole observed sepsis-ICU difference comes from 6 of 46 patients with very high ISG levels, and it is not significant in any draw.\n - The generator instead shifts the typical sepsis patient, both in the paired test (draws 1-3) and in the bulk-only set.\n- 'Sepsis and HC do not differ' is contradicted by:\n - the paired calibration (wrong sign in 5 of 5 draws);\n - the observed cells;\n - the bulk's classic ISGs;\n - Ye 2026 (PDF p.6 and Fig. 4d).\n- ICU < HC is real. But the generator calibrates it poorly (about 20% of the observed effect; only draws 2/3-4 carry it). It is already published (Ye 2026 Fig. 4d and the MOFA factor 3 on PDF p.14), and it is fully explained by the bulk.\n- The lens wrongly counts draw 0 as carrying sepsis vs ICU.\n- The 'cleaner than bulk' separation is not seen in the real paired cells.\n\n**Verdict.** Support is partial, so the verdict is 'qualified' rather than 'confirmed'. One sub-claim is refuted outright: 'sepsis and HC do not differ'. A narrower statement survives, with weak strength: 'Bulk-only cells reproduce the published ISG-monocyte deficit in ICU controls vs HC. They also give a higher mean in sepsis than in ICU controls, which in the paired truth comes from a minority (6 of 46) of sepsis patients with very high ISG levels, not a shift of the typical patient.'\n\n**Scope.** No GPU work was run. Every script and table is in the lens directory. Nothing was written to the paper archive; whether an archive update is needed is for the coordinating agent to decide.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v01_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v01_paired_contrasts.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v01_paired_stacked.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v01_controls_per_sample.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v02_depth_equalized.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v02_depth_eq_results.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v03_distribution.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v03_group_quantiles.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v03_quantreg_tail.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v03_train_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v03_loo_icu.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v04_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v04_bulk_contrasts.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v04_partial_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v05_ifi27_source.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v05_within_mono_isg.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v06_rule_without_ifi27.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v06_variants.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v07_tail_and_source.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v07_tail_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v07_source_vs_icu.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v08_bulk_mapping_paired.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v08_offset.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v08_bulk_mapping.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v09_paper_pages.sh", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v09_paper_pages.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/ye_p9_fig4d.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/ye_p9_fig4i.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v10_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v10_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v10_gated_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v11_icu_comparability.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v11_icu_comparability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v12_observed_controls.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-2_calib/v12_observed_controls.csv" + ] + } + ] + }, + { + "id": "SEP-CLIN-4", + "lens": "sep_clinical", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 9, + "theme": "interferon", + "synthesis_strength": "weak: bulk-derived, not BH-significant", + "reportable_statement": "Generated cells reproduce, in direction only, the lower interferon signal of abdominal sepsis (ISG-high T cells, mean-of-draws logit -0.71, p 0.0027, q 0.067). The bulk ISG score shows it more strongly and removes it entirely. The monocyte half does not hold.", + "carrying_draws": { + "ISG-high T": "1,3 by gate, but 3 is depth-confounded and 1 depends on the 10 Emergency-Center samples; none under combined controls" + }, + "honesty": { + "calibration": "ISG-high T ratio 0.35-0.69, Spearman 0.28-0.66; sepsis-vs-HC sign 5/5 on raw means, 2/5 on the logit scale; source contrast sign 5/5 on paired tests but observed p 0.14-0.30 and shared samples.", + "replication": "5/5 negative; logit OLS 4/5 p<0.05; beta-binomial 1/5; quasi-binomial 0/5.", + "observed": "15 vs 31: -0.84 (p 0.10), rank-ANCOVA p 0.044; vanishes after bulk ISG adjustment.", + "beyond_bulk": "None: bulk 10-ISG -0.65 to -0.72 z (p 4e-4); generated effect after bulk adjustment -0.18 (p 0.33)." + }, + "verdict_reasons": [ + "mean driven by three high-ISG 'other' samples", + "depth confound in draw 3", + "flexible age and site attenuate", + "multiplicity (one of four one-vs-rest contrasts; omnibus q 0.37)" + ], + "new_vs_cohort_papers": "Largely reported: Ye 2026 Fig. 4i (abdominal monocytes lowest IFN-alpha; pulmonary highest); Scicluna 2025 (interferon subtype mostly pneumonia).", + "next_steps": [ + "Pre-specified omnibus test; depth-matched state rule; external cohort with infection-source labels." + ], + "key_sources": [ + "verify/SEP-CLIN-4_recompute/v10_summary.csv", + "verify/SEP-CLIN-4_calib/v01_paired_by_draw.csv", + "verify/SEP-CLIN-4_calib/v02_bulk.csv" + ] + }, + "finding": { + "id": "SEP-CLIN-4", + "cohort": "ye2026_sepsis", + "claim": "Within adult sepsis, generated cells from abdominal-source samples have fewer ISG-high T cells (0.3% vs 1.4% of T cells) and ISG-high monocytes (4.6% vs 9.9%) than other sources. The cell-level effect does not pass BH and is weaker than the same signal in the bulk.", + "kind": "state", + "feature": "state | ISG-high T; state | ISG-high monocyte (positive control)", + "comparison": "Abdominal (22) vs other sources (46: respiratory 27, urinary 14, skin 5), adult sepsis, bulk-only", + "effect_bulkonly": "Five-draw mean: ISG-high T 0.32% vs 1.42%, adjusted logit -0.71 (SE 0.24). ISG-high monocyte 4.6% vs 9.9%, -0.59 (SE 0.25).", + "stats": "Model: OLS with HC3 errors, mean of draw logits ~ abdominal + age + sex, n=68. ISG-high T p=0.0027, BH q=0.067, so it fails q<0.05; rank-ANCOVA q=0.11. ISG-high monocyte p=0.019, q=0.17. The 4-source omnibus F-test for ISG-high T gives p=0.022, q=0.37. Site-adjusted q=0.081. Without the 10 Emergency-Center samples q=0.23. Gated on draws 1 and 3: -0.73, p=0.005; this is after selection.", + "draws_same_sign": "ISG-high T 5/5, 4/5 with p<0.05. ISG-high monocyte 5/5, 3/5 with p<0.05. Carrying draws for ISG-high T: 1,3.", + "paired_calibration": "calibration.csv, ISG-high T: ratio 0.66/0.53/0.35/0.69/0.38, Spearman 0.28/0.55/0.49/0.65/0.66; the sepsis-vs-HC and sepsis-vs-ICU signs both agree 5/5. Source is not a calibration.csv contrast. Same-model abdominal-vs-rest coefficient on the paired test samples, gen vs obs: -0.52/-0.53, -0.60/-0.75, -0.57/-0.81, -1.08/-0.80, -0.80/-0.85. That is 5/5 with similar size.", + "paired_observed_concordance": "Observed paired adult sepsis, 15 abdominal vs 31 other: ISG-high T 0.92% vs 4.66% (logit -0.84, p=0.10; rank-ANCOVA p=0.044). ISG-high monocyte -0.95 (p=0.10). Within-monocyte ISG CPM -0.88 log2 (p=0.030); within-T -0.50 (p=0.048). Concordant and marginal.", + "beyond_bulk": "No. The bulk shows it more strongly. Bulk-only 10-ISG score -0.76 z (p=4e-4, q=0.004); IFI44L -1.69 log2 (q=3e-4); MX1 -0.99 (q=3e-4); RSAD2, IFIT1 and SIGLEC1 also lower. Paired bulk -0.72 (p=0.01). After adjusting for the bulk ISG score, the cell effect disappears (p=0.32).", + "novelty": "Consistent with Ye 2026 and largely reported there. Pulmonary-sepsis monocytes have the highest 'interferon alpha response' (PDF p.8, Fig. 4i,j). The MARS1 immunosuppressive endotype is higher in abdominal and skin sepsis (p.12).", + "strength": "weak", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s02_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s03_paired_calib_long.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s04_observed_assoc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s07_mono_isg_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s08_bulk_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s06_feature_vs_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s10_evidence.txt" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a1a026e4b683bbc2b", + "verdict": "qualified", + "recomputed": "Everything was recomputed from the raw per-cell files (cells.parquet + markers.npz, ≥4 of 10 ISGs detected in AIFI-L1 T cells or monocytes). It matches states.parquet exactly: 610/610 bulk-only rows, 385/385 paired-gen rows and 82/82 observed rows for ISG-high T.\n\nGroup sizes: n=68 adult-sepsis bulk-only samples (22 abdominal vs 46 other).\n\nISG-high T (logit OLS, HC3, + age + sex)\n- Per draw: -0.827 / -0.719 / -0.768 / -0.732 / -0.480, p = 0.017 / 0.012 / 0.0063 / 0.015 / 0.060. All 5 draws are negative and 4/5 have p<0.05.\n- Mean of the draw logits: -0.705 (SE 0.235), p=0.0027, BH q=0.067 over 50 features. Stacked model with draw fixed effects and sample-clustered SE: -0.705, p=0.0016; draw heterogeneity p=0.51.\n- Five-draw mean fraction: 0.325% vs 1.421%. Medians are only 0.22% vs 0.42%. Three 'other' samples drive the mean: Res-S130 19.5%, Skin-S99 17.4%, Uri-S95 6.5%. Without them the 'other' mean is 0.51%.\n- Rank tests: Mann-Whitney p=0.0058, or 0.014 without those three; rank-ANCOVA p=0.0043.\n- Gated draws 1 and 3: -0.726, p=0.0047.\n- 4-source omnibus p=0.022. Versus respiratory, abdominal is -0.55 (p=0.039); urinary and skin are higher than respiratory, not significant.\n\nISG-high monocyte\n- Mean of draw logits: -0.590 (SE 0.252), p=0.019, q=0.17. Five-draw means 4.59% vs 9.93%; medians 2.5% vs 4.3%.\n\nConfounders (mean-of-draws coefficient, p, draws with p<0.05)\n- Sex-flag samples dropped: -0.720, p=0.0025, q=0.062.\n- Site: adding site gives -0.729, p=0.0033, q=0.081. Dropping the 10 Emergency-Center samples gives -0.610, p=0.018, q=0.23, 2/5 draws.\n- Donors: no repeated donors, and no bulk-only donor is in any draw's training set.\n- T-cell count (log N): -0.707, p=0.003.\n- Label confidence: T cells with confidence ≥0.9 only: -0.659, p=0.004. Adding mean confidence: -0.749. ISG-high T cells have normal label confidence and are mostly memory CD4/CD8.\n- Age modelled flexibly: age²: -0.556, p=0.014, 1/5 draws, q=0.35. Spline or quintiles: -0.51 to -0.54. Age 45–85 subset: -0.65, p=0.014, 3/5 draws.\n- Depth: in draw 3, abdominal T cells are shallower (log-library coefficient -0.19, p=0.0007; 49.6% vs 35.5% of T cells under 1,000 UMIs).\n - Adding log library: -0.752 overall, but draw 3 falls to -0.50 (p=0.14).\n - Per-cell equalisation to a common depth: at 1,500 UMIs -0.51 (p=0.014); at 1,000 UMIs -0.35 (p=0.07). Under the same equalisation the sepsis-vs-HC contrast gets stronger (0.07 to 0.34), so this drop is not just lost sensitivity.\n - Indirect depth standardisation (observed/expected): -0.611 (p=0.0097); draw 3 -0.45 (p=0.13); gated draws 1,3 -0.55 (p=0.035).\n- Depth (observed/expected) + site + age², combined: -0.461 (p=0.053), 0/5 draws; gated draws 1,3: -0.355 (p=0.19).\n- Fully adjusted logit model (site + age² + library + count + confidence): -0.583 (p=0.035), 2/5 draws.\n- Beta-binomial per draw: -0.49 / -0.79 / -0.52 / -0.29 / -0.27, LRT p = 0.062 / 0.033 / 0.050 / 0.31 / 0.33 (1/5 at p<0.05). Quasi-binomial: 0/5 (dispersion 32–90).\n- Adjusting for the bulk ISG score: -0.183 (p=0.33), 0/5 draws.\n- Monocyte under the same controls: + age² -0.41 (p=0.13); + site + age² -0.44 (p=0.12); fully adjusted -0.29 (p=0.28); observed/expected + site + age² -0.28 (p=0.22); beta-binomial 0/5.\n\nCalibration\n- calibration.csv, ISG-high T: ratio 0.66 / 0.53 / 0.35 / 0.69 / 0.38; Spearman 0.28 / 0.55 / 0.49 / 0.65 / 0.66; sepsis-vs-HC and sepsis-vs-ICU signs agree 5/5, but the generated differences are about 25–75% of the observed ones.\n- Gate (ratio 0.5–2, Spearman ≥0.3, signs agree) passes draws 1 and 3.\n- Within the adult-sepsis paired test samples, generated vs thinned-observed Spearman is 0.23 (p=0.14) / 0.49 / 0.68 / 0.62 / 0.66. Within the abdominal subgroup alone it is 0.03–0.34.\n- Abdominal-vs-other on the paired test samples: generated -0.52 / -0.60 / -0.57 / -1.08 / -0.80 (p 0.32 / 0.18 / 0.23 / 0.053 / 0.040); observed thinned -0.53 / -0.75 / -0.81 / -0.80 / -0.85 (p 0.14–0.30). Re-thinning with an independent seed gives the same.\n\nObserved paired cells (15 abdominal vs 31 other)\n- ISG-high T: -0.836 (p=0.10), rank-ANCOVA p=0.044. Means 0.92% vs 4.66%; medians 0.37% vs 0.60%.\n- Without Emergency-Center samples: -1.26 (p=0.027). Depth observed/expected: -0.675 (p=0.18). After adjusting for the paired bulk ISG score: +0.36 (p=0.23).\n- Within-T ISG CPM: -0.50 (p=0.048). Within-monocyte ISG CPM: -1.26 (p=0.009) on all 46 samples, or -0.88 (p=0.030) with the ≥10-monocyte filter.\n\nBulk\n- Bulk-only 10-ISG mean z: -0.648 (p=0.0004).\n- Genes: IFI44L -1.69 (q=1e-4), MX1 -0.99, IFIT1 -1.23, RSAD2 -0.89, SIGLEC1 -1.15; 9 of 10 genes q<0.05.\n- Paired bulk score: -0.66 (p=0.010).", + "issues": [ + "Headline fractions are mean-inflated: 0.3% vs 1.4% (T) and 4.6% vs 9.9% (monocyte) come mostly from three 'other' samples (Res-S130, Skin-S99, Uri-S95, at 19.5/17.4/6.5% ISG-high T and 64/62/58% ISG-high monocyte). Typical values are 0.22% vs 0.42% and 2.5% vs 4.3%, about 1.7–1.9x rather than 2–4x. The logit effect without them is -0.47 (p=0.016).", + "Depth confounds one of the two carrying draws. Draw 3 generates shallower abdominal T cells (p=0.0007), and the ≥4-of-10 rule depends on depth. Its effect fails after depth control: +library p=0.14, observed/expected p=0.13, equalised to 1,500 UMIs p=0.44, beta-binomial p=0.31. Draw 1 weakens under equalisation (p=0.15–0.40) and with confidence ≥0.99 (p=0.37).", + "Under reasonable controls combined (site + age² + indirect depth standardisation), the effect is borderline: five-draw mean -0.46 (p=0.053), 0/5 single draws at p<0.05, gated draws 1,3 -0.36 (p=0.19). So no draw carries a result that is both calibrated and robust to these controls.", + "Age is a partial confounder. The abdominal group has 8/22 patients aged 80 or older vs 5/46 in the others. Flexible age models cut the effect by 20–30% (q=0.35 with age²; 1/5 draws). The observed paired set shows no negative age curvature, but it has only 5 patients over 80, so this stays a sensitivity result.", + "Site: dropping the 10 Emergency-Center samples (5 abdominal, 5 urinary) gives q=0.23 with 2/5 draws; adjusting for site keeps it (q=0.081).", + "The ISG-high monocyte part does not hold up: p=0.12–0.28 under age/site/full adjustment, and 0/5 draws in the beta-binomial.", + "The result depends on the model: the beta-binomial gives 1/5 draws at p<0.05, the quasi-binomial 0/5, while the sample-cluster binomial and negative-binomial models (both dominated by the outlier samples) give p<0.01.", + "Calibration for this contrast is weak. Ratios are 0.35–0.69 (draws 2 and 4 under-generate about 2.7x). Draw 0's Spearman is 0.28 overall and 0.23 within sepsis (not significant). Within abdominal samples the generated values do not rank like the observed ones (Spearman 0.03–0.34). The paired-test abdominal-vs-other coefficients agree 5/5 in sign but come from overlapping subsets of the same 46 observed samples, and neither generated nor observed is significant in most draws (observed p 0.14–0.30).", + "The bulk explains it fully, as the claim already says: after adjusting for the bulk ISG score the cell effect is -0.18 (p=0.33). The generated cells add nothing beyond the bulk.", + "Minor wording errors in the claim. The bulk score of -0.76 z depends on the z reference (I get -0.65 z with the same p). The within-monocyte CPM of -0.88 needs the ≥10-monocyte filter (-1.26 without it). My v08 direct depth standardisation is artefact-prone and is superseded by v09; a note is added to its report." + ], + "reasoning": "Reproduction: every number in the claim reproduces exactly from my own recount of the raw per-cell files. That covers the per-draw coefficients, the mean of draw logits (-0.705, p=0.0027, q=0.067), the q values with site adjustment and without the Emergency-Center samples, the omnibus test, the gated estimate, the paired generated-vs-observed coefficients and the observed results. All five draws give a negative sign. The direction also agrees with the observed paired cells, with the bulk, and with Ye 2026 (pulmonary sepsis has the highest IFN-alpha response). I checked that last point against the paper text.\n\nThe finding holds after dropping sex-flag samples, adjusting for T-cell count and label confidence, adjusting for site, and in the age 45–85 subset. For the Imperial-specific checks: sepsis has no repeated donors, and no bulk-only donor trained any draw.\n\nIt weakens in several ways:\n- The headline fractions come mostly from three outlier samples in the 'other' group.\n- Flexible age adjustment and dropping the Emergency-Center samples each attenuate it and push q to 0.23–0.35.\n- Draw 3, one of the two calibration-gated carrying draws, makes shallower abdominal T cells. Its effect fails under any depth control, and the ≥4-of-10 rule is depth-sensitive. Draw 1 is fragile under equalisation and strict label confidence.\n- With site, flexible age and indirect depth standardisation together, the five-draw effect is borderline (p=0.053), no single draw reaches p<0.05, and the gated estimate is -0.36 (p=0.19).\n- The monocyte positive-control part fails under age/site adjustment.\n- Count models give conflicting answers (beta-binomial 1/5 draws).\n- Calibration for this contrast is weak: under-generation, near-zero ranking within the abdominal group, and paired-test source coefficients that are not significant in either the generated or the observed cells.\n- The bulk ISG score removes the effect entirely.\n\nWhy qualified rather than refuted: no single reasonable control removes the five-draw effect; it shrinks by about 15–50% (and the combined control leaves it borderline).\n\nWhy not confirmed: the calibrated carrying draws do not survive depth plus site plus age control, and calibration is weak.\n\nRecommendation: downgrade to \"generated cells reproduce, in direction only, the abdominal bulk ISG deficit\". Report it as bulk-derived, not as a cell-level or calibrated finding. Quote medians rather than means, and drop the monocyte part.\n\nAll work was CPU only. Every script and every number quoted here is in the verify directory, from v01 through v10 (v10_summary.csv collects them). Writes were limited to that directory, so this subagent made no archive update; the coordinator decides on the PBISC-Diffusion task-end maintenance step.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v10_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v10_summary.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/vh.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v01_meta_check.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v01_meta_check.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v02_recompute_states.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v02_sample_features.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v02_check_vs_states.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v03_bulkonly_models.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v03_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v03_balance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v03_report.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v03b_counts_age_influence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v03b_report.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v03c_betabinomial.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v03c_betabinomial.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v04_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v04_calibration_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v04_within_sepsis.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v05_observed_concordance.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v05_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v06_bulk_bh_gated.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v06_bh50.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v06_bulk_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v06_report.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v07_extra_checks.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v07_fullyadj.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v07_report.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v08_depth_strata.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v08_depth_strata.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v08_report.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v09_depth_indirect_std.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v09_depth_oe.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_recompute/v10_summary.py" + ] + }, + { + "verifier": "calib", + "agent_id": "aad9c3eb395c93280", + "verdict": "qualified", + "recomputed": "All numbers were recomputed from features/ye2026_sepsis (states.parquet, meta.csv, calibration.csv), the work/ cells and pb_l1 files, and the real bulk npz files. Scripts are v01–v07. Model: logit((n+0.5)/(N+1)), NaN when the parent has fewer than 10 cells, OLS with HC3 errors, + age + sex.\n\nBULK-ONLY (68 adult sepsis: 22 abdominal vs 46 other)\n- ISG-high T, per draw: −0.83 / −0.72 / −0.77 / −0.73 / −0.48 (p .017 / .012 / .006 / .015 / .060).\n - Mean-logit: −0.705 (SE 0.235, p=0.0027). This reproduces the claim exactly.\n - Draw as fixed effect, sample-clustered: −0.705 (p=0.0016).\n - Gated draws 1 and 3: −0.726 (p=0.0031).\n- ISG-high monocyte, per draw: −0.69 / −0.70 / −0.14 / −0.24 / −0.64 (p .021 / .004 / .61 / .52 / .014).\n - All five draws stacked: −0.51 (p=0.025).\n - sep_clinical gate (draws 0, 3, 4): −0.55 (p=0.035).\n - Stricter gate (draws 0, 3): −0.50 (p=0.080).\n\nPAIRED TEST CALIBRATION (adult sepsis, n=42–43 per draw; generated vs observed thinned to the draw's depth)\n- ISG-high T coefficient, generated vs observed: −0.52/−0.53, −0.60/−0.75, −0.57/−0.81, −1.08/−0.80, −0.80/−0.85.\n - Observed p in each draw's test set: .30 / .19 / .18 / .14 / .14.\n - Bootstrap 95% CI of generated minus observed: d0 [−1.03, 1.09], d1 [−0.56, 0.90], d2 [−0.61, 1.05], d3 [−1.07, 0.45], d4 [−0.68, 0.84].\n - Pooled (draw fixed effect, sample-clustered): observed −0.75 (p=0.12), generated −0.71 (p=0.023), difference +0.04 (95% CI −0.51 to 0.59).\n- ISG-high monocyte, pooled: observed −1.03 (p=0.048), generated −0.54 (p=0.087). The generated effect is about half the observed one.\n- calibration.csv, ISG-high T:\n - ratio .66 / .53 / .35 / .69 / .38; Spearman .28 / .55 / .49 / .65 / .66;\n - pooled ratio .52, Spearman .52;\n - the sign of the mean difference agrees 5/5 for sepsis vs HC and sepsis vs ICU.\n - My Spearman over all 77 test samples matches calibration.csv to 4 decimals.\n- Sepsis vs HC on the claim's own scales (v05):\n - ISG-high T, observed: sepsis is below HC (AUC .21–.25, Mann–Whitney p .018–.048; adjusted logit −0.71 to −0.95).\n - ISG-high T, generated logit: −0.09 / +0.51 / +0.86 / −0.12 / +0.18. Only 2 of 5 draws agree in sign (d0 and d3, both near zero).\n - ISG-high monocyte logit: 0 of 5 draws agree.\n- Generated ÷ observed level by source (ISG-high T): respiratory 0.17–0.38, abdominal 0.34–0.66, skin 0.51–1.29.\n\nDEPTH (v02)\n- Draw 3 makes the abdominal samples' cells shallower: T-cell median library −0.28 log2 (p=7e-4), monocyte −0.51 (p=1e-4).\n- Adjusting for depth shrinks draw 3's ISG-high T effect to −0.50 (p=0.14). Draw 1 holds at −0.84 (p=0.0024).\n- In the observed cells, the T-cell library is −0.19 log2 (p=0.20).\n\nEMERGENCY CENTER (v07)\n- Without the 10 Emergency-Center samples, ISG-high T per draw: −0.90 / −0.47 / −0.72 / −0.59 / −0.37 (p .019 / .14 / .035 / .051 / .16).\n- The real bulk ISG score without them: −0.79 (p=1e-4).\n\nWITHIN-TYPE ISG log2 CPM (bulk-only)\n- T: −0.31 / −0.37 / −0.18 / −0.21 / −0.33 (p .08 / .024 / .13 / .083 / .007).\n- Monocyte: −0.28 / −0.36 / +0.02 / −0.27 / −0.56.\n\nOBSERVED PAIRED (46 adult sepsis: 15 abdominal vs 31 other)\n- ISG-high T: 0.92% vs 4.66% (medians 0.37% vs 0.60%).\n - logit −0.84 (p=0.10); rank-ANCOVA p=0.044; Mann–Whitney p=0.096;\n - + Emergency-Center indicator −0.80 (p=0.081); without the Emergency-Center samples −1.26 (p=0.027);\n - within-T ISG CPM −0.50 (p=0.048);\n - + paired bulk ISG score +0.36 (p=0.22).\n- ISG-high monocyte: −0.95 (p=0.097); + bulk ISG score +0.05 (p=0.89); within-monocyte CPM −0.88 (p=0.030).\n- Observed ISG-high T by source: respiratory 6.3%, abdominal 0.92%, urinary 1.4%, skin 3.0%.\n\nREAL BULK\n- Bulk-only 10-ISG z-score: −0.72 (p=4e-4). All 10 genes are lower, 9 of 10 with q<0.05.\n- Paired bulk: −0.78 (p=0.010).\n- Bulk-only by source: respiratory +0.36, abdominal −0.49, urinary −0.05, skin +0.44.\n- Generated features after adjusting for the bulk score: T −0.18 (p=0.33), monocyte −0.04 (p=0.85); per-draw T p .24–.80.\n- On the 77 paired test samples, predicting the observed ISG-high T fraction:\n - Spearman: bulk score .86–.88 vs generated .28–.66;\n - R²: bulk .82–.86, generated .32–.62, both together .83–.88.\n- Bulk MS1 score, used as a severity proxy: abdominal +0.25 (p=0.19). The ISG effect after adjusting for it is −0.73 (p=6e-4).", + "issues": [ + "All of the claim's numbers reproduce exactly (−0.705, SE 0.235, p=0.0027; per-draw coefficients; observed −0.84, p=0.10). The direction is not contradicted anywhere: generated cells, observed cells, paired bulk and bulk-only bulk all put abdominal sepsis lowest.", + "The paired check of the abdominal contrast agrees 5/5 in sign, but it carries little weight. Within each draw's test set the observed effect is not significant (p .14–.30). The five checks reuse about 38 of the same 42 observed samples, so they are not independent. The pooled generated-minus-observed difference has a 95% CI of −0.51 to 0.59.", + "calibration.csv's 'sepsis vs HC agrees 5/5' holds only for raw mean differences. On the logit and rank scales the claim's tests use, observed sepsis is below HC for ISG-high T (AUC .21–.25). The generated sign agrees in only 2/5 draws (d0 and d3, both near zero). For ISG-high monocyte it agrees in 0/5 draws, so this 'positive control' feature fails its basic contrast on the claim's scale.", + "Feature-level calibration of ISG-high T is weak. The ratio is below 0.5 in draws 2 and 4, and the Spearman in draw 0 is 0.28. The generator compresses respiratory samples (generated/observed 0.17–0.38) and inflates skin (0.51–1.29). As a result the 'other' group in the generated cells is led by skin, while in the observed cells and in Ye 2026 respiratory is highest. Splitting abdominal from the rest hides this.", + "Both carrying draws are fragile. Draw 3 gives abdominal samples shallower cells (T −0.28 log2, p=7e-4). The ISG-high call needs ≥4 of 10 genes detected, so depth matters, and adjusting for depth drops draw 3 to −0.50 (p=0.14). Draw 1 drops to −0.47 (p=0.14) without the Emergency-Center samples. On the logit scale draw 1 also reverses the sepsis-vs-HC sign (+0.51 vs −0.95).", + "The claim does not name the carrying draws for the monocyte half. Under sep_clinical's gate they are 0 and 4; under a stricter gate only draw 0 (draw 3 is null). In the paired test the generated monocyte effect is about half the observed one (−0.54 vs −1.03).", + "Observed support is concordant but marginal: logit p=0.10, rank-ANCOVA p=0.044, within-T ISG CPM p=0.048. The difference in means (0.92% vs 4.66%) comes mostly from a few high-ISG samples, mostly respiratory; the medians are 0.37% vs 0.60%.", + "Nothing beyond the bulk, as the claim itself states. This is stronger than the claim says. The 10-ISG bulk score predicts the observed ISG-high T fraction far better than the generated cells do (Spearman .86–.88 vs .28–.66). Even in the real cells the abdominal difference disappears after adjusting for bulk ISG (+0.36, p=0.22). The state is a cellular readout of the global bulk interferon signature, so even perfect generation could not add information here.", + "Novelty is low, and the MARS1 citation is misapplied.\n- Ye 2026 Fig. 4i (PDF p.9, printed p.158; text on PDF p.8, printed p.157): abdominal monocytes have the lowest IFN-alpha score of the sepsis sources (median about 0.22, at the HC/IC level), and pulmonary the highest.\n- The MARS1 point (PDF p.12, printed p.161; Extended Data Fig. 9b) covers abdominal and skin sepsis. Skin is the most ISG-high source here (bulk z +0.44), so MARS1 does not follow the ISG pattern.\n- The cluster abundance in Fig. 4d (Mono_c05-CD14-IFITM1) is lowest in urinary sepsis (urinary < abdominal, Dunn P<0.05), which cuts against the monocyte half.\n- Ye 2026 does not report T-cell ISG by source.\n- Scicluna 2025 (Nat Med 31:4119): the interferon-signalling CTS3 subtype is 69% pneumonia, while CTS1 is enriched for intra-abdominal infection. So the bulk pattern is already known.", + "Confounders that cannot be measured here.\n- No severity, timing, pathogen or steroid data in this cohort.\n- Leite 2024 (PDF p.4, p.12): abdominal sepsis is more often in shock, and severity drives MS1 and lower interferon-response genes. The bulk MS1 proxy does not explain the effect here (−0.73 after adjustment).\n- Enrolment ran from October 2021 to April 2023, which includes the December 2022 COVID wave, and no viral exclusion is stated (Ye 2026, PDF p.17). The respiratory comparator may therefore include viral interferon drivers.", + "Multiplicity: abdominal vs rest is one of four one-vs-rest contrasts, chosen after seeing the lowest group. The pre-specified 4-source omnibus test gives p=0.022, q=0.37." + ], + "reasoning": "I recomputed every number independently from the primary feature files rather than from the sep_clinical tables. The claim's statistics reproduce exactly, and I checked the four honesty criteria in DATA_GUIDE.md.\n\n**1. Calibration: partial support.**\n- For this contrast the generated cells reproduce the observed abdominal effect in all 5 draws, with similar magnitude for T cells. The pooled difference is about 0, but its CI is wide.\n- The observed contrast inside each draw's test set is itself not significant, and the five checks share the same samples. This is one weak check that agrees, not five confirmations.\n- At the feature level, the ISG-high T ratio is below 0.5 in 2 of 5 draws.\n- calibration.csv's sign agreement holds only on raw means. On the logit and rank scales the claim actually uses, the generator fails the sepsis-vs-HC direction for ISG-high T in 3 of 5 draws and for ISG-high monocyte in 5 of 5.\n- The generator also distorts the source ordering: respiratory is compressed and skin inflated.\n- The carrying draws are fragile. Draw 3 is partly a depth artifact (the state call is detection-based). Draw 1 depends on the Emergency-Center samples.\n\n**2. Observed cells: concordant, not contradictory, but marginal.**\n- Effects run from p=0.044 to 0.10, and the within-T ISG expression is p=0.048.\n- Excluding the Emergency-Center samples strengthens the observed effect.\n- After adjusting for the real bulk ISG score, the observed effect itself vanishes.\n\n**3. Biology and novelty: consistent and largely known.**\n- Ye 2026 Fig. 4i already shows abdominal monocytes with the lowest IFN-alpha response score among sepsis sources, and respiratory the highest.\n- The consensus framework (Scicluna 2025) puts the interferon subtype CTS3 mostly in pneumonia.\n- The monocyte cluster abundance in Fig. 4d does not single out abdominal sepsis.\n- The claim's MARS1 argument does not support an ISG-specific pattern.\n- The T-cell version is not reported in Ye 2026, but it is fully derivable from the bulk.\n\n**4. Beyond bulk: nothing.**\n- The bulk 10-ISG score shows the effect more strongly and is robust to excluding the Emergency-Center samples.\n- It removes the generated effect completely.\n- It predicts the observed ISG-high fractions much better than the generated cells do.\n\n**Verdict: qualified.** Nothing contradicts the claim, so it is not refuted. It cannot be confirmed: the calibration for the feature and the source axis is partial, the observed support is marginal, and the generated cells add nothing over the bulk. The defensible reading: the real bulk ISG signature is lower in abdominal sepsis, and the LoRA-5 cells pass it on noisily as fewer ISG-high T cells and monocytes.\n\nGPU was not used, and nothing was written outside the verifier directory. This verifier changes no paper asset, so there is no archive delta; the archive workflow belongs to the coordinator.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v01_paired_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v01_paired_calibration.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v01_paired_by_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v01_paired_pooled.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v01_paired_beyond_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v01_calibration_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v01_bulk_isg_scores.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v02_bulkonly.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v02_bulkonly.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v02_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v02_stacked.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v02_depth.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v02_within_type_isg.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v02_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v03_observed.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v03_observed.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v03_observed_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v03_observed_source_means.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v03_observed_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v04_source_profile.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v04_source_profile.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v04_source_profile.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v04_source_coefs.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v05_contrast_scales.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v05_contrast_scales.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v05_contrast_scales.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v05_contrast_scales_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v06_severity_proxy.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v06_severity_proxy.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v06_severity_proxy.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v07_site_check.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v07_site_check.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v07_site_check.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/v08_literature_quotes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/ye2026_pages.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/ye2026_fig4i_ifn_crop.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/ye2026_fig4d_right_zoom.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/ye2026_fig4d_crop.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-4_calib/leite2024_pages.txt" + ] + } + ] + }, + { + "id": "SEP-POP-5", + "lens": "sep_populations", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 10, + "theme": "lymphocytes", + "synthesis_strength": "moderate for sepsis vs HC at cluster level (5/5); not sepsis-specific; bulk-derivable", + "reportable_statement": "A KLRB1+ GZMK+ IL7R+ CCL5+ memory T state (MAIT, memory CD4 and CD8) is a smaller share of T cells in bulk-only sepsis than HC in all five draws (-0.074/-0.060/-0.061/-0.038/-0.050), with paired calibration passing 5/5 and the same loss in observed cells. The generator also removes the state from ICU controls, which observed cells and Ye 2026 do not show.", + "carrying_draws": { + "cluster c9, sepsis vs HC": "0,1,2,3,4 (0.67-1.17x)", + "embedding-free (L2 MAIT, KLRB1+GZMK+) size": "0,1; 2-4 sign only", + "sepsis vs ICU": "none (0.17-0.41x)" + }, + "honesty": { + "calibration": "c9 ratio 0.86-1.08, Spearman 0.28-0.42 (within adult sepsis 0.07-0.28); calibration.csv l2 MAIT sign 5/5, ratio 0.52-0.55.", + "replication": "5/5, q<=2.2e-6; survives age, sex, site, sex_flag, depth GEE, label confidence.", + "observed": "Observed -0.074 (46 vs 7, p 3e-4), MAIT plus memory CD4; observed ICU controls at HC level (-0.006 to +0.008).", + "beyond_bulk": "No for ranking: a MAIT-core bulk score tracks observed shares at Spearman 0.44-0.59 (definition-dependent) vs 0.28-0.42 generated, and makes the generated share redundant in most draws; the raw 20-marker bulk score separates groups via T-cell content (pan-T alone AUC 0.052)." + }, + "verdict_reasons": [ + "not sepsis-specific in generated cells (ICU=0.40). Counterpart: 328 observed cells; concordance 0.90.", + "beyond_bulk": "No. Its 20-marker bulk score already separates sepsis from HC (AUC 0.056) and tracks the observed fraction as well as the generated fraction does (Spearman 0.37-0.42 vs 0.28-0.42).", + "novelty": "Ye 2026 reports reduced T_CD4_c04-GZMK, c05-GNLY and c06-FOXP3 (PDF p.2 = printed p.151, Fig. 2b). MAIT/KLRB1+ depletion in sepsis is reported elsewhere; no paper for it is in the reference folder, so this was not checked. Not new.", + "strength": "weak", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s07_candidate_detail.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s09_bulk_derivability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s12_sensitivity.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "ae8cd032026764868", + "verdict": "qualified", + "recomputed": "I wrote my own code from the raw compact files. From the discovery lens I read only the c9 labels (leiden_r2.0 == 9). The arithmetic matches the claim exactly.\n\nBULK-ONLY, c9 share of T cells, adult sepsis (68) vs adult HC (36):\n- d0: 0.0349 vs 0.1088 (-0.074, AUC 0.069, p 5.1e-13)\n- d1: 0.0373 vs 0.0972 (-0.060, AUC 0.112, p 8.2e-11)\n- d2: 0.0259 vs 0.0868 (-0.061, AUC 0.050, p 4.8e-14)\n- d3: 0.0264 vs 0.0646 (-0.038, AUC 0.162, p 1.5e-8)\n- d4: 0.0336 vs 0.0838 (-0.050, AUC 0.101, p 2.4e-11)\n- Same sign in 5/5 draws. p x 143 is at most 2.2e-6.\n- Sepsis vs ICU: -0.006 to -0.014, p 0.11 to 0.58.\n\nOWN CALIBRATION ROWS (each draw's paired TEST samples):\n- Ratio 0.86 / 1.08 / 0.96 / 1.01 / 0.95. Spearman 0.36 / 0.40 / 0.40 / 0.28 / 0.42.\n- Sepsis-HC, generated vs observed: -0.079/-0.067, -0.062/-0.059, -0.057/-0.075, -0.057/-0.076, -0.050/-0.074. Only 6-7 HC per draw.\n- Bootstrap 95% CIs exclude 0 on both sides in 5/5 draws. P(same sign) = 1.0. The gate passes 5/5.\n- Sepsis-ICU: generated/observed = 0.31 / 0.23 / 0.40 / 0.17 / 0.41.\n\ncalibration.csv 'l2 | MAIT': my recompute matches to 5 decimals. Ratio 0.52-0.55, Spearman 0.36-0.46, sepsis-HC sign agrees 5/5.\n\nOBSERVED PAIRED CELLS:\n- c9 sepsis vs HC: -0.074 (p 3.0e-4, 46 vs 7).\n - Leave-one-HC-out: max p 0.001.\n - Age + sex GLM: p 2.6e-8. Adding T library and L2 confidence: p 0.003.\n - Only the First Affiliated Hospital site (FAH): p 7e-4.\n- kNN-mapped view: -0.068 (p 5.5e-4).\n- Sepsis vs ICU: -0.068 (p 0.039). Leave-one-out p ranges 0.010-0.12.\n- Paediatric sepsis vs paediatric HC: -0.014 (p 0.39).\n\nCONFOUNDERS (c9; negative in every draw):\n- sex_flag samples dropped: q <= 3.1e-8.\n- FAH site only: q <= 7.7e-7.\n- Age + sex: log-odds -0.90 to -1.33.\n- Bulk technical metrics added: -0.87 to -1.16 (p <= 6.8e-7).\n- Cell-level GEE adjusting for library size and confidence: -0.89 to -1.34 (p <= 2.9e-10).\n- Cells restricted to library 1200-2400: diff -0.042 to -0.088.\n- Only T cells with L2 confidence >= 0.9: q <= 3.7e-8.\n- c9 within CD8-like T cells: q <= 2e-7.\n- Sample-level library + confidence + T-cell count (these are collinear with group): log-odds -0.37 to -0.79, p 1e-6 to 0.052.\n - Restricted to FAH with sex_flag dropped: p <= 0.035 in 5/5.\n - Rank OLS: p <= 0.048 in 5/5.\n- Every infection source is negative in 5/5 draws.\n- Bulk-only donors are unique and none is in the paired or training sets.\n\nINDEPENDENT DEFINITIONS:\n- KLRB1+ GZMK+ cells, depth-matched to 800 counts: -0.011 to -0.022 (q <= 5.1e-4, 5/5).\n - Calibration against thinned observed cells: ratio 0.40-0.69, Spearman 0.29-0.34, sign agrees 5/5.\n- AIFI L2 MAIT share of T cells: -0.007 to -0.022 (q <= 1.3e-5, 5/5).\n\nBULK COMPARISON:\n- The 20-marker bulk score (B20): AUC 0.054. A CD3 T-abundance score alone gives AUC 0.052.\n- T-normalised bulk ratios show no per-T-cell loss: B7_rel 0.63, KLRB1_rel 0.44, GZMK_rel 0.69, MAIT_rel 0.77.\n- Paired samples: rho(B20, observed) 0.38-0.43 vs rho(generated, observed) 0.28-0.42.\n - The generated fraction adds information beyond B20 in 4/5 draws (p 0.013-0.045).\n - It adds beyond a bulk MAIT-gene score (rho 0.44-0.53) in only 2/5.\n\nICU vs HC:\n- Bulk-only generated: -0.031 to -0.060 (p <= 0.002, 5/5).\n- Paired test samples: generated -0.021 to -0.059 vs observed -0.006 to +0.008.\n\nPROFILE CONCORDANCE: 0.68 over 1,098 genes (0.47-0.60 per bulk-only draw). Top-30 up-gene overlap 0.70-0.80.", + "issues": [ + "Not sepsis-specific in generated cells. Bulk-only adult ICU controls lose c9 against HC almost as much as sepsis does (-0.031 to -0.060, p <= 0.002 in 5/5 draws). On the paired test samples the model lowers c9 in ICU controls (generated ICU-HC -0.021 to -0.059), but the observed ICU controls sit at HC level (-0.006 to +0.008; n = 5-6 vs 6-7). Ye 2026 Fig. 3d also shows no HC vs ICU difference for MAIT. The generated sepsis-vs-ICU effect is only 0.17-0.41x the observed one, and its CI crosses 0 in 2/5 draws. The generated loss therefore reads as 'healthy vs hospitalised', not as a sepsis state.", + "The sepsis-HC calibration rests on 6-7 paired adult HC per draw, and Spearman over samples is only moderate (0.28-0.42). Bootstrap CIs still exclude 0 on both sides in 5/5 draws, so the sign is solid; the per-sample tracking is weak.", + "The 'beyond bulk' sentence is mis-argued. The bulk marker score separates the groups mainly because T cells are fewer in sepsis: a CD3 abundance score alone gives AUC 0.052, against 0.054 for B20. T-normalised bulk ratios show no per-T-cell loss (AUC 0.44-0.77; MAIT_rel and GZMK_rel are higher in sepsis), so the 'share of T cells' reading cannot be derived from a bulk ratio. The generated fraction adds beyond B20 in predicting the observed fraction in 4/5 draws. However, its per-sample ranking is no better than a bulk MAIT-gene score: it adds beyond that score in only 2/5 draws.", + "Novelty: the cohort paper itself shows this. Ye 2026 Fig. 3d (PDF p.7, printed p.156) shows T_CD8_c07-TRAV1-2 (MAIT) lower than HC in abdominal (*), urinary (**) and skin (*) sepsis. The text reports T_CD4_c04-GZMK as reduced (Fig. 2b). The candidate's 'not checked' should be corrected; 'Not new' still holds, more firmly.", + "The state definition is fragile. c9 is resolution-unstable (Jaccard 0.09 and 0.18 with the best r0.5 and r1.0 clusters). The claimed concordance of 0.90 depends on the gene set: I get 0.68 over 1,098 detected genes. The KLRB1+ GZMK+ marker rule keeps the sign 5/5, but in draw 4 its calibration is weak (generated -0.004 vs observed -0.062).", + "Adjusting at sample level for model-output covariates (median T library, mean L2 confidence, T-cell count) shrinks the log-odds by 40-70%, and draw 3 reaches only p = 0.052. These covariates nearly separate the groups (L2 confidence AUC 0.02-0.12), so this is over-adjustment. Cell-level depth controls leave the effect unchanged.", + "The loss does not replicate in observed paediatric cells: c9 -0.014, p 0.39 (15 sepsis vs 4 HC)." + ], + "reasoning": "The core claim reproduces exactly: the bulk-only sepsis < HC loss of the GZMK+ KLRB1+ IL7R+ CCL5+ T-cell state (c9), negative in all 5 draws.\n\nIt survives every confounder I tested:\n- age, sex, site, dropped sex_flag samples, bulk technical metrics and infection source;\n- library size, through cell-level GEE models, a depth window and depth deciles. Draw 4, where the sepsis T cells are deeper than the HC T cells, still shows the loss;\n- L2 label confidence;\n- T-cell number and composition (c9 within CD8-like T cells).\n\nIt also holds under independent definitions: a depth-matched KLRB1+ GZMK+ marker rule and the AIFI L2 MAIT share, both 5/5. It is not imprinted by an HC training donor: draws 2-4 had no adult-HC training donor and still show it.\n\nThe draw's own calibration passes 5/5 for sepsis vs HC (ratio about 1, sign agrees, bootstrap CIs exclude 0), so all 5 draws carry the result. Observed paired cells show the same loss, robust to leave-one-out, age and sex. So nothing refutes it.\n\nI rate it 'qualified' rather than 'confirmed' for three reasons:\n- The generated loss is not specific to sepsis. The model also lowers c9 in ICU controls, which are not lowered in observed cells or in Ye 2026 Fig. 3d.\n- The 'bulk shows the same, so the generated cells add little' argument is wrong in its mechanism. The absolute bulk score tracks lymphopenia, and T-normalised bulk ratios do not show the loss. Still, the per-sample ranking is no better than a bulk MAIT-gene score.\n- The novelty note missed that the cohort paper's Fig. 3d already shows MAIT depletion.\n\nAll four honesty items:\n- Calibration: good for the sign, moderate for per-sample tracking.\n- Draws: 5/5.\n- Observed agreement: yes.\n- Beyond bulk: partly (the per-T-cell direction yes, the per-sample ranking no).\n\nAll work was CPU-only, with memory capped at 6G per script and no GPU use. I wrote only under the verify dir. This verifier publishes no paper-archive update; the coordinator owns that.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/vc_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v01_extract.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v02_recompute_cluster.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v03_measures.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v04_tests_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v05_depth_cellnumber.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v06_observed_and_calib_uncertainty.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v07_bulk_and_technical.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v08_calcsv_meta_misc.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v09_profile_concordance.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v10_identity_markers.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v11_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/v12_ye2026_novelty_evidence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v11_summary_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v11_icu_vs_hc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v02_claim_check.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v04_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v04_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v04_confounder_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v05_gee.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v05_collinearity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v06_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v06_calib_bootstrap.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v07_bulk_auc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v07_bulk_vs_gen_paired.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v07_technical_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v08_calcsv_vs_mine.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v08_meta_checks.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v08_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v08_infection_source.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v09_profile_concordance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/tables/v12_ye2026_text_excerpts.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-5_recompute/figures/ye2026_fig3d_crop.png" + ] + }, + { + "verifier": "calib", + "agent_id": "a3cd7162224456135", + "verdict": "qualified", + "recomputed": "I recomputed everything from the lens's per-cell Leiden labels (v01) and matched the lens exactly. Bulk-only c9/T, sepsis vs HC, d0..d4: diff -0.074/-0.060/-0.061/-0.038/-0.050; AUC 0.069/0.112/0.050/0.162/0.101; MWU p identical to the lens (5.1e-13..1.5e-8, lens q 7.3e-11..2.2e-7).\n\nRobustness:\n- The result is unchanged with sex_flag dropped, with First-Affiliated-only sepsis (58 vs 36) and with MIN_PARENT 20.\n- An age+sex-adjusted quasi-binomial GLM gives logOR -0.90..-1.33 per draw (all p<1e-6).\n- With draw as a fixed effect, logOR is -1.12 (p<1e-10). The sepsis x draw Wald test is p=0.0004: the size differs between draws but the sign does not.\n- HC are age-matched (63.8 vs 65.2 y).\n\nPaired calibration (77 test samples per draw):\n- Ratio 0.86/1.08/0.96/1.01/0.95; Spearman 0.36/0.40/0.40/0.28/0.42; pooled ratio 0.97, Spearman 0.45.\n- Sepsis-HC, generated/observed: -0.079/-0.067, -0.062/-0.059, -0.057/-0.075, -0.057/-0.076, -0.050/-0.074. That is 0.67-1.17x, and all bootstrap CIs exclude 0, so all 5 draws carry sepsis Res > Uri ≈ ICU > HC. Urinary vs HC AUC 0.88/0.58/0.77/0.85/0.88; urinary vs ICU ~0.5.\n- **IEG adjustment in bulk-only:** absorbs most of the c10 effect in d2-4 (adjusted p 0.11/0.11/0.25); d0 keeps a residual (beta 0.67, p 1e-4).\n\n**(4) Beyond bulk**\n- **Group level:** in bulk-only, a 10-gene index (NR4A2, GADD45B, DUSP2, CD69 minus GNLY, GZMB, FGFBP2, PRF1, GZMH, CX3CR1) gives sepsis vs HC AUC 0.895 (p 4e-11). That matches generated c10 at 0.92/0.63/0.885/0.93/0.92 and shift at 0.91/0.79/0.93/0.92/0.93.\n - The TEMRA genes alone give AUC 0.066, i.e. strongly down.\n - The candidate's 20-marker score gives 0.519.\n- **Patient level, paired adults:** generated c10 adds beyond the index in d2/d3/d4 (p 0.003/0.007/0.001), not in d0 (0.10) or d1 (0.76).\n- **Patient level, within adult sepsis:** only d4 adds (p 0.025-0.036); d0-d3 give p 0.10-0.95.\n- **sex_flag:** dropping these samples leaves every draw's effect unchanged.", + "issues": [ + "The claim's '5/5 draws, all passing calibration' is overstated. In draw 1, c10 is near-absent (bulk-only 0.011 vs 0.002, AUC 0.63). Its paired contrast is 0.12x of observed (+0.008 vs +0.072, bootstrap CI [-0.006, 0.025]) and the ratio of 0.34 passes only at the edge. Draw 1's training donors have observed c10 of 0, 0, 0, 0 and 0.171. The c10 result is carried by draws 0, 2, 3, 4.", + "The 'away from TEMRA' half is weak in the observed data. Observed c13 sepsis−HC is -0.034 (MWU p 0.047, BH q 0.063 over the 4 features, permutation p 0.24, leave-one-HC-out max p 0.12), and its bootstrap CI includes 0 in every draw's test set. c10+c13 together do not change (observed +0.035, p 0.36). The calibrated sign for c13 in draws 0, 2, 3, 4 is matched to an observed contrast that is itself not robust.", + "The calibration and the observed contrast both rest on 7 adult HC: 6-7 per draw, and the same people in every draw. So the per-draw calibrations are not independent replications on the control side.", + "The effect is not sepsis-specific. Observed non-septic ICU controls already have c10 0.055 vs HC 0.004, and sepsis vs ICU is p 0.56. In bulk-only, ICU vs HC is AUC 0.76-0.81 in draws 0, 2, 3, 4, and ICU−HC calibration has the same sign in 5/5 draws. Ye Fig. 3e (printed p.156) gives Ro/e HC 0.14 vs ICU 0.7. The claim should read 'critically ill (septic and non-septic ICU) vs HC'.", + "The generator over-calls c10 in urinary sepsis. The 5 paired urinary patients have observed c10 0.003 (equal to HC; vs Abd/Res 0.088, p 0.007), but paired_gen gives them 0.044-0.103 in draws 0, 2, 3, 4. Ye also shows urinary depletion: Ro/e 0.3 and urinary < Abd/Res (** in Fig. 3d). So the bulk-only urinary c10 rise (AUC 0.77-0.88 vs HC) is not supported, and the pooled 'sepsis' contrast includes this mis-calibrated subgroup.", + "c10 vs c13 is mainly an immediate-early (IEG/AP-1) difference, and in observed cells that programme is T-cell-wide. 63% of the top-30 c10-vs-c13 genes are IEG/AP-1 genes. Adjusting for the IEG share of other non-cytotoxic T cells shrinks the observed sepsis effect on c10 by 65% (beta 1.16 to 0.41, p 0.12) and removes it for c13 (p 0.57). In the pseudobulk, the IEG rise is identical in Memory CD8 and Memory CD4 (CD8 minus CD4 p 0.73). The 'CD8 effector shift' is therefore not separable from SEP-POP-3, and the claim understates this as 'shares the programme'.", + "The sepsis-specific component in the observed cells is the cytokine side, which the generator lacks. In observed Memory CD8 pseudobulk, TNF/IFNG/CCL3/CCL4 beats ICU controls (AUC 0.80, p 0.015). The generator reproduces that contrast at only 0.04-0.33x. In c10-vs-c13, TNF is +0.21 vs +0.75 observed and IFNG +0.16 vs +0.37. These are the genes Ye uses to define T_CD8_c05 (printed p.153).", + "'Beyond bulk: Yes for c10' is refuted at the group level. In bulk-only, a 10-gene bulk index already separates sepsis from HC (AUC 0.895), as well as generated c10 or shift (0.89-0.93 in the good draws). The TEMRA genes alone give AUC 0.066. The candidate's 20-marker comparator (AUC 0.52) mixes genes that move up (IEG) and down (lymphopenia), so they cancel. At the patient level, the gain is modest and draw-dependent: among paired adults only draws 2-4 add (p 0.001-0.007), and within adult sepsis only draw 4 does (p 0.025).", + "Ye already reported this result for these bulk-only individuals. Ye's CIBERSORTx deconvolution of their in-house PBMC bulk RNA-seq (Methods, PDF p.20; Extended Data Fig. 9a 'This study', PDF p.39) shows T_CD8_c05-GNLY-CCL4 lowest in HC, low in ICU and urinary, and higher in Abd/Res/Skin. That bulk set includes the bulk-only individuals.", + "The paired 'observed' support is not independent of Ye 2026. The paired observed cells are a subset of Ye's own scRNA cohort, so 'matches the observed paired cells and Ye 2026' is one source of evidence, not two.", + "Minor: both states are about 25% gdT cells (L2), so 'CD8 effector' is approximate. GZMH is higher in c13 than in c10, so listing GZMH+ as a marker of the destination state is misleading.", + "Strength should be downgraded from 'moderate' to 'weak'. Ye already published the pattern in the same cells, in the same bulk (by deconvolution), by flow cytometry in 122 independent adults, and in external scRNA data (printed pp.153, 156, 161-162). Here it is reproduced in 4/5 draws without group-level gain over bulk." + ], + "reasoning": "**What holds.** The core direction survives every test:\n- **Calibration:** draws 0, 2, 3, 4 reproduce the paired sepsis−HC increase in c10 with the right sign, a level ratio of 0.93-1.96, Spearman 0.40-0.51 and a contrast magnitude of 0.74-1.84x. The bootstrap P(same sign) is 1.0. The c10/c13 balance (shift index) passes in the same draws.\n- **Observed cells:** c10 is higher in adult sepsis than adult HC (+0.070, p 0.0047, permutation p 0.024, robust to leaving out any one HC). The shift is the most robust observed feature (permutation p 0.0014).\n- **Biology:** the c10/c13 pair maps onto Ye's T_CD8_c05-GNLY-CCL4hi / T_CD8_c04-GNLY pair. All 17 of Ye's Fig. 3l marker genes are concordant in direction.\n- **Cluster-free check:** marker rules with observed cells thinned to generated depth agree.\n- **Sensitivity:** dropping sex_flag samples changes nothing.\n\nNothing in the calibration or observed data reverses the main direction, so the verdict is not 'refuted'.\n\n**Why it is only 'qualified'.** Several sub-claims fail, and the finding is narrower and less novel than stated.\n- **Draws:** in practice 4/5, not 5/5. Draw 1's training donors are c10-poor, and its contrast is 0.12x of observed with a CI that includes 0.\n- **The TEMRA half:** the c13 decrease is fragile in observed cells (permutation p 0.24). It matches only by sign.\n- **Sepsis specificity:** the effect is not sepsis-specific. Observed and generated ICU controls carry it too, and sepsis vs ICU is n.s.\n- **Urinary sepsis:** the generator assigns c10 to urinary patients who in observed cells (and in Ye: Ro/e 0.3, urinary < Abd/Res **) have essentially none. Part of the bulk-only 'sepsis' signal is therefore mis-calibrated.\n- **Patient level:** within adult sepsis the generated fraction barely tracks the observed fraction (Spearman 0.04-0.38), despite an observed reliability of 0.89. The per-patient values carry little beyond the group label.\n- **Biological reading:** c10 vs c13 is dominated by immediate-early (IEG/AP-1) genes. In observed cells the IEG rise is T-cell-wide: Memory CD4 equals Memory CD8. Adjusting for it removes most of the c10 and c13 effects. So this is the cytotoxic slice of SEP-POP-3, not a separate CD8-effector reprogramming. The only sepsis-vs-ICU-specific CD8 signal in observed cells is the TNF/IFNG/CCL3/CCL4 cytokine module, which the generator under-expresses and reproduces at only 0.04-0.33x. That module defines Ye's c05.\n\n**Novelty and 'beyond bulk'.**\n- **Already reported:** Ye 2026 reports the pattern on PDF p.4 / printed p.153 (text), PDF p.7 / printed p.156 (Fig. 3d,e,l), and PDF pp.12-13 / printed pp.161-162 (external scRNA, flow cytometry, RT-PCR). Ye's CIBERSORTx deconvolution of the same in-house bulk (ED Fig. 9a, PDF p.39; Methods PDF p.20) already shows it, and that bulk includes our bulk-only individuals.\n- **Group level:** a 10-gene bulk index separates sepsis from HC as well as the generated cells do (AUC 0.895).\n- **Patient level:** only modest, draw-dependent gains among paired adults (draws 2-4), and within sepsis only draw 4.\n\n**Which draws carry the result.**\n- c10 increase: draws 0, 2, 3, 4.\n- c13 decrease: draws 0, 2, 3, 4 by the basic gate; draws 2, 3, 4 by the strict gate.\n- Shift: draws 0, 2, 3, 4.\n- Draw 1 carries nothing.\n\n**Scope of what I ran.** Everything was CPU only under systemd-run with MemoryMax 6G; no GPU work was started. All scripts, logs and tables are in the SEP-POP-4_calib dir. Ye PDF pages were rendered to the session scratchpad only for reading. This verifier ran no archive update: no paper-archive delta was published, and the PBISC-diffusion maintenance step is left to the coordinator.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v08_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v08_pooled_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v08_observed_bh.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v01_fracs_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v01_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v01_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v01_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v01_fractions.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v02_source_calib_imprint.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v02_source_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v02_training_imprint.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v02_obs_reliability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v03_marker_states.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v03_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v03_ieg_coupling.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v04_rule_vs_cluster_and_adjust.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v04_adjusted.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v04_rule_vs_cluster.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v05_bulk_beyond.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v05_bulk_scores_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v05_bulk_beyond_paired.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v05_gen_auc_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v06_c10_vs_c13_ye_genes.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v06_c10_vs_c13.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v06_top_up.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v07_pseudobulk_programmes.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v07_pb_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v07_pb_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v08_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/v09_depth_by_group.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-4_calib/tables/v09_depth_by_group.csv", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s03_cells_T_cell.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s07_candidate_detail.csv" + ] + } + ] + }, + { + "id": "SEP-POP-3", + "lens": "sep_populations", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 12, + "theme": "lymphocytes / immediate-early program", + "synthesis_strength": "weak-to-moderate; handling suspect", + "reportable_statement": "An NR4A2+ GPR183+ AREG+ CCL5-low memory CD4 state with immediate-early genes is higher in bulk-only sepsis than HC in draws 0, 2, 3 and 4 (draw 1 makes almost none). Observed cells agree (all 7 HC at zero). The generator over-produces it in ICU controls 2-5x, it is one sample-level program shared across lineages, and a handling component is likely.", + "carrying_draws": { + "sepsis vs HC": "0,2,3,4 (4 most robust; 2 definition-dependent)", + "sepsis vs ICU": "none" + }, + "honesty": { + "calibration": "Own rows: ratio 1.52/0.13/0.98/0.67/0.98, Spearman 0.28/0.20/0.30/0.39/0.36 (within sepsis 0.09-0.38); rule-based Spearman 0.09-0.31; ICU-HC generated +0.06 to +0.14 vs observed +0.03.", + "replication": "4/5 draws (q<=6.6e-11); survives age, sex, site, depth band, label confidence.", + "observed": "Observed +0.142 (46 vs 7, p 0.0042); rule-based p 0.013-0.085; CEMC site higher within sepsis (0.24 vs 0.115, p 0.03).", + "beyond_bulk": "Overstated: a bulk score of the claim's own genes or bulk CCL5 alone gives AUC 0.92 (= generated); weak per-sample tracking beyond bulk (p 0.001-0.03)." + }, + "verdict_reasons": [ + "ICU over-production", + "observed support rests on 7 HC all at zero", + "bulk IEG batch between bulk-only and paired sets", + "lymphoid sc IEG invisible in same-sample bulk (rho 0.03-0.11)", + "imprint: training-donor level ranks draws (rho 0.9)" + ], + "new_vs_cohort_papers": "Known: Ye 2026 T_CD4_c02-GPR183-NR4A2 (abstract; PDF p.2; Fig. 2b; Fig. 7f flow; ED Fig. 9a). New only: cross-lineage coupling and donor imprint.", + "next_steps": [ + "Processing-time and storage metadata; dissociation-signature scoring; replication in fresh blood or an independent cohort." + ], + "key_sources": [ + "verify/SEP-POP-3_recompute/tables/v10_per_draw_summary.csv", + "verify/SEP-POP-3_calib/tables/v06_key_numbers.csv", + "verify/SEP-POP-3_calib/tables/v03_summary.csv", + "verify/SEP-POP-3_recompute/tables/v07_bulk_ieg_by_group.csv" + ] + }, + "finding": { + "id": "SEP-POP-3", + "cohort": "ye2026_sepsis", + "claim": "In bulk-only generated cells, an NR4A2+ GPR183+ AREG+ CCL5-low memory CD4 T state with immediate-early genes (FOS, JUN, CD69, NFKBIA, TNFAIP3) is a larger share of T cells in adult sepsis than in adult HC in the four draws that pass calibration (0, 2, 3, 4); draw 1 generates almost none of it. Observed paired cells show the same increase. The state moves together with matching immediate-early states in NK, B and CD8 cells across samples, so it is one sample-level program, and this analysis cannot tell in vivo biology from ex vivo handling.", + "kind": "state", + "feature": "T_cell|r2.0|c4 (about r1.0 c0; 86% nested), in L2 'Memory CD4 T cell' (62%). Generated vs observed lfc: NR4A2 +1.27/+1.19, GPR183 +0.46/+0.62, AREG +0.33/+0.67, CCL5 -1.76/-1.38. Share of T cells: bulk-only 8.7%, paired_gen 9.1%, observed 11.0%. Companion states of the same program: NK_cell|r1.0|c10 (FOS, NR4A2, CCL4, CCL3, IFNG) and B_cell|r2.0|c29 (CD83, NR4A2, CD69).", + "comparison": "Adult sepsis (68) vs adult HC (36), bulk-only, per draw; state fraction = cluster cells / T cells. Also vs all controls and vs ICU.", + "effect_bulkonly": "Sepsis vs HC: d0 0.117 vs 0.0085 (diff +0.108, AUC 0.91); d1 0.0020 vs 0.0007 (+0.001, AUC 0.51); d2 0.192 vs 0.021 (+0.171, AUC 0.95); d3 0.142 vs 0.0016 (+0.140, AUC 0.91); d4 0.113 vs 0.0018 (+0.111, AUC 0.89). Vs all controls +0.083/+0.001/+0.141/+0.116/+0.092. Vs ICU +0.03 to +0.08 (q 0.19-0.90, n.s.). Companions, sepsis vs HC: NK c10 +0.343/+0.037/+0.226/+0.148/+0.435 (q<0.05 in d0, 2, 3, 4); B c29 +0.183/+0.001/+0.100/+0.007/+0.138 (q<0.05 in d0, 2, 4).", + "stats": "MWU per draw; BH over 143 features. Sepsis vs HC q = 2.9e-10, 0.80, 2.9e-12, 6.6e-11, 4.5e-10. sex_flag samples dropped: unchanged (d1 0.49).", + "draws_same_sign": "5/5 (q<0.05 in 4/5; draw 1 ~0)", + "paired_calibration": "Own rows. Ratio 1.52/0.13/0.98/0.67/0.98; Spearman 0.28/0.20/0.30/0.39/0.36. Sepsis-HC diff, generated vs observed: +0.161/+0.136, +0.021/+0.151, +0.108/+0.146, +0.098/+0.147, +0.125/+0.139. Gate passes d0, 2, 3, 4; d1 fails (ratio 0.13). A draw's generated level follows its training donors' observed level (Spearman 0.9 across the 5 draws). Draw 1's donors (HC12, Uri-S134, Uri-S147, Res-S164, Res-S166) carry 0/0/0.20/0/0 of this state (s11). Companions: NK c10 gate passes 5/5 (ratio 0.44-1.49); B c29 passes d0, 2, 4 (fails d1, d3: ratio 0.18, 0.09). Related calibration.csv 'l2 | Memory CD4 T cell' (parent): HC contrast sign fails in d0-d2.", + "paired_observed_concordance": "Yes. Observed adult sepsis vs adult HC diff +0.142 (p=0.0042; 46 vs 7); kNN-mapped +0.141 (p=0.0025); vs all controls +0.116 (p=0.00045); vs ICU +0.112 (p=0.11; 46 vs 6). Counterpart: 880 observed cells; marker concordance 0.946. Companions, observed: NK c10 +0.312 (p=0.0075); B c29 +0.183 (p=0.019). In observed cells, the per-sample immediate-early score correlates across T, NK, B and monocytes (Spearman 0.70-0.89). The CD4 state fraction correlates with the NK (0.87), B (0.84) and CD8 CCL4+ (0.75) states. The observed immediate-early score is higher in sepsis than HC in T, NK and B cells (p 0.02-0.03) but not in monocytes; ICU controls are intermediate (n.s.).", + "beyond_bulk": "Yes, partly. The bulk score of its 20 markers separates sepsis from HC only modestly (AUC 0.72), against 0.89-0.95 for the generated fraction in draws 0, 2-4. The bulk score barely tracks the observed fraction (Spearman 0.01-0.06), while the generated fraction does (0.20-0.39); rank-OLS beyond bulk p = 0.010, 0.026, 0.007, 0.001, 0.001 in d0-d4. Within adult sepsis the gain is small (generated 0.09-0.38 vs bulk 0.12). The generated level is partly set by the training donors' single-cell data (imprint), not only by the bulk.", + "novelty": "Population known; coupling not checked in the paper. Ye 2026 reports T_CD4_c02-GPR183-NR4A2 (NR4A2, AREG up; CCL5, KLRB1, TNF down) enriched in abdominal, pulmonary and skin sepsis (abstract PDF p.1; PDF p.2-3 = printed p.151-152, Fig. 2). It also reports AP-1 (FOS/FOSB/JUN/JUNB) activation in B_c03-TCL1A-TNF (PDF p.6 = printed p.155, Ext. Data Fig. 6g,h) and NK_c03-PRF1-CCL4hi elevated (PDF p.8 = printed p.157). Two points are not in the paper: in this cohort's observed cells these states form one sample-level program across lymphoid lineages, and the LoRA-5 level of the program is imprinted by the training donors. Immediate-early genes are also a known ex vivo handling signature; ICU sampling vs healthy volunteers cannot be excluded.", + "strength": "moderate", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s07_candidate_detail.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s08_ap1_lineage_corr.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s08_ap1_cluster_corr.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s08_ap1_obs_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s09_bulk_derivability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s10_within_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s11_training_donor_imprint.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/tables/s13_marker_checks.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/s08_ap1_and_markers.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/s11_training_donor_imprint.py" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a2b77e44ee6f9f6a8", + "verdict": "qualified", + "recomputed": "All numbers come from my own code (v01-v11) run on the raw compact files. From the discovery lens I used only the cluster labels (s03_cells_T_cell.parquet), because the claim is about that exact cluster. My T-cell set matched theirs one-to-one in all 11 sets.\n\nBULK-ONLY, T_cell|r2.0|c4 share of T cells, adult sepsis (68) vs adult HC (36), draws 0/1/2/3/4:\n- Sepsis 0.117/0.002/0.192/0.142/0.113 vs HC 0.0085/0.0007/0.021/0.0016/0.0018.\n- Diff +0.108/+0.001/+0.171/+0.140/+0.111; AUC 0.91/0.51/0.95/0.91/0.89.\n- MWU p 7.0e-12/0.72/5.9e-14/1.2e-12/2.5e-11. BH q over the 143 features is 2.9e-10/0.80/2.9e-12/6.6e-11/4.5e-10, an exact match.\n- Sign is positive in 5/5 draws and significant in 4/5. Pooled shares match the claim: 8.7% bulk-only, 9.1% paired_gen, 11.0% observed.\n- The state is 61-67% L2 memory CD4. c4-vs-other-T lfc profile, bulk-only vs observed: r = 0.97 over 20 genes (NR4A2 +1.20/+1.19, CCL5 -1.68/-1.38).\n\nCONFOUNDERS (rank-OLS):\n- First Affiliated site only + age >= 50 + sex_flag dropped, adjusted for age, sex, log T share and log T median library: p 7.3e-8/0.73/3.3e-19/8.1e-12/7.1e-17.\n- Adding the L2 confidence of non-c4 T cells: p 4.8e-10/0.38/1.1e-14/9.7e-8/5.3e-6. With the all-T confidence instead, d4 drops to p = 0.09, but that covariate contains the state's own cells.\n- Per-cell depth: c4 cells are deeper than other T cells (+400-760 counts) in both generated and observed cells. Inside the draw's depth IQR the diffs are still +0.098/0.000/+0.186/+0.153/+0.084 (p <= 8e-8 except d1).\n- Per-cell L2 confidence >= 0.99: diffs +0.117/+0.002/+0.172/+0.133/+0.127.\n- Wave proxy (sample number): rho within sepsis <= |0.15|; early vs late HC p 0.13-0.66.\n- Site within bulk-only sepsis: p 0.24-0.90.\n- Repeated measures and training-donor leakage: none possible. One sample per donor, and no bulk-only donor is a training donor.\n\nINDEPENDENT DEFINITIONS:\n- Clustering-free rule A: CD8A = CD8B = 0, CCL5 = 0, NR4A2+, >= 2 of FOS/JUN/CD69/NFKBIA/TNFAIP3, after every cell is binomially thinned to 800 counts.\n - Bulk-only diff +0.076/+0.002/+0.075/+0.118/+0.073, p 6.8e-11/0.88/1.5e-12/3.7e-11/1.7e-8.\n - Observed +0.089 (p = 0.052; p = 0.013 at a 500-count target).\n- My own per-set Leiden clustering (r 2.0, cluster picked by a rule fixed in advance):\n - d0/d3/d4 give c4-like clusters (Jaccard 0.54/0.35/0.56), with diff +0.133/+0.080/+0.143 (p <= 1.3e-11).\n - d2 picks a CD8 cluster; d1 has no such state.\n - Observed cells alone give a memory-CD4 NR4A2+ GPR183+ AREG+ CCL5-low cluster, +0.072 (p = 0.006).\n\nCALIBRATION:\n- c4: ratio 1.52/0.13/0.98/0.66/0.98; Spearman 0.28/0.20/0.30/0.39/0.36; within adult sepsis only 0.22/0.21/0.09/0.30/0.38.\n- Paired sepsis-HC diff, generated vs observed: +0.161/+0.136, +0.021/+0.151, +0.108/+0.146, +0.098/+0.147, +0.125/+0.139. The gate passes d0, d2, d3, d4.\n- Rule A at matched depth: Spearman 0.20/0.09/0.09/0.22/0.31.\n- calibration.csv 'l2 | Memory CD4 T cell': the sepsis-HC sign fails in d0-d2 (observed diff only +0.005 to +0.010).\n- Companion gates confirmed: NK c10 passes 5/5; B c29 passes d0, d2, d4.\n\nICU SPECIFICITY:\n- Bulk-only ICU vs HC: +0.077/+0.002/+0.094/+0.078/+0.060, p 5.2e-5/0.94/2.5e-5/1.8e-6/2.2e-6.\n- Sepsis vs ICU: p 0.30/0.85/0.016/0.11/0.20.\n- Paired calibration, ICU-HC generated +0.141/+0.010/+0.085/+0.061/+0.081 vs observed +0.031/+0.031/+0.031/+0.031/+0.037. Sepsis-ICU generated +0.021-0.044 vs observed +0.103-0.121.\n- Rule A ICU-HC, generated +0.098/+0.004/+0.034/+0.050/+0.067 vs observed -0.001 to -0.005: the sign fails.\n\nOBSERVED (all 82 paired samples):\n- Sepsis 0.1425 vs HC 0.000; all 7 HC are exactly 0. p = 0.0042; kNN-mapped p = 0.0025; First Affiliated only p = 0.0074; vs ICU +0.112 (p = 0.11).\n- Within sepsis, the CEMC site is higher than First Affiliated: 0.24 vs 0.115 (p = 0.03).\n\nPROGRAM COUPLING (observed, depth-thinned IEG score):\n- Across T/NK/B/monocyte: Spearman 0.60-0.94. Within adult sepsis 0.77-0.92. Partial on group + depth 0.71-0.93.\n- State fractions, CD4 c4 vs NK c10 / B c29 / CD8 c10: 0.87/0.84/0.75 overall, 0.82/0.82/0.66 within sepsis.\n- Lymphoid single-cell IEG vs same-sample bulk IEG: Spearman 0.05-0.11. The observed c4 share vs bulk IEG: -0.002.\n\nBULK:\n- A bulk score of the claim's own genes (NR4A2, GPR183, AREG, FOS, JUN, CD69, NFKBIA, TNFAIP3 minus CCL5) separates bulk-only sepsis from HC with AUC 0.92, and bulk CCL5 alone also gives 0.92. That equals the generated AUCs of 0.89-0.95. An IEG-only bulk score gives 0.68, close to the discovery lens's 0.72.\n- Batch: bulk IEG6 is far higher in paired HC than in bulk-only HC (+1.3 SD, AUC 0.06). The sepsis-HC bulk IEG contrast is +0.40 SD in the bulk-only set and -0.35 SD in the paired set.\n- Generated vs observed beyond the bulk score (rank-OLS on paired test samples): p 0.006/0.033/0.008/0.001/0.001.\n\nIMPRINT:\n- Training-donor observed means 0.165/0.040/0.087/0.081/0.114. Spearman vs paired_gen means 0.9; vs bulk-only means 0.2.", + "issues": [ + "The bulk-only effect cannot tell sepsis from ICU controls. Bulk-only ICU controls (17) carry 55-70% of the sepsis level: ICU vs HC p 1.8e-6 to 5.2e-5 in d0, d2-d4. Sepsis vs ICU is not significant except in d2 (p = 0.016). In paired calibration the generator over-produces the state in ICU controls 2-5x (ICU-HC generated +0.06 to +0.14 vs observed +0.03). It under-reproduces sepsis-ICU (generated +0.02-0.04 vs observed +0.10-0.12). With the independent depth-matched rule, the ICU-HC sign fails in d0, d2, d3, d4 (generated + vs observed ~0). The bulk-only contrast is therefore an 'ICU patient vs check-up-center healthy' contrast at least as much as a sepsis contrast.", + "Sample-level calibration is weak and depends on how the state is defined. c4 Spearman is 0.20-0.39, and within adult sepsis only 0.09-0.38 (d2 0.09). With the clustering-free rule at matched depth, Spearman is 0.20/0.09/0.09/0.22/0.31, so d2 is near zero and d0/d3 are about 0.2. The gate passes in d0, d2, d3, d4 only under the discovery lens's own cluster and thresholds (Spearman >= 0.2).", + "The observed concordance rests on 7 adult HC, all exactly 0. The same contrast by the independent rule is p = 0.052 at an 800-count target and p = 0.013 at 500. Observed ICU controls number 6.", + "There is a bulk input batch or wave effect. Bulk-only HC (and ICU controls) have much lower bulk FOS, JUN, NR4A2, NFKBIA, CD69 and TNFAIP3 than paired HC (AUC 0.04-0.08). The bulk IEG sepsis-HC contrast flips sign between the bulk-only (+0.40 SD) and paired (-0.35 SD) sets. The bulk-only controls are therefore not exchangeable with the calibration controls on their input. The generated contrast does not appear to be driven by this: paired HC with high bulk IEG still get about 0.", + "The 'beyond bulk' claim is overstated. A bulk score built from the claim's own identity genes separates bulk-only sepsis from HC with AUC 0.92, and bulk CCL5 alone also gives 0.92. That equals the generated fraction (0.89-0.95). The discovery lens's AUC of 0.72 used IEG-only markers. What the generator adds is weak sample-level tracking in paired samples (rank-OLS p 0.001-0.03).", + "There are signs of a handling component, unresolved, as the claim itself says. HC were recruited at the Medical Examination Center, while patients came from the ICUs (Ye 2026 Methods). The observed state is higher in CEMC-site sepsis samples than in First Affiliated ones (0.24 vs 0.115, p = 0.03). The program is coupled across lineages, including monocytes, within sepsis (0.77-0.92). The lymphoid single-cell IEG program is invisible in the same sample's bulk IEG (Spearman 0.05-0.11).", + "The companion states show the same ICU over-production. NK c10 ICU-HC is generated +0.44/+0.29 vs observed +0.15/+0.02 in d0/d4, and its sepsis-ICU sign fails in d0 and d3. For B c29, the sepsis-ICU sign fails in d2.", + "Minor factual errors. In draw 1, the training donor carrying 0.20 is Res-S166, not Uri-S147. 'Generated level follows training donors (Spearman 0.9)' holds for the paired_gen test means only; for the bulk-only means it is 0.2.", + "Adjusting for the mean L2 confidence of all T cells weakens d4 (p = 0.09). That covariate includes the state's own lower-confidence cells. When rebuilt from non-c4 T cells, the effect stays (p <= 5e-6). Per-cell confidence filtering (>= 0.99) also leaves the effect intact.", + "Archive maintenance: this subagent wrote only inside its lens directory and made no change to the paper archive. The paper-archive update (POLICY.md) is left to the coordinating agent." + ], + "reasoning": "Reproduction: every quoted bulk-only, calibration, observed and companion number reproduces exactly. That includes the BH q over the 143 features. My T-cell set matched the discovery lens's one-to-one.\n\nRobustness: the sepsis > HC effect in draws 0, 2, 3, 4 survives every reasonable confounder control I tried:\n- age, sex, site (First Affiliated only) and dropping sex_flag samples;\n- T-cell share, T-cell median library and comparisons inside a depth band;\n- per-cell label confidence and non-c4 T-cell confidence;\n- the sample-number wave proxy.\n\nRepeated measures and training-donor leakage cannot occur in this cohort. The state is also re-derived without the discovery embedding: by a clustering-free, depth-matched marker rule (d0, d2-d4 and observed), and by my own per-set Leiden clustering (d0, d3, d4 and observed). Draw 1 does not generate the state. Sign is positive in 5/5 draws, significant in 4/5. So the core claim is not refuted.\n\nWhy not confirmed:\n1. Sample-level calibration is weak, Spearman 0.2-0.39. Under the independent depth-matched definition, d2 falls to 0.09.\n2. The generator puts the state into ICU controls at 2-5x the observed level. In bulk-only data, ICU controls sit close to sepsis, and the ICU-HC sign fails under the independent rule. The bulk-only finding therefore cannot be called sepsis-related rather than ICU/critical-illness or sampling-setting related.\n3. The observed support rests on 7 HC that are all zero.\n4. The bulk inputs carry a large IEG batch difference between the bulk-only and paired sets.\n5. A trivial bulk score of the claim's genes, or bulk CCL5 alone, matches the generated fraction's group separation (AUC 0.92), so 'beyond bulk' is overstated.\n6. Several observations point towards handling: HC came from a check-up center, the observed state is higher at the CEMC site, and the lymphoid IEG program is not visible in the same sample's bulk.\n\nDraws carrying the result: d0, d2, d3, d4 under the claim's own calibration gate. d4, then d3 and d0, are the most robust. d2 depends on how the state is defined. d1 fails (ratio 0.13).", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/common_v.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v01_recompute_feature.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v02_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v03_independent_marker_state.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v04_program_coupling.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v05_own_clustering.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v06_markers_wave_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v07_followups.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v08_sc_vs_bulk_ieg.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v09_calibration_csv_rows.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v10_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/v11_companion_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v10_per_draw_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v01_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v01_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v01_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v01_imprint.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v02_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v02_depth_conf_filters.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v03_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v03_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v04_ieg_lineage_corr.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v04_state_corr_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v05_own_cluster_profiles.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v05_own_cluster_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v06_bulk_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v06_bulk_derivability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v06_wave_source_site.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v07_bulk_ieg_by_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v07_nonc4_conf_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v07_q_check.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v07_observed_site_effect.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v08_sc_vs_bulk_ieg.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v09_calibration_csv_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_recompute/tables/v11_companion_calibration.csv" + ] + }, + { + "verifier": "calib", + "agent_id": "a36ba458ff54826bb", + "verdict": "qualified", + "recomputed": "I recomputed everything independently from the s03 cluster labels (v01) and checked it against a cluster-free marker rule (v02).\n\nCLUSTER T_cell|r2.0|c4 (cluster cells / T cells), BULK-ONLY, adult sepsis (68) vs adult HC (36):\n- d0: +0.108 [95% CI 0.084, 0.136], AUC 0.91, p 7e-12\n- d1: +0.001, p 0.72\n- d2: +0.171 [0.142, 0.199], AUC 0.95, p 6e-14\n- d3: +0.140 [0.107, 0.174], AUC 0.91, p 1e-12\n- d4: +0.111 [0.085, 0.137], AUC 0.89, p 3e-11\n- The result holds after dropping sex_flag samples, using FAH-only samples, and in a rank OLS adjusted for age and sex (p ≤ 2e-11 in d0, d2, d3, d4).\n- Sepsis vs ICU: +0.031 (p 0.30), -0.000, +0.077 (p 0.016), +0.062 (p 0.11), +0.051 (p 0.20).\n\nPAIRED CALIBRATION (77 test samples per draw; the HC arm is 42-43 sepsis vs 6-7 HC):\n- Ratio: 1.52 / 0.13 / 0.98 / 0.67 / 0.98.\n- Spearman: 0.28 (p 0.015) / 0.20 / 0.30 (0.008) / 0.39 (5e-4) / 0.36 (0.001).\n- Sepsis - HC, generated vs observed: +0.161/+0.136, +0.021/+0.151, +0.108/+0.146, +0.098/+0.147, +0.125/+0.139.\n- So the HC gate passes d0, d2, d3, d4. This matches the claim.\n- All 7 observed HC are exactly 0. Generated HC are 0.001-0.025.\n- ICU contrast: generated +0.02 to +0.04 vs observed +0.10 to +0.12. Generated ICU controls are 0.06-0.16, against 0.03 observed.\n- Within adult sepsis, generated vs observed Spearman: 0.22 (p 0.17), 0.21, 0.09 (p 0.58), 0.30 (p 0.056), 0.38 (p 0.013).\n- Pooled over draws: ratio 0.83, rho 0.39.\n\nOBSERVED (82 samples):\n- Sepsis vs HC: 0.142 vs 0.000, p 0.0042. Leaving out any one HC gives p ≤ 0.0077. FAH-only p 0.0074. Age/sex-adjusted p 0.0028. kNN-mapped view p 0.0025.\n- Sepsis vs ICU: 0.031, p 0.11.\n- 16 of 46 sepsis samples are at 0.\n- Within sepsis, the CEMC site is higher than FAH: 0.240 vs 0.115, p 0.03.\n- Urinary sepsis is 0.003 (n 5).\n\nCLUSTER-FREE RULE (NR4A2+ FOS+ CCL5-, at least 4 of 11 IEGs detected; observed thinned to each draw's generated T-cell depth):\n- Bulk-only sepsis - HC: +0.114 / +0.001 / +0.159 / +0.126 / +0.118. Depth-adjusted p ≤ 4e-6 in d0, d2, d3, d4.\n- The gate passes d0, d3, d4 only. d1 fails on ratio (0.13); d2 fails on rho (0.10).\n- Observed sepsis vs HC: 0.187 vs 0.074, p 0.081 raw and 0.034-0.085 thinned. Sepsis vs ICU: p 0.033.\n\nURINARY SEPSIS:\n- Paired-test generated vs observed: 0.141 / 0.091 / 0.031 / 0.053 vs 0.003.\n- Bulk-only urinary vs HC is significant in d0, d2, d3, d4.\n\nBULK VS SINGLE CELL (same 82 PBMC samples):\n- Per-gene Spearman between bulk and sc pseudobulk for FOS, JUN, JUNB, DUSP1, NR4A2, CD69, ZFP36 and NFKBIA is -0.01 to 0.15. That is the 0.5th-3.5th percentile of 3,014 expressed genes. The median gene is 0.54 and interferon genes are 0.80-0.97.\n- sc IEG score vs bulk IEG score: rho 0.03. Observed c4 fraction vs bulk IEG: rho -0.003. Observed c4 fraction vs sc IEG score: rho 0.88.\n- Bulk IEG score, sepsis vs HC AUC: 0.41 in paired samples, 0.71 in bulk-only. Bulk-only HC are 1.4 SD lower than paired HC (p 8e-6).\n\nCROSS-LINEAGE IEG COUPLING (observed):\n- Within adult sepsis: T-NK 0.94, T-B 0.85, T-monocyte 0.88.\n- Within controls: 0.76-0.81.\n\nBEYOND THE DISEASE LABEL:\n- p for the generated fraction given the sepsis label: 0.08 / 0.06 / 0.17 / 0.015 / 0.013.\n- A bulk sepsis axis learned on bulk-only samples alone separates paired sepsis from HC with AUC 0.97.\n\nTRAINING DONORS, d1:\n- HC12 0, Uri-S134 0, Uri-S147 0, Res-S164 0, Res-S166 0.200.", + "issues": [ + "Not new for this cohort. Ye 2026 already reports this subset (T_CD4_c02-GPR183-NR4A2) as enriched in abdominal, pulmonary and skin sepsis (abstract, PDF p.1 = printed p.150). It also reports NR4A2/AREG/ARID5A up, TNF/CCL5/KLRB1 down and AP-1 TF activity (PDF p.2 = p.151). Fig. 2b (PDF p.5 = p.154) marks Abd, Res and Skin above HC; urinary is n.s.; no difference vs ICU controls. Flow cytometry of NR4A2+ GPR183+ CCR7+ CD45RO+ CD4 T cells (Fig. 7f, PDF p.13 = p.162) gives Abd P=0.0205 and Res P=0.0036 vs HC, with urinary near HC. ED Fig. 9a (PDF p.39) shows the subset from deconvolution of the cohort's own PBMC bulk RNA-seq, low in HC. B_c03 FOS/FOSB/JUN/JUNB activation in Abd/Res/Skin but not urinary sepsis is on PDF p.6 (p.155). The bulk-only result replicates a known finding.", + "The generated state is not specific to sepsis. Paired-test ICU controls are over-generated: 0.16/0.11/0.06/0.09 in d0, d2, d3, d4 vs 0.03 observed. The sepsis-vs-ICU difference is reproduced at only 20-40% of its size. Urinary sepsis is generated at 0.141/0.091/0.031/0.053 vs 0.003 observed, which the paper also reports as not enriched. Bulk-only urinary vs HC is still significant in d0, d2, d3, d4. The generator outputs a generic acute-illness signal, not the source-specific biology.", + "The calibration is thin. The HC arm rests on 6-7 HC test samples, all observed at exactly 0. Spearman 0.28-0.39 comes mostly from group membership. Within adult sepsis, tracking is significant only in d4 (0.38, p 0.013); d3 is 0.30 (p 0.056) and d0/d2 are ns. Beyond the sepsis label, the generated fraction adds information only in d3 and d4.", + "The result depends on how the state is defined. With an embedding-free NR4A2+ FOS+ CCL5- rule on thinned observed cells, the gate passes d0, d3, d4 only; d2 fails (rho 0.10). The observed sepsis vs HC difference falls to p 0.03-0.09. Broad IEG detection is already in ~78% of observed T cells, HC included. The generator inflates the sepsis - HC difference in that detection rate 2-4-fold (generated HC T cells are quieter than observed HC).", + "The evidence points to an ex vivo component, not just 'cannot tell'. The IEG genes are the least bulk-vs-sc concordant genes in the 82 paired samples, even though both assays came from the same cryopreserved Ficoll PBMCs (PDF p.17; time to processing or thaw is not reported). The observed state tracks the sc IEG score (0.88) but not the bulk IEG score (-0.003). The IEG program is shared across the whole sample, including monocytes, and holds within controls. Observed levels also differ by collection site within sepsis (CEMC 0.24 vs FAH 0.12, p 0.03). Evidence for some in vivo biology: low ICU controls, and the paper's NR4A2 protein flow and external datasets.", + "The beyond-bulk statement needs rewording. Bulk IEG content carries none of the observed state (rho about 0) because the sc IEG variation is absent from bulk. The generated per-sample values mostly encode disease status (bulk sepsis axis AUC 0.97) plus imprint from the training donors. They add little per-sample information within sepsis.", + "Factual errors in the claim: (a) the 0.20 among draw 1's training donors belongs to Res-S166, not Uri-S147. This comes from a sample-order bug in s11. (b) The imprint Spearman of 0.9 holds for paired-test generation; for bulk-only generation it is 0.2. (c) '86% nested' means 86% of r1.0 c0 lies inside c4; only 61% of c4 lies in r1.0 c0.", + "The companion states are weaker than stated. The NK c10 bulk-only tests count only samples with at least 10 NK cells: d2 has 3 sepsis samples, d1 has 8, d3 has 18. The NK calibration HC arm has 2-7 HC (d4: 2). B c29 fails d1 and d3." + ], + "reasoning": "Verdict: qualified, not refuted. Calibration and the observed paired cells support the coarse direction (adult sepsis above HC) in d0, d2, d3, d4. Biology and novelty are only partly supported.\n\n(1) Calibration. The feature is not in calibration.csv, so I recomputed the equivalent paired check per draw (paired_gen vs observed on test samples). It matches the claim exactly: the HC-contrast gate passes d0, d2, d3, d4, and d1 fails on ratio. The weaknesses are listed under issues 3, 4 and 2: a 6-7 HC arm, weak within-sepsis tracking, sensitivity to the state definition, and over-generation in ICU controls and urinary sepsis. The parent calibration.csv row ('l2 | Memory CD4 T cell') fails the HC sign in d0-d2, as the claim says.\n\n(2) Observed support. The increase is robust in observed cells: p 0.0042, still ≤ 0.0077 leaving out any HC, and it survives site and age/sex adjustment. Observed ICU controls are low (0.03) and urinary sepsis is about 0. Those are patterns the generator does not reproduce.\n\n(3) Biology and novelty. The state and its sepsis > HC direction are in Ye 2026 (single-cell, flow and bulk deconvolution, with page numbers in issue 1), as is the parallel B-cell AP-1 activation. Only the per-sample correlation across lineages is new, and it is sample-wide, monocytes included. Together with near-zero bulk-vs-sc concordance for IEG genes, that favours a large ex vivo component.\n\n(4) Beyond bulk. The generator does reproduce a state that bulk IEG content cannot predict. But the reason is that the sc IEG program is not in the bulk at all. The generated values mostly encode disease status and imprint from the training donors.\n\nAll analyses were CPU-only under systemd-run with a 6G memory cap, and every write is under SEP-POP-3_calib. This verifier changed no paper assets, so there is no archive delta.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/v_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/v00_pdf_extract.sh", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/v01_recompute_cluster.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/v02_marker_rule.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/v03_bulk_vs_sc_ieg.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/v04_beyond_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/v05_distribution.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/v06_key_numbers.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v06_key_numbers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v01_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v01_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v01_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v01_bulkonly_source.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v01_train_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v01_levels_by_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v02_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v02_bulkonly_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v02_observed_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v02_lineage_coupling.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v03_gene_concordance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v03_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v04_beyond_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v04_source_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/tables/v05_distribution.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/pdf_text/ye2026.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/fig_crops/ye_p5_fig2b.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/fig_crops/ye_p13_fig7f_left.png", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-POP-3_calib/fig_crops/ye_p39_edf9a.png", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf" + ] + } + ] + }, + { + "id": "SEP-PROG-6", + "lens": "sep_programs", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 13, + "theme": "lymphocytes / immediate-early program", + "synthesis_strength": "artifact-suspect", + "reportable_statement": "Generated bulk-only lymphocytes reproduce the cohort's pan-cell-type immediate-early / CCL single-cell factor in draws 0, 3, 4 (CCL) and 0, 2, 3, 4 (NR4A). Its size per draw follows the training donors. The factor is most likely largely ex vivo handling; about 0.6 log2 of CD8/NK-specific CCL signal survives adjustment in observed cells.", + "carrying_draws": { + "CCL program (memory CD8, NK, gdT)": "0,4 strict; 3 lenient", + "NR4A program (memory CD4)": "0,2,3,4" + }, + "honesty": { + "calibration": "Memory CD8 CCL paired ratio 0.77/0.00/0.24/0.39/0.61, within-sepsis Spearman -0.06 to 0.33; calibration.csv has no program rows.", + "replication": "Memory CD8 5/5 same sign (4/5 q<0.05); gdT draw 1 significantly opposite.", + "observed": "+1.6 to +1.7 (46 vs 7); one sample-level factor across all cell types (rho 0.83-0.96); generic IEG adjustment removes ~60%.", + "beyond_bulk": "The claim that the bulk input lacks the signal is wrong: the 7 paired HC bulks sit in a high-IEG batch; bulk-only bulk NR4A +0.74 (p 5e-7)." + }, + "verdict_reasons": [ + "paired-bulk discordance confounded by a high-IEG HC bulk batch", + "training-donor imprint (Spearman 1.0)", + "within-sepsis fidelity weak", + "Ye validation used PMA/ionomycin-stimulated cells" + ], + "new_vs_cohort_papers": "Reported as biology in Ye 2026 (CCL4hi CD8/NK/NKT, PDF p.12; NR4A2+ CD4 TCM, PDF p.1, 14). New: same-sample bulk/sc discordance and HC bulk batch.", + "next_steps": [ + "Handling metadata and dissociation scores; do not report as biology until then." + ], + "key_sources": [ + "verify/SEP-PROG-6_recompute/v08_summary.csv", + "verify/SEP-PROG-6_recompute/v06_bulk_genes_vs_clean_HC.csv", + "verify/SEP-PROG-6_calib/v05_ieg_adjustment.csv", + "verify/SEP-PROG-6_calib/v03_training_donors.csv" + ] + }, + "finding": { + "id": "SEP-PROG-6", + "cohort": "ye2026_sepsis", + "claim": "In bulk-only sepsis vs HC, the generated memory CD8 T, CD56dim NK and γδ T cells express more CCL3/CCL4/CCL4L2/TNF/IFNG (memory CD8 +0.96 log2; strong in draws 0, 3, 4, weaker in draw 2, null in draw 1). The NR4A1/NR4A2/GADD45B/AREG immediate-early program rises in every lymphocyte type. The observed paired cells (+1.6 to +2.0) and the cohort paper show the same. But the same paired samples' bulk RNA-seq shows these genes flat or falling, so the generator appears to transfer a single-cell-specific signal, possibly from ex vivo handling, that its bulk input does not contain.", + "kind": "program", + "feature": "CCL3/CCL4/CCL4L2/TNF/IFNG score in memory CD8 T, CD56dim NK and gdT pseudobulks; NR4A2/NR4A1/GADD45B/HIF1A/AREG score (Ye 2026 'NR4A2' program) in all lymphocyte types", + "comparison": "adult sepsis vs adult HC, bulk-only (68 vs 36; NK 25–66), within draw", + "effect_bulkonly": "CCL program in memory CD8: +1.67, +0.06, +0.68, +0.94, +1.46 log2. CD56dim NK: +1.31, −0.02, +0.42, +0.63, +1.18. gdT: +1.43, −0.17, +0.04, +0.83, +1.47. NR4A2 program in memory CD4: +1.71, +0.22, +1.43, +1.84, +1.86, with similar values in every lymphocyte type.", + "stats": "Memory CD8: Welch p < 1e-20 in draws 0, 3, 4, p 2e-9 in draw 2, p 0.39 in draw 1. BH q < 1e-7 where significant. Robust to age+sex adjustment, same-hospital sepsis and sex-flag exclusion.", + "draws_same_sign": "memory CD8 5/5 (4/5 q<0.05); NK 4/5; gdT 4/5; NR4A2 program in memory CD4 5/5", + "paired_calibration": "CCL program in memory CD8: PASS in draws 0, 3, 4 (paired-test generated/observed 1.19/1.55, 0.66/1.67, 0.99/1.61); SIGN-ONLY in draws 1, 2 (ratio 0.00, 0.24). NK: PASS 0, 4. gdT: PASS 0, 3, 4. Per-sample Spearman 0.19–0.47 (bulk about 0). NR4A2 program in memory CD4: PASS 0, 2, 3, 4. calibration.csv composition rows keep the sepsis−HC sign for 'l2 | Memory CD8 T cell' (5/5, ratio 1.16–1.88), 'l2 | CD56dim NK cell' (5/5, ratio 0.42–0.90) and 'l2 | gdT' (4/5).", + "paired_observed_concordance": "Observed paired within-type change: +1.68 (memory CD8), +1.66 (NK), +1.62 (gdT), p < 1e-6. The rise is broad, though: memory B +2.30 and CD16 monocytes +1.86, and the generated cross-type pattern tracks it poorly (r 0.13–0.28). The NR4A2 program rises in every observed lymphocyte type (+1.4 to +2.2); the generated cells reproduce this (cross-type r 0.83–0.94 in draws 0, 2, 3, 4), so it is not CD4-specific.", + "beyond_bulk": "Apparently yes, but suspect. In the bulk-only samples the bulk CCL score falls (−0.57 log2, p 3e-4). On the paired test samples, both the raw bulk (−0.84 to −0.90) and a bulk/T-marker ratio (−0.09 to +0.04, p >= 0.78) miss the within-CD8 rise that draws 0, 3 and 4 recover. However, on the same paired samples the bulk RNA-seq does not show what the single cells show. Single-cell whole-sample pseudobulk vs bulk log2FC: CCL3 +1.33 vs −1.06, TNF +2.11 vs −0.87, NR4A1 +2.83 vs −0.28, CD83 +2.79 vs −0.07, FOS +0.91 vs −0.79. Spearman over 39 immediate-early/activation genes is 0.36. For the NR4A2 program, paired bulk −0.09 vs +2.03 within memory CD4. So the generator reproduces a single-cell-specific signal that its bulk input does not contain.", + "novelty": "Reported as biology in Ye 2026. CCL4hi proinflammatory CD8/NK/NKT subsets expand in abdominal and pulmonary sepsis, and RT-PCR of purified cells shows more CCL4, CCL3, TNF and IFNG (PDF p.12). NR4A2+ CD4 TCM cells carry exhaustion features (PDF pp.1, 14). The bulk vs single-cell discordance on the same samples, and the finding that the NR4A program is not CD4-specific, are new here.", + "strength": "artifact-suspect", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/lens_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s03_programs.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_draws.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_type_specificity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_types_matrix.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s08_immediate_early_check.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/immediate_early_check.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/de_beyond_bulk_examples_S_vs_HC.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/bulk_baselines.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/site_sensitivity.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "ac9a7f799e49dc282", + "verdict": "qualified", + "recomputed": "How I scored it: my own code. Per sample × L2 pseudobulk (at least 10 cells, 5,975 measured genes), the score is the mean log2(CPM+1) of the five program genes. Contrast is bulk-only adult sepsis (68) vs adult HC (36), Welch test within each draw. Values below are draws 0, 1, 2, 3, 4 in order.\n\nGenerated effect, CCL3/CCL4/CCL4L2/TNF/IFNG program:\n- Memory CD8: +1.97, -0.03, +0.79, +1.06, +1.74 (p 9e-18, 0.80, 7e-9, 5e-19, 4e-21).\n- CD56dim NK: +1.63, -0.42, +0.25, +0.70, +2.00.\n- gdT: +1.79, -0.78 (p 7e-5), +0.05, +1.07, +2.06.\n\nGenerated effect, NR4A program:\n- Memory CD4: +2.07, +0.26, +1.55, +2.45, +2.40 (5/5, all q<0.05, BH over 80 tests).\n- It is positive in 4–5 of 5 draws in every lymphocyte type, and also in CD14 monocytes: +1.44, -0.27, +0.67, +0.61, +1.39.\n\nConfounder controls on the memory CD8 CCL effect:\n- Same-hospital sepsis with sex_flag samples removed: +2.09, +0.05, +0.83, +1.09, +1.83.\n- OLS with age, sex, log2 cell number, log2 UMI per cell, mean L2 label confidence and a monocyte-transcript index: +1.91, +0.72, +0.88, +0.83, +1.28 (all p<4e-6).\n- Expression-matched control genes: +1.99, +0.26, +0.81, +1.19, +2.04.\n- Binomial thinning to one library size: +1.98, +0.43, +0.86, +1.14, +1.96.\n- Only cells with label confidence ≥0.95 and no LYZ/S100A8/S100A9/CST3 counts (200-cell subsample): +2.62, -0.68, +0.57, +1.12, +2.39.\n- Dropping the 4 bulk-only HC that fall in the high-IEG bulk batch: +2.01, +0.01, +0.76, +1.06, +1.80.\n- No bulk-only donor trained any draw (overlap with paired donors is 0).\n\nPaired calibration on each draw's test samples (42–43 sepsis vs 6–7 HC). Memory CD8 CCL, generated/observed difference (ratio):\n- d0: 1.42/1.56 (0.91)\n- d1: -0.06/1.68\n- d2: 0.46/1.72 (0.27)\n- d3: 0.80/1.68 (0.47)\n- d4: 1.12/1.61 (0.69)\n- Spearman over all test samples 0.19–0.34; within sepsis only -0.06 to 0.33.\n\nDraws that pass calibration:\n- Memory CD8 CCL: PASS in d0 and d4 under a strict gate (ratio 0.5–2); d0, d3 and d4 under the discovery gate (ratio 0.25–4).\n- NK CCL: PASS in d0 and d4.\n- gdT CCL: PASS in d0 and d4; d3 has ratio 0.30.\n- NR4A program in memory CD4: PASS in d0, d2, d3 and d4.\n\nObserved paired cells (46 sepsis vs 7 HC):\n- CCL program: memory CD8 +1.68, NK +1.67, gdT +1.63 (leave-one-HC-out range 1.52–1.77).\n- The CCL rise is broad: memory B +1.11, CD16 monocytes +1.20, MAIT +1.43.\n- NR4A program in lymphocytes: +1.58 to +2.18.\n\nDraw sizes follow the training donors. The draws' sepsis training donors have observed memory-CD8 CCL scores above the HC median of +2.80 (d0), +2.02 (d4), +1.63 (d3), +1.22 (d2), +0.83 (d1). That is the same order as the generated effects.\n\nBulk:\n- In the bulk-only cohort the CCL program falls (-0.57, p 3e-4), which reproduces the claim. But the NR4A program rises (+0.74, p 5e-7), and the CCL program relative to cytotoxic-cell markers rises (+1.03).\n- All 7 paired HC bulks belong to a cluster of 11 early-ID HCs with high immediate-early gene (IEG) levels. Their core-IEG score is 9.05–9.65, against 6.26–8.51 for the other 32 HC. Paired HC minus bulk-only HC is +1.30 for IEGs vs -0.06 genome-wide (Mann–Whitney p 1e-19). The top genes are FOSB, JUN, NR4A2, CCL3, EGR1 and FOS.\n- Bulk-only sepsis vs the 32 HC outside that cluster: NR4A2 +0.98, NR4A1 +0.49, GADD45B +0.64, AREG +1.92, FOS +1.14, CCL3 +0.66, core IEG +0.82.\n- Paired sepsis bulk vs the same 32 HC: CCL3 +1.59, TNF +0.68, NR4A1 +1.46, NR4A2 +2.29, CD83 +1.61, FOS +1.62. Against the 7 paired HC (the claim's comparison) the same genes read -1.06, -0.87, -0.28, -0.95, -0.07, -0.79.\n- The single-cell vs bulk discordance on the paired samples reproduces exactly: CCL3 +1.33 vs -1.06, TNF +2.11 vs -0.87, NR4A1 +2.83 vs -0.28. IEG residual -1.14 vs -0.02 for other genes.", + "issues": [ + "Core inference is confounded. The 'same-sample bulk is flat or falling' argument uses the 7 paired HC as reference, and all 7 bulks belong to a high immediate-early gene (IEG) bulk batch (11 HCs with IDs ≤16, including 4 bulk-only HC). Against the 32 HC outside that batch, the same paired sepsis bulks show CCL3 +1.59, TNF +0.68, NR4A1 +1.46, NR4A2 +2.29, CD83 +1.61 and core IEG +1.50. So the single-cell vs bulk discordance comes from the HC reference (high IEG in their bulk, low in their single cells), not necessarily from a sepsis-specific single-cell artifact.", + "The generator's actual input does contain the NR4A/IEG rise, and the claim leaves this out. In bulk-only bulk, the NR4A program is +0.74 (p 5e-7); against the 32 HC outside the batch, NR4A2 is +0.98, AREG +1.92, FOS +1.14 and core IEG +0.82. For CCL only the NK/CD8-specific genes fall (CCL4 -0.76, CCL4L2 -1.09, IFNG -0.74), which matches lymphocyte loss (GNLY -2.39); relative to cytotoxic markers the CCL program rises +1.03. 'Beyond bulk' is not established.", + "Calibration is weaker than claimed. Under a strict gate (generated/observed ratio 0.5–2) the CCL program passes only in d0 and d4 for memory CD8, NK and gdT. d3 passes only under the lenient 0.25–4 gate (memory CD8 0.47, gdT 0.30). d2 ratios are 0.27 (CD8), 0.24 (NK) and about 0 (gdT). Within-sepsis per-sample Spearman is only -0.06 to 0.33, so the generator reproduces the group shift, not sample-level variation.", + "calibration.csv has no expression-program rows, so it is not the trust layer for this feature. The composition 'ratios' the claim cites (1.16–1.88 for memory CD8, 0.42–0.90 for NK) are level ratios (gen_mean/obs_mean), not contrast ratios. The NK composition contrast in d4 is about 0 (-0.0002 vs -0.084), and gdT d0 has the opposite sign.", + "Draw 1 contradicts rather than being null. gdT is significantly negative (-0.78, p 7e-5 with my score; -0.17 in the discovery score) and NK is -0.42. The sign for memory CD8 in d1 depends on the score (-0.03 vs +0.06), so '5/5 same sign' is fragile.", + "Draw effect sizes are imprinted by the 5 LoRA training donors. The sepsis training donors' observed memory-CD8 CCL excess over the HC median ranks d0>d4>d3>d2>d1 (+2.80 to +0.83), the same order as the generated effects. Draw 1 trained on two urinary donors with low CCL.", + "The generated NR4A rise is not lymphocyte-specific. CD14 monocytes rise +1.44/+0.67/+0.61/+1.39 in d0/d2/d3/d4 against +0.64 observed. The cross-type agreement (r 0.75–0.90 over 13 types) is mostly lymphocyte vs monocyte; among the 10 lymphocyte types r is 0.07–0.68.", + "The generator misses the source specificity. The generated memory-CD8 CCL rise is present in urinary sepsis (d0 +1.68, d3 +0.80, d4 +1.38). Observed paired urinary sepsis is +0.50 (n=5, not significant), and Ye 2026 find CCL4hi enrichment in abdominal and respiratory sepsis, not urinary.", + "Ye 2026 RT-PCR and ELISA were done on purified CD8/NK/NKT cells after 4 h PMA/ionomycin stimulation (Methods: RNA extracted from stimulated cells). This shows induced capacity, not baseline ex vivo expression.", + "Magnitudes depend on the score. With log2(CPM+1) the observed rises are memory B +1.11 and CD16 monocytes +1.20, against the claim's +2.30 and +1.86 (edgeR prior-count score). The breadth of the rise holds either way.", + "Within paired sepsis, per-sample single-cell IEG, CCL and NR4A scores do not track bulk (Spearman 0.04–0.21). The generator does partly read bulk IEG level: the 4 high-IEG bulk-only HC get higher generated NR4A in memory CD4 (+0.56, p 0.03 in d1; +0.51, p 0.01 in d4). Whether the program is biology or handling remains open." + ], + "reasoning": "I recomputed everything with independent code from the raw files: pb_l2.npz, cells.parquet, subsample.npz, meta.csv, the bulk npz files and the panel mask. All of it ran on CPU in a systemd scope capped at MemoryMax=6G; no GPU work was started.\n\n**The generated effect reproduces.** The within-type CCL3/CCL4/CCL4L2/TNF/IFNG rise in memory CD8, NK and gdT cells, and the NR4A program rise across lymphocytes, come out with the same draw pattern as claimed: strong in d0, d3 and d4, weaker in d2, and null to negative in d1.\n\n**It survives every confounder control I tried:**\n- age and sex;\n- same-hospital sepsis only, and sex_flag samples removed;\n- cell number, UMI per cell and mean label confidence;\n- a monocyte-transcript contamination index, plus cells with confidence ≥0.95 and no monocyte transcripts;\n- expression-matched control genes and binomial thinning;\n- removing the high-IEG batch HCs.\n\nDonor leakage cannot apply: sepsis is cross-sectional and no bulk-only donor trained any draw. The observed paired cells agree (+1.63 to +1.68), robustly to leaving out any one HC.\n\n**Where it falls short:**\n- **Calibration.** Under a strict gate it passes only in d0 and d4; d3 passes only under the lenient gate. Sample-level tracking within sepsis is weak. The effect size in each draw follows its five training donors. calibration.csv holds no program rows, and the composition 'ratios' quoted in the claim are level ratios, not contrast ratios.\n- **The 'beyond bulk / single-cell-specific artifact' conclusion does not hold.** The paired-bulk comparison relies on 7 HC bulks that all sit in a high-IEG bulk batch, separated from the other HCs by IEGs only (genome-wide shift about 0). Against the HCs outside that batch, the bulk of the very same sepsis samples shows the IEG, NR4A, CCL3 and TNF rise. The bulk-only cohort's own bulk, which is what the generator reads, shows an NR4A/IEG rise too.\n\nThe generated within-type program is therefore real as a model output and is calibrated in d0 and d4 (d3 marginal). Whether it is in vivo biology or ex vivo handling is unresolved, and the claim that the bulk input lacks the signal is contradicted. I rate it qualified rather than refuted because the effect itself reproduces and survives.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/vcommon.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v01_bulkonly_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v01_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v01_tests_wide.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v01_balance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v01_source.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v01_donor_overlap.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v02_calibration_observed.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v02_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v02_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v02_observed_by_source.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v02_crosstype.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v03_bulk_checks.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v03_bulk_program.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v03_sc_vs_bulk_summary.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v03_sc_vs_bulk_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v04_label_confidence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v04_label_confidence.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v05_bulk_wave.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v05_bulk_wave_tests.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v05_bulk_scores_per_sample.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v05_hc_paired_vs_bulkonly_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v06_ieg_batch_effects.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v06_bulk_genes_vs_clean_HC.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v06_generator_batch_response.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v06_S_vs_HC_without_batchHC.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v06_paired_sepsis_sc_vs_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v06_batch_members.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v07_training_donors.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v07_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v08_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v08_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_recompute/v08_carrying_draws.csv" + ] + }, + { + "verifier": "calib", + "agent_id": "a5bb8f311c4a0212a", + "verdict": "qualified", + "recomputed": "I recomputed everything independently with my own code (vlib.py and v01–v06). CPU only, inside systemd-run scopes capped at 6G.\n\nBULK-ONLY RESULT (adult sepsis − adult HC, log2; draws 0/1/2/3/4)\n- CCL score, memory CD8: +1.68 / +0.06 / +0.69 / +0.94 / +1.48. Welch p 1.4e-20 / 0.40 / 1.8e-9 / 6e-21 / 6e-24.\n- CCL score, CD56dim NK: +1.32 / −0.03 / +0.42 / +0.64 / +1.19.\n- CCL score, gdT: +1.44 / −0.18 / +0.04 / +0.83 / +1.49. Draw 1 is significantly opposite (q 0.03).\n- NR4A score, memory CD4: +1.70 / +0.22 / +1.43 / +1.85 / +1.87.\n- Summing the program counts before taking the log gives the same pattern.\n\nPAIRED-TEST CALIBRATION (generated diff / observed diff, ratio)\n- Memory CD8 CCL: 1.19/1.55 (0.77), 0.00/1.67 (0.00), 0.41/1.71 (0.24), 0.66/1.67 (0.39), 0.99/1.61 (0.61). Gate: PASS in draws 0, 3, 4.\n- NK CCL: PASS in draws 0 and 4 (also draw 3 with the summed score).\n- gdT CCL: PASS in draws 0, 3, 4.\n- NR4A in memory CD4: PASS in draws 0, 2, 3, 4 (ratios 0.61–0.85). Draw 1 is SIGN-ONLY here and FAIL in 4 other types.\n- Dropping any one test HC changes the observed difference only within 1.45–1.65.\n- Spearman, generated vs observed, memory CD8 CCL (draws 0–4):\n - all test samples: 0.34 / 0.19 / 0.22 / 0.34 / 0.31;\n - adult sepsis only (n 42): 0.20 / 0.17 / −0.03 / 0.16 / 0.33;\n - after removing each group's mean: 0.19 / 0.22 / 0.03 / 0.16 / 0.22. Draw 1, which fails the contrast, scores as high as the passing draws.\n\nOBSERVED PAIRED ADULTS (46 sepsis vs 7 HC)\n- CCL: +1.68 (memory CD8), +1.66 (NK), +1.62 (gdT). Dropping one HC at a time: 1.52–1.76. Mann-Whitney p 0.002–0.003.\n- NR4A: +1.39 to +2.16 in every lymphocyte type; +0.64 (CD14 monocytes) and +0.90 (CD16 monocytes).\n- All 7 HC sit below the sepsis 25th percentile.\n\nONE SAMPLE-LEVEL FACTOR IN THE OBSERVED CELLS (46 sepsis samples)\n- CCL in memory CD8 correlates with NR4A in memory CD4, naive B and CD14 monocytes at rho 0.89–0.92.\n- It correlates with a generic immediate-early score (FOS, JUN, EGR1, DUSP1 and similar) in CD4 T, B and monocytes at rho 0.83–0.90.\n- All of these correlate with the sample's median per-cell library size at rho −0.44 to −0.57.\n- Holding that generic immediate-early score fixed in the same cells shrinks the effects:\n - CCL: memory CD8 1.68 → 0.67 (p 0.008); NK 1.66 → 0.69; gdT 1.62 → 0.56 (p 0.06);\n - NR4A in memory CD4: 2.06 → 0.84;\n - CCL in B cells: falls to not significant.\n- After the same adjustment, generated draws 0, 3 and 4 keep 0.50–0.77 in memory CD8.\n\nBULK\n- Bulk-only samples, sepsis − HC:\n - NR4A program +0.74 (p 5e-7);\n - immediate-early score +0.58 (p 0.002);\n - CCL raw −0.57, but relative to CD8 / NK / T markers +0.76 / +0.89 / +0.79 (p ≤ 9e-5).\n- Paired samples (46 vs 7): CCL −0.90, NR4A −0.09.\n- The 7 paired adult HC vs the 36 bulk-only adult HC, in bulk:\n - immediate-early score +1.73, NR4A program +1.39, CCL program +1.07 (p ≤ 1e-4);\n - 29 of the 50 most-raised genes are immediate-early or activation genes: CXCL8 +4.33, FOSB +3.88, JUN +3.22, NR4A2 +2.90, CCL3 +2.35;\n - HBB is lower by 4.18.\n- The same paired vs bulk-only gap: ICU controls +1.67; sepsis +0.52.\n- Paired sepsis vs all 43 HC: immediate-early score +0.82 (p 0.001), NR4A program +1.07.\n- Per-sample agreement between single cells and bulk on the same 82 paired samples (Spearman): FOS 0.02, JUN 0.02, DUSP1 −0.01, NR4A1 −0.04, TNF 0.00. Lineage marker genes agree at 0.48–0.90.\n\nDOES THE GENERATOR READ THE BULK? (same draw, one shared normalisation)\n- Generated paired-test HC sit above generated bulk-only HC on the immediate-early score: +0.37–0.42 (draw 0), +0.66–0.91 (draw 1), 0 (draw 2), +0.45–0.53 (draw 3), +0.56–0.66 (draw 4); p ≤ 0.034 in draws 0, 1, 3, 4.\n- For NR4A in memory CD4 the gap is +0.31 (draw 0), +0.39 (draw 1) and +0.47 (draw 4). That puts the bulk-derived share of the bulk-only NR4A contrast at roughly 10–15% in draws 0 and 4.\n\nTRAINING-DONOR IMPRINT (5 draws, descriptive only)\n- Spearman between a draw's bulk-only CCL effect and its 4 training sepsis donors' observed score minus the HC median: 1.0 for memory CD8 and NK, 0.9 for gdT.\n- Draw 0's training sepsis donors sit at the 52nd–94th percentile of observed sepsis; three of draw 1's sit at the 9th–33rd and the fourth at the 72nd.", + "issues": [ + "Wrong for the bulk-only inputs: 'its bulk input does not contain [the signal]'. In the 68 vs 36 bulk-only samples (the generator's actual input), sepsis bulk already rises versus HC: NR4A program +0.74 (p 5e-7); immediate-early score +0.58 (p 0.002); FOS +0.89, DUSP1 +1.34, EGR1 +1.09, NR4A2 +0.64. Relative to CD8 / NK / T markers the CCL program also rises: +0.76 / +0.89 / +0.79 (p ≤ 9e-5). The claim quotes only the raw bulk-only CCL score (−0.57) and gives the marker-ratio result only for the paired samples.", + "The paired bulk is 'flat or falling' because of its 7 controls. Their bulk carries an ex vivo activation-type signature compared with the 36 bulk-only HC: immediate-early score +1.73, NR4A program +1.39, CCL program +1.07 (p ≤ 1e-4). Of the 50 most-raised genes, 29 are immediate-early or activation genes (CXCL8 +4.33, FOSB +3.88, JUN +3.22, NR4A2 +2.90, CCL3 +2.35), and HBB is lower by 4.18. The paired ICU controls show the same gap (+1.67); paired sepsis much less (+0.52). Compared with all 43 HC, paired sepsis bulk rises (immediate-early +0.82, p 0.001; NR4A +1.07). So the same-sample bulk vs single-cell discordance reflects handling in both assays, not a purely single-cell signal.", + "The generator partly reads the bulk's immediate-early content. Generated paired-test HC sit +0.37 to +0.91 above generated bulk-only HC on the immediate-early score (draws 0, 1, 3, 4; p ≤ 0.034). Roughly 10–15% of the bulk-only NR4A contrast in draws 0 and 4 is bulk-derived. Most of the effect is still imprint (next item).", + "How strong each draw's effect is follows its training donors. Across the 5 draws, the Spearman between the bulk-only CCL effect and the 4 training sepsis donors' observed score minus the HC median is 1.0 (memory CD8, NK) and 0.9 (gdT). Draw 0's training sepsis donors sit at the 52nd–94th percentile of observed sepsis; three of draw 1's sit at the 9th–33rd. So the effect size is the training donors' single-cell state, not a property of the bulk-only patients.", + "The observed rise is one sample-level factor, not a CD8/NK program. Within the 46 sepsis samples, CCL in memory CD8 correlates rho 0.89–0.92 with NR4A in memory CD4, naive B and CD14 monocytes, and rho 0.83–0.90 with a generic immediate-early score; all correlate rho −0.44 to −0.57 with lower per-cell library size. Holding the generic score fixed in the same cells removes about 60% of each effect: CCL 1.68→0.67 (memory CD8), 1.66→0.69 (NK), 1.62→0.56 (gdT, p 0.06); NR4A in memory CD4 2.06→0.84; CCL in B cells becomes not significant.", + "Calibration limits:\n- the same 6–7 HC serve every draw;\n- the generated contrast reaches only 0.39–0.77 of the observed for CCL and 0.61–0.85 for NR4A in memory CD4;\n- within adult sepsis, the generated vs observed Spearman for memory CD8 CCL is −0.03 to 0.33 (n 42), so the quoted 0.19–0.47 mostly reflects the sepsis/HC split; after removing group means, draw 1 (which fails the contrast) scores 0.22, as high as the passing draws;\n- calibration compares against the same single-cell assay, so it cannot tell biology from handling.", + "Paper support is weaker than quoted:\n- The RT-PCR and ELISA (PDF p.12) were run on magnetically purified CD8, NK and NKT cells after 4 h of PMA/ionomycin stimulation (Methods, PDF p.21; ED Fig. 10 legend, PDF p.42 'following short-term stimulation'). They show stimulated capacity, not resting transcript levels.\n- Flow cytometry followed 4 h of culture with brefeldin A (PDF p.21).\n- All adult single-cell HC comparisons use the same 7 HC (Fig. 1b, PDF p.3; ED Fig. 10 legend, PDF p.42).\n- The paper reads NR4A1/NR4A2 and AP-1 (JUN, JUND, FOSB) as exhaustion regulators (PDF p.2) and never discusses handling artefacts.\n- Plasma Olink CCL3/CCL4/TNF are raised in sepsis (PDF p.12): in vivo inflammation, but the cell source is unspecified.\n- COMBAT 2022 found a pan-cell-type AP-1 (FOS/JUN/ATF) module that was high in COVID-19 but 'distinct from influenza and sepsis' (PDF p.6). That cohort's cryopreserved PBMCs were run in batches balanced across groups (PDF p.48).", + "Novelty is partly overstated. 'The NR4A program is not CD4-specific' is partly in the paper already:\n- MOFA factor 2's top genes include CCL3, CCL4, TNF, IFNG and NR4A2, driven by CD8, NK, NKT and CD14-IL1B subsets (PDF p.14);\n- the paper calls NR4A2 'a broader regulator of immune exhaustion across lymphocyte lineages' (PDF p.14);\n- its CCL4-high CD8 subset also up-regulates GADD45B (PDF p.4).\nThe genuinely new parts are the per-sample independence of single-cell and bulk immediate-early levels, and the handling signature in the paired controls' bulk.", + "Minor points:\n- Draw 0 memory CD8 Welch p is 1.3e-20 (lens) / 1.4e-20 (mine), not < 1e-20.\n- gdT in draw 1 is significantly opposite (−0.18, q 0.03), not merely null.\n- Draw 1's NR4A cross-type r is also 0.87; its effects are just tiny.\n- The paired-test bulk/T-marker ratio for draw 1 is −0.18 (p 0.62), outside the quoted −0.09 to +0.04.\n- The calibration.csv rows the claim cites are cell-type proportions, not expression; they say nothing about the programs." + ], + "reasoning": "Verdict: qualified. The generated contrast holds where the claim says it does, and the 'artifact-suspect' label is well founded (the evidence is stronger than the claim states). But the 'beyond bulk' explanation needs rewriting.\n\n1. Calibration. The program scores are not in calibration.csv, so I recomputed the paired check myself on each draw's test samples.\n- The direction is reproduced by draws 0, 3 and 4 for CCL in memory CD8 and gdT, by draws 0 and 4 for NK (also draw 3 with the summed score), and by draws 0, 2, 3 and 4 for NR4A in CD4 cells.\n- Draw 1 fails (generated difference about 0, or FAIL in several types). Draw 2 keeps only the sign for CCL.\n- Magnitude is under-reproduced: 0.39–0.77 of the observed for CCL.\n- Per-sample fidelity within sepsis is weak, and the reference is the same 6–7 HC in every draw.\n- Calibration compares against the same single-cell assay, so it confirms that the generator copies that assay, artefacts included.\n\n2. Replication across draws: memory CD8 5/5 same sign, with 4/5 significant at q < 0.05. NK and gdT 4/5 (gdT draw 1 is significantly opposite). NR4A in memory CD4 5/5. The draws that carry the result are 0, 3, 4 (CCL) and 0, 2, 3, 4 (NR4A).\n\n3. Observed paired cells support the direction: +1.6 to +2.1, robust to dropping any one of the 7 HC. But the rise is a single sample-level factor shared by every cell type, including monocytes and B cells (rho 0.83–0.96). It tracks lower per-cell library size and is about 60% explained by a generic immediate-early score. The cohort paper reports CCL4-high CD8/NK/NKT cells and NR4A2+ CD4 central memory cells as biology (PDF pp.1, 2, 4, 12, 14). Its RT-PCR/ELISA validation was on PMA/ionomycin-stimulated cells (PDF pp.21, 42). COMBAT found a pan-cell-type AP-1 module that is distinct from sepsis (PDF p.6). Together these favour handling over in vivo biology for most of the signal; a CD8/NK-specific remainder of about 0.6 log2 survives the adjustment.\n\n4. Beyond bulk.\n- On paired samples, single-cell and bulk immediate-early levels are independent per sample (rho ≈ 0, versus 0.48–0.90 for lineage markers). That supports a single-cell-specific component.\n- The paired bulk 'falls' because the 7 paired HC (and the 6 paired ICU controls) carry an ex vivo activation-type signature in their bulk, relative to the 36 bulk-only HC.\n- The bulk-only bulk, which is the input that matters here, already shows a sepsis rise in NR4A and immediate-early genes, and a lineage-relative CCL rise. The generator partly reads it (about 10–15% of the effect in draws 0 and 4).\n- Most of each draw's effect is an imprint of its 4 training sepsis donors' single-cell state (Spearman 1.0 across draws).\n- So 'transfers a learned single-cell pattern' holds, but 'its bulk input does not contain the signal' is false for bulk-only samples, and the paired discordance is confounded by bulk handling.\n\nSuggested rewording: generated bulk-only lymphocytes reproduce the cohort's pan-cell-type immediate-early/CCL single-cell factor in draws 0, 3, 4 (CCL) and 0, 2, 3, 4 (NR4A). Its size per draw follows the training donors. The factor is most likely ex vivo handling; it is not a CD8/NK-specific biological finding.\n\nNo GPU was used. I wrote only to my verify directory. There is no archive delta (verifier scratch outputs only).", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/vlib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v01_programs_calib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v01_programs_calib.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v01_bulkonly.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v01_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v01_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v01_observed_hc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v02_bulk_checks.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v02_bulk_checks.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v02_bulk_groups.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v02_bulk_program_scores.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v02_sc_vs_bulk_per_gene.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v02_gen_vs_bulk_input.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v03_training_imprint.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v03_training_imprint.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v03_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v03_draw_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v04_crossset_and_sample_factor.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v04_crossset_and_sample_factor.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v04_crossset_levels.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v04_observed_sample_factor.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v04_observed_covariates.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v05_ieg_adjustment.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v05_ieg_adjustment.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v05_ieg_adjustment.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v06_bulk_set_effect.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v06_bulk_set_effect.log", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v06_bulk_scores_by_set_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v06_bulk_pairedHC_vs_bulkonlyHC_top.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/v06_bulk_sepsis_vs_allHC.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-6_calib/paper_text/", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/COMBAT2022_Blood_atlas_COVID19_severity.pdf" + ] + } + ] + }, + { + "id": "SEP-PROG-5", + "lens": "sep_programs", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 14, + "theme": "lymphocytes / B lineage", + "synthesis_strength": "weak: direction only", + "reportable_statement": "Plasmablast genes rise in the generated B-lineage pseudobulk of bulk-only sepsis in all five draws, through more plasma cells. Only the direction is calibrated: plasma cells are under-generated and the log-scale size is inflated by a near-zero floor in generated HC.", + "carrying_draws": { + "direction": "0,1,2,3,4", + "partly calibrated": "3 (weakest effect); 0 marginal" + }, + "honesty": { + "calibration": "Paired Spearman 0.03-0.28 (bulk 0.67-0.71); log ratio 0.81-2.41 but linear-CPM 0.13-0.73x; l2 Plasma cell ratio 0.15-0.46.", + "replication": "5/5 positive, 4/5 q<0.05 (draw 3 null on means).", + "observed": "Observed +1.02 (46 vs 7, p 0.022; MWU 0.056); plasma fraction +0.0097 (p 0.0002).", + "beyond_bulk": "None: bulk +0.94 log2 (+1.38 relative to B markers) and ranks samples far better." + }, + "verdict_reasons": [ + "zero-plasma floor in generated HC", + "composition, not expression", + "draw 1/2/4 SIGN-ONLY due to Spearman, not overshoot", + "MKI67 attribution misstated for draw 4" + ], + "new_vs_cohort_papers": "Known: Ye 2026 B_c06-MZB1 / B_c07-MKI67 increased and tied to survival using this cohort's bulk (PDF p.6, p.19; ED Fig. 6e).", + "next_steps": [ + "Use linear-scale contrasts; fix plasma-cell under-generation before any magnitude claim." + ], + "key_sources": [ + "verify/SEP-PROG-5_recompute/v05_summary_by_draw.csv", + "verify/SEP-PROG-5_recompute/v04_zero_plasma_floor.csv", + "verify/SEP-PROG-5_calib/v03_per_draw_summary.csv" + ] + }, + "finding": { + "id": "SEP-PROG-5", + "cohort": "ye2026_sepsis", + "claim": "Plasmablast genes (JCHAIN, MZB1, TNFRSF17, TXNDC5, PRDM1, IRF4) rise in the generated B-lineage pseudobulk of bulk-only sepsis vs HC (+1.9 log2 mean; 5/5 draws), matching the observed direction (+1.0). However, the program calibrates in draws 0 and 3 only, plasma-cell abundance is under-generated (0.15–0.46×), per-sample tracking is poor, and the bulk already shows the rise.", + "kind": "program", + "feature": "plasmablast score (JCHAIN, MZB1, TNFRSF17, TXNDC5, PRDM1, IRF4) in the six-type 'B cell' pseudobulk, which includes AIFI-L2 plasma cells", + "comparison": "adult sepsis vs adult HC, bulk-only (66–68 vs 36), within draw", + "effect_bulkonly": "+3.13, +1.86, +1.55, +0.40, +2.62 log2 (draws 0–4; mean +1.91, d 1.5). All 6 genes rise in 5/5 draws (draw 3 small).", + "stats": "Welch p 4.9e-19, 2.0e-11, 9.7e-11, 0.12, 3.1e-12. Within-draw BH q < 0.05 in 4/5. Mann–Whitney p <= 0.002 in 5/5. Stable with age+sex adjustment, same-hospital sepsis and sex-flag exclusion (site_sensitivity.csv).", + "draws_same_sign": "5/5 (4/5 q<0.05)", + "paired_calibration": "programs_calib.csv: PASS in draws 0 and 3 only. Draws 1, 2 and 4 are SIGN-ONLY: they overshoot (ratio 1.97, 1.22, 2.35) and track samples poorly (per-sample Spearman 0.03–0.28, against 0.67–0.71 for the bulk). calibration.csv row 'l2 | Plasma cell': sepsis−HC sign kept 5/5, but ratio 0.15–0.46 and Spearman 0.21–0.39. Row 'Plasmablast (markers)': ratio 0.09–0.24. Within-plasma-cell DE cannot be calibrated: 2–16 test samples have >=10 generated plasma cells, and none of them are HC.", + "paired_observed_concordance": "Observed paired B-cell pseudobulk +1.02 (p 0.022, 46 vs 7). In observed cells the bulk rise of JCHAIN/MZB1/TXNDC5/TNFRSF17 is carried by plasma cells, as in all 5 draws.", + "beyond_bulk": "None. The bulk plasmablast genes rise +0.94 log2 (+1.38 relative to B markers), and the bulk ranks samples far better. The plasma-cell attribution equals the reference-atlas answer. The bulk MKI67 rise is carried by plasmablasts in observed cells (7.2 of 12.1 CPM); the generated cells assign it to proliferating T cells or show no rise (draws 0, 3).", + "novelty": "Known. Ye 2026 PDF p.6: B_c06-MZB1 plasma cells and B_c07-MKI67 plasmablasts increased across infection sources and associated with poor survival.", + "strength": "weak", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s03_programs.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_draws.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/program_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/bulk_baselines.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_genes_fine.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/attribution_lineage_fine.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/site_sensitivity.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a495437a74c38db77", + "verdict": "qualified", + "recomputed": "Score is the six plasmablast genes in the six-type 'B cell' pseudobulk (this gate holds every AIFI-L2 plasma cell). Contrast is bulk-only adult sepsis vs adult HC, 68 vs 36 (66 vs 36 in draw 3), within each draw.\n\nReproduction:\n- Discovery metric (S3, edgeR-like prior), my own code, reproduced exactly: +3.13, +1.86, +1.55, +0.40, +2.62 (draws 0–4). Welch p 4.9e-19, 2.0e-11, 9.7e-11, 0.124, 3.1e-12. Mann–Whitney p 9e-14, 2.5e-9, 7.6e-9, 0.0022, 4.8e-12.\n- S1, mean of log2(CPM+1): +3.59, +2.42, +1.88, +0.17 (p 0.65), +2.97.\n- S2, log2 of the summed program CPM: +3.56, +2.39, +1.87, +0.11 (p 0.79), +3.14.\n- Sign is positive in 5/5 draws under all three scores. Draw 3 is null on means (Welch p 0.12–0.79) but shifted on ranks (Mann–Whitney p 0.002–0.005).\n- Per gene under S1, draw 3: JCHAIN −0.14, MZB1 −0.09, TXNDC5 −0.02, TNFRSF17 +0.06, IRF4 +0.18, PRDM1 +1.05. So \"all 6 genes rise in 5/5 draws\" depends on the normalisation.\n\nConfounder controls, S1, draws 0/1/2/3/4:\n- OLS with age and sex: 3.55, 2.39, 1.85, 0.22 (p 0.56), 2.88.\n- Plus log B-cell count, log library per B cell, bulk library size and bulk no-feature fraction: 2.64, 1.91, 1.94, 0.78 (p 0.08), 2.34.\n- Same hospital and no sex_flag: 3.57, 2.70, 2.09, 0.16, 3.13.\n- 1:1 age-matched: 3.67, 2.66, 1.64, 0.48, 3.04.\n- B-gate cells with label confidence ≥ 0.9: 3.90, 2.62, 2.03, 0.22, 3.18.\n- Depth-matched by thinning to 50k counts: 3.99, 2.53, 2.14, 1.25, 3.28.\n- 31 vs 31 matched on the bulk no-feature fraction: 2.98, 2.04, 1.50, 0.49 (p 0.25), 2.48.\n- Permutation within age-tertile × sex strata: p 0.0005 in draws 0, 1, 2, 4; p 0.71 in draw 3.\n- Random 6-gene sets matched on expression: the plasmablast set lies above the null maximum in draws 0, 1, 2, 4 (p 0.001) and in observed (p 0.008); in draw 3, p 0.20.\n- Draw-centred, sample-level summary: +2.22, p 2e-13.\n\nDecomposition:\n- The rise is carried by plasma cells. Plasma cells per 1,000 generated cells, sepsis vs HC: 12.4 vs 0.48, 3.8 vs 0.74, 3.3 vs 0.52, 4.1 vs 2.83 (Mann–Whitney p 0.53), 11.6 vs 1.37. Observed paired: 11.4 vs 1.66.\n- Non-plasma B cells change little: +0.24, −0.35, +0.27, −1.08, +0.50. In observed cells they fall (−0.44).\n- Zero-plasma floor: bulk-only HC with zero generated plasma cells are 53%, 28%, 47%, 3%, 25%. Observed: 0/7 adult HC and 2/46 adult sepsis. In draw 3 the floor runs the other way: 18% of sepsis samples have zero plasma cells. Restricted to samples with ≥1 plasma cell: +2.29, +1.80, +1.13, +1.14, +2.18.\n\nObserved paired, 46 vs 7: +1.02 (S1/S3, Welch p 0.021), +1.06 (S2, p 0.028). Mann–Whitney p 0.056–0.059.\n\nPaired calibration of this score (S3):\n- Spearman generated vs observed: 0.23, 0.16, 0.03, 0.28, 0.17. Bulk vs observed: 0.67–0.71. Partial Spearman given the bulk scores: −0.30, −0.22, −0.10, +0.03, −0.23.\n- Ratio generated/observed contrast: 2.41, 1.97, 1.22, 0.81, 2.35, on 6–7 HC.\n- Discovery gate: PASS in draws 0 and 3, the rest SIGN-ONLY. Bootstrap probability of PASS: 0.51 (draw 0) and 0.60 (draw 3). A stricter gate (ratio 0.5–2, Spearman ≥ 0.3 with p < 0.05) passes no draw.\n\ncalibration.csv rows, draws 0–4:\n- 'l2 | Plasma cell': ratio 0.38, 0.25, 0.15, 0.46, 0.43; Spearman 0.39, 0.22, 0.21, 0.26, 0.33; sepsis−HC sign 5/5.\n- 'Plasmablast (markers)': ratio 0.16, 0.12, 0.09, 0.24, 0.23.\n- 'l1 | B cell': ratio 1.20, 0.56, 0.98, 0.55, 1.00.\n\nBulk alone: B1 (6-gene bulk mean) +0.94 (p 8.7e-8); B2 (B1 relative to B markers) +1.38 (p 4e-10). Both hold with age, sex and bulk QC adjustment.\n\nMKI67: observed 21.4 vs 9.3 CPM, of which plasma cells carry 7.18 of the 12.06 difference. Generated: no rise in draws 0 and 3; proliferating T cells lead in draws 1 and 2; plasma cells lead in draw 4 (1.46 of 2.80).", + "issues": [ + "Only one draw carries a calibrated result: draw 0 (discovery gate PASS and a significant bulk-only effect), and it overshoots the observed contrast 2.4–2.7×. Draw 3 is the other PASS and the only draw with a ratio near 1 (0.81), but its bulk-only effect is not significant by Welch (p 0.12 / 0.65 / 0.79 by score) and its plasma-cell abundance contrast is null (Mann–Whitney p 0.53).", + "The two PASS gates are fragile. Bootstrap probability of PASS is 0.51 (draw 0) and 0.60 (draw 3). Spearman passes the 0.20 cut by a small margin (0.23, 0.28) against 0.67–0.71 for the bulk alone. The observed test contrast rests on 6–7 HC (p 0.014–0.054). A stricter gate (ratio 0.5–2, Spearman ≥ 0.3 with p < 0.05) passes no draw.", + "New caveat, not in the claim: a zero-plasma floor inflates the effect size. 53/28/47/3/25% of bulk-only HC (draws 0–4) have zero generated plasma cells, against 0/7 observed adult HC (and 5/6 paired-generated HC in draw 0). Dropping samples with zero plasma cells shrinks the effect by about 25–40% in draws 0, 1, 2 and 4. Draw 3's null mean comes from the reverse floor: 18% of its sepsis samples have zero plasma cells. The +1.9 log2 mean, about twice the observed +1.0, is partly this artefact; the direction survives.", + "'All 6 genes rise in 5/5 draws' depends on the normalisation. Under log2(CPM+1), draw 3 has JCHAIN −0.14, MZB1 −0.09 and TXNDC5 −0.02; only PRDM1 is significant there.", + "The observed paired confirmation is itself marginal: 46 vs 7, Welch p 0.021, but Mann–Whitney p 0.056.", + "Nothing beyond the bulk. The bulk shows +0.94 / +1.38 and ranks observed samples far better (Spearman 0.67–0.71 against 0.00–0.28). Partial Spearman of generated vs observed given the bulk is negative in 4/5 draws. The generated group effect stays after adjusting for the bulk score, but the calibration shows this is overshoot, not information.", + "Draw heterogeneity follows the training donors. Draw 0 trained on Res-S150 (58 plasma cells per 1,000 observed cells) and Abd-S66 (16). Draw 3 trained on a paediatric HC (PHC05, 6.5 per 1,000), so it generates plasma cells in HC (2.8 per 1,000) and the contrast shrinks.", + "Group-level technical difference in the bulk: HC have a higher htseq no-feature fraction (0.170 vs 0.156, p 2.7e-6) in both paired and bulk-only samples. Matching or adjusting on it leaves the effect intact, but any group-specific processing batch cannot be separated from sepsis vs HC.", + "Minor: the MKI67 side statement is incomplete. In draw 4 the generated MKI67 rise (+2.8 CPM) is carried mainly by plasma cells (1.46), as in observed cells. Test samples with ≥10 generated plasma cells include paediatric HC (1–2) in draws 1, 3 and 4; there are no adult HC.", + "Not applicable or checked clean: donor repeated measures (one sample per donor) and training-donor leakage (no bulk-only sample is a training donor in any draw). Age, sex, site and sex_flag are balanced and do not change the result." + ], + "reasoning": "Method: I read the data guide and wrote independent code from the raw compact files: pb_l1.npz, cells.parquet, markers.npz, meta.csv, calibration.csv, bulkonly_raw.npz, conditions_raw_fixed.npz and the bulk h5ad obs (QC only). I did not reuse the discovery scripts. Everything ran CPU-only under systemd-run with a 6G memory cap.\n\nRecompute:\n- The six-type 'B cell' gate contains every AIFI-L2 plasma cell in all three sets.\n- markers.npz and lib_measured reproduce the pb_l1 B-gate sums exactly.\n- The discovery's edgeR-like score is reproduced to the reported digits.\n- Simpler normalisations give larger effects in draws 0, 1, 2 and 4, but leave draw 3 null on means. So the effect reproduces, and its sign is positive in 5/5 draws.\n\nConfounders: age, sex, hospital, sex_flag, age matching, sex strata, infection source, B-cell number, library size per B cell, depth-matched thinning, label confidence ≥ 0.9, bulk library size, the bulk no-feature fraction (a batch proxy, also matched), min-cell thresholds and a random gene-set specificity null. None removes the effect in draws 0, 1, 2 and 4. The effect is carried by plasma-cell composition within the B gate, as claimed and as in observed cells.\n\nWhy qualified, not confirmed:\n1. Calibration is thin. Under the discovery's own lenient gate, only draw 0 is both calibrated and significant, with a 2.4× overshoot. Both PASS gates are coin-flips under bootstrap, and no draw passes a stricter ratio and Spearman gate.\n2. The only draw with a ratio near 1 (draw 3) shows no mean effect and no plasma-abundance contrast.\n3. A zero-plasma floor in generated HC, absent in observed HC, inflates the log-scale effect. The +1.9 vs +1.0 overshoot is partly an artefact, though the direction survives removing zero-plasma samples.\n4. The observed confirmation rests on 7 HC (Mann–Whitney p 0.056).\n5. Nothing is added beyond the bulk, which already shows the rise and ranks samples better.\n\nMost of these caveats are already in the candidate's text ('weak'). My additions are the HC floor, the fragility of the gates, the normalisation dependence of draw 3 and the per-gene statement, and the draw-4 MKI67 correction.\n\nHonesty items:\n- (1) Calibration: weak; draw 0 only.\n- (2) Replication: sign 5/5, significant in 4/5, magnitude driven by training donors.\n- (3) Observed agreement: direction agrees (+1.0, p 0.02) and the plasma-cell carrier agrees.\n- (4) Beyond bulk: none.\n\nWhat can be supported: the generator reproduces in direction the known, bulk-visible rise of plasmablast genes in sepsis and assigns it to plasma cells. The effect size is not reliable.\n\nArchive: I wrote only under my lens directory and changed no paper evidence. The coordinator owns any POLICY.md update.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v05_summary_by_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_effect.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_effect_by_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_genes_by_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_confounders.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_cellnumber_library.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_bulk_alone.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_bulk_and_covariates_by_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_gen_vs_bulk_corr.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_drawcentred_sample_level.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v01_per_sample_scores.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v02_cells.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v02_decomposition.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v02_consistency_checks.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v02_per_sample.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v03_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v03_calibration_bscore.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v03_calibration_bootstrap.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v03_calibration_leave_one_HC_out.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v03_calibration_csv_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v03_plasma_coverage_and_training_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v03_observed_plasma_per_1000_by_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v04_extra.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v04_zero_plasma_floor.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v04_specificity_random_sets.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v04_nofeature_matched.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v04_domain_shift.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v04_observed_paired_scores.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v04_mki67_attribution.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v05_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_recompute/v00_inspect.py" + ] + }, + { + "verifier": "calib", + "agent_id": "a90fbb329335a91c3", + "verdict": "qualified", + "recomputed": "I recomputed everything with my own code, without importing lens_common. It matches sep_programs exactly.\n\n**The B gate holds all plasma cells.** Every AIFI-L2 plasma cell falls in the six-type 'B cell' gate in all 11 sets (v00_gate_crosstab.csv). The per-cell markers and the pb_l1 sums agree exactly.\n\n**Bulk-only effect (B-gate score from the 6 genes, adult sepsis minus adult HC, 68 vs 36):**\n- Differences are +3.13, +1.86, +1.55, +0.40 and +2.62 log2 (draws 0–4).\n- Welch p is below 1e-10 except draw 3 (p 0.12). Mann–Whitney p is at most 0.0022 in every draw.\n- Draw 3 is significant once pseudobulks need at least 20 cells: +0.67, p 0.008.\n- With same-site sepsis only and sex-flag samples removed: +3.13, +2.06, +1.73, +0.40, +2.76.\n\n**Paired calibration (77 test samples; 42–43 sepsis vs 6–7 HC):**\n- Generated contrast 2.69, 1.79, 1.36, 0.83, 2.31; observed 1.11, 0.91, 1.11, 1.03, 0.98. The sign agrees in 5/5 draws.\n- Ratio on the log scale: 2.41, 1.97, 1.22, 0.81, 2.35. Ratio on the linear-CPM scale: 0.26, 0.34, 0.13, 0.39, 0.73.\n- The generated HC level sits 0.9–3.6 log2 below the observed HC level, so the log-scale ratio is inflated by a near-zero HC floor.\n- Spearman, generated vs observed: 0.23, 0.16, 0.03, 0.28, 0.17. Bulk vs observed: 0.67–0.71. Partial Spearman given bulk: 0.02, −0.07, 0.10, 0.12, 0.09.\n- The generated score tracks its own bulk input poorly in bulk-only samples: Spearman 0.06–0.28.\n\n**Composition rows:**\n- calibration.csv 'l2 | Plasma cell': sign 5/5; level ratio 0.38, 0.25, 0.15, 0.46, 0.43; Spearman 0.39, 0.22, 0.21, 0.26, 0.33; contrast ratio 0.75, 0.36, 0.21, 0.44, 0.72.\n- calibration.csv 'Plasmablast (markers)': level ratio 0.09–0.24.\n- Plasma share of the B gate: contrast ratio 0.59, 1.06, 0.31, 0.66, 1.04; Spearman 0.09–0.28, against 0.77–0.82 for the bulk; partial given bulk −0.43, −0.30, −0.26, +0.01, −0.41.\n\n**B gate without plasma cells (4 plasma-specific genes):**\n- Bulk-only: +0.13, −0.06, +0.06, −0.58, +0.38.\n- Observed: −0.49 (p 0.0008).\n- The calibration sign is wrong in draws 0 and 4.\n\n**Observed paired adults (46 vs 7):**\n- Program score +1.02 (Welch p 0.022, MW p 0.056).\n- Plasma cells as a fraction of all cells: +0.0097 (p 0.0002, MW p 0.0065).\n- Plasma share of the B gate: +0.063 (p 0.0012).\n- Paired bulk, 6 genes: +0.94 (p 3e-5).\n\n**Bulk-only bulk:**\n- 6 genes +0.94 (d 0.93, p 9e-8); 4 genes +1.40.\n- Relative to B-cell markers: +1.38 (6 genes) and +1.84 (4 genes).\n\n**Per gene:** all 6 genes rise in the B gate in 5/5 draws, with the same sign as observed. Observed TNFRSF17 is not significant (+0.35 to +0.55).\n\n**Test samples with at least 10 generated plasma cells:**\n- Adult sepsis: 13, 8, 2, 12, 18. Adult HC: 0 in every draw. Paediatric HC: 0, 1, 0, 2, 1.\n\n**Training donors explain the draw differences:**\n- Draw 0 includes Res-S150: plasma cells are 5.8% of its cells, the 94th percentile.\n- Draw 4 includes Skin-S46: 2.3%, 88th percentile.\n- Draw 3 includes the paediatric HC PHC05, whose B gate is 8.4% plasma cells. Draw 3's generated HC plasma share of the B gate is 0.052, against 0.005–0.024 in the other draws. That raised HC level shrinks draw 3's contrast.", + "issues": [ + "Why draws 1, 2 and 4 are SIGN-ONLY is described wrongly. They fail on Spearman (0.16, 0.03, 0.17 < 0.20), not because they overshoot. Draw 2's ratio of 1.22 is close to 1, and draw 0, which passes, overshoots most (2.41).", + "The size of the effect cannot be read. '+1.9 log2 vs +1.0 observed' comes from a near-zero floor in generated HC: generated HC sits 0.9–3.6 log2 below observed HC. On a linear-CPM scale every draw undershoots the observed contrast (0.13–0.73×). Only the direction is calibrated.", + "This is plasma-cell composition, not an expression programme. With plasma cells removed from the B gate, the 4-gene score changes by +0.13, −0.06, +0.06, −0.58 and +0.38, while observed falls by 0.49. The calibration sign is wrong in draws 0 and 4.", + "Which draws carry it: under sep_programs' lenient ratio band (0.25–4), draw 0 is PASS with q < 0.05. Draw 0's calibration is marginal (Spearman 0.23; ratio 2.41 on the log scale, 0.26 on the linear scale). Under the guide's criteria only draw 3 is partly calibrated (Spearman 0.28, ratio 0.81), and it carries the weakest effect: +0.40 (Welch p 0.12; MW p 0.002), or +0.67 (p 0.008) when pseudobulks need at least 20 cells.", + "Per sample, the generated cells add nothing. The generated score tracks its own bulk input poorly (Spearman 0.06–0.28). Given the bulk, the partial Spearman is about 0 for the programme score and negative for the plasma share of the B gate (−0.26 to −0.43). The bulk ranks the observed plasma fraction at Spearman 0.86–0.87.", + "Observed support for the programme score is modest with only 7 HC (Welch p 0.022, MW p 0.056). The observed plasma fraction (MW p 0.0065) and the paired bulk (p 3e-5) are robust.", + "The MKI67 sub-statement is inaccurate. Draw 4 assigns the MKI67 rise to plasma cells (share 0.52), not to T cells. MKI67 attribution fails paired calibration in all 5 draws.", + "'None of them are HC' is true for adult HC only: 1–2 paediatric HC test samples have at least 10 generated plasma cells in draws 1, 3 and 4. '2–16' requires at least 10 plasma cells in both the generated and observed sets; counting generated cells alone gives 2–18 adult sepsis samples.", + "The finding is less novel than the claim says. Ye 2026 already tied the plasma-cell (B_c06-MZB1) and plasmablast (B_c07-MKI67) marker signatures to poor survival using this cohort's own bulk mRNA-seq (Methods, PDF p.19, n = 174 adults; ED Fig. 6e, PDF p.33–34, p = 0.019 and 0.034). It also ran CIBERSORTx on the same in-house PBMC bulk (PDF p.19–20).", + "Small point: the largest Mann–Whitney p is 0.0022, not ≤ 0.002." + ], + "reasoning": "**1. Calibration.** The trust layer supports only the direction.\n- The sepsis > HC sign is reproduced in 5/5 draws for three rows: the B-gate programme score, the 'l2 | Plasma cell' composition row and the 'Plasmablast (markers)' row.\n- No draw is well calibrated on magnitude or on sample ranking. Plasma cells are under-generated (0.15–0.46×). Spearman is 0.03–0.28. Partial Spearman given bulk is about 0 or negative.\n- The log-scale 'overshoot' is an artefact of the near-zero HC floor; on a linear scale every draw undershoots.\n- The generated rise comes entirely from more plasma cells inside the B gate (every AIFI-L2 plasma cell sits in that gate). Non-plasma B cells do not rise consistently.\n- Gating per draw:\n - Draw 0 passes the pre-registered gate only marginally.\n - Draw 3 is the only draw I would call partly calibrated, and it carries the weakest effect: significant by rank test, or once pseudobulks need at least 20 cells.\n - Draws 1, 2 and 4 are direction-only.\n- Nothing in the calibration contradicts the direction, so this is not refuted.\n\n**2. Observed data.** The paired observed cells agree in direction.\n- B-gate score +1.02 (p 0.022; Mann–Whitney p 0.056, with only 7 HC).\n- Plasma-cell fraction +0.0097 (p 0.0002).\n- Paired bulk +0.94 (p 3e-5).\n- In observed cells the bulk rise of JCHAIN, MZB1, TXNDC5 and TNFRSF17 is carried by plasma cells, as in all 5 draws.\n\n**3. Biology and novelty.** Already reported in this cohort.\n- Ye 2026, Nat Immunol 27:150, journal p.155 (PDF p.6): \"B_c06-MZB1 and B_c07-MKI67 were increased across multiple infection sources … High expression of B_c06-MZB1 and B_c07-MKI67 markers correlated with poor survival\". The supporting figure is ED Fig. 6d–e (PDF p.33–34).\n- That survival analysis used this cohort's bulk mRNA-seq (Methods, PDF p.19).\n- COMBAT 2022 (Cell 185, p.924) reports plasmablast clonal expansions in sepsis. Its p.932 notes that terminally differentiated plasmablasts are absent in sepsis, which differs from COVID.\n- Ye 2026's reference 37 (Nascimento 2021, Immunity) reports CD39+ plasmablast expansion in sepsis.\n\n**4. Beyond the bulk.** Nothing.\n- The bulk shows the rise itself (+0.94 log2; +1.38 relative to B markers).\n- The bulk ranks samples much better than the generated cells.\n- The plasma-cell attribution is what a reference atlas would already say.\n\n**Verdict.** The claim's own summary stands: weak, known, nothing beyond the bulk. It is qualified rather than confirmed because calibration supports only the direction. Several details need correcting: why draws 1, 2 and 4 are SIGN-ONLY, how to read the effect size, that the effect is composition rather than expression, which draws carry the result, the MKI67 attribution in draw 4, and the novelty.\n\nI published nothing to the paper archive; the coordinator decides on any archive delta. I ran CPU only, inside systemd-run with MemoryMax=6G and CUDA_VISIBLE_DEVICES empty, and wrote only to my own directory.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v00_consistency.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v00_gate_crosstab.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v00_consistency.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v01_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v01_bulkonly_effects.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v01_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v01_observed_effects.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v01_group_levels.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v01_plasma_counts.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v01_train_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v01_scores_long.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v02_genes_and_inputs.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v02_gene_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v02_gen_vs_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v02_mincells.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v02_draw3_quantiles.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v03_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-5_calib/v03_per_draw_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/features/ye2026_sepsis/calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/programs_calib.csv", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/COMBAT2022_Blood_atlas_COVID19_severity.pdf" + ] + } + ] + }, + { + "id": "SEP-CLIN-3", + "lens": "sep_clinical", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 15, + "theme": "age", + "synthesis_strength": "weak: bulk-level observation", + "reportable_statement": "Generated B cells (all subsets) decline with age in bulk-only adults (~11-13% lower odds per decade), matching the bulk B-cell genes and the direction of observed paired adults (p 0.05-0.12). Not naive-specific and not beyond bulk.", + "carrying_draws": { + "naive B": "1,3,4 (guide gate); 0,1,4 with adult-only level calibration", + "B": "1,3,4" + }, + "honesty": { + "calibration": "No age contrast in calibration.csv; paired age slopes same sign 5/5 but observed slope non-significant in every test set; draw 4 overstates 2.5x.", + "replication": "Naive B 5/5 negative, 4/5 p<0.05; B 3/5 p<0.05.", + "observed": "Naive B -0.0153/yr (p 0.12), B -0.0148/yr (p 0.052); depends on 10 EMC samples.", + "beyond_bulk": "None: bulk naive-B and pan-B scores fall with age (p 0.002-0.003, also in HC alone); generated slope after bulk adjustment p 0.15-0.21." + }, + "verdict_reasons": [ + "pan-B, not naive-specific", + "underpowered observed reference", + "best-calibrated draws carry the weakest slopes", + "screen q~0.03 also hits a negative-control state" + ], + "new_vs_cohort_papers": "Not in Ye 2026 (no within-adult age analysis). Not textbook: healthy-adult flow finds stable naive B (Caraux 2010; Blanco 2018).", + "next_steps": [ + "Age calibration row; flow B subsets; note the generator attenuates the dominant true age effect (naive CD8 T) to 27-55%." + ], + "key_sources": [ + "verify/SEP-CLIN-3_recompute/v05_gated_estimates.csv", + "verify/SEP-CLIN-3_calib/v03_paired_age_calib.csv", + "verify/SEP-CLIN-3_calib/v04_mediation.csv" + ] + }, + "finding": { + "id": "SEP-CLIN-3", + "cohort": "ye2026_sepsis", + "claim": "Across the 121 adult bulk-only samples, generated B cells, mostly naive B cells, decline with age: about 13% lower odds per decade after adjusting for group and sex. The bulk naive-B signature declines in the same way, so the cells add nothing beyond the bulk.", + "kind": "composition", + "feature": "l2 | Naive B cell; l1 | B cell", + "comparison": "Age (continuous) in all adults, adjusted for group and sex; also within adult sepsis", + "effect_bulkonly": "Logit slope per year: naive B -0.0136 (SE 0.0045), B -0.0121. Spearman of the five-draw mean fraction with age: -0.32 and -0.28. Within sepsis (n=68), naive B -0.0152 per year.", + "stats": "Model: OLS with HC3 errors, mean of draw logits ~ age + sex + group, n=121. Naive B p=0.0024, q=0.030; B p=0.0023, q=0.030 (BH over 50 features); rank-ANCOVA q=0.029 for both; Spearman q=0.006 and 0.013. Within sepsis, naive B p=0.016, q=0.087. Without the sex-flag or Emergency-Center samples q=0.039. Gated on draws 1-4: naive B slope -0.0148, p=0.0019.", + "draws_same_sign": "Naive B 5/5, 4/5 with p<0.05 (d0 -0.009, d1 -0.019, d2 -0.007 with p=0.06, d3 -0.017, d4 -0.016). B 5/5, 3/5 with p<0.05. Carrying draws 1,3,4 for both.", + "paired_calibration": "calibration.csv has no age contrast. Overall calibration: B cell ratio 1.20/0.56/0.98/0.55/1.00, Spearman 0.83/0.59/0.75/0.59/0.82; naive B ratio 2.06/0.86/1.88/0.64/1.42 (d0 fails), Spearman 0.73/0.42/0.55/0.59/0.61. Same-model age slope on each draw's paired test samples, gen vs obs, naive B: -0.016/-0.017, -0.018/-0.017, -0.011/-0.012, -0.012/-0.016, -0.025/-0.010. That is 5/5 same sign with similar size; B is also 5/5.", + "paired_observed_concordance": "Observed paired adults, n=59: B -0.0148 per year (p=0.052), naive B -0.0153 per year (p=0.12). Within sepsis (n=46), naive B -0.018 per year (p=0.083). Same direction and similar slope, not significant.", + "beyond_bulk": "Nothing beyond the bulk. The bulk-only naive-B marker score (TCL1A, FCER2, IL4R, BACH2, PLPP5) falls by 0.016 z per year (p=0.002, q=0.007; paired bulk p=0.007). The B score falls by 0.018 (q=0.012). Spearman between the bulk score and the generated fraction is 0.76-0.87. After adjusting for the bulk score, the cell slope is -0.003 (p=0.21).", + "novelty": "Not in the Ye 2026 text. Gong 2025 (rare-disease-states PDF, p.3) names naive CD8 T cells as the main compositional age change in healthy adults and reports 'few baseline changes' in other compartments. Age-related B-cell decline is generic immunosenescence; I did not check it against further literature.", + "strength": "weak", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s02_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s03_paired_calib_long.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s04_observed_assoc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s06_feature_vs_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s06_bulk_assoc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s10_evidence.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s12_gated.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "af1e92e6009085789", + "verdict": "qualified", + "recomputed": "I rebuilt everything from composition.parquet, states.parquet, meta.csv, calibration.csv, cells/labels parquet, pb_l1/pb_l2.npz, bulkonly_raw.npz and conditions_raw_fixed.npz, using my own code (v01-v08).\n\n1. Headline (121 adult bulk-only samples; OLS HC3 on the mean of the five draw logits, ~ age + sex + group):\n - Naive B: -0.01358/yr (SE 0.00447), p=0.0024, q=0.030 over 50 features. Odds ratio per decade 0.873 (-12.7%).\n - B: -0.01205/yr, p=0.0023, q=0.030 (OR per decade 0.886).\n - Stacked model (draw fixed effect, clustered by sample): p=0.0013 and 0.0012.\n - Rank-ANCOVA q=0.029 for both. Five-draw-mean Spearman -0.319 (q=0.006) and -0.275 (q=0.014).\n - Within sepsis (n=68), naive B: -0.0152/yr, p=0.016, q=0.087.\n - Gated on draws 1-4: -0.0148/yr, p=0.0019.\n - Without the sex-flag samples q=0.039; without EMC q=0.039; without both q=0.058.\n - All match the claim.\n\n2. Per-draw slopes (OLS):\n - Naive B: d0 -0.0086 (p=0.024), d1 -0.0193 (0.0008), d2 -0.0072 (0.059), d3 -0.0171 (0.0002), d4 -0.0157 (0.023). 5/5 negative, 4/5 p<0.05.\n - B: d0 -0.0068 (0.08), d1 -0.0160 (0.002), d2 -0.0055 (0.10), d3 -0.0206 (<0.001), d4 -0.0113 (0.037). 3/5 p<0.05.\n - Quasi-binomial models (overdispersion 19-70) give the same signs; naive B is p<0.05 in 5/5 draws.\n\n3. Confounder controls (naive B, mean logit):\n\n| Control | Slope/yr | p |\n|---|---|---|\n| Drop sex_flag samples | -0.0131 | 0.004 |\n| Use the bulk sex call | -0.0139 | 0.002 |\n| Drop EMC | -0.0137 | 0.004 |\n| Add EMC covariate | -0.0143 | 0.002 |\n| HC + sepsis only (ICU controls are 15 years younger) | -0.0166 | 0.002 |\n| Age ≥ 40 | -0.0160 | 0.003 |\n| Huber regression | -0.0128 | 0.003 |\n| Drop top-5 Cook's distance | -0.0146 | 0.0008 |\n| Bulk QC covariates | -0.0114 | 0.003 |\n| Bulk contamination scores | -0.0128 | 0.007 |\n| QC + contamination | -0.0100 | 0.011 |\n| Generated library size, per draw (stacked) | -0.0146 | 0.0004 |\n| Draw's own 'Other' fraction (stacked) | -0.0109 | 0.005 |\n| Everything together (stacked) | -0.0094 | 0.005 |\n| Labels with confidence ≥ 0.9 only | -0.0145 | 0.0014 |\n\n - Bulk QC covariates: log assigned reads and the htseq fractions.\n - Bulk contamination scores: platelet, red-cell and granulocyte marker scores plus the mitochondrial fraction.\n - \"Everything together\": no sex-flag samples, EMC, QC, contamination and library size; 5/5 draws still negative, 2/5 with p<0.05.\n - Leave-one-out slopes range from -0.0152 to -0.0122.\n - Slope by group: HC -0.0224 (p=0.016), sepsis -0.0152 (p=0.016), ICU -0.0042 (n=17, p=0.62). The age-by-group interaction is not significant (p=0.26).\n - Generated cell number is 1,500 for every sample.\n - Not applicable to sepsis: each donor has one sample, bulk-only donors are disjoint from paired and training donors, and test sets contain no training donors.\n\n4. Composition:\n - B among T+B+NK cells also falls: -0.0106/yr, p=0.0002.\n - T, NK and monocyte shares do not change with age. 'Other' rises (+0.0103/yr, p=0.003).\n - Naive B as a share of the B lineage: -0.0039/yr (p=0.034) in bulk-only generated cells. It is flat in observed paired cells (-0.0019, p=0.77) and in paired generated cells (-0.0042, p=0.24).\n - Memory and effector shares within B are flat.\n\n5. Calibration:\n - calibration.csv matches what the claim quotes. B: ratio 1.20/0.56/0.98/0.55/1.00, Spearman 0.83/0.59/0.75/0.59/0.82. Naive B: ratio 2.06/0.86/1.88/0.64/1.42, Spearman 0.73/0.42/0.55/0.59/0.61.\n - Draw 1 fails the \"sepsis vs all controls\" sign for both features.\n - Paired adult test samples, age slope generated vs observed, naive B: d0 -0.016/-0.017, d1 -0.018/-0.017, d2 -0.011/-0.012, d3 -0.012/-0.016, d4 -0.025/-0.010.\n - Draw 4 differs from observed (slope-difference p=0.041 for naive B, 0.036 for B).\n - Partial Spearman of generated vs observed after removing sex and group, naive B: 0.68/0.31/0.40/0.52/0.55.\n - Generated-on-observed logit gain is only 0.34-0.65 for naive B.\n - Given the observed composition, the pooled generated slope on paired samples is still -0.0109/yr (p=0.040).\n\n6. Calibration-gated bulk-only estimates:\n - Naive B: draws 2,3,4 (ratio and Spearman gate) -0.0133, p=0.003; draws 2,3 (ratio, Spearman and paired age slope) -0.0122, p=0.0008. No draw passes the strict gate together with the paired-age check.\n - B: draws 0,2,4 (strict gate) -0.0079, p=0.037; draws 0,2 (strict plus paired age) -0.0062/yr, p=0.055, sepsis-only p=0.24.\n\n7. Observed paired adults (n=59):\n - Naive B -0.0153/yr (p=0.12); B -0.0148/yr (p=0.051).\n - Within sepsis (n=46), naive B -0.0176 (p=0.083).\n - Without the 10 EMC samples (n=49): naive B -0.0077 (p=0.49), B -0.0099 (p=0.27). Non-EMC sepsis: -0.0096 (p=0.42).\n\n8. Bulk:\n - Naive-B marker score (TCL1A, FCER2, IL4R, BACH2, PLPP5, YBX3): -0.0132 z/yr (p=0.0035). Paired set: -0.0164 (p=0.006).\n - Pan-B score: -0.0180 (p=0.003).\n - Generated slope after adjusting for the naive-B score: -0.0041 (p=0.15); after the pan-B score: -0.0022 (p=0.27).\n - On paired test adults, the bulk score tracks observed naive B better than every draw's generated fraction (Spearman 0.78-0.83 vs 0.33-0.74).\n - Adding the generated fraction to the bulk score does not help (p=0.12-0.85), while the bulk score adds strongly to the generated fraction (p<1e-6).\n - Observed composition explains 84% of the bulk pan-B age slope but only 53% of the naive-B score slope.\n - Within observed naive B cells, naive-marker expression does not change with age (p=0.76). B-cell UMI per cell does not change with age.\n\n9. Context: the \"Platelet signal in monocytes\" negative-control state also rises with age (+0.0106/yr, p=0.0014, q=0.030). It remains at p=0.013 after adjusting for bulk platelet, red-cell and granulocyte scores.", + "issues": [ + "Not refuted. Every number and per-draw sign reproduces exactly from the raw files, and the effect survives every confounder control tested. Sex-flag samples, bulk sex call, EMC site, the younger ICU controls, leverage, bulk QC, contamination, generated library size and label confidence all leave it standing; at most it shrinks about 30% (to -0.009 to -0.010/yr) and stays negative in 5/5 draws.", + "Calibration for an age contrast is weak and indirect. calibration.csv has no age contrast. The observed paired slope that the generated slopes 'match' is not significant itself (naive B p=0.12, B p=0.051). The five per-draw comparisons reuse about 54 of the same samples, so they are not independent confirmations.", + "The observed paired age slope depends on the 10 EMC sepsis samples. Without them it is naive B -0.0077 (p=0.49) and B -0.0099 (p=0.27); in non-EMC sepsis -0.0096 (p=0.42). The generated bulk-only slope does not depend on EMC.", + "Draw 4 exaggerates the age slope on paired samples (gen -0.025 vs obs -0.010; difference p=0.041 for naive B, 0.036 for B). Draw 1 fails the 'sepsis vs all controls' sign for both features and tracks naive B weakly (partial Spearman 0.31, calibration Spearman 0.42).", + "Significance is carried mostly by the worst-calibrated draws (1 and 3; B ratio 0.55-0.56). Under a strict gate (ratio 0.67-1.5, Spearman >= 0.6, all contrasts agree, paired age slope reproduced) no naive-B draw qualifies. For B only draws 0 and 2 qualify, giving about half the slope (-0.0062/yr, p=0.055; sepsis-only p=0.24). Moderate gates still give p of about 0.001.", + "'Mostly naive B' means only that naive B is the largest B subset. The decline covers the whole B compartment (effector B p=0.0005, memory B p=0.0075). The naive share within B is flat in observed paired cells (p=0.77) and in paired generated cells (p=0.24). Within-B composition is not tracked across people (partial Spearman about 0 to 0.23).", + "Part of the generated age slope is not explained by true composition. The logit gain on observed composition is only 0.34-0.65, yet generated and observed slopes are the same size. After conditioning on observed composition the pooled generated slope is still -0.011/yr (p=0.04). The bulk naive-B score keeps 47% of its age slope given observed composition, so the matching slope sizes are partly coincidental.", + "Nothing beyond bulk is confirmed, and more strongly than the claim says. A plain bulk marker score tracks observed naive B better than any draw's generated fraction, and the generated fraction adds nothing to it (p=0.12-0.85). Only draws 1 and 3 keep a slope after bulk adjustment, and those are the poorly calibrated draws.", + "Fragility. Dropping both the sex-flag and EMC samples moves q to 0.058. Within sepsis q is 0.087, and the discovery lens reports rank-based within-sepsis q of 0.18-0.28.", + "Screen context. A designated negative-control state ('Platelet signal in monocytes') also rises with age at the same q (0.030), and the 'Other' cell fraction rises too (p=0.003). So q of about 0.03 alone does not separate biology from artifact in this age screen. Adjusting for 'Other' and for bulk contamination does not remove the B effect.", + "Minor: the claim's 'about 13% per decade' is the naive-B figure (12.7%); B is 11.4%. The B Spearman q is 0.014 in my recompute versus 0.013 quoted, which is rounding or family construction." + ], + "reasoning": "I recomputed the effect independently from the raw compact files using my own scripts. All quoted numbers reproduce exactly: slopes, SEs, p, BH q over 50 features, rank-ANCOVA, Spearman, the within-sepsis estimates, the per-draw signs and p-values, the draws 1-4 gated estimate, the sensitivity q-values and the paired same-model slopes. The effect is therefore real in the generated data and is not an arithmetic or pipeline artifact.\n\nI then tried to break it with the confounders on the checklist:\n- Sex: dropping the sex-flag samples or using the bulk sex call.\n- Site: dropping EMC or adding it as a covariate.\n- Age imbalance: the ICU controls are 15 years younger, so I fitted HC + sepsis only, age ≥ 40, a quadratic, per-group slopes and the interaction.\n- Leverage: Huber regression, Cook's distance and leave-one-out.\n- Bulk library size and quality, from the htseq metrics.\n- Bulk platelet, red-cell, granulocyte and mitochondrial contamination.\n- Generated library size, and the constant generated cell number.\n- Label confidence: counts restricted to labels with confidence ≥ 0.9.\n- Compositional closure: 'Other', lymphocyte-relative B and monocyte adjustment.\n- Per-cell RNA content and per-cell marker expression in observed B cells.\n\nTraining-donor leakage and repeated measures do not apply to this sepsis cohort, and I checked that explicitly. Nothing removed the effect; the strongest combined adjustment shrank it by about 30% with 5/5 draws still negative. So it is not refuted.\n\nIt is not fully confirmed either. Calibration for an age contrast is weak:\n- The observed paired slope is non-significant and rests largely on 10 EMC samples.\n- Draw 4 exaggerates the slope, and draw 1 fails a group-contrast sign and tracks naive B weakly.\n- The draws that carry the effect (1, 3) are the least calibrated. Under a strict calibration-plus-age gate the effect halves for B (p=0.055) and has no qualifying draw for naive B.\n- The naive-specific reading is not supported by observed cells.\n- The generated cells add nothing beyond a plain bulk marker score, as the claim itself says; on paired samples the bulk score is a better predictor of the observed fraction than any draw.\n\nThe finding stands as a weak, bulk-derived decline of the whole B compartment with age, consistent with generic immunosenescence. The claim's own 'weak' rating is appropriate, so the verdict is qualified.\n\nAll scripts, tables and logs are in my lens directory. I wrote nothing elsewhere and started no GPU work. This verifier published no paper-archive delta; that step belongs to the coordinator.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v01_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v01_targets.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v01_all_features.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v01_perdraw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v01_quasibinomial.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v02_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v02_confounders.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v02_variants.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v02_covariates_vs_age.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v02_influence.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v02_sample_table.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v03_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v03_calibration.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v03_calib_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v03_paired_perdraw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v03_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v03b_emc.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v03b_emc.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v03b_emc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v04_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v04_bulk.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v04_beyond_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v04_tracking.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v04_gene_slopes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v04_bulk_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v05_gated.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v05_gated.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v05_gates.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v05_gated_estimates.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v06_variant_q_gain.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v06.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v06_variant_q.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v06_gain.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v07_expr_vs_comp.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v07.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v07_within_type_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v08_rna_content.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_recompute/v08.txt" + ] + }, + { + "verifier": "calib", + "agent_id": "a91ef6b5d2b44900a", + "verdict": "qualified", + "recomputed": "My own code (it does not import the sep_clinical helpers) reproduces every headline number exactly.\n\nBULK-ONLY, 121 adults. Model: logit ~ age + sex + group, OLS with HC3 errors.\n- Mean of the draw logits:\n - Naive B: -0.0136/yr (SE 0.0045), p=0.0024, q=0.030, about 12.7% lower odds per decade.\n - B: -0.0121/yr, p=0.0023, q=0.030, about 11.4% lower odds per decade.\n - Rank-ANCOVA q=0.029 for both.\n- Alternative models, naive B / B:\n - draws stacked with draw as a fixed effect and SE clustered by sample: p=0.0013 / 0.0012;\n - quasi-binomial: p=0.0068 / 0.012.\n- Naive B per draw, d0..d4 (p): -0.0086 (0.024), -0.0193 (0.0008), -0.0072 (0.059), -0.0171 (0.0002), -0.0157 (0.023). Only d1 and d3 pass BH q<0.05 over the 50 features.\n- B per draw: -0.0068, -0.0160, -0.0055, -0.0206, -0.0113.\n- Naive B by group:\n - HC only (n=36): -0.022, p=0.016;\n - sepsis only (n=68): -0.015, p=0.016;\n - ICU controls only (n=17): -0.004, p=0.62;\n - age x sepsis interaction: p=0.64.\n- Leave-one-out: maximum p 0.0054.\n\nPAIRED AGE CALIBRATION. Each draw's adult test samples (n 54-55); same model on generated and observed cells.\n- Naive B, generated vs observed slope:\n - d0: -0.0159 vs -0.0173\n - d1: -0.0181 vs -0.0170\n - d2: -0.0108 vs -0.0119\n - d3: -0.0118 vs -0.0165\n - d4: -0.0252 vs -0.0099\n - The observed test-set slope is never significant (p 0.08-0.32; bootstrap P(observed slope < 0) 0.83-0.96).\n - The bootstrap CI of the generated-minus-observed slope covers 0 in all draws; d4 is borderline (-0.030 to +0.0003).\n- B, generated vs observed slope: -0.0181/-0.0166, -0.0183/-0.0147, -0.0072/-0.0135, -0.0113/-0.0157, -0.0198/-0.0112. Observed p 0.034-0.15.\n- Level calibration on adult test samples only (calibration.csv also includes 22-23 children):\n - naive B ratio 1.84/0.58/1.58/0.46/1.01, Spearman 0.74/0.33/0.49/0.57/0.58;\n - B ratio 1.35/0.50/1.01/0.51/1.01, Spearman 0.82/0.53/0.67/0.51/0.79.\n\nGATED slope (mean logit over the draws that pass each gate):\n- Naive B:\n - A, the guide's calibration.csv rows plus same-sign age slope: draws 1234, -0.0148, p=0.0019, carrying draws 1,3,4.\n - B, the same with adult-only level calibration: draws 0124, -0.0127, p=0.0063, carrying draws 0,1,4.\n - C, gate A plus generated/observed slope ratio within 0.5-2: draws 123, -0.0145, p=0.0006.\n - D, strict level (adult ratio 0.67-1.5, Spearman ≥0.5): draw 4 only, -0.0157, p=0.023.\n- B: gates A-C give all five draws, -0.0121, p=0.0023, carrying 1,3,4. Gate D gives draws 0,2,4, -0.0079, p=0.048.\n\nOBSERVED paired adults (n=59).\n- Naive B: -0.0153/yr, p=0.117, bootstrap CI -0.034 to +0.004, rank-ANCOVA p=0.18, Spearman -0.095. Within sepsis (n=46): -0.0176, p=0.083.\n- B: -0.0148, p=0.052. With \"Other\" removed from the denominator: -0.0150, p=0.045.\n- Naive share of B: 0.000, p=1.0. Memory B: -0.0138, p=0.096. Effector B: +0.0077, p=0.37. Naive CD8 T: -0.049, p=2e-11, Spearman -0.62.\n- Power to detect the bulk-only slope: 28% (naive B) and 35% (B).\n- Leave-one-out p range: naive B 0.038-0.29; B 0.010-0.15.\n- Fraction scale: the observed naive-B decline is -0.0092 per decade; the generated draws give -0.0032 to -0.0077 per decade.\n\nBULK.\n- Naive-B score (TCL1A, FCER2, IL4R, BACH2, PLPP5):\n - bulk-only: -0.0162 z/yr, p=0.002;\n - paired: -0.0196, p=0.007;\n - bulk-only HC alone: -0.035, p=0.0009.\n- Pan-B score: -0.0181 (p=0.003); in bulk-only HC alone -0.038 (p=0.002).\n- Naive-minus-pan proxy for the naive share: +0.0064, p=0.075.\n- The generated naive-B slope after adjusting for the bulk score: -0.0034 (p=0.21). R² of the generated fraction on the bulk score: 0.64 (0.79 with the pan-B score).\n- On the adult paired test samples:\n - Spearman of observed naive B with the bulk score: 0.74-0.79.\n - Spearman of observed naive B with the generated fraction: 0.33-0.74.\n - Given the bulk score, the generated fraction adds nothing: p 0.37-0.80 in four draws; in d1 its coefficient is negative (p=0.037).\n\nCONTEXT.\n- Generated bulk-only effector B: -0.0097 (p=0.0005). Its paired generated sign is opposite to observed in 4/5 draws.\n- Generated naive CD8 T: -0.0106 (p=0.026). Paired generated slopes are 27-55% of the observed slope.\n- Other q<0.05 age hits in the screen, generated vs observed slope:\n - l1 Other: +0.0103 vs -0.0047;\n - l2 Platelet: +0.0122 vs -0.0104;\n - platelet-in-monocyte negative control: +0.0106 vs -0.0219.", + "issues": [ + "The decline is not naive-specific; it is a pan-B decline. The naive share of B has a generated slope of -0.003/yr (p=0.06; d3 has the opposite sign; adult paired calibration Spearman only 0.06-0.31) and an observed slope of 0.000 (p=1.0). Generated memory B (-0.010, p=0.0075) and effector B (-0.0097, p=0.0005) fall too. 'Mostly naive B' is true only because naive B is the largest B subset.", + "Observed support is in the same direction but underpowered and fragile. Paired adults (n=59): naive B -0.0153/yr (p=0.12, bootstrap CI -0.034 to +0.004, rank-ANCOVA p=0.18, Spearman -0.095); B -0.0148 (p=0.052). The test has 28% / 35% power to detect the bulk-only slope, and leave-one-out p ranges from 0.038 to 0.29 for naive B. This does not contradict the claim, but it does not confirm it either.", + "The 5/5 same-sign paired age calibration compares one observed estimate five times: the adult test sets overlap (54-55 of the 59 adults each). The observed naive-B test-set slope is not significant in any draw (p 0.08-0.32). Draw 4 overstates it 2.5-fold (generated -0.025 vs observed -0.010; CI of the difference -0.030 to +0.0003).", + "calibration.csv computes the level check over 77 test samples, 22-23 of them children, whose B fractions are much higher. Using adults only, naive-B ratios are 1.84/0.58/1.58/0.46/1.01: d3 fails and d0 passes. The carrying draws then become 0,1,4 instead of 1,3,4. The gated slope stays significant under all four gates tried (p 0.0006-0.023).", + "For total B, the strongest slopes come from d1 and d3, which generate only about half the observed B level (adult ratio 0.50-0.51). The draws whose level matches best (d0, d2, d4) give weaker slopes (-0.007, -0.006, -0.011). Gated on those three draws, the slope is -0.0079 (p=0.048), carried by d4 alone.", + "Carrying draws 1,3,4 rests on unadjusted per-draw p<0.05. After BH correction over the 50 features, only d1 and d3 reach q<0.05 (naive B q 0.019 and 0.002; B q 0.022 and <0.001).", + "The novelty framing is wrong: a naive-B decline is not textbook immunosenescence in healthy adults.\n- Flow studies found no age correlation for total or naive B in adults, while memory B and plasma cells decline: Caraux 2010, Haematologica 95:1016 (n=106); Blanco 2018, JACI 141:2208 (n=215, ages 0-90; abstract-level only).\n- Gong 2025 reports 'few baseline changes' outside T cells (p.3) and an expansion of CD27- effector ('age-associated') B cells with age, r=0.21 (p.5).\n- The generator does the opposite: generated effector B falls with age (-0.0097/yr). Its paired generated sign is opposite to observed in 4/5 draws, and the observed slope is +0.008.", + "Ye 2026 has no within-adult age analysis. Its age axis is adults vs children: p.2 composition paragraph; p.3 Fig. 1e, B cells 26.5% of PBMC in children vs 13.5% in adults; p.6 B subsets vs HC; p.17 HC matched by age and sex. So the pattern is unreported for this cohort, but it is a bulk-visible pattern.", + "Nothing beyond the bulk, and the bulk does better.\n- On the adult paired test samples, a 5-gene bulk naive-B score tracks the observed naive-B fraction better than the generated cells in all 5 draws (Spearman 0.74-0.79 vs 0.33-0.74).\n- Given the bulk score, the generated fraction adds nothing: p 0.37-0.80 in four draws; in d1 its coefficient is negative (p=0.037).\n- Bulk B genes already fall with age in the 36 bulk-only HC alone (pan-B score -0.038 z/yr, p=0.002).", + "A screen-level q≈0.03 is weak evidence here. Of the five age hits with q<0.05 in the generated screen, three (l1 Other +0.010, l2 Platelet +0.012, and the platelet-in-monocyte negative control +0.011) and the effector-B decline have the opposite sign in the observed paired cells. B and naive B are the only hits whose direction matches the observed cells.", + "The generator weakens this cohort's strongest real age effect. Observed naive CD8 T falls by -0.049/yr (p=2e-11, Spearman -0.62). Generated paired slopes are only 27-55% of that, with the difference CI excluding 0 in all five draws. Bulk-only generated naive CD8 T (-0.0106, p=0.026) ranks below B, even though the bulk naive-CD8 score (t=-5.2) is stronger than the bulk B score (t=-3.1). The B decline must not be presented as the leading generated age signal." + ], + "reasoning": "My independent code reproduces every headline number exactly, and the bulk-only effect is statistically robust. It holds when draws are stacked with draw as a fixed effect (p=0.0013) and in a quasi-binomial model (p=0.0068). It survives leave-one-out (maximum p 0.0054), the sex-flag and Emergency Medical Center exclusions, adjustment for infection source within sepsis, and restriction to ages 40-85. It is present in HC alone (n=36, -0.022/yr) and in sepsis alone (n=68, -0.015/yr), but not in the 17 ICU controls; the age x sepsis interaction has p=0.64. Nothing in the data refutes it. It is only partly supported, though, and the claim needs rewording.\n\n(1) Calibration. calibration.csv has no age contrast, so I fitted the same model on each draw's adult paired test samples (n 54-55), generated vs observed.\n- The feature level passes for B in all draws and for naive B in four (d0 ratio 2.06).\n- The age slope has the same sign in 5/5 draws, with similar size in 4/5; d4 overstates it 2.5-fold.\n- The observed reference is weak: the five test sets are almost the same 59 adults, and the observed naive-B slope is not significant in any of them.\n- The level check in calibration.csv mixes in 22-23 children. With adults only, the naive-B gate changes (d3 fails at ratio 0.46, d0 passes at 1.84), and the carrying draws move from 1,3,4 to 0,1,4.\n- Still, the gated naive-B slope is significant under every gate I tried (p 0.0006-0.023).\n- Under the guide's gate, naive B is gated to draws 1-4 and carried by draws 1, 3 and 4; B is gated to all five draws and carried by 1, 3 and 4. Only d1 and d3 survive BH correction.\n- For B, the draws whose level matches best (d0, d2, d4) carry weaker slopes (-0.0079, p=0.048 when gated on them). On the fraction scale, generated declines are no larger than the observed one, so the effect is not overstated in absolute terms.\n\n(2) Observed paired cells. The direction and size agree: naive B -0.0153/yr (p=0.12) and B -0.0148/yr (p=0.052; p=0.045 with \"Other\" removed from the denominator). But the test has only 28-35% power, and leave-one-out p ranges from 0.038 to 0.29. This is partial support, not confirmation. The naive share of B shows no age trend in observed cells (p=1.0), so a naive-specific reading has no observed support.\n\n(3) Biology and novelty.\n- Ye 2026 does not analyse age within adults (pp. 2, 3, 6 and 17 checked); its age axis is adults vs children. The finding is new only as an unreported pattern in this cohort.\n- It is not textbook immunosenescence. Healthy-adult flow studies (Caraux 2010, n=106; Blanco 2018, n=215) report stable naive B in adults, with memory B and plasma cells declining. Gong 2025 reports few non-T compositional changes (p.3) and an expansion of age-associated CD27- effector B cells (r=0.21, p.5).\n- The generator instead lowers every B subset, effector B included, which goes against the observed direction and against Gong. This pan-B scaling is what you would expect if the generator follows the bulk B-gene signal.\n- The generator also weakens the cohort's dominant true age effect (naive CD8 T, observed -0.049/yr) to 27-55% of its size. So the generated age ranking is distorted.\n- In the same age screen, three other q<0.05 hits (Other, Platelet, and the platelet-in-monocyte negative control) have the wrong sign against observed cells. That makes screen-level q a weak guide here. B and naive B are the exceptions whose direction matches.\n\n(4) Beyond the bulk. Nothing, as the candidate says, and the bulk does better. The bulk B and naive-B scores fall with age in bulk-only adults (p=0.002-0.003), in paired adults (p=0.007) and in the 36 HC alone (p≤0.002). After adjusting for the bulk score, the generated slope is -0.0034 (p=0.21). On paired adults, the bulk score predicts the observed naive-B fraction better than the generated cells in every draw, and the generated cells add no information given the bulk.\n\nVerdict: qualified. Calibration is consistent but rests on an underpowered observed reference. Observed cells agree in direction but not significantly. The naive-specific and 'generic immunosenescence' framings are not supported. The pattern is fully explained by the bulk.\n\nSuggested wording: \"In bulk-only adults, generated B cells (all B subsets; most B cells are naive) fall with age, by about 11-13% lower odds per decade. Under the guide's gate this is carried by draws 1, 3 and 4 (0, 1 and 4 with adult-only calibration). The bulk B-cell genes already show the same decline, and the observed paired cells point the same way (p=0.05-0.12, underpowered). This is a bulk-level observation, not a generated-cell finding.\"\n\nProcess notes: all work was CPU-only, run under a systemd scope with a 6 GiB memory limit; no GPU was used. Writes went only to my lens directory. No archive update was made: that maintenance step belongs to the coordinating agent.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v_common.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v01_explore.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v01_explore.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v02_bulkonly_age.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v02_age_models.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v02_age_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v02_main_allfeatures.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v03_paired_age_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v03_paired_age_calib.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v03_paired_age_calib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v03_observed_age.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v04_bulk_derivability.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v04_bulk.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v04_bulk_assoc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v04_mediation.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v04_bulk_vs_gen_tracking.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v05_pdf_search.sh", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v05_pdf_hits.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v05_external_refs.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v05_ye2026_pages.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v05_gong2025_pages.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v06_gating_influence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v06_gating_influence.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v06_gating.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v06_scale.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v07_negative_control_age.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v07_negative_control_age.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v08_evidence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-3_calib/v08_evidence.csv" + ] + } + ] + }, + { + "id": "SEP-PROG-4", + "lens": "sep_programs", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "qualified", + "synthesis": { + "rank": 16, + "theme": "method: cell-type-specific content beyond bulk", + "synthesis_strength": "moderate: the main evidence of cell-type-specific content beyond a bulk broadcast; small effect", + "reportable_statement": "After removing both gene and cell-type means, the generated within-type sepsis-vs-HC log fold-change pattern agrees with observed paired cells in all five draws (double-centred r 0.35/0.21/0.17/0.21/0.26 bulk-only; 0.13-0.21 on paired tests; permutation null <=0.018), while a pure bulk broadcast gives about 0. It is modest (about 5-16% of the reliable observed interaction variance), concentrated in well-expressed and monocyte genes, and mostly between lineages.", + "carrying_draws": { + "double-centred interaction": "0,1,2,3,4 (draw 0 strongest, draw 2 weakest)", + "sepsis vs ICU": "none" + }, + "honesty": { + "calibration": "Paired double-centred r 0.215/0.177/0.133/0.176/0.154 (null <=0.016); robust to normalisation (0.08-0.14 with log2(CPM+1)), IEG removal, training-donor-free truth.", + "replication": "5/5 positive.", + "observed": "Observed split-half reliability 0.65-0.70 (Spearman-Brown); generated captures a modest part.", + "beyond_bulk": "Yes, modestly: bulk broadcast double-centred r -0.09 to 0.05; but the claim's single-centred r is matched by a bulk broadcast (0.22-0.30)." + }, + "verdict_reasons": [ + "headline single-centred metric is normalisation-fragile and matched by bulk baselines", + "'by construction' statement false", + "gene-level direction gains only for ensembles; monocyte-normalised bulk beats every draw in monocytes", + "no sepsis-vs-ICU signal" + ], + "new_vs_cohort_papers": "Methodological; the underlying biology is in Ye 2026.", + "next_steps": [ + "Adopt the double-centred statistic with a bulk-broadcast null as the primary beyond-bulk metric for all cohorts.", + "Estimate which gene sets carry it; repeat for sepsis vs ICU with more controls." + ], + "key_sources": [ + "verify/SEP-PROG-4_recompute/v15_summary.csv", + "verify/SEP-PROG-4_recompute/v13_double_centred_calibration.csv", + "verify/SEP-PROG-4_calib/v07_summary_per_draw.csv", + "verify/SEP-PROG-4_calib/v04_reliability.csv" + ] + }, + "finding": { + "id": "SEP-PROG-4", + "cohort": "ye2026_sepsis", + "claim": "The bulk-only generated cells carry a cell-type-specific sepsis response. After removing each gene's mean change across 12 L2 types (the part a whole-PBMC bulk could give), the generated and observed type-specific logFC still correlate (r 0.15–0.32 over 2,790 genes in all 5 draws; permutation null <= 0.02). This survives removing handling and immediate-early genes. Gene by gene, the generated cells beat the raw bulk at predicting the observed within-lymphocyte direction, but most of that edge comes from single-cell-specific genes. In monocytes a monocyte-normalised bulk is as good or better.", + "kind": "other", + "feature": "Type-specific part of the within-type sepsis logFC: gene-wise logFC minus its mean over the 12 main AIFI-L2 types, over 2,790 genes. Also the per-gene within-type direction compared with bulk baselines.", + "comparison": "adult sepsis vs adult HC. Bulk-only generated (per draw) vs observed paired cells, and paired-test generated vs observed.", + "effect_bulkonly": "Centred logFC correlation, bulk-only generated vs observed paired: r = 0.32, 0.17, 0.23, 0.29, 0.28 (draws 0–4). Without 39 immediate-early/handling genes: 0.28, 0.15, 0.20, 0.26, 0.25. On assay-consistent genes (single-cell whole-sample change has the same sign as the bulk): 0.29, 0.15, 0.20, 0.26, 0.24. Per gene, over observed-significant genes, within-lymphocyte direction: generated (carrying draws) 0.84–0.94 sign agreement vs raw bulk 0.75–0.86 (r 0.61–0.80 vs 0.35–0.61). On assay-consistent genes: 0.82–0.94 vs 0.85–0.95. Monocyte types: generated 0.78–0.92, raw bulk 0.81–0.90, monocyte-normalised bulk 0.92–0.95.", + "stats": "Permutation null with type labels shuffled within gene (100–200 permutations): 97.5% quantile <= 0.02 and permutation p = 0.005–0.01 in every draw and gene subset. Direction comparisons are descriptive (sign agreement over 383–1,833 observed q<0.05 genes per type).", + "draws_same_sign": "5/5 (all positive; draw 1 weakest)", + "paired_calibration": "Paired-test centred r (same samples): 0.16, 0.18, 0.18, 0.27, 0.17; without immediate-early genes 0.13–0.26; assay-consistent 0.14–0.27; null <= 0.017. Type-level within-type logFC calibration (calib_types.csv), sepsis vs HC, PASS draws per type: CD14 0,1,3,4; CD16 1,3,4; intermediate 0,1,3; memory CD4 0,2,3,4; CD56dim NK 0,2,3,4; memory CD8, naive CD4, naive CD8, Treg, gdT, naive B 0,3,4; memory B 0,4. PASS rows have r 0.31–0.62, bulk-broadcast r 0.25–0.75, partial r given bulk 0.10–0.58, top-100 sign agreement 0.72–0.99.", + "paired_observed_concordance": "Without any gene-level calibration gate, genes significant in every carrying draw agree with the observed direction in 93–100% of cases across types. Genes moving opposite to the bulk agree in 80–100%. Example: 45 CD14 genes (CSF3R, TFEC, RXRA, LRRK2, KCTD12, ...) fall within monocytes while rising in the bulk; 42 of 45 agree with observed, 33 at observed q<0.05.", + "beyond_bulk": "Yes for the type-specific pattern; a bulk broadcast carries none of it by construction. For gene-level within-type direction the picture is mixed. In lymphocytes the generated cells beat the raw bulk, but the edge largely disappears on assay-consistent genes, so it is mostly single-cell-specific signal. In monocytes a monocyte-normalised bulk is better, and it also recovers 40 of the 45 CD14 genes that move opposite to the bulk.", + "novelty": "Methodological; not reported in the cohort paper.", + "strength": "moderate", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s01_calibration_types.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/calib_types.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/calib_types_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/calib_genes.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s01b_specificity.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/calib_specificity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/broadcast_index.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s01c_specificity_subsets.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/calib_specificity_subsets.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/specificity_bulkonly_vs_observed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s02_bulkonly_de.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/de_within_type.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s07_de_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/de_type_summary_S_vs_HC.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/de_gene_summary_S_vs_HC.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s09_concordance_by_bulk_class.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/concordance_by_bulk_class.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/cd14_opposite_to_bulk_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/s10_direction_vs_bulk_baselines.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/direction_vs_bulk_baselines.csv" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a4ee351bc7e3e9537", + "verdict": "qualified", + "recomputed": "I wrote my own code from pb_l2.npz, meta.csv and the bulk npz files. The claimed numbers reproduce exactly (v01, v06, v05, v09).\n- Single-centred r, bulk-only vs observed, draws 0–4: 0.316, 0.169, 0.234, 0.288, 0.279. Type-shuffle null q97.5 is 0.015–0.020.\n- Same r without IEGs: 0.147–0.282 with the discovery's list, 0.143–0.278 with my broader list. Assay-consistent genes: 0.151–0.289.\n- Paired calibration: 0.159, 0.183, 0.176, 0.270, 0.174.\n- Per-type PASS lists are identical to the claim.\n- The carrying-draw ensemble direction table and the 45 CD14 opposite-to-bulk genes are identical: 42 of 45 agree with observed, 33 at observed q<0.05, and the monocyte-normalised bulk matches 40 of 45.\n\nProper null (v04, v10, v12): one common bulk log2FC per gene, pushed through each set's real per-type expression profile and sample depths.\n- This broadcast gives centred r with observed of 0.22, 0.25, 0.23, 0.24, 0.30 through the generated structure, and 0.41 through the observed structure.\n- In the paired calibration it gives 0.13, 0.24, 0.23, 0.28, 0.26.\n- Its correlation comes entirely from type-level offsets: across the 12 types, its offsets match the observed offsets at r 0.85–0.94.\n\nWith gene AND type means removed (double-centred), the broadcast falls to −0.09 to 0.05. The generated signal stays:\n- bulk-only vs observed: 0.346, 0.213, 0.174, 0.208, 0.260 (null ≤ 0.018);\n- paired calibration: 0.215, 0.177, 0.133, 0.176, 0.154 (null ≤ 0.016);\n- partial r given both broadcasts: 0.355, 0.234, 0.218, 0.249, 0.254 (null ≤ 0.017).\n\nDouble-centred r by expression level:\n- gene × type cells with observed CPM ≥ 100: 0.42–0.64;\n- cells with CPM < 100: 0.12–0.23.\n\nLineage-centred r (only differences between subtypes of one lineage): generated 0.09–0.22, broadcast 0.03–0.07.\n\nConfounder controls (single-centred / double-centred, ranges over draws 0–4):\n- truth limited to each draw's test samples: 0.153–0.286 / 0.173–0.316;\n- age and sex adjusted: 0.167–0.317;\n- Emergency-Center sepsis samples removed: 0.187–0.307;\n- sex_flag samples removed: 0.165–0.311;\n- all of these combined: 0.166–0.277 / 0.170–0.310;\n- every pseudobulk thinned to equal depth: 0.267–0.418 / 0.225–0.387;\n- simulated depth differences alone: −0.26 to 0.03;\n- label confidence ≥ 0.9 (200-cell subsample): 0.180–0.229, against 0.151–0.230 with all cells;\n- other-lineage (ambient) cells removed: 0.172–0.299;\n- at least 20 cells per pseudobulk: 0.236–0.357;\n- log2(CPM+1) normalisation: −0.063 to 0.213 / 0.110–0.251.\n\nA csSAM bulk-deconvolution baseline gives only 0.05–0.07.\n\nGene-level direction per draw (truth = that draw's test samples, PASS rows only):\n- lymphocyte rows where the single draw beats the raw bulk: 5/9 (draw 0), 0/2 (draw 2), 5/8 (draw 3), 8/9 (draw 4);\n- on assay-consistent genes: 1/28;\n- monocyte rows where a draw beats the monocyte-normalised bulk: 0/10.\n\nSingle-draw concordance with observed is 0.66–0.95. For opposite-to-bulk genes it is 0.31–0.81, with 7 of 38 rows below 0.55.", + "issues": [ + "The claim says a bulk broadcast carries none of the centred signal by construction. That is false for its own statistic. One common bulk change per gene, run through each type's real expression level and depth (edgeR prior shrinkage, CPM closure, depth), gives centred r with observed of 0.22–0.30 through the generated structure and 0.41 through the observed structure. The generated cells do not beat this in draws 1, 2 and 4 (bulk-only: 0.169 vs 0.253, 0.234 vs 0.232, 0.279 vs 0.303). In the paired calibration they do not beat it in draws 1–4. The type-label permutation null (≤0.02) destroys this shared structure, so it is the wrong null.", + "The broadcast correlation is entirely type-level offsets (offset r 0.85–0.94). The defensible statistic is the double-centred r (gene and type means removed), where the broadcast is about 0. On that statistic the result holds in 5/5 draws: 0.17–0.35 bulk-only, 0.13–0.21 paired calibration, null ≤ 0.018. The claim should be restated on this statistic or on the partial r given the broadcast (0.22–0.36).", + "The claimed single-centred statistic depends on normalisation. With log2(CPM+1) or edgeR prior 0.5, draw 1 turns negative (−0.063 / −0.048) and draw 2 drops to 0.09–0.15. The double-centred r stays positive (0.11–0.25).", + "The signal sits mostly in well-expressed genes. In gene × type cells with observed CPM 10–100 (about 80% of the 30,153 cells), the single-centred generated r (0.08–0.22) is below the broadcast (0.26–0.35) in every draw. Double-centred r is 0.12–0.23 there, against 0.42–0.64 at CPM ≥ 100.", + "\"Cell-type-specific\" is mostly monocyte vs lymphocyte. Within lineages, the subtype part is weak: lineage-centred r 0.09–0.22 (draw 1 0.09; T lineage 0.08–0.21).", + "The generated type-offset pattern depends on the draw: it matches the observed offsets at r 0.15, −0.14, 0.61, 0.71, 0.48 (draws 0–4).", + "The gene-level edge over the raw bulk in lymphocytes appears only when draws are averaged over carrying draws. Per draw, against each draw's test-sample truth, the draw beats the raw bulk in 8/9 PASS lymphocyte rows in draw 4, about half in draws 0 and 3, and 0/2 in draw 2. On assay-consistent genes the raw bulk beats the single draw in 27 of 28 rows, so the edge is wholly single-cell-specific per draw. In monocytes the monocyte-normalised bulk beats every draw (0/10).", + "The 93–100% concordance, and 80–100% for opposite-to-bulk genes, needs the intersection of all carrying draws. For genes significant in one draw, agreement is 0.66–0.95, and for opposite-to-bulk genes 0.31–0.81 (7 of 38 rows below 0.55, near chance).", + "The 45 CD14 opposite-to-bulk genes are largely derivable from bulk. A monocyte-normalised bulk predicts 40 of 45; a closure-only broadcast with no type-specific change predicts 32 of 45. With the carrying draws' training donors removed, 41 of 45 still agree in sign, but the share at observed q<0.05 falls from 0.73 to 0.42 (only 5 HC remain).", + "The composition rows of calibration.csv for these types are mixed. The sepsis−HC sign is wrong for Treg in draws 1–4, CD16 monocyte in draws 2–4, memory CD4 in draws 0–2, naive CD8 in draws 0, 1 and 3, and L1 monocyte in draws 3–4. Per-sample Spearman is 0.01–0.83, and draw 1's intermediate-monocyte ratio is 7.6. Depth-equalisation and depth-only simulation controls show these cell-number differences do not create the effect.", + "The result does not extend to sepsis vs ICU control: centred r is 0.07–0.09 bulk-only and 0.03–0.06 in the paired calibration." + ], + "reasoning": "My code reproduces every claimed number exactly: the centred r per draw, the null, the IEG and assay-consistent subsets, the paired calibration, the PASS lists, the direction tables and the CD14 gene list. The usual confounders do not remove the effect: age, sex, site, sex_flag samples, leakage from training donors into the observed truth, depth or cell-number imbalance, label confidence, other-lineage (ambient) genes and minimum cells. Imperial-only confounders do not apply: sepsis donor IDs are unique and no bulk-only donor trained a draw. A csSAM bulk-deconvolution baseline does not recover the pattern (r 0.05–0.07).\n\nThe main weakness is the null. I built a pure bulk broadcast: one common log2FC per gene, the same in every type, pushed through each set's real per-type expression profiles and depths. It reaches the same single-centred r as the generated cells (0.22–0.30 through the generated structure, 0.41 through the observed structure), all through type-level offsets. So the claim's statistic, with its type-shuffle null, does not by itself show beyond-bulk type-specific content. The claim's wording that \"a bulk broadcast carries none of it by construction\" is wrong.\n\nWith gene and type means both removed, the broadcast falls to about 0, while the generated cells keep r 0.17–0.35 in all 5 draws. That holds in the paired calibration (0.13–0.21) and under every control and normalisation I tried (0.11–0.41). So a gene-level type-specific sepsis signal beyond the bulk does exist. It is concentrated in well-expressed genes and mostly between lineages.\n\nThe gene-level direction and concordance claims hold only for ensembles or intersections across draws. Per draw they are mixed: the lymphocyte edge over bulk is clear only in draw 4 and absent on assay-consistent genes, and the monocyte-normalised bulk matches or beats every draw in monocytes.\n\nThe core finding survives; its stated statistic, null and beyond-bulk framing need correcting, hence qualified.\n\nAll work was CPU only, run under systemd-run with a 6G cap. I wrote only under my verify directory and did not run the paper-archive update.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/p4lib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v15_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v15_calibration_rows.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v01_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v01_core.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v02_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v02_confounders.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v03_artifact_models.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v03_artifacts.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v04_broadcast_null.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v04_broadcast_null.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v05_direction_vs_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v05_direction_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v05_direction_carrying.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v06_calib_types.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v06_calib_types.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v06_pass_lists.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v07_label_confidence.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v07_label_confidence.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v08_cssam_baseline.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v08_cssam.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v09_concordance.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v09_concordance.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v09_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v09_cd14_opposite.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v10_broadcast_null_sensitivity.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v10_broadcast_sensitivity.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v11_lineage_vs_broadcast.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v11_lineage_vs_broadcast.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v12_double_centred.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v12_double_centred.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v13_double_centred_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v13_double_centred_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v14_double_centred_norms.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v14_double_centred_norms.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v16_direction_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v16_direction_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v00_inspect.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v00_cells_lib_by_group.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_recompute/v00_confidence_by_group.csv" + ] + }, + { + "verifier": "calib", + "agent_id": "a40f8a8f6d5bfa0fa", + "verdict": "qualified", + "recomputed": "I recomputed everything independently from the raw pb_l2.npz files and used none of the candidate's parquet files. The candidate's numbers reproduce exactly.\n\n**Candidate's metric (sepsis vs HC, gene-centred r over 2,790 genes)**\n- Paired test, draws 0–4: 0.159, 0.184, 0.176, 0.270, 0.174. Null q975 at most 0.016.\n- Bulk-only vs all-paired observed: 0.316, 0.169, 0.234, 0.288, 0.279.\n\n**Training-donor overlap (leakage-free truth)**\n- The candidate's observed truth contains each draw's own training donors (HC07 trained draw 0, HC12 trained draw 1).\n- With the truth restricted to each draw's test samples: 0.286, 0.153, 0.232, 0.286, 0.278.\n\n**Normalisation fragility**\n- With log2(CPM+1) instead of the edgeR-style prior, the same metric gives:\n - paired: 0.044, -0.005 (p_perm 0.75), 0.104, 0.185, 0.130;\n - bulk-only: 0.131, -0.071, 0.085, 0.158, 0.214.\n- Cause: per-type global offsets make up 14–20% of the observed gene-centred variance.\n- The generated offsets track the observed ones inconsistently. Paired offset r by draw: -0.47, 0.44, 0.72, 0.86, 0.47.\n\n**Bulk-only baselines with no model (against the observed gene-centred values)**\n- Lineage-marker-ratio bulk (E4): r 0.37–0.41.\n- Bulk broadcast pushed through the per-type logCPM pipeline (E3): r 0.39–0.43.\n- Both beat the generated cells (0.15–0.29) on the candidate's own metric.\n\n**Gene × type interaction (gene and type means both removed)**\n- edgeR, all 12 types:\n - paired: 0.215, 0.177, 0.133, 0.176, 0.154 (null at most 0.016);\n - bulk-only vs leakage-free truth: 0.316, 0.193, 0.173, 0.204, 0.256.\n- log2(CPM+1): paired 0.144, 0.081, 0.096, 0.136, 0.102; bulk-only 0.226, 0.094, 0.109, 0.177, 0.205.\n- The E3 baseline drops to about 0 (-0.02 to 0.02) once type means are removed.\n- Lymphoid types only: paired 0.08–0.17, bulk-only 0.09–0.21.\n- Monocyte types only: paired 0.08–0.15, bulk-only 0.12–0.25.\n\n**Observed ceiling and draw consistency**\n- Observed split-half reliability, sepsis vs HC: gene-centred 0.54 (Spearman-Brown 0.70); interaction 0.48 (Spearman-Brown 0.65, ceiling about 0.80).\n- So the generated cells explain about 5–16% of the reliable observed interaction variance.\n- The five bulk-only draws agree with each other at r 0.36–0.54.\n\n**Sepsis vs ICU controls**\n- Paired interaction r 0.011–0.036.\n- The observed truth itself is unreliable: split-half 0.11 (Spearman-Brown 0.19). This contrast cannot be tested here.\n\n**What drives the agreement**\n- The top 50 genes carry 32–42% of the interaction covariance; without them r is 0.11–0.28.\n- The 37 immediate-early genes carry 14–21%; without them r is 0.15–0.32.\n- HLA-II genes carry 1–2%; ribosomal genes about 1%.\n- The leading contributors rise specifically in CD14 monocytes (CLU, SLC39A8, CD82, FMN1, ACSL1, PLAC8), while immediate-early genes (NR4A1, FOSB, JUN, EGR1, DUSP5, RHOB) fall in monocytes.\n\n**Named genes (interaction values; draws matching the observed sign)**\n\n| Genes | Type | Draws matching observed |\n|---|---|---|\n| HLA-DRA, CD74, HLA-DPA1 (down) | CD14 monocyte | 5/5 |\n| CLU, PLAC8, ALOX5AP (up) | CD14 monocyte | 5/5 |\n| S100A8, S100A12 | CD14 monocyte | 4/5 (draw 1 opposite) |\n| NR4A1, NR4A2, AREG | memory CD4 T | 5/5 |\n| CCL3, TNF | memory CD8 T | 5/5 |\n| CCL3, CCL4 | CD56dim NK | 2/5 |\n| IFITM1, ISG15 | CD14 monocyte | 1/5 |\n\n**Gene-level direction, per draw, leakage-free, carrying draws only**\n\n| Lineage | Generated sign agreement | Raw bulk | Monocyte-normalised bulk | Generated wins (sign) | Generated wins (r) | Assay-consistent genes, wins (sign; r) |\n|---|---|---|---|---|---|---|\n| Lymphoid | 0.65–0.90 | 0.75–0.87 | – | 18/28 rows (draw 4 8/9, draw 0 5/9, draw 3 5/8, draw 2 0/2) | 28/28 | 3/28; 13/28 |\n| Monocyte | 0.70–0.92 | 0.82–0.91 | 0.92–0.95 | 2/10 vs raw, 0/10 vs normalised | 0/10 | – |\n\n**CD14 genes moving opposite to the bulk**\n- 41 of the 45 genes fall within monocytes while rising in the bulk; the other 4 move the other way.\n- Against the leakage-free truth: generated agrees for 42–43/45, monocyte-normalised bulk 39–40/45, raw bulk 2–3/45.\n\n**Selected monocyte genes, sepsis vs ICU controls**\n- CLU and PLAC8 rise within CD14 monocytes vs ICU controls in all 5 draws (observed CLU +1.41, p 0.05; PLAC8 +0.86, p 0.04).", + "issues": [ + "The headline metric (gene-centred r) mixes in per-type global offsets, which are a normalisation/closure effect. They are 14-20% of the observed centred variance, and the generated offsets agree only inconsistently (r from -0.47 to 0.86 across draws). The metric is therefore normalisation-fragile: with log2(CPM+1), draw 1 is null or negative (paired -0.005; bulk-only -0.071) and draw 0 falls to 0.044.", + "The claim that 'a bulk broadcast carries none of it by construction' is misleading. Two simple bulk-only predictors with no model beat the generated cells on the candidate's own metric: a lineage-ratio bulk (r 0.37-0.41) and the bulk broadcast passed through the per-type logCPM pipeline (r 0.39-0.43), against 0.15-0.29 for the generated cells. The beyond-bulk statement only holds for the double-centred gene x type interaction, where the bulk baselines are about 0.", + "Leakage: the bulk-only vs observed truth (46 vs 7) includes each draw's own training donors. HC07 trained draw 0 and HC12 trained draw 1. The inflation is minor: draw 0 0.316 -> 0.286, draw 1 0.169 -> 0.153 on test-only truth.", + "The surviving effect is modest. The interaction r is 0.13-0.22 in the paired test and 0.17-0.32 for bulk-only, against an observed reliability ceiling of about 0.80, so about 5-16% of the reliable observed type-specific variance. It is concentrated: the top 50 genes carry 32-42% of the covariance, mostly CD14-monocyte genes, and 37 immediate-early (handling-type) genes carry 14-21%. 'Moderate' overstates it; weak-to-moderate fits better.", + "It is a sepsis-vs-healthy pattern, not a demonstrated sepsis-specific one. Against ICU controls the paired interaction r is 0.011-0.036. The observed sepsis-vs-ICU truth is itself unreliable (split-half Spearman-Brown 0.19), so sepsis specificity cannot be tested.", + "'Generated beats raw bulk in lymphocytes' holds in correlation (28/28 carrying draw x type rows) but only partly in sign agreement when checked per draw on leakage-free truth: 18/28 rows. Draw 4 wins 8/9 types and draw 2 wins 0/2. On assay-consistent genes the raw bulk wins on sign in 25/28 rows. In monocytes the raw bulk beats the generated cells in 8/10 carrying rows and the monocyte-normalised bulk in 10/10, which is stronger than 'as good or better'.", + "The biology is only partly reproduced. In CD14 monocytes, HLA-II loss and the MS1-like rise of CLU, PLAC8 and ALOX5AP appear in 5/5 draws, as do NR4A1/NR4A2/AREG in memory CD4 T and CCL3/TNF in memory CD8 T. But the fall of IFITM1/ISG15 in monocytes (Ye 2026 Mono_c05-CD14-IFITM1 reduced, journal p.155) is reproduced in only 1/5 draws, S100A8/S100A12 reverse in draw 1, and CCL3/CCL4 in NK hold in only 2/5 draws.", + "Novelty is methodological only. Ye 2026 already reports the underlying cell-type-specific biology: monocyte HLA-DR loss and MS1-like Mono_c03-CD14-RETN (journal pp.155, 157, 163), the NR4A2/AREG CD4 TCM program (p.151) and CCL3/CCL4/TNF CD8/NK (p.150). It uses bulk RNA-seq only for deconvolution validation (p.161). Reyes 2020 (author manuscript p.3, Fig. 2e) reports CLU and PLAC8 in MS1 cells as sepsis markers.", + "Minor wording: of the '45 CD14 genes that fall within monocytes while rising in the bulk', 41 do; 4 rise within monocytes while falling in the bulk. The 42/45 observed agreement holds on leakage-free truth (42-43/45). The generated cells beat the monocyte-normalised bulk there (39-40/45), but on a gene set selected on the generated side." + ], + "reasoning": "**Calibration.** The feature is not in calibration.csv, so I built the paired check myself: generated vs observed on each draw's own test samples. It supports sepsis vs HC in all 5 draws. The candidate's gene-centred r (0.16-0.27) and my double-centred r (0.13-0.22) both sit well above a type-label permutation null of at most 0.016. Every draw carries the sepsis-vs-HC result. Ranked by the interaction, draw 0 is strongest and draw 2 weakest. No draw carries a sepsis-vs-ICU result, because the observed truth for that contrast is unreliable.\n\n**What the headline metric really measures.** The candidate centres each gene but not each cell type, so per-type global offsets stay in. These offsets are normalisation/closure effects, not biology. Because of them the metric flips sign for draw 1 under log2(CPM+1), and a bulk-only lineage-ratio or pipeline-broadcast predictor reproduces it better than the generated cells do. The claim that the bulk carries none of this pattern therefore does not hold for the metric as reported.\n\n**What survives.** After removing both gene and type means, the gene x type interaction remains:\n- positive in all 5 draws;\n- robust to normalisation;\n- robust to removing training donors from the truth, to dropping sex-flag samples and to dropping immediate-early genes;\n- present within lymphoid types alone and within monocyte types alone;\n- not explained by any bulk-derived baseline (all about 0).\n\n**Observed data and biology.** The observed paired cells show a reliable type-specific pattern (split-half Spearman-Brown 0.65-0.70). The generated cells capture a modest part of it, driven mainly by CD14-monocyte genes. HLA-II loss and the MS1-like CLU/PLAC8/ALOX5AP gain agree with Ye 2026 and Reyes 2020; the interferon-monocyte fall does not reproduce.\n\n**Gene-level direction.** Checked per draw on leakage-free truth, the edge over the raw bulk in lymphocytes is draw-dependent and largely limited to single-cell-specific genes. In monocytes, simple bulk baselines are better than the generated cells.\n\n**Verdict.** Calibration, observed agreement and biology support the claim only in part, so it is qualified rather than confirmed or refuted.\n\nAll runs were CPU-only under a 6 GiB systemd scope. Nothing was written outside the verify directory, and I made no paper-archive update; that is left to the coordinating agent.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v01_type_specific_verify.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v01_results.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v01_split_half.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v01_centred_cells.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v02_double_centring.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v02_double_centring.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v02_type_offsets.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v03_direction_and_biology.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v03_direction_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v03_direction_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v03_gene_contrib.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v03_contrib_classes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v03_named_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v03b_named_selected.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v03_named_genes_selected.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v04_reliability_and_consistency.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v04_reliability.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v04_reliability_iters.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v04_cross_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v05_cd14_opposite_leakfree.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v05_cd14_opposite_leakfree.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v06_clu_plac8_icu.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v06_clu_plac8_icu.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v07_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v07_summary_per_draw.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-PROG-4_calib/v07_direction_wins.csv", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf", + "/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/rare-disease-states/Reyes2020_Immune_cell_signature_bacterial_sepsis.pdf" + ] + } + ] + }, + { + "id": "SEP-CLIN-6", + "lens": "sep_clinical", + "final": "qualified", + "verdict_recompute": "qualified", + "verdict_calib": "confirmed", + "synthesis": { + "rank": 17, + "theme": "artefact", + "synthesis_strength": "artefact (demonstrated); not biology", + "reportable_statement": "The male excess of 'Proliferating NK' labels (generated 1.09% vs 0.62%; observed 1.72% vs 0.88%) is a labeller artefact: the frozen scANVI labeller takes RPS4Y1, USP9Y and UTY as inputs. Zeroing them removes the observed per-NK male excess (+0.451 to -0.007, p 0.97). In generated cells a second layer remains: the generator makes more NK cells in men than the bulk supports.", + "carrying_draws": { + "biological result": "none", + "label excess": "0-4 (4/5 p<0.05)" + }, + "honesty": { + "calibration": "Feature ratio 0.41-0.72, Spearman 0.23-0.51; sepsis-vs-HC sign 3/5 (claim said 4/5); only draw 3 passes the gate and it distorts the sex contrast most.", + "replication": "Label excess 5/5; chrY explains 20-82% of the generated excess per draw.", + "observed": "Label excess +0.60 (p 0.0016) but marker-defined cycling NK show none (-0.17, p 0.63).", + "beyond_bulk": "Bulk shows no sex difference in NK or proliferation scores." + }, + "verdict_reasons": [ + "mechanism demonstrated by counterfactual relabelling", + "generator NK-by-sex prior beyond bulk", + "draws 0 and 2 (5 male training donors) leak chrY into women's cells (31-35%)", + "other L2 labels (transitional B, CD8aa, naive B) also shift with chrY inputs" + ], + "new_vs_cohort_papers": "Not reported; flagged so no lens reports L2-label sex contrasts as biology.", + "next_steps": [ + "Relabel without chrY (and XIST) inputs before any sex analysis; sex-balanced training sets." + ], + "key_sources": [ + "verify/SEP-CLIN-6_calib/v09_counterfactual_models.csv", + "verify/SEP-CLIN-6_recompute/v07b_models.csv", + "verify/SEP-CLIN-6_recompute/v11_other_labels.txt", + "verify/SEP-CLIN-6_calib/v04_paired_sex_calibration.csv" + ] + }, + "finding": { + "id": "SEP-CLIN-6", + "cohort": "ye2026_sepsis", + "claim": "'Proliferating NK cell' labels are about 1.8x more frequent in men than in women in bulk-only generated cells (1.09% vs 0.62% of cells), and observed paired cells show the same. This is most likely a labeller artifact rather than more proliferating NK cells in men.", + "kind": "composition", + "feature": "l2 | Proliferating NK cell (AIFI L2 label from the frozen f01common scANVI)", + "comparison": "Men (78) vs women (43), all bulk-only adults, adjusted for age and group; also within sepsis", + "effect_bulkonly": "Five-draw mean: 1.09% (men) vs 0.62% (women) of cells. Adjusted logit +0.53 (SE 0.10); within sepsis +0.56.", + "stats": "Model: OLS with HC3 errors, mean of draw logits ~ male + age + group, n=121: p=1.5e-7, BH q=7.6e-6; rank-ANCOVA q=1.4e-5. Within sepsis q=0.041.", + "draws_same_sign": "5/5, with 4/5 at p<0.05. The calibration gate passes only draw 3.", + "paired_calibration": "calibration.csv: ratio 0.63/0.41/0.48/0.72/0.52, Spearman 0.27/0.51/0.26/0.50/0.23. Sepsis vs HC same sign 4/5; sepsis vs ICU 5/5. Sex coefficient on the paired test samples, gen vs obs: 0.71/0.55, 0.45/0.62, 0.47/0.59, 1.04/0.59, 0.10/0.70, i.e. 5/5.", + "paired_observed_concordance": "Observed paired adults (40 men, 19 women): 1.72% vs 0.88% (logit +0.60, p=0.0016, q=0.016); within sepsis +0.65 (q=0.044). However, marker-defined cycling NK cells do not follow it. Defined as NK cells with at least 2 of MKI67, TOP2A, TYMS, CENPF, UBE2C and BIRC5, they show no adjusted male excess in observed cells (-0.17, p=0.63). In generated cells they are lower in men (-0.26 to -0.63; 4/5 draws p<0.05).", + "beyond_bulk": "The bulk does not show it. Neither the NK marker score (p=0.30), the proliferation score (p=0.87) nor their sum (p=0.65) differs by sex, in bulk-only or paired bulk. The labeller's 1,006 input genes include the chrY genes RPS4Y1, USP9Y and UTY, and lack MKI67, TOP2A, TYMS, UBE2C and BIRC5; only CENPF and STMN1 of 16 cycling genes are inputs. Only 1-9% of 'Proliferating NK' labels carry at least 2 cycling genes (observed: 6-7%). The association therefore belongs to the label, not to proliferation, and most likely not to biology.", + "novelty": "Not reported in Ye 2026. Not pursued further because it is artifact-suspect. Flagged so that other lenses do not report it as biology.", + "strength": "artifact-suspect", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s02_summary.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s04_observed_assoc.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s06_feature_vs_bulk.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s07_artifact_checks.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s07_prolif_nk_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s07_ygenes_pb.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s07b_query_genes.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical/s10_evidence.txt" + ] + }, + "verdicts": [ + { + "verifier": "recompute", + "agent_id": "a43bb567dffb964dd", + "verdict": "qualified", + "recomputed": "The numbers were recomputed from raw cells.parquet and markers.npz. Per-sample counts match composition.parquet exactly (1,077 sample-draw rows).\n\nBulk-only adults (78 men, 43 women):\n- Five-draw mean PN share: 1.090% vs 0.616% (ratio 1.77).\n- Mean-logit OLS (HC3), male +0.531 (SE 0.101), p 1.5e-7; BH q 7.6e-6 (rank 1 of 50 features). Rank-ANCOVA p 2.9e-7.\n- Within sepsis +0.555, p 8.2e-4, q 0.041.\n- Per draw (d0..d4): +0.27 (p 0.052), +0.41 (p 0.0067), +0.54 (p 7.7e-5), +1.14 (p 1e-11), +0.30 (p 0.013). That is 5/5 positive and 4/5 at p<0.05. Within sepsis only 3/5 reach p<0.05.\n- Stacked model with draw as a fixed effect and sample-clustered errors: +0.531, p 7e-8.\n\nObserved adults (40 men, 19 women): 1.72% vs 0.88%, +0.596, p 0.0016, q 0.016. Within sepsis +0.651, p 0.0035, q 0.044.\n\nCalibration (checked against calibration.csv, all values identical):\n- Ratio 0.63/0.41/0.48/0.72/0.52; Spearman 0.27/0.51/0.26/0.50/0.23.\n- Same-sign flags: sepsis vs HC 1,1,0,1,0 = 3/5 (the claim says 4/5); sepsis vs ICU 5/5; sepsis vs all controls 3/5.\n- Strict gate (all three signs agree, Spearman ≥0.3, ratio 0.5–2) passes only draw 3. A lax gate (Spearman ≥0.2, ratio 0.4–2.5) passes draws 1 and 3.\n- Sex coefficient on paired test adults, gen/obs: 0.71/0.55, 0.45/0.62, 0.47/0.59, 1.04/0.59, 0.10/0.70. Same sign 5/5, but generated p<0.05 only in d0, d1, d3; d4 p 0.54.\n\nDecomposition on paired test samples (gen/obs):\n- NK share of all cells: 0.49/0.21, 0.51/0.22, 0.35/0.26, 0.81/0.20, 0.30/0.21. Observed is non-significant in every draw.\n- PN share of NK: 0.25/0.40, 0.09/0.48, 0.07/0.41, 0.23/0.46, -0.23/0.59. Observed p ≤ 0.025 in every draw.\n\nBulk-only five-draw mean: NK share +0.378 (p 0.001); PN share of NK +0.216 (p 0.012); PN adjusted for NK share +0.240.\nObserved: NK share +0.227 (p 0.30); PN share of NK +0.451 (p 0.0097).\n\nConfounders, PN effect (five-draw mean / observed):\n- Drop the 3 sex_flag samples: 0.526 / 0.596.\n- Adjust for site (Emergency Medical Center): 0.525 / 0.614.\n- First Affiliated hospital only: 0.529 / 0.482 (p 0.014).\n- Adjust for infection source: 0.512 / 0.482 (p 0.023).\n- Age: quadratic, male×age interaction, 40–85 only: unchanged.\n- Adjust for median library size: 0.534 / 0.545.\n- Huber robust regression: 0.481 / 0.582.\n- Permutation within group (20,000): p 5e-5 / 0.0034.\n- Adjust for log cell number (observed): 0.595.\n- Labels with confidence ≥0.95 only: 0.538 / 0.583.\n- PN labels carrying ≥2 cycling genes: -0.05 (p 0.58) / +0.03 (p 0.91).\n- Marker-defined cycling NK as a share of NK: generated -0.26 to -0.63 (4/5 p<0.05); observed -0.17 (p 0.63).\n\nLabeller perturbation (CPU, frozen scANVI): labels reproduce exactly before perturbing (100% of 589,478 observed cells and 138,400 subsample cells).\n- All observed cells, chrY inputs (RPS4Y1, USP9Y, UTY) zeroed: men's PN share of NK falls from 0.198 to 0.139; women stay at 0.136. PN-share-of-NK coefficient +0.451 → -0.007 (p 0.97). Overall PN effect +0.596 → +0.200 (p 0.29), which equals the NK-share effect (+0.22, p 0.31). Men/women ratio 1.96 → 1.33.\n- Generated, 200-cell subsamples: chrY explains 44/57/50/20/82% of the male excess in d0..d4. Five-draw mean +0.56 → +0.39 (p 9e-6). After zeroing, the within-NK effect is null in d0, d1, d2, d4 but stays at +0.78 (p 1e-4) in d3.\n- Giving women's cells male chrY counts raises their observed PN share from 0.79% to 1.18%.\n- Generated women's cells carry chrY counts in 34.6% (d0) and 30.8% (d2) of cells; observed women 0%.\n\nBulk:\n- Bulk-only (men vs women): NK-specific score +0.27 (p 0.067); cycling score -0.03 (p 0.87); sum p 0.33.\n- Paired bulk: NK p 0.56, cycling p 0.97.\n- Adjusting generated NK share for the bulk NK score leaves d0..d2 non-significant (0.11, 0.08, 0.17) but d3 0.54 (p 5e-7) and d4 0.27 (p 6e-4).\n\nLabeller reference: PN class precision 0.54 and recall 0.92 on held-out reference data; 188 of 444 PN calls are CD56dim NK cells. The reference PN class is 51% male, and chrY counts per 10k are equal in PN and CD56dim cells of men (55.2 vs 55.1).", + "issues": [ + "The claim misstates calibration: sepsis vs HC same sign is 3/5 in calibration.csv (draws 2 and 4 disagree), not 4/5. Sepsis vs all controls is also 3/5 (draws 0 and 4 disagree). Calibration of this feature is weak overall: ratio 0.41-0.72, Spearman 0.23-0.51.", + "The only draw that passes the strict gate (draw 3) is the least faithful on the sex contrast. On the same paired test samples it overstates the sex effect (1.04 vs 0.59) and the NK-share sex effect (0.81 vs 0.20). ChrY explains only 20% of its excess, and a within-NK male excess remains after chrY removal (+0.78, p 1e-4). Passing the group-contrast gate does not carry over to the sex contrast.", + "On paired test samples the generated sex effect reaches p<0.05 in only 3/5 draws (0, 1, 3). Draw 4 is null (0.10 vs observed 0.70).", + "The generated and observed effects have different makeups. Observed: a shift within NK cells (PN share of NK +0.45, p 0.01) with NK abundance non-significant (+0.23, p 0.30). Generated: mostly more NK cells in men (+0.38, p 0.001; 5/5 draws, exaggerated vs observed on paired test). Adjusting for NK share halves the generated effect (0.53 to 0.24). 'Observed cells show the same' holds for the label frequency, not for the mechanism.", + "The artifact interpretation is confirmed directly for observed cells. Re-running the frozen labeller with its 3 chrY inputs zeroed abolishes the within-NK male excess (+0.45 to -0.01, p 0.97). The remaining overall effect (+0.20, p 0.29) is just the non-significant NK-abundance difference. For generated cells the labeller explains only 20-82% per draw (five-draw mean 0.56 to 0.39, still p 9e-6). The rest is a generator-level sex difference in NK abundance plus a draw-3-specific residual. So the generated effect is a two-layer artifact (labeller plus generator), not only a labeller artifact.", + "Generator chrY leakage: draws 0 and 2 were trained on 5 male donors and 0 female donors. They put chrY counts into 35% and 31% of women's generated cells; observed women have 0%. This dilutes any sex contrast of chrY-sensitive labels in those draws and breaks sex fidelity to the bulk.", + "The bulk check depends on the gene set. An NK-specific bulk score (KLRF1, NCR1, SH2D1B, KIRs, NCAM1, KLRC1) shows a borderline male excess in bulk-only samples (+0.27, p 0.067; paired p 0.56). Adjusting for it removes the generated NK sex difference in draws 0-2 but not in draws 3 (0.54) or 4 (0.27). Bulk proliferation shows no sex difference (p 0.87), as the claim says.", + "'Cycling NK lower in men in generated cells' (cycling NK as a share of NK, -0.42) is a denominator effect from the inflated male NK count. As a share of all cells, cycling NK is flat (-0.05, p 0.55).", + "The chrY input also shifts other AIFI L2 labels in men's cells: Transitional B -24 to -45%, Proliferating T -13 to -15%, Memory CD4 T -11%, Effector B -8 to -11%, CD8aa +17 to +26%, Naive B +7 to +18%. Any sex contrast of L2 labels from this labeller is suspect, not only Proliferating NK.", + "Within sepsis only 3/5 draws reach p<0.05 (draw 0 p 0.07, draw 1 p 0.25).", + "The labeller's PN class is weak even on its own reference held-out data (precision 0.54, recall 0.92; false PN calls are mostly CD56dim NK). The reference PN class is not sex-skewed, so the chrY sensitivity comes from the classifier, not from the reference's sex mix.", + "Limitation of this check: generated cells were re-labelled only on the stored 200-cell subsamples, because the full generated rates file (4.9 GB) does not fit the 6 GB memory cap. Per-draw numbers after chrY removal are less precise than the full-data observed result.", + "Not applicable for sepsis: each donor has one sample, so there are no repeated measures, and no bulk-only donor trained any draw. Training donors are excluded from the paired test sets by construction.", + "Which draws carry the result: bulk-only p<0.05 in draws 1-4 (draw 0 p 0.052); strict calibration gate only draw 3; paired-test sex direction reproduced at p<0.05 in draws 0, 1 and 3. None of them supports biology. After chrY removal, the only within-NK excess left is in draw 3, and it is generator-specific." + ], + "reasoning": "I recomputed everything with my own code from cells.parquet, markers.npz, meta.csv and calibration.csv. My per-sample counts matched composition.parquet exactly. The headline effect, the per-draw signs, the observed effect and all four BH q-values reproduced (q values come from a family of 50 features). The effect also survived every standard confounder control: age (linear, quadratic, interaction, restricted range), site, infection source, dropping the sex_flag samples, library size, cell number, label confidence, robust regression, dropping outliers, and a within-group permutation test. So it is not refuted.\n\nThe claim's mechanism (chrY genes as labeller inputs) was only inferred. I tested it directly by re-running the frozen f01common scANVI on CPU with CUDA hidden; it reproduced the stored labels 100%. With RPS4Y1, USP9Y and UTY zeroed across all 589,478 observed cells, the within-NK male excess disappeared completely and the overall effect became non-significant. Injecting male chrY counts into women's cells raised their PN rate. This supports the artifact reading for observed cells more strongly than the discovery agent's evidence did.\n\nI chose qualified rather than confirmed for four reasons:\n1. Calibration is weak, and one stated calibration figure is wrong (sepsis vs HC is 3/5, not 4/5).\n2. The generated effect is only partly the labeller artifact. About half is a generator-level excess of NK cells in men, which is exaggerated relative to observed on the same paired samples. Draw 3 also keeps a within-NK residual.\n3. The one strictly calibrated draw (draw 3) is the most inflated on the sex contrast.\n4. Male-only-trained draws leak chrY into women's generated cells.\n\nThe claim's conclusion stands, and is strengthened for the observed cells: this is not biology and should not be reported as such. But the mechanism behind the generated data is broader than \"labeller artifact\".\n\nNo GPU work was started; all Python ran in memory-capped systemd scopes. Outputs are only in my lens dir. This verifier published no archive delta; the coordinating agent owns any registration.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v12_key_numbers.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v12_key_numbers.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/vlib.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v01_samples.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v01_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v02_recompute.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v02_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v02_means.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v03_calibration.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v03_calib_recomputed.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v03_sex_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v04_confounders.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v04_confounders.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v05_decomposition.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v05_paired_decomp.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v05_train_donors.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v06a_genes.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v06b_reference.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v07_perturb.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v07_cells.parquet", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v07b_analyse.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v07b_rates.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v07b_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v08_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v08_bulk_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v09_observed_full_relabel.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v09_observed_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v09b_observed_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v10_bh_family.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_recompute/v11_other_labels.txt" + ] + }, + { + "verifier": "calib", + "agent_id": "a92a3f78416902ae1", + "verdict": "confirmed", + "recomputed": "The verdict confirms the artifact flag. It does not confirm any biology.\n\nBULK-ONLY, 121 adults (78 men, 43 women). Model: OLS with HC3 errors, logit ~ male + age + group.\n- 'Proliferating NK' (PN) as a share of all cells, per draw (d0–d4):\n - coefficients +0.27, +0.41, +0.54, +1.14, +0.30;\n - p 0.052, 0.0067, 7.7e-5, 1e-11, 0.013;\n - 5/5 draws positive, 4/5 at p<0.05.\n- Five-draw mean logit: +0.531 (SE 0.101, p 1.5e-7). Five-draw mean shares: 1.09% (men) vs 0.62% (women).\n- Within sepsis: +0.555 (p 8e-4).\n- NK share of all cells: +0.22, +0.30, +0.27, +0.73, +0.37 (mean logit +0.38, p 0.001).\n- PN as a share of NK cells: +0.08, +0.25, +0.28, +0.52, −0.09.\n\nCALIBRATION\n- calibration.csv for this feature, per draw:\n - ratio 0.63, 0.41, 0.48, 0.72, 0.52 (pooled 0.55);\n - Spearman 0.27, 0.51, 0.26, 0.50, 0.23 (pooled 0.35).\n- Disease contrasts, draws with the same sign as observed:\n - sepsis vs HC 3/5 (the candidate wrote 4/5);\n - sepsis vs ICU 5/5;\n - sepsis vs all controls 3/5.\n- Gate (ratio ≥ 0.5, Spearman ≥ 0.3, contrast sign kept): only draw 3 passes.\n- My own paired sex check (paired_gen vs observed, on each draw's adult test samples), generated / observed:\n - PN of all cells: 0.71/0.55, 0.45/0.62, 0.47/0.59, 1.04/0.59, 0.10/0.70. Same sign 5/5; generated p<0.05 only in d0, d1, d3.\n - PN of NK cells: 0.25/0.40, 0.09/0.48, 0.07/0.41, 0.23/0.46, −0.23/0.59. Same sign 4/5; per-sample Spearman 0.06–0.18.\n - NK of all cells: 0.49/0.22, 0.51/0.22, 0.35/0.26, 0.81/0.20, 0.30/0.21. Same sign 5/5, but generated is 1.4–4.0× the observed.\n\nOBSERVED paired adults (n=59: 40 men, 19 women)\n- PN of all cells: 1.72% vs 0.88%, +0.596 (p 0.0016). Within sepsis +0.651 (p 0.0035).\n- PN of NK cells: +0.451 (p 0.0097). NK of all cells: +0.227 (p 0.30).\n- Marker-defined cycling NK, of NK cells: −0.169 (p 0.63). MKI67+ NK: −0.171 (p 0.59).\n- Of all PN labels, 5.5% carry at least 2 cycling genes, against 0.02% of CD56dim labels. 96.5% of truly cycling NK cells are labelled PN.\n- PN makes up 15.4% of NK cells. In the 7 adult HC the per-sample median is 19.7%, against 2.0% in the healthy AIFI atlas.\n\nCOUNTERFACTUAL RE-LABELLING with the frozen labeller\n- Unchanged cells reproduce the stored labels 100% in all 11 sets.\n- Observed cells, PN of all / PN of NK:\n - stored labels: +0.596 / +0.451;\n - three chrY inputs (RPS4Y1, USP9Y, UTY) set to zero: +0.200 (p 0.29) / −0.007 (p 0.97);\n - women given men's chrY counts: +0.203 / −0.008;\n - control, three expression-matched autosomal genes (CLIC3, HLA-DQA1, GBP1) set to zero: +0.556 / +0.420.\n- Observed PN rates:\n - men 1.70% → 1.16% when chrY is zeroed; women unchanged at 0.90%;\n - women rise to 1.35% when given men's chrY counts;\n - men/women ratio 1.88 → 1.28 (chrY zeroed), 1.25 (women given chrY), 1.86 (control).\n- Bulk-only, 200-cell subsamples, PN of all cells per draw:\n - stored 0.45, 0.52, 0.44, 1.24, 0.16;\n - chrY zeroed 0.30, 0.33, 0.26, 1.09, −0.04;\n - control 0.44, 0.57, 0.48, 1.24, 0.24.\n\nHEALTHY ATLAS (AIFI, 92 adults, full-transcriptome labels)\n- PN of NK cells: −0.02 (p 0.88).\n- NK of all cells: +0.30 (p 0.007).\n- PN of all cells: +0.26 (p 0.02), entirely through NK abundance.\n\nLABELLER\n- 1,005 model genes, including RPS4Y1, USP9Y and UTY. No XIST.\n- Cycling genes it sees: only CENPF, STMN1 and CENPM.\n- Its own held-out validation for PN: precision 0.536, recall 0.919. 7.3% of true CD56dim cells are called PN.\n\nBULK\n- NK score by sex: +0.13 (p 0.33) in bulk-only; +0.08 (p 0.67) in paired.\n- Proliferation score by sex: −0.03 (p 0.85) in bulk-only; −0.05 (p 0.78) in paired.\n- Generated NK share by sex: +0.39 (p 3e-4), and still +0.31 (p 1.5e-5) after adjusting for the bulk NK score.", + "issues": [ + "Mechanism now shown directly. The three chrY genes the labeller takes as input (RPS4Y1, USP9Y, UTY) cause the male excess. In observed cells, zeroing them or giving women men's chrY counts removes the per-NK excess completely (+0.451 -> -0.007 / -0.008). The per-cell excess falls to +0.20 (p 0.29), which equals the NK-abundance gap (+0.23, p 0.30). Zeroing three expression-matched autosomal genes leaves the excess in place (+0.56 / +0.42). This is stronger than the candidate's 'most likely'.", + "Why the labeller links chrY to PN: in its training reference, men's true PN cells carry more total RNA than CD56dim cells (603 vs 441 counts on the model genes). Their absolute chrY counts are therefore higher (RPS4Y1 3.10 vs 2.16), though the share per 1,000 counts is about the same (5.6 vs 5.4). The PN class is not male-enriched in training (48.4% male vs 46.9% for CD56dim), so the sex link comes from an absolute-count proxy that only exists in men.", + "In bulk-only generated cells the chrY artifact explains only part of the 1.8x. After zeroing chrY, PN of all cells stays at +0.26 to +0.33 in draws 0-2 and +1.09 in draw 3. This residual mostly comes from the generator making more NK cells in men. Generated NK share is +0.31 (p 1.5e-5) even after adjusting for the bulk NK score, while the bulk itself shows no sex difference in NK. In paired calibration the generated NK sex effect is 1.4-4.0x the observed one.", + "Draw 3 is the only draw that passes the feature gate, yet it distorts the sex contrast most. In paired calibration its PN effect is 1.04 vs 0.59 observed and its NK effect 0.81 vs 0.20. It also generates shallower NK cells in men (median 941 vs 1,183 counts, p<0.001). In generated cells the labeller calls shallow NK cells PN more often (logit slope about -1 per log2 library). The result is a per-NK excess of +0.80 that remains after chrY removal (p 0.001). No draw carries a biological result.", + "Draws 0 and 2 were trained on 5 male donors and write chrY counts into women's cells (median 2.2 and 2.0 per 1,000 counts vs about 0 in the other draws). This weakens the artifact's sex contrast in those draws.", + "Natural test (n=1, anecdotal): Res-S34 is male in the metadata but its bulk has no chrY. Its generated PN share is low in every draw, at the 0-41st percentile among men.", + "Minor factual errors in the candidate text: (1) The labeller has 1,005 model genes; GPX1 is on the 1,006-gene query list but is dropped. (2) CENPM is also a cycling input. (3) calibration.csv gives sepsis vs HC the same sign in 3/5 draws, not 4/5. Sepsis vs all controls is also 3/5.", + "Feature calibration is weak: generated cells carry about half the observed PN share (ratio 0.41-0.72) and per-sample Spearman is 0.23-0.51. Per-NK Spearman for the sex contrast is only 0.06-0.18. Absolute PN levels should not be used for anything.", + "Cross-lens warning: other lenses should not report the generated NK-cell fraction by sex as biology either. It goes beyond the bulk and is over-amplified compared with observed cells." + ], + "reasoning": "I tried to refute SEP-CLIN-6 and could not. The candidate's numbers reproduce exactly, and new direct tests turn its \"most likely an artifact\" into a demonstrated labeller artifact. The male excess of 'Proliferating NK' (PN) labels is not more proliferating NK cells in men.\n\n1. Calibration\n- The feature is in calibration.csv, but its calibration is weak: ratio about 0.55 and Spearman about 0.35. Disease-contrast signs agree in 3/5, 5/5 and 3/5 draws. Only draw 3 passes the gate.\n- The sex contrast is not in calibration.csv, so I computed it on each draw's adult test samples. PN as a share of all cells keeps its sign in 5/5 draws, but magnitudes range from 0.15x to 1.8x of observed.\n- The structure of the effect differs between generated and observed cells:\n - generated cells reproduce only 17–63% of the observed per-NK excess (one draw has the opposite sign);\n - they over-amplify the male NK-abundance excess by 1.4–4.0x.\n- Draw 3, the only gated draw, over-amplifies most.\n\n2. Observed data\n- Observed paired adults show the label-level excess: +0.60 (p 0.0016) for PN of all cells, +0.45 (p 0.0097) for PN of NK.\n- Marker-defined cycling NK cells show no excess: −0.17 (p 0.63).\n- A label-only artifact looks exactly like this:\n - only 5.5% of PN labels carry at least 2 cycling genes;\n - PN makes up 15–20% of NK cells, about 10x the healthy-atlas share and about 7x the share of the authors' MKI67+ NK cluster;\n - the labeller's own validation precision for PN is 0.54.\n- Counterfactual re-labelling with the frozen scANVI on CPU (unchanged cells reproduce the stored labels 100%) settles the mechanism:\n - zeroing the three chrY input genes erases the per-NK excess (−0.007, p 0.97);\n - giving women men's chrY counts does the same (−0.008);\n - an expression-matched autosomal control leaves the excess in place.\n- What remains is only the NK-abundance gap, which matches healthy biology.\n\n3. Biology and novelty\n- Ye 2026 does not report it:\n - p.8 describes NK_c04-MKI67 (MKI67, STMN1) with no sex analysis;\n - p.17 uses sex only to match controls;\n - p.36 (Extended Data Fig. 7g) lists NK_c04-MKI67 as 182 of 8,210 NK cells, about 2.2%.\n- In Gong 2025 (AIFI atlas), sex is a covariate only (p.3, p.14). Its own obs data (92 adults) show:\n - men have more NK cells (+0.30, p 0.007);\n - the same PN share of NK (−0.02, p 0.88).\n- So \"more proliferating NK cells in men\" is not supported anywhere. \"More NK cells in men\" is known biology and is not new.\n\n4. Beyond bulk\n- The bulk shows no sex difference in NK or proliferation scores, in bulk-only or paired samples.\n- The generated label excess adds nothing biological beyond the bulk. It combines:\n - the labeller's chrY artifact, active in all draws;\n - in draw 3, shallower generated male NK cells that the labeller pushes towards PN;\n - a generator NK-abundance-by-sex prior that goes beyond the bulk's NK markers.\n- None of these is proliferation.\n\nDraws: the label excess appears in 5/5 draws (4/5 at p<0.05). Only draw 3 passes the calibration gate, and it is the draw that distorts the sex contrast most. No draw carries a biological result.\n\nThe verdict is \"confirmed\" in this sense: the artifact flag stands and is now proven, and the finding must not be reported as biology. The refinements and the small factual corrections are listed under issues.\n\nEverything ran on CPU only, with no GPU work. Each run was capped at 6G memory with no swap. All writes are under my lens directory. There is no archive delta from this verifier; the coordinator owns the paper-archive workflow.", + "files": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v01_labeller_genes.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v01_labeller_genes.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v01_model_genes.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v02_reference_sex.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v02_reference_sex.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v02_atlas_donor_pn.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v03_relabel.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v03_relabel/", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v04_analysis.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v04_analysis.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v04_paired_sex_calibration.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v04_bulkonly_sex.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v04_observed_sex.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v04_observed_samples.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v05_pdf_pages.sh", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v05_YE_pages.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v05_GO_pages.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v06_relabel_controls.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v06_matched_genes.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v07_reference_chrY.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v07_reference_chrY.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v08_bulk_check.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v08_bulk_check.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v09_counterfactual_summary.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v09_counterfactual_summary.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v09_counterfactual_models.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v09_counterfactual_rates.csv", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v10_sexflag_natural_test.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v10_sexflag.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v10b_sexflag_bulk.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v10b_sexflag_bulk.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v11_draw_donors.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v11_draw_donors.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v12_nk_depth_by_sex.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v12_nk_depth_by_sex.txt", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v13_labeller_validation.py", + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/verify/SEP-CLIN-6_calib/v13_labeller_validation.txt" + ] + } + ] + }, + { + "id": "SEP-CLIN-1", + "lens": "sep_clinical", + "final": "refuted", + "verdict_recompute": "qualified", + "verdict_calib": "refuted", + "synthesis": { + "rank": 18, + "theme": "ICU controls (refuted)", + "synthesis_strength": "refuted as a generated-cell finding", + "reportable_statement": "Not reportable. What survives is a bulk-level, already reported fact: adult ICU controls are lymphopenic vs HC (NK clearest; observed -1.24, p 0.017). 'ICU controls show the same MAIT and memory CD8 depletion as sepsis' and 'MAIT loss beyond T-cell loss' are generator artefacts.", + "refuted_subclaims": [ + "ICU = sepsis for MAIT and memory CD8", + "MAIT-specific loss (MAIT/T -1.06)", + "memory CD8 sepsis-vs-ICU direction" + ], + "carrying_draws": { + "MAIT ICU/cells.parquet and observed/cells.parquet. They match composition.parquet exactly: 0 mismatches over 21,350 rows. Every sample has 1,500 cells in every draw. Unless noted otherwise, models are OLS with HC3 errors on logit((n+0.5)/(N+1)) ~ group + age + sex.\n\n**Bulk-only adults, ICU control (17) vs HC (36)**\n- Five-draw mean fractions, ICU / HC / sepsis: MAIT 0.469 / 1.353 / 0.402 %; NK 3.64 / 9.56 / 4.43 %; memory CD8 11.55 / 20.51 / 13.99 %.\n- Mean-logit coefficients:\n - MAIT -1.285 (SE 0.192, p=2.35e-11, BH q over 50 features 5.9e-10)\n - NK -1.261 (SE 0.202, p=4.65e-10, q=4.6e-9)\n - CD56dim NK -1.417 (p=1.5e-9)\n - memory CD8 -0.556 (SE 0.141, p=8.3e-5, q=3.0e-4)\n - MAIT/T -1.061 (p=4.0e-8)\n- A stacked model with draw fixed effects and sample-clustered SE gives the same coefficients.\n- Per draw, d0 to d4 (quasi-binomial GLMs agree):\n - MAIT -1.60 / -1.86 / -1.20 / -0.83 / -0.93, all p≤2e-4\n - NK -0.80 / -2.62 / -1.29 / -1.40 / -0.19 (d4 p=0.20)\n - memory CD8 -0.44 / -0.90 / -0.70 / -0.04 / -0.71 (d3 p=0.80)\n\n**Bulk-only sepsis (68) vs ICU (17)**\n- MAIT -0.076, 95% CI [-0.45, +0.30], p=0.69\n- NK +0.21, p=0.29\n- memory CD8 +0.15, p=0.19 (MWU p=0.056; d4 p=0.016)\n\n**Confounder variants, MAIT coefficient (ICU vs HC)**\n- Unadjusted -1.12. Age quadratic -1.27. Age spline -1.24.\n- Age restrictions: 32–80 -1.31; 40–80 -1.13 (p=1.2e-8, 10 vs 33); 50–80 -1.10.\n- Age-sex matching: 1:1 -1.21; 1:2 -1.13.\n- Sex: males -1.35; females -1.19; sex_flag samples dropped -1.30.\n- Bulk QC: adding the multimapping fraction (the only imbalanced QC metric, p=5e-4) gives -1.12. Adding all 7 bulk QC metrics gives MAIT -1.04, NK -0.96, memory CD8 -0.38 (p=0.012).\n- Generated library -1.11. Adding the l1 Other logit -1.18. Adding the monocyte logit -1.32.\n- High-confidence MAIT labels only: ≥0.8 gives -1.51, ≥0.95 gives -1.65.\n- Leave-one-ICU-out range -1.39 to -1.23. Huber -1.31. Permutation p=2e-4 (the floor).\n- Age interaction: effect -1.61 at age 40 and -1.17 at age 65 (interaction p=0.11).\n\n**Calibration**\n- The calibration.csv rows recompute from raw cells (max |diff| 4e-6).\n- MAIT: ratio 0.53 / 0.53 / 0.55 / 0.55 / 0.55; Spearman 0.46 / 0.36 / 0.43 / 0.38 / 0.45.\n- MAIT sepsis-vs-ICU fraction difference, generated vs observed: -0.0023 / -0.0017 / -0.0030 / -0.0001 / -0.0009 vs -0.008 to -0.011.\n- Paired-test ICU vs HC, generated / observed coefficient by draw:\n\n| Feature | d0 | d1 | d2 | d3 | d4 |\n|---|---|---|---|---|---|\n| MAIT | -1.78 / -0.73 | -1.68 / -0.65 | -0.93 / -0.78 | -1.26 / -0.78 | -0.64 / -0.66 |\n| NK | -0.38 / -1.25 | -1.78 / -1.26 | -0.29 / -1.39 | -0.94 / -1.39 | -0.03 / -1.16 |\n| MAIT/T | -1.53 / -0.22 | -0.70 / -0.04 | -0.62 / -0.31 | -1.10 / -0.31 | -0.15 / -0.18 |\n\n- Gate: ratio 0.5–2, Spearman ≥0.3, contrast signs agree, and bulk-only p<0.05. Draws that carry the result:\n - MAIT: 0–4. With a magnitude check (generated/observed paired coefficient between 0.5 and 2), only 2, 3, 4 carry it; pooled over those three the effect is -0.99 [-1.37, -0.60].\n - NK: 0 and 3 (with the magnitude check, only 3).\n - memory CD8: 0, 1, 2, 4.\n\n**Where ICU sits between HC and sepsis**\nPosition on the logit scale = (ICU−HC)/(SEP−HC), where 0 is the HC level and 1 is the sepsis level.\n- MAIT, paired observed: 0.47 / 0.39 / 0.45 / 0.50 / 0.35.\n- MAIT, paired generated: 0.77 / 0.75 / 0.58 / 1.08 / 0.72.\n- MAIT, bulk-only generated: 0.74 / 1.00 / 0.99 / 0.66 / 1.09.\n- MAIT, bulk input: SLC4A10 0.56; MAIT4 score 0.72.\n- MAIT/T: observed 0.21; bulk SLC4A10−T 0.06; bulk-only generated 0.67–1.32.\n\n**Observed paired adults**\n- ICU (6) vs HC (7):\n - MAIT 1.74 vs 2.84 % (-0.78, p=0.056; MWU 0.29)\n - NK 6.5 vs 18.6 % (-1.39, p=0.017; MWU 0.014)\n - memory CD8 -0.60 (p=0.29)\n - MAIT/T 4.5 vs 5.4 % (-0.31, p=0.60)\n - memory CD8 / T -0.07 (p=0.86)\n- Sepsis (46) vs ICU: MAIT -0.61 (p=0.086); MAIT/T -0.85 (p=0.035); NK +0.17 (p=0.70); memory CD8 -0.22 (p=0.60).\n\n**Bulk input, bulk-only ICU vs HC (log2 CPM)**\n- SLC4A10 -1.42 (t=-3.9)\n- MAIT4 score -1.02 (t=-5.7)\n- NK score -1.88 (t=-7.3)\n- T score -0.91 (t=-4.4)\n- MAIT4−T -0.11 (p=0.65)\n- SLC4A10−T -0.51 (p=0.19)\n\n**Bulk-marker baseline**\nA single-marker regression fitted on the paired adults' observed cells, then applied to the bulk-only adults:\n- ICU vs HC: MAIT -1.06 / -1.09; NK -1.28; memory CD8 -0.32; MAIT/T -0.43 (p=0.19).\n- Sepsis vs ICU, memory CD8: -0.18.", + "issues": [ + "Partly holds: the ICU-control < HC reduction in generated MAIT, NK and memory CD8 reproduces exactly and survives every confounder control. It is bulk-derivable, though. A single-marker bulk baseline fitted to the paired adults' observed cells gives the same ICU-vs-HC effect for MAIT (-1.06 to -1.09 vs generated -1.28) and NK (-1.28 vs -1.26). For this contrast the generated cells add nothing beyond the bulk.", + "Not supported: the sub-claim that MAIT/T is lower (-1.06) and that the cells assign the loss to MAIT beyond general T-cell loss. The bulk input shows no MAIT-specific signal (MAIT4−T -0.11, p=0.65; SLC4A10−T -0.51, p=0.19; this proxy tracks observed MAIT/T at Spearman 0.81 in paired adults). Observed paired MAIT/T for ICU vs HC is -0.31 (p=0.60), so ICU MAIT/T sits near the HC level (position 0.21). On the same paired test samples the generator inflates this contrast 2x to 17x in d0–d3 (e.g. d0 -1.53 vs -0.22; d1 -0.70 vs -0.04).", + "Not supported for MAIT or memory CD8: 'same depletion as sepsis' / 'none of the three differs from sepsis'. On the same paired test samples, the generator puts ICU closer to sepsis than the observed cells do in 5/5 draws (position 0.58–1.08 generated vs 0.35–0.50 observed). The sepsis-vs-ICU MAIT fraction difference is under-generated in 5/5 draws (1–34% of observed). The null has n=17 and a CI of [-0.45, +0.30], which excludes the observed paired estimate of -0.61. Non-significance here is not equivalence.", + "Memory CD8, sepsis vs ICU: the generated direction (+0.15, sepsis higher) is opposite to the observed paired cells (-0.22), the bulk CD8 genes (-0.41) and the bulk baseline (-0.18). In calibration.csv the sepsis-vs-ICU sign fails in d2, d3 and d4.", + "NK: ICU close to sepsis is consistent with the bulk input (position 1.02) and the observed cells (1.24), but no draw reproduces the paired sepsis-vs-ICU NK contrast (0/5). The ICU=0.9) = 0.13. HC/ICU fold 1.63.\n- MAIT/T 5.4 / 4.5 / 1.9%. ICU vs HC p 0.59. Sepsis vs ICU -0.83 (p 0.042, permutation 0.018).\n- MAIT/lymphoid 3.52 / 3.45 / 1.21%. Sepsis vs ICU p 0.033 (permutation 0.010).\n- NK 18.6 / 6.5 / 8.3%. ICU vs HC -1.24 (p 0.017, exact permutation 0.012); ICU is at the sepsis level.\n- Memory CD8 15.8 / 11.2 / 8.9%; no comparison is significant.\n- Lymphoid 81.7 / 57.0 / 73.0%. ICU vs HC p 0.002; ICU is below sepsis (p 0.003).\n\nBulk (log2 CPM):\n- SLC4A10, ICU vs HC: bulk-only -1.45 (p 4e-5), PI 0.90. In the paired set PI is 0.64.\n- NK-specific genes, bulk-only ICU vs HC: -1.87, PI 1.00.\n- SLC4A10 minus T-cell genes: bulk-only -0.40 (p 0.26); paired -0.10.\n\nBulk-implied effects: the line observed MAIT logit ~ SLC4A10 in paired adults has slope 0.75, R2 0.65 and no ICU offset. Applied to the bulk-only bulk it gives:\n- MAIT ICU vs HC -1.08 (generated -1.24); sepsis vs ICU -0.11 (generated -0.03).\n- NK -1.27 (generated -1.32).\n- MAIT/T -0.33 (generated -1.01).\n- MAIT/lymphoid -0.54 (generated -1.02).", + "issues": [ + "ICU vs HC calibration (I recomputed it because calibration.csv does not have this contrast): the MAIT sign agrees in 5/5 draws, but the size is inflated 2.4x in draws 0 and 1 and 1.8x in draw 3. Only draws 2, 3 and 4 carry ICUsepsis) = 0.92. Sepsis is lower than ICU for MAIT/T (p 0.042) and MAIT/lymphoid (p 0.033). The observed ICU-vs-HC fold is 1.6, not about 3. ICU controls are also more lymphopenic overall than sepsis (lymphoid 57% vs 73%, p 0.003), so 'same lymphoid depletion' is wrong in both directions depending on the compartment.", + "The cohort paper (Ye 2026) already shows part of this and contradicts the MAIT part. Extended Data Fig. 1e (PDF p.23, Dunn tests): NK cells are lower in ICU controls than HC (**). Fig. 3d (PDF p.7 = journal p.156): the MAIT cluster T_CD8_c07-TRAV1-2 is not significantly lower in ICU controls than HC, and is lower in urinary sepsis than in ICU controls (*). Extended Data Fig. 10d/g (PDF p.41; text p.163): MOFA factor 3, with MAIT among its top-10 loadings, is reduced in ICU controls (**) but sits between HC and sepsis. Methods, PDF p.17: ICU controls are 'critically ill patients admitted to the ICU without sepsis or bacteremia'.", + "Prior literature: Grimaldi 2014 (Intensive Care Med 40:192) reports MAIT counts of 4.3/mm3 in non-infected critically ill patients vs 31.1/mm3 in healthy subjects, and 1.3/mm3 in severe bacterial infection (P<0.0001 for all comparisons). So MAIT loss in ICU controls is known, and it contradicts 'ICU = sepsis'. Forel 2012 (PLoS One 7:e50446): absolute NK counts were reduced similarly in sepsis and non-septic SIRS (69 and 111 vs 203/mm3), with a normal NK share of lymphocytes.", + "Nothing beyond the bulk for the parts that hold. The bulk-only bulk itself shows SLC4A10 -1.45 log2 in ICU (p 4e-5), with ICU near sepsis (PI 0.90), and NK-specific genes -1.87 (PI 1.00). Through the observed paired calibration line, the bulk implies MAIT -1.08 (generated -1.24) and NK -1.27 (generated -1.32). The generated selective MAIT losses (MAIT/T and MAIT/lymphoid, about -1.0) are 2-3x the bulk-implied -0.33 and -0.54.", + "NK and memory CD8 calibration for this contrast is weak. For NK only draw 3 passes both the calibration.csv gate and the within-2x check. Draws 2 and 4 reproduce at most 14% of the observed paired effect (-0.17 and +0.07 vs -1.25 and -1.11), and the generator understates ICU NK depletion in 4/5 draws. Memory CD8 ICU vs HC is not significant in observed cells (p 0.24) and its share of T cells is normal (26.0% vs 27.1%). The generated memory CD8 carries the same ICU offset at fixed bulk (-0.47 to -0.83 vs about 0 observed).", + "Honest caveat: in the bulk, the bulk-only ICU group looks sicker than the 6 paired ICU controls (SLC4A10 PI 0.90 vs 0.64; granulocyte genes +1.15 vs +0.29 log2). The one draw whose paired PI is unbiased (d2: 0.54 vs 0.53) also places the bulk-only ICU controls at the sepsis level (PI 1.06). So MAIT as a share of PBMC may truly be near sepsis in the bulk-only ICU group. That is a bulk-level fact the cells only restate, and observed cells exist for only 6 ICU controls." + ], + "reasoning": "I call this refuted as a generated-cell finding. What survives is narrower, already visible in the bulk, and already reported.\n\n1. Calibration. MAIT passes the feature-level gate (ratio 0.52-0.55, Spearman 0.36-0.46) and the ICU HC (+0.54, +1.75). Draws 2-4, trained with paediatric or ICU controls (PIC01, PHC05, IC23), give the reverse (-0.56, -0.55, -0.20).", + "ISG-high monocytes, sepsis vs HC: observed paired sepsis is below HC after adjustment (-0.85, p=0.031; rank-ANCOVA p=8e-5), as Ye 2026 p.6 also reports for IFITM1 monocytes. The adjusted paired-test generated coefficient has the wrong sign in 5/5 draws, and bulk-only shows no difference (-0.05, q=0.76). Calibration fails for this contrast.", + "Sepsis vs ICU control: apart from ISG-high monocytes, no generated feature passes BH. Observed paired differences are not reproduced (all q>=0.62): cycling monocytes (q<1e-4), C1Q-high monocytes (q=0.027) and the CD14-monocyte share (q=1e-4). Cycling and C1Q-high monocytes are under-generated (ratio 0.05-0.11 and 0.05-0.52).", + "Strong sepsis-vs-HC and ICU-vs-HC hits in poorly calibrated features are artifact-suspect: ILC, progenitor, HSPC, erythrocyte, plasma cell/plasmablast and the platelet-in-monocyte negative control (q<1e-10). ILC, progenitor and HSPC have Spearman <=0.26; plasmablast markers have ratio 0.09-0.24. Generated HSPC/progenitors go up while the bulk HSPC score (CD34, PRSS57, SPINK2, SMIM24) goes down (t=-3.8).", + "The negative control 'Platelet signal in monocytes' is significant for ICU vs HC (q<1e-4), sepsis vs HC, age (q=0.03) and abdominal source (q=0.067). It follows bulk platelet content, and its calibration fails (Spearman 0.15-0.23). Some generated state differences reflect bulk contaminant signal.", + "Effector B decline with age (q=0.024) fails calibration: the paired age-slope sign agrees in only 1/5 draws, and observed cells go the opposite way (+0.008 per year).", + "MS1 increase with age (q=0.058) fails calibration: the paired age-slope sign agrees in 0/5 draws, and observed is +0.0115 (p=0.56). It is visible only as a bulk IL1R2 age trend (q=0.004)." + ], + "notes": "Scope. Sepsis clinical lens, bulk-only set. The 121 adult bulk-only samples (68 sepsis, 36 HC, 17 ICU) were analysed; the single paediatric ICU control was excluded. Features: 50 in total (6 l1, 29 l2 composition types, 15 states). Comparisons: sepsis vs HC, sepsis vs ICU, ICU vs HC, the 4-source omnibus plus one-vs-rest within sepsis, age (within sepsis, within controls, all adults) and sex (all adults, within sepsis).\n\nStatistics. The main model is OLS with HC3 errors on logit((n+0.5)/(N+1)), adjusted for age and sex, with group added where relevant. It was fitted per draw and on the per-sample mean of the 5 draw logits; the latter equals a stacked model with draw as a fixed effect. Supporting fits: arcsine-sqrt, rank-ANCOVA, and unadjusted Mann-Whitney / Kruskal-Wallis / Spearman. BH was applied across the 50 features within each comparison. The five-draw mean from bulkonly.parquet is reported only as a summary. Sensitivity checks: dropping the sex-flag samples, dropping the Emergency-Center samples, a site covariate, an age 40-80 window, and 1:1 age/sex matching. ICU controls are younger (mean 49 vs 64 years).\n\nCalibration gate. A draw passes for a feature and comparison when four conditions hold: the same-model coefficient on its paired test samples has the right sign; its calibration.csv contrast sign agrees, where that contrast exists; its Spearman is at least 0.3; and its ratio is between 0.5 and 2. A draw carries a result when it passes the gate and its own bulk-only effect has the same sign with p<0.05. ICU vs HC, age, sex and source have no calibration.csv contrast, so they were calibrated by fitting the same model to each draw's paired test samples, generated vs observed (s03).\n\nObserved concordance. The same models were fitted to the 59 adult paired samples. Composition and label states used the observed set; marker states used both native and observed_thin_d cells. The paired control groups are small (7 HC, 6 ICU), so most observed checks are directional only.\n\nBulk check. Marker scores and single-gene CPM from bulkonly_raw.npz and the paired conditions_raw_fixed.npz, plus cell-level models adjusted for the matching bulk score.\n\nMain message. The extra bulk-only controls mainly power control contrasts. ICU controls share sepsis-like MAIT/NK/memory-CD8 depletion; this is mostly visible in the bulk, with all draws calibrated for MAIT. ICU controls also have the lowest ISG-high monocytes; this is partly beyond the crude bulk ISG score and calibrated in draws 0, 3 and 4. Infection source, sex and age give nothing that survives calibration and goes beyond the bulk. The generator attenuates canonical ageing signals (naive CD8, MAIT).\n\nArtifact warning. The striking male excess of 'Proliferating NK' labels also appears in observed cells, but it is a label artifact. The labeller uses chrY genes and lacks MKI67/TOP2A, and marker-defined cycling NK cells show no sex difference.\n\nCaveats. Interpretations involving the ICU group could be confounded by ICU diagnosis, medication (e.g. steroids) and sequencing batch, none of which is available.\n\nHousekeeping. CPU only, run in memory-limited user scopes; no GPU was used. Writes went only to the lens dir. Paper-archive policy (pbisc-diffusion): this subagent published no archive delta, because its write scope is the lens dir. The coordinator may register analysis/wf/sep_clinical as analysis evidence if any finding enters the paper. Literature page numbers come from s09_pdf_pages.txt (PDF pages, not journal pages). Lens dir: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_clinical (scripts s00-s12, shared helpers s_common.py, quoted numbers collected in s10_evidence.txt)." + }, + { + "lens": "sep_populations", + "agent_id": "a96530c9b5aa6d0d3", + "negative_results": [ + "A plain joint embedding (log-normalised, batch-aware HVGs, balanced PCA, kNN) separates observed from generated cells almost completely: the generated-neighbour share of observed cells is 0.10-0.42 vs 0.92 if mixed, and several Leiden clusters are >98% observed. The generator almost never emits about 270-410 genes per lineage that are detected in >=5% of observed cells (<1% of generated cells). Co-clustering therefore needs per-domain standardisation on genes detected in >=5% of both domains. After that, the generated-neighbour share of observed cells is 0.76-0.85 vs 0.89-0.94 expected. So co-clustering is evidence of matching relative position, not of identical expression (s01, s02, s03 tables).", + "IL1B+ CXCL8+ CCL3+ inflammatory CD14 monocytes (Ye Mono_c02-CD14-IL1B; Monocyte r2.0 c12). Bulk-only sepsis > HC in draws 0, 2, 4 (q<=0.009), but the observed paired contrast is not significant (+0.080, p=0.66). The generator also under-expresses IL1B in this state (lfc +0.72 vs +2.00 observed). Not a finding.", + "ISG-high monocytes (Ye Mono_c05-CD14-IFITM1-like; Monocyte r2.0 c9). Significant only in draw 0 (q=0.008); observed +0.040 (p=0.39). Ye's reported reduction is reproduced in neither domain.", + "VCAN/CSF3R classical monocytes (Monocyte r2.0 c3). Bulk-only lower in sepsis in 5/5 draws (q<0.05 in 4) with the gate passing 5/5, but observed paired is not significant (-0.036, p=0.26; vs all controls -0.019, p=0.39). The association exists only in generated cells.", + "Platelet-complexed monocytes (PPBP/PF4/NRGN; Monocyte r1.0 c11). Bulk-only higher in sepsis in 5/5 draws (q<0.05 in d1-d4), but calibration Spearman is only 0.01-0.33 (gate passes d1, d2, d4), the generated share is 1.5-2x observed, and observed paired is not significant (+0.014, p=0.38). Artefact-suspect.", + "Sepsis vs adult ICU controls: no candidate state is significant in bulk-only cells in a calibrated draw. The observed paired ICU contrasts have only 6 ICU samples and are mostly not significant, except the KLRB1+GZMK+ CD8/MAIT-like state (-0.068, p=0.039). For the MS1-like and HLA-II-high states the ICU contrast sign fails calibration in most draws. Every reported effect separates sepsis and critical illness from healthy controls, not sepsis from ICU controls.", + "Observed-only granulocyte-like cluster (Monocyte r1.0 c18: MMP9, IL1R2, CSF3R, ALPL, CYP4F3, NAMPT): 65 observed cells (2.0% of observed monocyte-gated cells) and 0 generated cells. It is a generator blind spot, consistent with calibration.csv LDG ratio 0.01-0.09. In observed cells it appears in 3/46 adult sepsis and 0/7 HC samples (p=0.51, 200-cell subsample).", + "Draw-specific (imprinted) monocyte clusters: r1.0 c9 (TMEM176B/FCN1/PLBD1; 31.7% of draw-1 bulk-only monocytes vs <=2% in other draws) and c13 (FCGR3A/MS4A7; 16.8% vs <=0.4%), and r2.0 c10, c20, c21, c22 (max-draw fraction 0.69-0.94). They come from draw 1's training donors (observed c13 fractions 0.34 and 0.10 in two of them). Not reportable for bulk-only inference.", + "Generated-only micro-clusters at r=1.0: 71 clusters of <=30 cells with no observed cell (8 monocyte, 24 T, 15 NK, 24 B), holding 592 generated cells (0.15-1.3% of generated cells per lineage). Their markers are long, low-expressed genes (RPRD2, RICTOR, PDE4D, NFKB1 ...). Artefact-suspect noise; none was tested.", + "Library-size-driven clusters with few markers, present in both domains: Monocyte r1.0 c2 / r2.0 c2 (AUC of log counts 0.05), r2.0 c1 and c11, T r1.0 c8 (0.10), NK r1.0 c15 (0.13), NK r2.0 c1. Not populations.", + "Doublet-like mixed-lineage clusters are over-produced by the generator: monocytes with T/NK genes (Monocyte r1.0 c5: 13% of generated vs 7.5% of observed monocytes), T cells with S100A8/S100A9/LYZ/HLA-II (T r1.0 c10: 4.6-4.9% vs 2.8%), NK cells with myeloid genes (NK r1.0 c22: 2.9-3.1% vs 1.0%).", + "No cycling MKI67+ STMN1+ MPO+ ELANE+ monocyte cluster in bulk-only (adult) generated cells. This is consistent with Ye 2026, where Mono_c06-CD14-MKI67 is paediatric-only (PDF p.10 = printed p.159), and with the calibration ratio of 0.05-0.11 for the cycling monocyte state.", + "C1Q+/HLA-II+ FCGR3A+ monocytes (COMBAT reports C1Q+ non-classical monocytes rising in sepsis; Monocyte r2.0 c16): no consistent effect (q<0.05 in 2/5 draws; observed -0.003, p=0.60).", + "The immediate-early (AP-1/NR4A2) states in CD4 T, CD8 T, NK and B cells must not be counted as independent populations. In observed cells, per-sample immediate-early scores correlate 0.70-0.89 across lineages. Their generated level is imprinted by the training donors (draw 1 makes almost none). Ex vivo handling cannot be excluded." + ], + "notes": "Method (all CPU, MemoryMax=6G; MemAvailable checked >=48 GiB before every embedding step; no GPU work).\n\n1. The 200-cell subsamples were pooled: observed 16,400 cells, paired_gen 5 x 15,400, bulkonly 5 x 24,400; 215,400 cells in total.\n2. Within each six-type parent (Monocyte, T, NK, B), a joint embedding was built:\n - seurat_v3 HVGs (batch = set), restricted to genes detected in >=5% of both observed and paired_gen cells (709-794 genes);\n - per-domain z-scoring: observed cells with observed statistics; every generated cell with pooled paired_gen statistics (matched samples, so bulk-only vs paired differences are kept);\n - PCA fit on a balanced set, kNN 15, Leiden via igraph (r = 0.5 / 1 / 2).\n3. Observed cells were assigned to clusters by co-clustering, and also by majority vote of their 15 nearest generated neighbours.\n4. Tests:\n - per-sample state fraction (cluster / parent, >=10 parent cells) and abundance;\n - two-sided MWU per draw;\n - BH across the 143 cluster features (r=1.0 and r=2.0 clusters with >=150 cells) within each draw x contrast x measure.\n5. Gate: each draw was gated by its own paired-test calibration row for the new feature. The gate requires the same contrast sign, Spearman >=0.2 and a ratio between 1/3 and 3. My cluster features are not in calibration.csv, so equivalent rows were computed in s05_calibration.csv; related calibration.csv rows are quoted.\n6. Beyond bulk: 20-marker bulk scores from bulkonly_raw / conditions_raw_fixed, rank-OLS of observed fraction on bulk score plus generated fraction, and within-sepsis tracking.\n\nMain caveats:\n- Calibration contrasts rest on only 6-7 paired HC and 6 ICU samples.\n- Observed state fractions come from 200 cells per sample.\n- Co-clustering needed per-domain standardisation.\n- Every positive finding separates sepsis or critical illness from healthy controls; none is sepsis-specific vs ICU controls.\n- Three of the six findings are reproductions of Ye 2026 subsets: MS1/RETN, HLA-DRA, GNLY vs GNLY-CCL4hi CD8. A fourth (NR4A2+ CD4) matches Ye's T_CD4_c02 and adds the cross-lineage, donor-imprinted reading.\n- The only clear 'beyond bulk' gains are for the immediate-early CD4/CD8/NK/B states. Those states are one sample-level program that may be ex vivo handling.\n\nLarge intermediates (PCs) are in the scratchpad and can be regenerated with s01/s03.\n\nFigure: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/figures/s14_forest.png. Scripts s00-s14 plus common_sp.py and embed_lib.py, tables/ and logs/ are all in /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_populations/.\n\nThis subagent published no archive delta (PBISC-Diffusion POLICY.md task-end update); the coordinator should register the lens outputs if they enter paper evidence." + }, + { + "lens": "sep_programs", + "agent_id": "a46ace97f8183869e", + "negative_results": [ + "Interferon (ISG) program in CD14 monocytes is not reproduced. Observed paired monocytes fall in sepsis vs HC: −1.38 log2, p 0.001; all 7 HC score 4.85–6.90 vs a sepsis median of 3.78 (isg_check_observed.csv). This agrees with Ye 2026's reduced Mono_c05-CD14-IFITM1, and the bulk shows it too: bulk-only −0.34 raw, −1.29 relative to monocyte markers. Generated monocytes do not fall: bulk-only mean +0.06, positive in 4/5 draws. Paired-test calibration FAILs in 5/5 draws (generated +0.05 to +0.87 vs observed −1.23 to −1.64), while on the same samples a monocyte-normalised bulk gets −1.18 to −1.51. The cross-type ISG pattern correlates negatively in 3/5 draws. Draw 4's bulk-only −0.72 is not supported by its own paired test (+0.05).", + "T-cell exhaustion (LAG3, HAVCR2, TIGIT, CTLA4, TOX; PDCD1 is not on the panel) is not reproduced. Generated cells show decreases (memory CD8 −0.13, Treg −0.24 log2) where observed cells show small increases (+0.21, p 0.05; +0.13, ns). Calibration FAILs in draws 3–4 for memory CD8, and the cross-type r is −0.71 to −0.87 in draws 1, 3, 4. A bulk/T-marker ratio (+0.32 to +0.34 on the paired test samples) points the way the observed cells do.", + "T activation (CD38, HLA-DRA, HLA-DRB1, MKI67, ICOS) is not reportable. Generated cells raise it in every T subset (memory CD4 +0.46, naive CD4 +0.36, naive CD8 +0.37, Treg +0.23) where observed cells show nothing (−0.05 to −0.17). Only memory CD8 matches (+0.27 vs +0.47; PASS in draws 1 and 4 only), and with different driver genes: observed CD38/MKI67, generated ICOS/HLA-DR.", + "Sepsis vs adult ICU control: within-type logFC calibration fails for every AIFI-L2 type in every draw (51 FAIL, 9 WEAK, 0 PASS; r −0.04 to 0.22; top-100 sign agreement 0.35–0.80). Only the six-type T cell passes (draws 0, 2, 3, 4), and even there no gene is significant in all carrying draws. Program-level ICU contrasts are UNINFORMATIVE: observed paired-test p >= 0.1 with 5–6 ICU controls per draw. No within-type sepsis-vs-ICU finding is reportable.", + "Within-plasma-cell DE cannot be calibrated: only 2–16 paired test samples per draw have >=10 generated plasma cells, none of them HC. Bulk-only sepsis vs HC within plasma cells cannot be tested because too few HC have >=10 plasma cells.", + "Attribution of the bulk-decreased genes fails. Observed paired cells place the 30 most decreased bulk genes (broadly expressed lymphocyte genes) in NK-cell loss (29/30, composition-driven). Generated cells place them in CD4/CD8 T cells: 1–11 of 30 match observed, vs 11/30 for the reference atlas. They are significantly worse than the reference in draws 1, 2, 4 (McNemar p 0.002–0.008), and the result holds at top 50/100. This fits the composition calibration: generated NK loss is too small (l2 | CD56dim NK sepsis−HC −0.0002 to −0.067 vs observed −0.071 to −0.089) and memory CD8 loss too large (−0.048 to −0.196 vs −0.067 to −0.083).", + "MKI67 attribution fails. The observed bulk rise (+1.11 log2) is carried by plasma cells/plasmablasts (7.2 of 12.1 CPM). Generated cells assign it to proliferating T cells, or show no rise at all in draws 0 and 3 (−0.05, −0.08 log2).", + "Draw 1 fails broadly. Lymphocyte types are only WEAK/FAIL (r 0.09–0.29), the CCL/TNF and NR4A programs are near zero, and attribution matches observed for only 7 of 30 up genes. Draw 1 over-labels intermediate monocytes (calibration.csv l2 | Intermediate monocyte: ratio 7.55, Spearman 0.06, wrong sepsis−HC sign).", + "Monocyte within-type DE adds little beyond the bulk. On paired test samples, generated CD14 logFC tracks the observed logFC less well than the bulk does (r 0.28–0.47 vs bulk 0.66–0.74). On observed-significant CD14 genes, a monocyte-normalised bulk gets the within-cell direction right more often (0.93 vs 0.89). It also recovers 40 of the 45 CD14 genes that fall within monocytes while rising in the bulk.", + "Per-sample levels of monocyte programs: the whole-PBMC bulk ranks samples better than the generated cells. Spearman with the observed score, bulk vs generated: HLA-II 0.68–0.78 vs 0.51–0.64; MS1 0.83–0.88 vs 0.62–0.73; S100 0.71–0.79 vs 0.22–0.60; ISG 0.89 vs 0.36 (mean).", + "The NR4A2 'CD4 exhaustion' program (Ye 2026) is not CD4-specific. Observed cells show it rising in every lymphocyte type (+1.4 to +2.2), and the generated cells reproduce that broad pattern, so no CD4-specific claim is possible. The paired bulk of the same samples does not rise (−0.09).", + "IL1B inflammatory monocyte program (IL1B, CXCL8, CCL3, CCL4) cannot be calibrated. Generated CD14 monocytes rise +0.89 log2 (5/5 draws), but the observed paired effect is not significant (+0.55, p 0.33). Not reportable.", + "S100 cross-type pattern and subset ordering are wrong. Observed S100 also rises in T/B pseudobulks (probably ambient RNA), and observed CD14 rises more than CD16 (+1.55 vs +0.89). Generated cells show CD16 > CD14 (+1.50 vs +0.71); the cross-type r is about 0." + ], + "notes": "Lens dir: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/sep_programs/. Every script is kept there (lens_common.py; s01, s01b, s01c, s02, s03, s04, s04b, s04c, s04d, s05–s12), together with every table quoted.\n\nRun conditions: CPU only, under systemd-run (MemoryMax=6G, MemorySwapMax=0, CUDA_VISIBLE_DEVICES='', 2 threads). No subsample.npz was loaded; no GPU work.\n\nMethod:\n- Per-sample × type pseudobulks from pb_l2.npz / pb_l1.npz, on the 5,975 measured panel genes.\n- edgeR-style log2 CPM with prior.count 2; a pseudobulk is used only with >=10 cells. A 20-cell minimum was rerun as a sensitivity check; all calls were unchanged.\n- Genes per type are those expressed in observed cells of that type (pooled CPM >= 10).\n- Every contrast is tested within a draw (Welch; BH within draw × type × contrast for genes, and within draw over all program tests). Draws are combined by sign count and mean effect.\n- Each draw is gated by its own paired-test calibration. Gates were written before the results were seen:\n - Type-level within-type logFC: PASS needs r >= 0.30 and top-100 sign agreement >= 0.70.\n - Program level: PASS needs an informative observed effect (p < 0.10), the same sign, Spearman >= 0.2 and a generated/observed ratio between 0.25 and 4.\n\nCarrying draws (type-level PASS, sepsis vs HC):\n- CD14 monocytes 0,1,3,4; CD16 monocytes 1,3,4; intermediate monocytes 0,1,3.\n- Memory CD4 0,2,3,4; CD56dim NK 0,2,3,4.\n- Memory CD8, naive CD4, naive CD8, Treg, gdT and naive B 0,3,4; memory B 0,4.\n- Plasma cells cannot be calibrated.\n\nDraws 0, 3 and 4 carry most lymphocyte results. Draw 1 carries monocytes only.\n\nAdded after reading the cohort paper (Ye 2026, Nat Immunol 27:150), before their generated results were seen: the Ye 2026 programs (NR4A2; CCL3/CCL4/TNF/IFNG; IL1B monocyte), the coarse-lineage attribution check, the top-50/100 attribution lists, the bulk-ratio baselines, and the immediate-early check.\n\nCaveats:\n1. Each draw's paired test set has only 6–7 adult HC and 5–6 ICU controls, so observed contrasts are noisy. The observed 'truth' is the paired cohort, which shares about 94% of its samples with each test set, so for programs the calibration and the observed-concordance items are not independent. The cross-cohort comparisons (bulk-only generated vs observed paired) are the independent checks.\n2. Generated magnitudes are attenuated for most features (about 0.4–0.8× for HLA-II, and only 0.03–0.14× for MS1 abundance outside draw 0). Absolute levels were never compared across draws.\n3. The main extra content of the generated cells beyond the bulk is localisation: monocyte attribution of bulk-up genes, a type-specific response pattern, and a within-lymphocyte direction. Much of the lymphocyte advantage is single-cell-specific signal that the paired bulk does not show (immediate-early/activation genes). This makes lymphocyte 'activation' results artifact-suspect, even where they match the cohort paper.\n4. Sepsis vs ICU control is not reportable for within-type programs.\n\nArchive: this subagent may write only in the lens dir, so it did not run the PBISC-Diffusion archive update. There is no archive delta from this lens; the coordinating agent decides whether to register these tables." + } + ], + "provenance": { + "source": "/home/sj_server_1/.claude/projects/-home-sj-server-1/cd3948fe-c9dc-46c9-874d-211078d8ba7a/subagents/workflows/wf_e137f359-578/journal.jsonl", + "source_sha256": "7b072b9239e487bee767b35c3f912680c52b5015abeb262e4c4668cbb26946e7", + "workflow": "wf_e137f359-578", + "n_journal_rows": 80, + "note": "verbatim discover results (findings, negative_results, notes) and verify results", + "scripts": [ + "synthesis_ye2026_sepsis/y00_counts.py", + "synthesis_ye2026_sepsis/y01_calibration_summary.py", + "synthesis_ye2026_sepsis/y02_extract_graded.py", + "synthesis_ye2026_sepsis/y03_build_findings_json.py" + ], + "gpu": "none (CPU only, systemd-run MemoryMax=6G)", + "archive": "no archive delta from this synthesis agent; registration under PBISC-Diffusion POLICY.md is left to the coordinator" + } +} \ No newline at end of file