Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261006T033406Z-bulkonly-lora5-k1000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_20261006/analysis/DATA_GUIDE.md

@@ -0,0 +1,171 @@
+# Data guide for the bulk-only LoRA-5 analysis (2026-10-06)
+
+Plan root: `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006` (= `P`).
+Cohorts (`DS`): `ye2026_sepsis` and `imperial_eskd`.
+
+## What was generated
+
+- Bulk-only samples are bulk RNA-seq samples that have no single-cell partner.
+  - Ye 2026 sepsis: 122 samples (68 adult sepsis, 36 adult HC, 17 adult ICU control, 1 paediatric ICU control).
+  - Imperial ESKD COVID: 164 samples (143 COVID+, 19 COVID−, 2 recovery). The cohort is longitudinal, with 45 donors
+    that have no single cells at all.
+- "LoRA-5" models: each few-pair draw `j = 0..4` is the final PBISC model plus LoRA r16, fine-tuned on the pairs of
+  5 donors (3,000 steps). The 5 draws use 5 different donor sets.
+  - Each draw generated 1,500 cells per bulk-only sample from that sample's real bulk.
+  - Generation used K = 1,000, seed 20260927 and the benchmark's sampler flags.
+- Generated cells were read out once with Poisson (seed 20260916) and labelled with the benchmark's frozen
+  f01common scANVI model:
+  - `celltype_l1_pred`: the six-type gate (T cell, B cell, Monocyte, NK cell, DC, Other);
+  - AIFI L1 / L2 labels (29 L2 types) and the L2 confidence.
+
+## Truth for calibration (paired samples)
+
+- Each draw also generated cells for its paired TEST samples. These are the usable paired samples of the donors it
+  did not train on: sepsis 77 per draw; Imperial about 142 per draw, which includes other visits of non-training
+  donors.
+- Those samples have observed single cells, so every feature can be checked there: generated vs observed.
+- An observed set contains every usable paired sample: sepsis 82, Imperial 157.
+
+## Files (all CPU-readable, small)
+
+`P/analysis/features/DS/`
+
+- `meta.csv`: one row per sample.
+  - `set` is `paired` (observed sc exists) or `bulkonly`. Other columns: `sample_id`, `donor_id`, `group`, `case`.
+  - Clinical columns, sepsis: `age`, `sex`, `infection_source`, `site`.
+  - Clinical columns, Imperial:
+    - `severity_peak` (mild/moderate vs severe/critical over the illness);
+    - `severity_at_sample` (WHO at the sample);
+    - `days_since_pos_swab`;
+    - `fatal`, `age`, `sex`, `wave`.
+  - `sex_flag`: the bulk sex call disagrees with the metadata, i.e. a possible sample swap. Exclude these in sensitivity
+    analyses.
+  - `paired_test_d<j>`: the sample is a test sample of draw j.
+  - `donor_trained_d<j>` (Imperial): the sample's donor was one of draw j's five training donors. Its other visits
+    were seen in training.
+- `composition.parquet`: cell-type fractions per sample.
+  - Columns: `set`, `draw`, `sample_id`, `level`, `type`, `n`, `n_cells`, `frac`.
+  - `set` is `observed` (draw −1), `paired_gen` (draw j, test samples only) or `bulkonly` (draw j).
+  - `level` is `l1` (the six-type gate) or `l2` (AIFI L2).
+- `states.parquet`: rare / disease states from marker rules.
+  - Columns: `set`, `draw`, `sample_id`, `state`, `parent_n`, `state_n`, `frac`, `thin`.
+  - The rules are in `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/rare_states_screen_20261004/tools/states.json`.
+    A marker rule is ≥ k of the genes detected in a parent type; a label rule is AIFI L2 labels.
+  - `observed_thin_d<j>`: observed cells binomially thinned to draw j's generated parent median library. Compare
+    marker states generated vs observed only through this set, because generated cells are shallower.
+- `calibration.csv`: per feature and draw, plus `draw == "pooled"`. Over each draw's paired TEST samples it gives:
+  - `gen_mean`, `obs_mean`, `ratio` (generated / observed), `spearman` (over samples);
+  - per contrast, `gen diff` / `obs diff` and whether the signs agree.
+  - Contrasts, sepsis: adult sepsis vs adult HC; adult sepsis vs adult ICU control; sepsis vs all controls.
+  - Contrasts, Imperial: COVID+ vs COVID−; severe/critical vs mild/moderate among COVID+.
+  - **This is the trust layer. A bulk-only pattern in a feature whose calibration fails, i.e. wrong contrast
+    direction, near-zero Spearman or ratio far from 1, is not a finding.**
+- `bulkonly.parquet`: per bulk-only sample and feature, the mean, SD, min and max over the five draws.
+  - Feature names: `l1 | <type>`, `l2 | <type>`, `state | <name>`.
+
+`P/work/DS/{bulkonly/draw<j>, paired_gen/draw<j>, observed}/`
+
+- `cells.parquet`: one row per cell.
+  - Columns: `sample_id`, `celltype_l1_pred`, `aifi_l1`, `aifi_l2`, `aifi_l2_confidence`, `lib_measured`, `lib_score`.
+- `markers.npz`: per-cell counts of the state marker genes (`markers`, `marker_symbols`).
+- `pb_l2.npz` / `pb_l1.npz`: per sample × cell type summed integer counts on the 6k panel.
+  - Keys: `sums [samples, types, genes]`, `n_cells`, `samples`, `types`, `gene_ids`, `gene_symbols`.
+  - Use them for within-cell-type pseudobulk DE. Normalise per sample × type (CPM or similar) and require enough cells.
+- `subsample.npz`: 200 cells per sample as CSR integer counts on the 6k panel (`data`, `indices`, `indptr`, `shape`).
+  - `cell_row` = row index into that set's `cells.parquet`. Use it for embeddings and clustering.
+- `RECEIPT.json`: provenance and sha256.
+
+Depth caveat: generated cells have a different library size than observed cells. Use per-cell or per-pseudobulk
+normalisation. For marker states, compare only through `observed_thin_d<j>`.
+
+## Draw heterogeneity (read before any analysis)
+
+The five draws are five different 5-donor LoRA models. Each imprints its training donors' composition and states, so
+absolute levels and even contrast magnitudes differ a lot between draws.
+
+- **Composition.** Sepsis bulk-only monocytes make up 8 % to 38 % of cells depending on the draw (six-type labels
+  pooled over samples).
+- **MS1 monocytes, paired calibration.**
+  - Draw 0 reproduces the level: ratio 0.98, sepsis − HC difference 0.167 generated vs 0.151 observed. Its training
+    donors carry MS1 cells.
+  - Draws 1–4 generate MS1 at 0.03–0.14× of observed, though the sign is kept.
+- **Monocyte share.** The sepsis − HC difference is reproduced by draws 0–1, but is near zero or negative in draws 3–4.
+
+Therefore:
+
+1. Analyse every contrast **within each draw**: draw as a fixed effect, or a per-draw test followed by a sign / effect
+   meta-summary. Never compare absolute levels across draws.
+2. Weight or gate each draw by **its own** calibration row for that feature and contrast (`calibration.csv`, rows
+   `draw == j`). A bulk-only effect is reportable for a feature only in the draws that reproduce that feature's paired
+   contrast. Say which draws those are.
+3. The five-draw mean in `bulkonly.parquet` mixes calibrated and uncalibrated draws. Use it only as a summary, never
+   as the sole evidence.
+4. The draw-specific training donors are in
+   `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/DS/fewshot_n5/DRAWS.json`.
+   Use them to explain a draw's behaviour, for example draw 0's training donors are sepsis samples rich in MS1 cells.
+
+## Bulk inputs (for bulk-derivability checks)
+
+- Bulk-only condition: `P/data/DS/bulkonly_raw.npz`.
+  - Keys: `bulk_counts [samples, 20097 canonical genes]` (CPM-scaled condition counts), `sample_ids`, `gene_ids`.
+- Paired raw bulk: `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/DS/conditions_raw_fixed.npz`.
+  Same keys, every paired row, including the unusable ones.
+- Gene symbols: the canonical gene table
+  `/home/sj_server_1/gblee/datalake/paired-pbmc-bulk-sc/processing/strict_qc_v1_20260913/reference/canonical_genes.tsv`
+  (gene_id, gene_symbol).
+
+## Cohort papers (PDFs, readable with the Read tool; `pages` for long ones)
+
+Folder: `/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/` (see `INDEX.md`).
+
+- Ye 2026 sepsis: `paired-cohorts/Single-cell multi-omic landscape reveals anatomical-specific immune features in adult and pediatric sepsis.pdf`.
+- Imperial bulk: `paired-cohorts/Gisby2022_NatCommun_multiomics_LRRC15_COVID19_haemodialysis.pdf`.
+- Imperial single cell: `paired-cohorts/Stephenson2025_CellGenom_temporal_multiomics_COVID19_ESKD.pdf`.
+- MS1 / HLA-DR-low / COMBAT and others: `rare-disease-states/` (check `INDEX.md` for exact file names).
+
+## Rules for every analysis agent
+
+- **CPU only.** Never import torch with CUDA, never start GPU work; set `CUDA_VISIBLE_DEVICES=''`. This host's GPU is
+  shared and guarded.
+- **Memory.** Run Python as
+  `systemd-run --user --scope -q -p MemoryMax=6G -p MemorySwapMax=0 env CUDA_VISIBLE_DEVICES= OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 MKL_NUM_THREADS=2 <python> script.py`.
+  Before any step that loads `subsample.npz` of several sets or builds embeddings, check that `MemAvailable` (in
+  `/proc/meminfo`) is ≥ 48 GiB, and wait if it is lower.
+- **Python.**
+  - `/home/sj_server_1/miniforge3/envs/cxg_corpus/bin/python` has scanpy, statsmodels, sklearn, scipy, umap,
+    pynndescent, seaborn, but no leiden / igraph.
+  - For Leiden clustering use `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/immicro_scsemi_comparison_20260927/runtime_env/bin/python`,
+    which has scanpy and igraph: `sc.tl.leiden(..., flavor="igraph", n_iterations=2, directed=False)`.
+- **Writes.** Only under `P/analysis/wf/<your lens>/`. Save every number you report, with the script that made it.
+  Never modify other files.
+- **Statistics.**
+  - Sample-level tests; donor-aware for Imperial (repeated visits: mixed model or donor aggregation).
+  - BH correction across the features you test.
+  - Effects per draw: report in how many of the 5 draws the effect has the same sign.
+  - Sensitivity: drop `sex_flag` samples and, for Imperial, samples whose donor trained the draw.
+- **Honesty.** Generated cells are model predictions. A claim is only as strong as:
+  1. its paired calibration;
+  2. its replication across draws;
+  3. agreement with the paired observed cells, where the covariate exists there;
+  4. whether it says more than the bulk alone already says.
+  Report all four for every claim.
+
+## Lessons from the sepsis run (added 2026-10-06 after its completeness review; apply them to every new analysis)
+
+- **Shared noise across draws.** Every draw was sampled with seed 20260927, so cell slot c of sample i starts from
+  the same noise in all five draws. "k of 5 draws agree" is therefore not five independent replications. In sepsis,
+  shared noise explained 23–33 % of cross-draw agreement for rare types and under 5 % for major types. Say so whenever
+  draw agreement is used as evidence, and prefer per-draw calibration gates over agreement counts.
+- **Trust layer population.** `calibration.csv` uses every paired test sample of a draw. When an analysis is
+  restricted (adults only, COVID+ only, an illness window), recompute the calibration on the same restriction.
+- **Does calibration transfer?** Paired and bulk-only samples can differ within a group (sepsis HC: immediate-early
+  genes +2.4 log2 in paired vs bulk-only bulk). Before relying on calibration, compare the bulk profiles of paired
+  and bulk-only samples within each group for the genes of the feature.
+- **statsmodels bug.** In statsmodels 0.14.6, a GLM with a two-column count response and `scale='X2'` gets the
+  dispersion wrong (standard errors up to 18x too small). Use proportions with `var_weights`, a beta-binomial or a
+  rank model instead, and never quote those quasi-binomial p-values.
+- **Headline numbers** are gated per-draw estimates, never the five-draw mean of logits.
+- **Population ratios.** Z-scoring each domain separately before co-clustering pushes cluster-level ratios toward 1.
+  Confirm every population with an embedding-free marker definition and an independent re-clustering.
+- **Missing analyses to cover:** DC and NK compartments, and the bulk-signature baseline for every
+  outcome/severity association (does the generated feature add to a bulk score?).

paper-releases/pbisc-diffusion/updates/20261006T033406Z-bulkonly-lora5-k1000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_20261006/analysis/wf/REPORT_imperial_eskd.md

@@ -0,0 +1,1046 @@
+# Bulk-only LoRA-5 discovery: synthesis for the Imperial ESKD COVID-19 cohort
+
+Cohort imperial_eskd, 2026-10-06. This report combines four lenses (clinical, longitudinal, programs, populations), their 21 candidate findings and 42 adversarial verdicts.
+
+- **Paths** are relative to /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/ unless they start with /.
+- **Synthesis scripts and tables** are in synthesis_imperial_eskd/ (i00–i05).
+- **findings_imperial_eskd.json** holds every finding, both verbatim verdicts, the lens negative results and my annotations.
+- **Grading rule:** refuted if any verifier refuted; confirmed only if both confirmed; otherwise qualified.
+- **"Draws"** means the draws whose own paired calibration reproduces the feature's contrast, and which therefore carry the result.
+
+## Bottom line
+
+- **No finding is confirmed.** 19 are qualified and 2 are refuted (synthesis_imperial_eskd/i02_grades.csv). The best result is IMP-LONG-1 (ISG-high monocytes fall over days), which got one "confirmed" verdict. It is a positive control.
+- **The generated cells reproduce known shifts, in direction.** The bulk-only set is 164 samples from 76 donors; 45 of those donors have no single cells. The reproduced shifts are:
+  - an interferon response in every lineage in COVID+ samples, fading over the first weeks after the positive swab;
+  - monocyte HLA class II loss and a monocyte S100/CD163 programme with severity;
+  - more B-lineage cells in COVID+ samples, and more plasma cells at higher WHO grade;
+  - more platelets with severity.
+- **None of this is new.** All of it is in Stephenson 2025, Gisby 2022 or other COVID-19 cohorts, and the bulk already shows every one.
+- **Wave is the main trap.**
+  - 126 of the 143 bulk-only COVID+ samples are from wave 1.
+  - But 62 of the 113 paired COVID+ samples, which are the truth, are from wave 2. Wave 2 had a different single-cell processing site and thaw protocol, and dexamethasone treatment.
+  - Several observed contrasts exist only in wave 2. Recalibrating on wave 1 changes which draws carry a result.
+- **Fresh bulk vs frozen cells.**
+  - The bulk was made from fresh PBMC. It contains low-density granulocyte (LDG) and S100 alarmin RNA that the frozen single cells lack.
+  - The generator turns this RNA into extra classical monocytes and fewer T cells (IMP-CLIN-5).
+  - That inflates or creates several monocyte, T-subset and platelet associations.
+- **Interferon-linked labelling errors.** The frozen labeller:
+  - calls ISG-high memory B cells "naive" (IMP-LONG-2, refuted);
+  - labels mostly ISG-high, non-dividing NK cells as "Proliferating NK" (IMP-LONG-3);
+  - labels CD8B+ memory-like CD8 cells as "CD8aa" (IMP-CLIN-4, refuted).
+- **Fatal outcome adds nothing beyond peak severity** in any finding.
+  - The paired fatal truth comes from 8 donors, 7 of them from wave 2.
+  - The fatal labels disagree between data sources.
+- **Little beyond the bulk.** Only the COVID+ vs COVID− cell-type-specific pattern (IMP-PROG-2) survives a donor-disjoint, wave-1 test, and it captures only 4–8% of the reliable observed signal. It replicates the sepsis result SEP-PROG-4 for disease vs control only.
+- **No new population.** The artefacts found are:
+  - a generated-only lineage-mixed "T cell" cluster;
+  - over-produced DN T cells;
+  - a spurious CD14 monocyte expansion in severe samples.
+- **Trust-layer bug.** calibration.csv's severity contrast puts all 10 paired recovery samples into the severe arm, and 8 of them belong to mild/moderate donors (synthesis_imperial_eskd/i03b_recovery_samples.csv).
+- **Refuted:**
+  - "CD8αα T cells fall with severity" (IMP-CLIN-4);
+  - "the memory share of B cells rises over the illness" (IMP-LONG-2).
+
+## 1. What was generated and how much to trust it
+
+### 1.1 Generated data
+
+Source for this section: synthesis_imperial_eskd/i00_counts.txt and its CSV tables.
+
+**Bulk-only samples: 164 from 76 donors.**
+- 143 COVID+ (58 donors), 19 COVID− (19 donors), 2 recovery.
+- Single cells:
+  - 45 donors (97 samples) have no single cells at all;
+  - 47 donors (100 samples) have no usable paired sample.
+- COVID+ by peak severity:
+  - 70 mild/moderate (30 donors);
+  - 73 severe/critical (28 donors).
+- 20 COVID+ samples come from 12 fatal donors.
+- Wave: 126 of the 143 COVID+ samples (88%) and all 19 COVID− samples are from wave 1.
+- Timing: median day since the first positive swab is 12. 16 samples are after day 21, beyond the paired range (the latest paired COVID+ sample is at day 20).
+- One sample has a sex flag: C137_positive_2.
+
+**Draws.**
+- Five LoRA-5 models, each fine-tuned on 5 donors.
+- Each generated 1,500 cells per bulk-only sample: 246,000 per draw, 1,230,000 in total (i00_cells_by_set_draw.csv).
+
+**Calibration cells.**
+- 1,500 cells for each of the 142–145 paired test samples per draw: 1,072,500 in total.
+- Each test set covers 52 donors (i00_paired_test_by_draw.csv):
+  - 31–32 COVID−;
+  - 102–104 COVID+ (36–41 severe/critical, 62–66 mild/moderate, 7 fatal donors);
+  - 8–10 recovery samples;
+  - 64–69 samples are wave 1.
+
+**Observed truth.**
+- 157 paired samples: 34 COVID−, 113 COVID+ from 35 donors, 10 recovery.
+- 514,260 cells (226–11,388 per sample).
+
+### 1.2 Training donors and how the draws differ
+
+| Draw | Training donors (paired profile) | Donors processed with the 2021 protocol |
+|---|---|---|
+| 0 | C106, C237 (COVID−); C126 (mild); C141 (severe, fatal); C58 (mild, with a 2020 control) | C141, C58 |
+| 1 | C142, C164 (COVID−); C65 (severe); C31 (mild); C101 (severe, fatal*) | C101 |
+| 2 | C164, C85 (COVID−); C101 (severe, fatal*); C60 (severe); C187 (mild, with a 2020 control) | C101, C187 |
+| 3 | C240, C251 (COVID−); C31 (mild); C108 (severe, fatal*, with a 2020 control); C40 (severe) | C108 |
+| 4 | C152 (COVID−); C137 (2020 control and recovery); C33 (severe, fatal); C170 (mild, with a 2020 control); C36 (mild) | C137, C33, C170 |
+
+- Source: synthesis_imperial_eskd/i00_training_donors.csv.
+- *Fatal in the bulk metadata, but the single-cell metadata says "no" for C101 and "indirect" for C108 (verify/IMP-PROG-6_calib/v00_fatal_label_crosscheck.txt).
+- Draws 1 and 2 share C101 and C164; draws 1 and 3 share C31.
+- Each draw's training donors also contribute 5–11 bulk-only samples (i03b_trained_donor_bulkonly_samples.csv). The verifiers excluded these in sensitivity analyses.
+
+**Six-type composition** (i00_bulkonly_l1_composition_by_draw_group.csv, i00_observed_l1_composition_by_group_wave.csv):
+- Generated monocyte share in bulk-only COVID+ samples is 0.20, 0.16, 0.26, 0.28 and 0.19 in draws 0–4.
+- Observed paired COVID+ monocyte share is 0.083 in wave 1 and 0.228 in wave 2.
+
+| COVID+ minus COVID− | Generated (every draw) | Observed, wave 1 | Same direction? |
+|---|---|---|---|
+| Monocytes | +0.02 to +0.10 | −0.08 | no |
+| T cells | −0.06 to −0.15 | +0.03 | no |
+| B cells | +0.03 to +0.07 | +0.05 | yes |
+
+So the generator reverses the wave-1 direction of both monocytes and T cells.
+
+**Per draw** (i01_calibration_long.csv, i01_draw_summary.csv, and the verifier tables in section 2):
+- **Draw 0:**
+  - 10 COVID passes on all test samples and 10 on wave-1 tests.
+  - Plasma cells at 0.23x of observed (its training donors are plasma-poor: 0.28% vs 1.16%).
+  - Platelets at 3.66x (C141 has 7.6% platelets). MS1 monocytes at 0.04x.
+- **Draw 1:**
+  - CD56dim NK and proliferating NK at 0.40x.
+  - Trained on C65, a severe donor with 8.5% plasma cells.
+  - Creates a WHO-plasma-cell effect in wave 2 that observed cells lack.
+- **Draw 2:**
+  - Largest generated monocyte excess in COVID+ (+0.10).
+  - Wrong COVID and severity signs for cDC2 (Spearman 0.12).
+  - Creates a false cDC2 rise in wave 2.
+- **Draw 3:**
+  - ISG-high T at 0.29x and CD8aa at 0.37x.
+  - Often the weakest draw for COVID contrasts (for example, CD14 HLA-II for COVID is not significant).
+- **Draw 4:**
+  - Three of its five training donors were processed with the 2021 protocol.
+  - Fewest level passes: 16 of 50 on all tests and 6 of 50 on wave-1 tests. On wave-1 samples its per-sample Spearman is −0.15 for T cells and −0.09 for monocytes.
+  - Best MS1 level (0.37x).
+  - The only draw at the observed CD163+ monocyte level (1.04–1.09x), copying C33 (74% CD163+).
+
+### 1.3 The trust layer (calibration.csv)
+
+I summarised analysis/features/imperial_eskd/calibration.csv: 50 features (35 composition, 15 states) and 2 contrasts. My own implementation reproduces the file to within 5e-6 (i01_reproduce_check.csv).
+
+**Gates.** These were fixed before I looked at the summary (i01_calibration_summary.py):
+- **Level:** ratio 0.5–2 and Spearman ≥ 0.3.
+- **Informative:** on the same test samples, the observed contrast has a donor-level Mann-Whitney p < 0.10.
+  - If not, the code is NULL, because sign agreement is then a coin flip.
+  - It is NULL_genfx if the generated contrast is nevertheless significant (p < 0.05), meaning the generator invents an effect.
+- **PASS:** informative, level, same sign, and generated/observed contrast between 0.5 and 2.
+- **DIR (direction only):** informative, ratio 1/3–3, Spearman ≥ 0.2, and same sign.
+- **FAIL:** anything else.
+
+**Restrictions.** I applied the gates to three restrictions of each draw's test samples:
+- R0: calibration.csv as published;
+- R1: severity among COVID+ only, which fixes the recovery-sample bug;
+- R2: wave-1 samples only, the population of 88% of the bulk-only samples.
+
+**Per draw** (i01_draw_summary.csv). In each cell, "inf" counts the features whose observed contrast is informative. The PASS / DIR / FAIL / NULL counts add up to 50. In the NULL column, the number in brackets is how many of those are NULL_genfx.
+
+| Draw | Level pass, all / wave 1 (of 50) | COVID, all tests: same sign; inf; PASS / DIR / FAIL / NULL (genfx) | Severity, official: inf; PASS / DIR | Severity, COVID+ only: inf; PASS / DIR / FAIL | COVID, wave 1: inf; PASS / DIR / FAIL |
+|---|---|---|---|---|---|
+| 0 | 21 / 23 | 43; 23; 10 / 5 / 8 / 27 (12) | 6; 3 / 1 | 16; 2 / 4 / 10 | 25; 10 / 2 / 13 |
+| 1 | 24 / 22 | 45; 24; 7 / 9 / 8 / 26 (8) | 4; 2 / 1 | 15; 4 / 4 / 7 | 23; 7 / 9 / 7 |
+| 2 | 25 / 20 | 40; 24; 8 / 6 / 10 / 26 (18) | 12; 8 / 2 | 18; 7 / 4 / 7 | 25; 11 / 5 / 9 |
+| 3 | 25 / 25 | 39; 23; 7 / 8 / 8 / 27 (10) | 5; 1 / 2 | 13; 2 / 6 / 5 | 26; 9 / 8 / 9 |
+| 4 | 16 / 6 | 47; 22; 7 / 9 / 6 / 28 (15) | 9; 2 / 5 | 15; 3 / 6 / 6 | 27; 5 / 9 / 13 |
+
+**Pooled** (i01_pooled.csv):
+- 24 of 50 features meet the level gate.
+- 8 features have Spearman ≥ 0.5: ISG-high monocytes 0.82, ISG-high T 0.78, memory CD8 0.60, six-type B 0.59, naive B 0.58, gdT 0.57, memory B 0.54, proliferating NK 0.53.
+- 12 features have Spearman < 0.2.
+- The COVID sign holds in 5/5 draws for 31 features; the official severity sign holds in 5/5 for 24.
+
+**What is reproduced** (i01_codes_by_feature.csv):
+- **COVID+ vs COVID−, all test samples:**
+  - PASS in at least 3 draws:
+    - ISG-high monocytes (5/5);
+    - proliferating NK (4 PASS, 1 DIR);
+    - proliferating T (L2: 4 + 1; label: 3 + 2);
+    - six-type B (3 + 2);
+    - ISG-high T (3 PASS; FAIL in draw 3);
+    - plasmablast markers (3 PASS; FAIL in draw 0).
+  - Direction in 5/5 draws: naive B (2 PASS + 3 DIR); CD16 monocytes, effector B and pDC (1 + 4 each).
+- **COVID+ vs COVID−, wave 1 only:**
+  - PASS in 5/5: naive B, proliferating NK, ISG-high monocytes.
+  - PASS in 4/5: six-type B and proliferating T.
+  - Direction in 4–5/5: ISG-high T, effector B, DN T and plasma cells.
+- **Severity among COVID+:**
+  - ISG-high monocytes: PASS 5/5, also in wave 1.
+  - Six-type "Other" (platelets and others): 3 PASS + 2 DIR.
+  - Proliferating T (label): 2 + 3.
+  - Platelets: 2 + 2 (FAIL in draw 0).
+  - ISG-high T: 2 + 2 (FAIL in draw 3).
+  - CD8aa: 1 + 4.
+  - MAIT: 4 DIR + 1 FAIL.
+  - Wave 1 only: cDC2 3 PASS (2 NULL); plasma cells 2 + 2 (FAIL in draw 0); CD8aa 2 + 2.
+
+**What is not reproduced:**
+- **FAIL in all five draws:** transitional B (ratio 0.08–0.20), C1Q-high monocytes (0.06–0.60, Spearman ≤ 0.15), cycling monocytes (0.03–0.12).
+- **MS1 monocytes:** ratio 0.04–0.37; FAIL in 4/5 draws for every contrast (DIR only in draw 4).
+- **Wrong direction in wave 1, COVID+ vs COVID−:** six-type monocytes (FAIL 4/5), CD14 monocytes (5/5), intermediate monocytes (3/5) and memory CD4 T (4/5).
+  - Observed wave-1 COVID+ samples have fewer monocytes and more memory CD4 T cells. The generated cells show the opposite.
+- **Severity among COVID+:** CD56bright NK, pDC and cDC1 FAIL in 4/5 draws; cycling monocytes 5/5; MS1 4/5; cDC2 3/5.
+- **Levels far off** (pooled ratio):
+  - too low: LDG 0.02, cycling monocytes 0.08, transitional B 0.12, cycling activated T 0.15, MS1 0.16, C1Q-high 0.25;
+  - too high: DN T 2.9x, ILC 2.0x, effector B 1.9x, gdT 1.6x, memory B 1.5x, platelets 1.5x (2.8–16.5x on wave-1 samples).
+- **Effects the generator invents.** For COVID+ vs COVID−, only 22–24 features per draw have an observed contrast at all. The generator creates a contrast the truth lacks (NULL_genfx in 5/5 draws) for:
+  - six-type T cells;
+  - platelets;
+  - the "platelet signal in monocytes" negative control.
+
+**Limits of the trust layer:**
+1. **Severity row bug.**
+   - All 10 paired recovery samples carry severity_peak = severe/critical and so enter the severe arm. Eight of them belong to donors whose illness was mild/moderate (i03b_recovery_samples.csv).
+   - With them, an observed severity contrast is detectable for only 4–12 features per draw. Restricted to COVID+ samples, it is detectable for 13–18.
+2. **Wave mismatch.**
+   - The paired COVID+ truth is 55% wave 2 (62 of 113 samples).
+   - Within wave 1, observed COVID contrasts are smaller, and some reverse (monocytes, memory CD4 T).
+   - Draw 4, trained on three wave-2-processed donors, tracks wave-1 samples worst.
+3. **Few outcome donors.**
+   - Each test set holds about 13–14 severe/critical donors and 7 fatal donors, 6 of the fatal ones from wave 2.
+   - Fatal outcome cannot be calibrated in wave 1: there is 1 fatal paired donor in wave 1.
+4. **Calibration and observed truth are not independent.** Each draw's test set is 142–145 of the 157 observed samples.
+5. **Does calibration transfer?** I compared paired and bulk-only bulk RNA within groups for 21 programme scores (imp_programs/transfer_set_shift.csv):
+   - **COVID− and mild/moderate COVID+:** they agree (no q below 0.12).
+   - **Severe/critical:** bulk-only samples carry more granulocyte RNA (+0.69 log2, p 0.015) and less ISG (−0.82, p 0.02) than paired ones.
+   - **Fatal:** bulk-only samples carry far more granulocyte RNA (+1.37, q 0.011) and fewer T-cell transcripts (−1.05, q 0.027).
+   - So severity and fatal calibration transfers poorly for monocyte, T-cell and platelet features. It does transfer for interferon features: the ISG genes do not differ within any time window (verify/IMP-LONG-1_calib/v04_d_transfer_genes.csv).
+6. **Shared noise.** All draws used the same sampling seed.
+   - Within COVID+ bulk-only samples, shared noise explains a median 22% of cross-draw agreement for L2 types below 1% abundance, and 4% for others (imp_clinical/s14_shared_noise.txt).
+   - State calls at the same cell slot agree across draws far above chance. For ISG-high CD14 monocytes the odds ratio is 39–86 at the same slot vs 5–14 at shifted slots (imp_populations/tables/s12_shared_noise.csv).
+   - So "k of 5 draws" is not k independent replications. The per-draw gates are the evidence.
+
+**Checks I ran** (i03_checks.txt):
+- No Imperial lens or verifier script uses the statsmodels `scale='X2'` quasi-binomial path. All 9 mentions are comments stating it was avoided; one was flagged automatically and is a docstring line.
+- statsmodels MixedLM with lbfgs can return degenerate fits (imp_clinical/s20_mixedlm_lbfgs_bug.txt). The lens and verifier code refit those fits with other optimizers.
+
+## 2. Findings
+
+### 2.1 Confirmed
+
+None. IMP-LONG-1 had one "confirmed" verdict, from the recompute verifier, and is the strongest result.
+
+### 2.2 Qualified findings at a glance
+
+| ID | What survives | Draws | Observed paired cells | Beyond bulk | New | Strength |
+|---|---|---|---|---|---|---|
+| IMP-PROG-4 | ISG programme up in every testable lineage; SIGLEC1 up in monocytes (COVID+ vs COVID−) | 0–4 in 8 types; ≥ 3 draws in 13 of 15 | yes, strong | no | no | strong positive control |
+| IMP-LONG-1 | ISG-high monocytes fall over days since the swab | 0–4 (T cells: 0–2) | yes | no | no | strong positive control |
+| IMP-POP-4 | ISG-high sub-states up in COVID+ in every lineage | myeloid 0–4; NK/B/T 0–2 | yes | no | no | strong positive control (myeloid) |
+| IMP-PROG-1 | HLA-II loss in CD14/intermediate monocytes and cDC2 with severity | severity 0–4; COVID 0,1,2,4; fatal 0,1,2,4 | yes | location only | no | moderate (severity) |
+| IMP-PROG-3 | S100/MCEMP1/CD163/MS1 genes up in CD14 monocytes | 0–4 (CD14 only) | yes | no | no | moderate; size inflated |
+| IMP-CLIN-1 | More B-lineage cells, mostly labelled naive, in COVID+ | direction 0–4; size 3–4 (all tests), 0–4 (wave 1) | yes; wave 1 only nominal | partly | no | moderate |
+| IMP-PROG-2 | Cell-type-specific COVID pattern beyond a bulk broadcast | 0–4 (COVID only) | yes, small share | yes, modestly | method only | moderate (COVID) |
+| IMP-POP-1 | More CD14 monocytes that are both ISG-high and HLA-DR-low | 0–4 | yes, but not a coupled state | no | no | moderate as a fraction |
+| IMP-CLIN-2 | Plasma cells rise with WHO grade (between patients, wave 1) | direction 1–4; size 3–4 | wave 1 only | no | no | moderate |
+| IMP-PROG-5 | Bulk cell-cycle genes attributed to plasma cells | 1–3 | yes | no; deconvolution does it | no | positive control |
+| IMP-POP-3 | Fewer HLA-II-high CD14 monocytes | wave 1: 1, 2, 4 | WHO yes; severity weak | no | no | weak (dilution) |
+| IMP-CLIN-3 | More platelets with severity | both waves: size 1–2, direction 3–4; wave 1: none | yes | partly, uncalibrated | not as a cell fraction here | moderate echo |
+| IMP-POP-2 | More CD163+ CD14 monocytes with severity | 1, 3 (4 confounded) | yes | no | no | weak |
+| IMP-LONG-3 | "Proliferating NK" label share falls over days (ISG-high NK) | size 1–2 | yes | no | no | weak |
+| IMP-LONG-4 | cDC2 share rises over days | 3–4 (all waves) or 0–1 (wave 1) | wave 1, between donors | no | partly | weak |
+| IMP-PROG-6 | Glucocorticoid programme in fatal patients' monocytes | 1–4, uncalibrated beyond severity | wave-2 truth | no | no | artefact-suspect |
+| IMP-CLIN-5 | T cells down, CD14 up with severity is a mapping artefact | CD14 artefact 0–4; T artefact 2–3 | observed shows neither | - | the artefact is new | artefact |
+| IMP-LONG-5 | Platelet decline over days is a generator artefact | artefact shown for 0 | observed flat | - | - | artefact |
+| IMP-POP-5 | Generated-only lineage-mixed "T" cluster | none | none | - | - | artefact |
+
+### 2.3 Details
+
+For each finding I give:
+- what survives, and the draws that carry it;
+- the four honesty items: (1) calibration, (2) replication across draws, (3) observed paired cells, (4) beyond the bulk;
+- the verdict reasons;
+- whether it is new relative to the cohort papers;
+- the sources.
+
+#### A. Interferon (the cohort's positive controls)
+
+**IMP-PROG-4: ISG programme in every lineage and SIGLEC1 in monocytes.** Lens imp_programs; verdicts qualified / qualified.
+
+- **What survives.** In bulk-only COVID+ samples the generated cells raise the 10-gene ISG programme in every testable lineage (verify/IMP-PROG-4_recompute/v08_key_numbers.csv):
+  - CD14 monocytes +1.24, +0.98, +1.14, +1.32, +0.96 log2 (draw 0 group means 7.08 vs 5.84);
+  - CD16 monocytes +0.89 to +1.13;
+  - memory CD8 +0.90 to +1.11;
+  - Treg +0.23 to +0.40.
+- SIGLEC1 also rises in CD14 monocytes (+0.86 to +1.59), CD16 monocytes and intermediate monocytes.
+- **Draws.**
+  - All of 0–4 in 8 types: CD14, CD16 and intermediate monocytes, CD56dim NK, memory CD4, memory CD8, cDC2, gdT.
+  - At least 3 draws in 13 of 15 types.
+  - MAIT and pDC: none; they can be tested in only 4 and 2 draws.
+  - SIGLEC1: monocytes 0–4; cDC2 0, 1, 3, 4.
+- **(1) Calibration.**
+  - The programme gate passes 5/5 in most types. Ratio 0.31–1.07 (Treg lowest); CD14 per-sample Spearman 0.81–0.89 (imp_programs/programs_calib.csv).
+  - The calibration.csv rows agree: ISG-high monocyte ratio 0.65–0.81, Spearman 0.78–0.85; ISG-high T 0.29–0.60 and 0.72–0.81; both signs 5/5.
+- **(2) Replication.** 5/5 draws. It survives:
+  - dropping the sex-flag sample and training-donor samples;
+  - wave 1 only, and one sample per donor;
+  - equal per-cell depth, label confidence ≥ 0.9, and myeloid-free T/NK cells (verify/IMP-PROG-4_recompute/v05_clean_lymphoid_tests.csv).
+  - Diabetes and ethnicity shrink it by 20–30%, exactly as they shrink the bulk.
+- **(3) Observed.** +0.93 (MAIT) to +2.14 (CD16 monocytes), p ≤ 0.004 in every type.
+- **(4) Beyond bulk.** None.
+  - Bulk ISG rises by +1.38 (p 1.6e-4) and tracks observed type-level ISG better than the generated cells in all 15 types.
+  - The generated score adds to the bulk ISG score in 1 of 71 likelihood-ratio tests (verify/IMP-PROG-4_calib/v03_bulk_lr.csv).
+- **Failed sub-claims.**
+  - **"The cross-type ordering follows the observed one."**
+    - On a common sample set, the observed 9-gene response is flat (0.63–0.88, CD14 among the lowest), and bulk-only vs observed r is −0.60 to +0.44 (verify/IMP-PROG-4_calib/v04_common_sample_profile_corr.csv).
+    - The claimed r of 0.74–0.86 comes from two things: lymphopenic COVID+ samples drop out of the T- and B-cell estimates, and IFI27 rises only in myeloid cells.
+    - Within samples, generated cells make 5–9 of 13 types differ from the reference, where observed cells make none; Treg goes the opposite way (v02_interaction_calib_summary.csv).
+  - **"The severity ISG difference does not carry over."** It is confounded by sampling time: bulk-only severe samples have a median of day 15 vs 9.5 for mild.
+    - Adjusted for days since swab, CD14 severity ISG is +0.92 to +1.31 (p 0.0013–0.0075) in 5/5 draws, and +0.85 to +1.28 within 14 days.
+    - The bulk ISG difference is +1.09 (p 0.005) (verify/IMP-PROG-4_calib/v05_severity_robust.csv).
+- **Other issues.**
+  - Generated COVID+ lymphoid pseudobulks carry myeloid transcripts: 23 of 49 tests vs 0 of 10 observed.
+  - The bulk-only effect is 0.18–0.73x the observed paired effect.
+  - Gisby 2022 p.2 concerns SIGLEC1, not IFN pathways.
+- **New?** No.
+  - Stephenson 2025: IFN-alpha/beta responses across innate and adaptive cells (PDF p.4); stronger early in severe disease (PDF p.8–9).
+  - Gisby 2022: SIGLEC1 in the bulk (p.2); monocyte SIGLEC-1 and type I IFN in severe disease (pp.7–8).
+  - Non-ESKD cohorts (Combes 2021, Arunachalam 2020) report the opposite severity direction.
+
+**IMP-LONG-1: ISG-high monocytes fall over days since the positive swab.** Lens imp_longitudinal; verdicts confirmed / qualified.
+
+- **What survives.** In bulk-only COVID+ samples up to day 21 (127 samples, 57 donors), the generated ISG-high share of monocytes falls by −0.238, −0.186, −0.219, −0.160, −0.172 logit/day (p 8e-13 to 5e-17) (verify/IMP-LONG-1_recompute/v10_numbers.txt). It falls in every subset:
+  - wave 1: −0.168 to −0.246;
+  - the 31 COVID+ donors without single cells: −0.194 to −0.283;
+  - within donors: −0.166 to −0.254.
+- Window means drop from 0.43–0.48 (days 0–7) to 0.09–0.18 (after day 21).
+- **Draws.** Monocytes 0–4. ISG-high T cells 0–2 (0, 1, 2, 4 in wave 1).
+- **(1) Calibration.**
+  - calibration.csv: ratio 0.65–0.81, Spearman 0.78–0.85, both contrast signs 5/5. It passes 5/5 in my COVID and COVID+-only severity restrictions, including wave 1.
+  - The time contrast, recomputed on COVID+ paired tests (verify/IMP-LONG-1_calib/v01_temporal_calibration.csv):
+    - ratio 0.65–0.83, Spearman 0.84–0.89, within-donor r 0.88–0.94;
+    - generated slope −0.193 to −0.272 vs observed −0.251 to −0.301;
+    - wave-1 slope ratio 0.72–1.00.
+  - Calibration transfers: the ISG bulk genes do not differ between paired and bulk-only samples within any time window (smallest q 0.97).
+- **(2) Replication.** 5/5 draws.
+  - It survives age, sex, ethnicity, the sex flag, training donors, onset-based time, WHO grade, fatal outcome, depth (thinning to 800 counts) and label confidence.
+  - Residuals still correlate across draws (r 0.31–0.69): shared noise.
+- **(3) Observed.**
+  - Weekly means 0.70 → 0.47 → 0.19 (COVID− 0.14); slope −0.282 to −0.286 (p 4.5e-28).
+  - By wave: wave 1 −0.222, wave 2 −0.340 (interaction p 0.018). Within donors −0.29.
+  - Bulk-only wave-1 slopes are 0.76–1.20x the observed wave-1 slope.
+- **(4) Beyond bulk.** None for the trend.
+  - Adjusting for the 10-gene bulk ISG score removes it in observed cells (−0.003 to −0.011, p ≥ 0.58) and in generated cells.
+  - The bulk score tracks the observed fraction better (Spearman 0.90–0.92 vs 0.84–0.89).
+- **Why qualified.**
+  - ISG-high T is calibrated only in draws 0–2 (level 0.29 in draw 3, 0.44 in draw 4).
+  - The decline is front-loaded, not steady; it follows the convex bulk ISG curve.
+  - The window after day 21 is uncalibrated.
+  - Bulk-only visits of donors who also have paired visits fall about half as fast (bulk-driven).
+  - The steeper decline in severe patients (observed −0.142, p 0.006; Stephenson Fig. 4B) is absent in bulk-only cells, because severe bulk-only donors are first sampled later (median day 9 vs 4) (verify/IMP-LONG-1_calib/v08_sevxday_bulkscore.csv).
+- **New?** No. Stephenson 2025 Fig. 4A–B (PDF p.8–9) and PDF p.14; Gisby 2022 PDF p.8.
+
+**IMP-POP-4: ISG-high sub-states in every lineage.** Lens imp_populations; verdicts qualified / qualified.
+
+- **What survives.** COVID+ bulk-only samples carry larger ISG-high shares (≥ 4 of 10 ISGs detected) than COVID− samples (verify/IMP-POP-4_recompute/tables/v09_summary.csv; verify/IMP-POP-4_calib/tables/v03_summary.csv):
+  - CD14 monocytes +0.171 to +0.200 (0.23–0.26 vs 0.03–0.08);
+  - CD16 monocytes +0.159 to +0.190;
+  - NK +0.055 to +0.081;
+  - non-plasma B +0.042 to +0.065;
+  - ISG+ cytotoxic T +0.021 to +0.049;
+  - DC +0.136 to +0.176.
+- **Draws.**
+  - CD14, CD16 and DC (marker rules without embeddings): 0–4.
+  - cDC2: 0, 1, 3, 4.
+  - NK, B, T: 0–2. Their ratios of 0.50–0.71 sit near the gate. With equal per-cell depth, T is carried by draw 2 only and B by draws 1–2.
+- **(1) Calibration.**
+  - CD14 ratio 0.60–0.78, Spearman 0.74–0.83; NK 0.37–0.70; B 0.36–0.71; T 0.23–0.72; DC 0.62–0.79. Signs 5/5 everywhere.
+  - On wave-1 tests, ISG+ cytotoxic T is over-produced (1.25–1.42x in draws 0, 2, 4).
+- **(2) Replication.** 5/5 draws in every lineage (q ≤ 0.011).
+  - It survives the confounders, including rank-ANCOVA for age, sex, wave, ethnicity, diabetes and bulk depth.
+  - Shared noise: same-slot odds ratio 42–94 vs 7–16 at shifted slots.
+- **(3) Observed.**
+  - CD14 +0.365 (p 1.2e-6), CD16 +0.382, NK +0.227, B +0.187, T +0.122, DC +0.344.
+  - Within donors, 9/9 to 11/11 rise from their pre-infection sample.
+- **(4) Beyond bulk.** None. The bulk ISG score predicts the observed fractions better (0.80–0.92 vs 0.69–0.88).
+- **Why qualified.**
+  - "Severity reproduced only in monocytes" is too strong: within 10 days of the swab, NK and T cells rise with severity in 5/5 draws (gated draws p 0.014–0.036), and B cells in 3/5 (verify/IMP-POP-4_recompute/tables/v06_severity_timing.csv).
+  - The lymphoid calibration depends on the thinning method.
+  - Count errors: T severity is 1/5 (draw 4), not 0/5; NK fatal 3/5.
+  - One COVID− control (C105_negative) was taken 7 days before the first positive swab and has the highest bulk ISG of all COVID− samples. That makes the mean-scale contrasts conservative.
+- **New?** No. Stephenson 2025 PDF p.4, Fig. 4 (PDF p.8); Gisby 2022 pp.7–8. Combes 2021 found the opposite severity direction in non-ESKD cohorts.
+
+**IMP-LONG-3: the "Proliferating NK" label share falls over days.** Lens imp_longitudinal; verdicts qualified / qualified.
+
+- **What survives.** The share labelled "Proliferating NK" falls by −0.073, −0.039, −0.056, −0.089, −0.064 logit/day, as observed.
+  - The label mostly marks ISG-high NK cells that are not dividing: only 7–9% of labelled cells carry any cycling transcript.
+  - Truly cycling NK cells stay flat (0.58 / 0.58 / 0.61% of cells; −0.007/day, p 0.55) (verify/IMP-LONG-3_calib/v05_label_markers.csv, v05c_native_slopes.csv).
+  - So this is the interferon-stimulated NK decline.
+- **Draws.**
+  - Direction 0–4.
+  - Size 1–2 only. Draws 0, 3, 4 are 1.8–2.6x too steep on the same samples (p ≤ 3e-4). In draws 3–4 this is explained by generated library depth rising over time (verify/IMP-LONG-3_recompute/v04_depth_adjusted.csv).
+- **(1) Calibration.** calibration.csv ratio 0.40–1.04, Spearman 0.45–0.59. COVID PASS in 4/5 draws (5/5 in wave 1).
+- **(2) Replication.** 5/5 negative, but the decline is carried by donors without single cells. Bulk-only visits of donors who also have paired visits show no decline (−0.050 to +0.020) (verify/IMP-LONG-3_recompute/v07_scdonors.csv).
+- **(3) Observed.** −0.045/day (p 2e-7). In wave 1 there is no week-1 peak (0.026, 0.027, 0.012).
+- **(4) Beyond bulk.** None.
+  - Bulk ISG removes the slope and tracks the observed share better (0.54–0.63 vs 0.43–0.54).
+  - A three-score linear model on bulk RNA, trained on the same 5 donors, beats the generated cells in 4/5 draws (verify/IMP-LONG-3_calib/v10_bulk_baseline.csv).
+- **New?** No. Stephenson 2025 Fig. 4A (PDF p.8) and text (PDF p.9).
+
+#### B. Monocyte HLA class II
+
+**IMP-PROG-1: HLA class II loss in monocytes and cDC2, not in B cells.** Lens imp_programs; verdicts qualified / qualified.
+
+- **What survives.** Generated CD14 and intermediate monocytes of bulk-only patients express less HLA class II with severity, consistent with observed paired cells (verify/IMP-PROG-1_recompute/v02_effects.csv):
+  - CD14 −0.29, −0.31, −0.34, −0.24, −0.39 log2;
+  - intermediate monocytes −0.21 to −0.42 in every draw and contrast.
+- COVID+ vs COVID− (CD14 −0.15 to −0.50) and fatal (−0.13 to −0.46) go the same way but are weaker.
+- **Draws.**
+  - CD14: severity 0–4 (also in wave 1); COVID 0, 1, 2, 4; fatal 0, 1, 2, 4.
+  - Intermediate monocytes: 0–4 in all contrasts.
+  - cDC2: severity in 4/5 draws, fatal in 5/5, COVID in none.
+- **(1) Calibration.** HLA-II is not in calibration.csv. The paired check gives (verify/IMP-PROG-1_calib/v01_calib.csv, v02_tmm_calib.csv):
+  - CD14 PASS in all 15 draw × contrast cells, also with TMM normalisation;
+  - ratios: COVID 0.77–1.55, severity 0.54–1.04, fatal 0.47–0.85; Spearman 0.40–0.54.
+  - But the wave-1 COVID truth (−0.16 to −0.30, p 0.07–0.27) is uninformative in 4/5 draws.
+  - The fatal truth comes from 7 of 8 wave-2 donors.
+  - B-cell HLA-II has no per-sample skill (Spearman −0.14 to 0.17).
+- **(2) Replication.** 5/5 same sign.
+  - It survives age, sex, the sex flag, training donors, days, leave-one-donor-out, label confidence and depth.
+  - The generator inflates the CD14 fraction in cases (log2 +0.45 to +1.69 vs observed −0.28 to +0.93). After adjusting for that fraction, CD14 holds in 3/5 (COVID), 4/5 (severity) and 1/5 (fatal) draws.
+  - Intermediate monocytes and cDC2 stay 5/5 (verify/IMP-PROG-1_recompute/v05_adjusted.csv).
+- **(3) Observed.**
+  - CD14 −0.43 (COVID, p 0.001), −0.58 (severity, p 0.001; wave 1 −0.70), −0.53 (fatal, p 0.003).
+  - B cells are not unchanged: in wave-1 COVID+ samples, naive B HLA-II rises (+0.22, p 0.0009). In fatal patients, naive B falls (−0.23 to −0.45) under three other normalisations.
+- **(4) Beyond bulk.** Localisation only, of a known effect.
+  - Bulk HLA-II falls by −0.52, −0.63 and −0.74 log2 (COVID, severity, fatal).
+  - Bulk HLA-II minus monocyte markers tracks observed CD14 HLA-II better (0.71–0.74 vs 0.43–0.48).
+  - None of 45 "adds to bulk" tests survives BH correction (verify/IMP-PROG-1_recompute/v07_tracking.csv).
+- **Why qualified.**
+  - B-cell sparing is not calibrated.
+  - The CD14-minus-B contrast holds only under the lens's normalisation.
+  - COVID is in effect a wave-1 comparison with a weak truth; fatal is nested in severity.
+  - The loss is largest at days 0–7, not progressive.
+- **New?** No. Stephenson 2025 Fig. 3C (PDF p.6) and Fig. 4C (PDF p.8–9); Gisby 2022 Fig. 3c (PDF p.6); Arunachalam 2020 (mDC HLA-DR, p.7–8). Stephenson p.9 also reports HLA effects in B cells.
+
+**IMP-POP-1: more CD14 monocytes that are both ISG-high and HLA-DR-low.** Lens imp_populations; verdicts qualified / qualified.
+
+- **What survives.** The share of CD14 monocytes that are both ISG-high and HLA-DR-low (below the COVID− 25th percentile) is higher in bulk-only COVID+, severe/critical and higher-WHO samples (verify/IMP-POP-1_recompute/tables/v07_verdict_numbers.csv, v04_windows.csv).
+  - Severity: +0.061 to +0.096 on donor means; +0.11 to +0.16 within 14 days (5/5).
+  - It is calibrated in 5/5 draws.
+- But it is the overlap of two separate shifts, an ISG expansion and a global HLA-II decline. It is not one coupled cell state.
+- **Draws.** 0–4 for COVID, severity and WHO grade as a fraction. None for fatal within severe. None in the 47 donors without a paired sample (severity p 0.12–0.53).
+- **(1) Calibration.** Recalibrated on COVID+ tests: ratio 0.72–1.20, Spearman 0.81–0.85, signs 5/5. But the joint state is not calibrated (verify/IMP-POP-1_calib/tables/v02_component_calibration.csv):
+  - the HLA-low share is over-produced 1.25–1.44x with Spearman 0.17–0.50;
+  - HLA-low among ISG+ cells has Spearman −0.01 to 0.39.
+- **(2) Replication.** 5/5 draws on donor means (severity p 0.0034–0.020).
+  - Unadjusted sample-level severity is 0/5, because severe bulk-only samples are later (day 15 vs 9.5).
+  - Shared noise: same-slot odds ratio 15–31 vs 3–6.
+- **(3) Observed.** COVID +0.149 (p 3e-7); severity +0.180 (p 0.0011; wave 1 +0.155); WHO +0.094 per grade. Against a coupled state:
+  - the HLA-low shift is larger in ISG− cells (+0.215, p 0.003) than in ISG+ cells (+0.073, p 0.29);
+  - the composite's AUC equals that of the product of its two parts (0.839 vs 0.835) (v02_decomposition_contrasts.csv, v04_auc_composite_vs_marginals.csv).
+- **(4) Beyond bulk.** None.
+  - A bulk ISG10-minus-HLA2 score gives severity p 0.0066, fatal p 0.0004, WHO p < 1e-4.
+  - Bulk HLA2 alone gives severity p 1e-4.
+- **New?** No.
+  - Stephenson 2025 Fig. 5B (IFN-stimulated CD14 monocytes expanded in severe disease) and Fig. 5C (lower HLA-DR protein on them), PDF p.11.
+  - Both cohort papers say the COVID effects are not ESKD-specific (Stephenson p.12; Gisby p.8). That contradicts the lens's "ESKD differs from non-ESKD" framing.
+
+**IMP-POP-3: fewer HLA-II-high CD14 monocytes.** Lens imp_populations; verdicts qualified / qualified.
+
+- **What survives.** The HLA-II-high share is lower in COVID+, severe and higher-WHO bulk-only samples (severity −0.021 to −0.054) (verify/IMP-POP-3_recompute/tables/v04_feature_calibration.csv).
+  - In bulk-only cells this is mostly dilution by a CD14 expansion that the generator invents: +6.8 to +12.9 percentage points in severe samples vs +0.6 observed (p 0.65).
+  - The absolute number of HLA-II-high CD14 cells per 1,000 cells actually rises with severity (+0.4 to +4.0) (v03_parent_share.csv, v07_quantiles.csv).
+- **Draws.** On wave-1 tests: severity and WHO grade 1, 2, 4; COVID 1, 4; fatal none.
+- **(1) Calibration.** COVID+ ratio 0.53–0.72, Spearman 0.25–0.46. It fails at a stricter threshold (p90) in draws 1–3 (verify/IMP-POP-3_calib/tables/v04_gates.csv).
+- **(2) Replication.**
+  - 5/5 same sign.
+  - Adjusted for the CD14 share: severity 2/5, or 4/5 with demographics; COVID 2/5; fatal 0/5.
+- **(3) Observed.** COVID −0.075 (p 0.0013, but wave 1 −0.039, p 0.12); WHO −0.047 per grade (p 3e-5); severity −0.071, not significant after adjustment at this threshold.
+- **(4) Beyond bulk.** None. Bulk HLA-DRA/DRB1 alone separates severity (AUC 0.80) and tracks the observed share better (0.48–0.56 vs 0.32–0.47).
+- **New?** No. Stephenson 2025 Fig. 3C; Gisby 2022 Fig. 3c (bulk HLA-DRA falls with WHO grade in both waves); COMBAT 2022 p.922; Schulte-Schrepping 2020 p.1427.
+
+#### C. Monocyte alarmin, CD163 and glucocorticoid programmes
+
+**IMP-PROG-3: S100/MCEMP1/CD163/MS1 programme in monocytes.** Lens imp_programs; verdicts qualified / qualified.
+
+- **What survives.** Generated CD14 monocytes of severe/critical bulk-only patients express more of the Stephenson 2025 severity programme (CD163, MCEMP1, PLAC8, S100A8/9/12) than those of mild/moderate patients: +0.39 to +0.55 log2. They also express more MS1 and S100 genes. The same holds for COVID and fatal.
+  - It is a gene-level shift. The discrete MS1 cell state is under-generated (0.04–0.37x).
+- **Draws.**
+  - CD14: 0–4 in all contrasts. In draw 2 the severity effect is explained by the CD14 count.
+  - CD16: none, once corrected for classical-identity genes.
+  - MS1 in CD16 for fatal: draws 2 and 4.
+  - Intermediate-monocyte MS1 for severity: 0/5.
+- **(1) Calibration.**
+  - Paired CD14 check PASS 5/5 for all contrasts, also on wave 1 (severity ratio 0.56–0.82; Spearman 0.36–0.77).
+  - With the CD14 count held fixed the ratio drops to 0.29–0.64 (verify/IMP-PROG-3_recompute/v08_count_adjusted.csv).
+  - The calibration.csv MS1 state fails (ratio 0.04–0.37).
+- **(2) Replication.**
+  - 5/5 in all 27 programme × cell type × contrast cells.
+  - Generated CD14 count and programme score are coupled (Spearman 0.43–0.78 vs observed 0.02–0.35). Adjusted for the count, CD14 severity shrinks to 0.12–0.34 log2, about 0 in draw 2 (v07_coupling.csv).
+- **(3) Observed.**
+  - CD14 severity +0.90 (p < 1e-4); MS1 +1.06 to +1.20; S100 +0.59.
+  - Fatal within severe/critical: +0.11 (p 0.72).
+  - ALOX5AP, CYP1B1 and RETN do not rise in observed CD14 cells.
+  - The observed CD16 rise follows classical-identity genes, and disappears after adjusting for them (verify/IMP-PROG-3_calib/c02_genes_summary.csv, c02_adjusted.csv).
+- **(4) Beyond bulk.** Little.
+  - The bulk programme score rises by +0.76 to +1.07 and tracks observed cells better.
+  - The one "adds to bulk" result (S100 in draws 0, 3, 4) fails on each draw's own test donors (verify/IMP-PROG-3_calib/c04_beyond_bulk.csv).
+- **Why qualified.**
+  - The size is inflated by spurious CD14 over-production.
+  - "All draws, all three monocyte subsets" is wrong.
+  - The fatal calibration does not transfer: bulk-only fatal programme scores are 0.47–0.60 log2 higher than paired fatal ones.
+  - TNF loss, Stephenson's top monocyte severity gene, is not reproduced (0/5).
+- **New?** No. Stephenson 2025 PDF p.6; Gisby 2022 PDF p.5 (module tB, attributed to LDGs); Schulte-Schrepping 2020 p.11; Silvin 2020 p.9; COMBAT 2022 p.8.
+
+**IMP-POP-2: more CD163+ CD14 monocytes with severity.** Lens imp_populations; verdicts qualified / qualified.
+
+- **What survives.** The CD163+ share of CD14 monocytes rises with severity, WHO grade and fatal outcome: +0.043, +0.056, +0.029, +0.127, +0.119 for severity (draw 2 p 0.088) (verify/IMP-POP-2_recompute/tables/v07_summary_by_draw.csv). This echoes a real observed association (+0.217, p 2.6e-4).
+- **Draws.**
+  - Robust in 1 and 3.
+  - Draw 4 is calibrated but confounded: adjusting for the CD14 share removes it (+0.119 → +0.024), and its training donor C33 has 74% CD163+ cells.
+  - Draws 0 and 2 are borderline.
+- **(1) Calibration.**
+  - Ratio 0.51/0.55/0.53/0.67/1.09, contrast 0.44–0.83x.
+  - Draws 0–2 fail with a depth-robust CP10k level (0.39–0.44).
+  - On wave-1 tests, draws 1, 3 and 4 pass (verify/IMP-POP-2_calib/tables/v01_calibration.csv).
+- **(2) Replication.** 5/5 positive. In the 31 COVID+ donors without single cells, only draw 4 is significant (p 0.034).
+- **(3) Observed.** Severity +0.217 (wave 1 +0.213, p 0.029, fragile to dropping one donor); WHO +0.136 per grade (p 1e-5). The described state is not supported:
+  - observed CD163+ cells are not enriched for low HLA-II (37% vs 38.5%; within-sample odds ratio 0.71–0.79);
+  - FKBP5, RNASE2 and PLBD1 do not co-occur with CD163 within samples (v02_within_sample_or.csv).
+- **(4) Beyond bulk.**
+  - Generated cells: none (likelihood-ratio p 0.41–0.98).
+  - Observed cells: the raw fraction adds to the bulk for severity (permutation p 0.021), but not after depth matching.
+- **Why qualified.**
+  - "Not dexamethasone-induced" is not shown. Wave-1 severe monocytes carry a glucocorticoid programme (ADAMTS2 +2.9, FKBP5 +0.95 log2), and FKBP5 absorbs about half of the wave-1 effect (v04_w1_gc_adjust.csv).
+  - Fatal adds nothing beyond severity.
+- **New?** No. Stephenson 2025 Fig. 3C and dex-monocytes (PDF p.11); Schulte-Schrepping 2020 p.1427; COMBAT 2022 p.922 (FKBP5/CD163 module).
+
+**IMP-PROG-6: glucocorticoid / dex-monocyte programme in fatal patients.** Lens imp_programs; verdicts qualified / qualified.
+
+- **What survives.** Fatal bulk-only patients' CD14 monocytes carry more of this programme: +0.21, +0.36, +0.40, +0.65, +0.52 log2 (draws 1–4 q ≤ 0.005). The bulk shows the same (+1.49).
+- **Draws.**
+  - CD14 fatal: 1–4 (draw 0 fragile).
+  - Within severe/critical: none.
+  - Wave 1 after adjusting for WHO grade: none.
+- **(1) Calibration.**
+  - The gate passes 5/5 but compressed (0.26–0.47x).
+  - In each test set, 6 of 7 fatal donors are from wave 2, where "fatal" equals "severe/critical".
+  - Within severe/critical the check can be computed and fails: −0.02 to 0.17x, with draw 4 of the wrong sign (verify/IMP-PROG-6_calib/v01_calib.csv).
+- **(2) Replication.**
+  - 5/5 positive.
+  - Removed within wave 1 by adjusting for WHO grade at sampling (+0.01 to +0.21, p ≥ 0.12).
+  - Each draw's effect size follows its fatal training donor's dex score (Spearman 0.87–0.97, n = 5) (verify/IMP-PROG-6_recompute/v03h_imprinting.csv).
+- **(3) Observed.** Fatal +0.94 (p < 0.001); +0.53 (p 0.10) after adjusting for severity at sampling and wave.
+  - The fatal labels disagree across sources: the bulk metadata lists 7 wave-2 deaths, Gisby 2022 reports 4, and the single-cell metadata says "no" for C101, C146 and C147.
+  - The core glucocorticoid targets (TSC22D3, ZBTB16, PER1, KLF9, DDIT4) rise only in wave-2 fatal bulk (verify/IMP-PROG-6_calib/v02_bulk_gc_genes.csv).
+- **(4) Beyond bulk.** None (likelihood-ratio p 0.06–0.95).
+- **Verdict reasons.** This is a critical-illness monocyte programme, calibrated against dexamethasone-era truth. The generator misses the TSC22D3 rise, and the glucocorticoid core also rises in T, NK and B cells.
+- **New?** No. Dex-monocytes in wave 2: Stephenson 2025 PDF p.11. Glucocorticoids given to 8 of 17 wave-2 patients: Gisby 2022 PDF p.2. Precedents: COMBAT 2022 p.8; Bernardes 2020 p.15.
+
+#### D. B lineage and plasma cells
+
+**IMP-CLIN-1: more B-lineage cells, mostly labelled naive, in COVID+.** Lens imp_clinical; verdicts qualified / qualified.
+
+- **What survives.** Generated cells of bulk-only COVID+ patients contain more naive-labelled B cells than those of COVID− patients: logit +0.73, +0.71, +1.31, +0.98, +0.93 (1.8–2.7x; p 2e-8 to 8e-5). Total B rises +0.56 to +1.01 (verify/IMP-CLIN-1_recompute/v02_primary_focus.csv).
+  - It also holds in wave 1 and in the 45 donors without single cells.
+  - Restated by both verifiers: it is a B-lineage increase made mostly of naive-labelled cells, not a naive-specific shift.
+- **Draws.**
+  - Direction 0–4.
+  - Size 3–4 on all test samples.
+  - Size 0–4 on wave-1 test samples (i01: PASS 5/5, ratio 0.62–1.47), where the truth itself is only nominal.
+- **(1) Calibration.**
+  - calibration.csv: ratio 0.45–0.86, Spearman 0.54–0.62, contrast 0.27–0.69x (gates DIR, DIR, DIR, PASS, PASS).
+  - Wave-2 batch: 0.16–0.49x.
+  - Transitional B is under-generated (0.06–0.25x) (verify/IMP-CLIN-1_calib/v01_calib_rows.csv).
+- **(2) Replication.** 5/5 draws, q < 1e-3; robust to most confounders.
+  - Weakens with five bulk quality-control covariates (draw 3 p 0.31).
+  - Shared noise explains 6.6% of cross-draw agreement.
+- **(3) Observed.**
+  - +1.02 (p 7e-8) over all samples, which mixes processing batches; wave 1 +0.79 (p 0.037).
+  - In wave 1 the naive share within B does not change (+0.18, p 0.59); total B rises (+0.69, p 0.0014) (verify/IMP-CLIN-1_recompute/v05_observed.csv).
+- **(4) Beyond bulk.** Partly.
+  - Naive-B, pan-B and IGHD bulk markers do not move with COVID.
+  - But a ridge model on bulk genes, trained on paired samples, reproduces the shift (+0.76, p 1e-4) and tracks naive B better (0.73 vs 0.54–0.62). The generated shift left after it is +0.33 to +0.75 (verify/IMP-CLIN-1_calib/v03_ridge_pred.csv).
+  - Draws 0–3 partly turn the bulk plasmablast signal into naive B.
+- **Synthesis note.** IMP-LONG-2 showed that the labeller calls ISG-high switched memory B cells "naive". The week-1 within-B naive enrichment (+0.83, p 8e-6) coincides with peak interferon, so part of the "naive" attribution is probably that labelling error.
+- **New?** No.
+  - Stephenson 2025 PDF p.3 and Fig. 2E (PDF p.5): naive B higher in week 1 COVID+.
+  - Other cohorts report naive B reduced (COMBAT 2022 p.10) or unchanged (Mathew 2020 p.8).
+  - New only: bulk-only donors and donors without single cells reproduce it.
+
+**IMP-CLIN-2: plasma cells rise with WHO grade.** Lens imp_clinical; verdicts qualified / qualified.
+
+- **What survives.** Generated plasma cells rise with WHO grade at sampling among COVID+ patients: +0.51, +0.62, +0.44, +0.46, +0.40 logit per step, mostly between patients. Within wave 1 they also rise with peak severity (+0.79 to +1.42) (verify/IMP-CLIN-2_recompute/v03_assoc.csv).
+- **Draws.**
+  - WHO grade: direction 1–4, size 3–4 (gates FAIL, DIR, DIR, PASS, PASS).
+  - Wave 1: 1–2 PASS, 3–4 DIR.
+  - Draw 0 fails on level (0.21x).
+- **(1) Calibration.**
+  - calibration.csv Spearman 0.20–0.43.
+  - The WHO slope on COVID+ tests is reproduced: generated +0.33 to +0.71 vs observed +0.52 to +0.55 (|difference| ≤ 0.19).
+  - In wave 2 the per-sample tracking collapses (0.06–0.39) (verify/IMP-CLIN-2_calib/v05_gates.csv).
+- **(2) Replication.** 5/5 positive. Draw 3 is not significant at donor level (p 0.088–0.11).
+- **(3) Observed.** WHO +0.50 (p 0.007); wave 1 +0.69 (p 0.003); wave 2 +0.16 (p 0.56).
+- **(4) Beyond bulk.** None.
+  - A plasma-gene bulk score tracks plasma cells better (0.63–0.71 vs 0.20–0.45).
+  - Given that score, observed plasma cells have no WHO effect left (+0.11 SD, p 0.24), while draws 0, 1 and 4 keep one (+0.23 to +0.47). That residual is excess the truth does not have (verify/IMP-CLIN-2_recompute/v07_beyond_bulk_paired.csv).
+- **New?** No. Stephenson 2025 Fig. 3A (antibody-secreting B cells up at severe sampling; BH p 0.00031); Gisby 2022 module tL (p.5) and CD38-high plasmablasts (p.7).
+
+**IMP-PROG-5: bulk cell-cycle genes attributed to plasma cells.** Lens imp_programs; verdicts qualified / qualified.
+
+- **What survives.** Draws 1–3 attribute ten COVID+-up bulk cell-cycle genes to plasma-cell expansion, as observed cells do. The genes are MKI67, TOP2A, BUB1, ASPM, RRM2, TYMS, MYBL2, CENPF, TPX2 and KIF11.
+  - 29 of 30 gene × draw cells are composition-driven.
+  - The generated plasma share rises in every draw (p ≤ 6e-4) (verify/IMP-PROG-5_recompute/v08_summary.csv).
+- **Draws.** 1–3. Draw 0 gets 1 of 10 genes; draw 4 gets 4 of 10.
+- **(1) Calibration.**
+  - A constant "always plasma" guess also scores 10/10, so the check does not discriminate.
+  - Generated plasma cells are real plasma cells, but mostly not dividing (0.17–0.44x the observed cycling fraction). They carry 5–13% of the observed level of these genes (verify/IMP-PROG-5_calib/v03_identity.csv, v03_cc_level.csv).
+- **(2) Replication.** 10/10 genes in draws 1–3 under every sensitivity subset.
+- **(3) Observed.** Plasma cells carry all 10 genes (share 0.70–0.97).
+- **(4) Beyond bulk.** None. Deconvolution of the bulk with a reference atlas gives 9–10 of 10 (v04_deconv_summary.csv), and Gisby 2022 made the same attribution with CIBERSORTx.
+- **Other issues.** The gene list was hand-picked. Four hallmark cell-cycle genes in the top 30 were left out (BRCA1, LMNB1, MXD3, STMN1).
+- **New?** No. Gisby 2022 Fig. 4a "Plasma cell module", whose hub genes include RRM2, MKI67, TPX2 and BUB1 (PDF p.5, p.7).
+
+#### E. Platelets and dendritic cells
+
+**IMP-CLIN-3: more platelets with severity.** Lens imp_clinical; verdicts qualified / qualified.
+
+- **What survives.** Generated platelets are 1.5–2.9x more frequent in severe/critical than in mild/moderate COVID+ patients (+1.77, +1.25, +1.19, +0.71, +1.04 logit) and rise with WHO grade between patients. Observed cells agree (verify/IMP-CLIN-3_recompute/v02_family.csv).
+- **Draws.**
+  - Both waves: size in 1–2, direction in 3–4. Draw 0 fails (level 4.2x).
+  - Wave 1: no draw passes on level. Generated platelets are 2.4–19x too many, although the relative contrast is reproduced (0.74–1.82x) (verify/IMP-CLIN-3_recompute/v05_calibration.txt).
+- **(1) Calibration.** calibration.csv severity gates FAIL, PASS, PASS, DIR, DIR, the same among COVID+ only (i01).
+- **(2) Replication.** 5/5.
+  - Draw 3 is fragile.
+  - Adjusting for bulk platelet, LDG and neutrophil scores plus wave removes 48–72% of the effect.
+- **(3) Observed.** Peak severity +1.76 (p 2e-4); wave 1 +1.28 (p 0.008); within donors WHO +0.64.
+  - Observed platelets are about 10x more frequent in wave-2 samples (imp_clinical/s16_platelet_by_wave.csv).
+- **(4) Beyond bulk.** Partly.
+  - It goes beyond a platelet-gene bulk score, but that score is weak in bulk-only samples.
+  - Beyond granulocyte and platelet scores together, it is only partly reproduced and uncalibrated: paired truth keeps +0.80 to +0.97, generated draws 1–4 only +0.11 to +0.54.
+  - Generated platelets track bulk neutrophil RNA (Spearman 0.38–0.73); observed platelets do not (0.15–0.24).
+  - A 9-gene neutrophil score alone predicts severity at donor level as well (AUC 0.767 vs 0.62–0.76) (verify/IMP-CLIN-3_calib/v06_tracking.txt, v09_fullbulk.txt).
+- **Failed sub-claims.**
+  - "Does not track severity within a patient": the within-patient effect is positive in 5/5 draws and nominally significant in 3.
+  - "Platelets make up most of 'Other'": they are 30–42% of it in draws 1–3.
+- **New?** Not reported as a cell fraction for this cohort. Stephenson 2025 does not mention platelets; Gisby 2022 reports only pathways (p.5, p.10). The association is established in other cohorts' PBMC: COMBAT 2022 p.4, Bernardes 2020, Hottz 2020.
+
+**IMP-LONG-4: cDC2 share rises over days.** Lens imp_longitudinal; verdicts qualified / qualified.
+
+- **What survives.** Generated cDC2 share rises in bulk-only COVID+ samples (wave 1 +0.049 to +0.087 logit/day). This coincides with the observed wave-1 rise (+0.055, p 0.006) (verify/IMP-LONG-4_recompute/v09_summary.txt).
+- **Draws.** 3–4 under the all-wave time gate, 0–1 under the wave-1 gate; no draw passes both.
+- **(1) Calibration.**
+  - calibration.csv Spearman 0.12–0.34; draw 2 has the wrong contrast signs.
+  - Draws 0–2 create a rise in wave 2, where observed cDC2 is flat (+0.084 to +0.120, p ≤ 0.016) (verify/IMP-LONG-4_calib/v05_windows.csv, v05_gc.csv).
+- **(2) Replication.** 5/5 positive; 4/5 in donors without single cells.
+- **(3) Observed.** The rise is mostly between donors (within donors +0.049, p 0.13). It is part of a myeloid-wide rise: cDC2 relative to monocytes falls in wave 1 (−0.045, p 0.027), while generated cells raise it in 4/5 draws (verify/IMP-LONG-4_recompute/v06_compartments.csv).
+- **(4) Beyond bulk.** None. A 3-gene bulk score (CD1C, FCER1A, CLEC10A) tracks the truth better (0.42–0.48 vs 0.03–0.35) (verify/IMP-LONG-4_calib/v04_tracking.csv).
+- **New?** Partly reported. Stephenson 2025 Fig. 4A shows DC3 depleted in weeks 1–2 and back near baseline by week 3; Gisby 2022 p.6 reports fewer cDC2 by flow cytometry.
+
+#### F. Beyond the bulk (method)
+
+**IMP-PROG-2: a cell-type-specific COVID response that a bulk broadcast cannot give.** Lens imp_programs; verdicts qualified / qualified.
+
+- **Method.** The statistic "double-centres" the within-type logFC matrix: it removes each gene's mean and each cell type's mean, so a bulk change copied to every cell type leaves nothing behind.
+- **What survives.** For COVID+ vs COVID−, the generated double-centred pattern agrees with observed cells, beyond a bulk broadcast and csSAM deconvolution:
+  - all samples: L2 r 0.104–0.141;
+  - bulk-only donors with no paired sample: r 0.105–0.156;
+  - strict test (donor-disjoint, wave 1, sex flag dropped, training donors removed): r 0.05–0.09, p 0.005 in 5/5 draws (verify/IMP-PROG-2_recompute/v12_strict.csv).
+- **Draws.** COVID 0–4. Severity: none independent of shared patients. Fatal: none.
+- **(1) Calibration.** Paired-test r: COVID 0.146–0.196, severity 0.094–0.148, fatal 0.071–0.087. For fatal, a bulk broadcast matches it in 4/5 draws.
+- **(2) Replication.** 5/5 for COVID. The draws' matrices correlate 0.25–0.30 because of the shared seed.
+- **(3) Observed.** The observed pattern is noisy (split-half reliability 0.26 for COVID). The generated pattern captures 4–8% of its reliable variance, not "a quarter" (verify/IMP-PROG-2_calib/v05_share.csv).
+- **(4) Beyond bulk.** Yes, modestly.
+  - Partial r after pipeline broadcasts and reference-atlas predictors: 0.10–0.13 (p 0.003, donor-disjoint, 5/5) (v08_disjoint_beyond_bulk.csv).
+  - The pattern is interferon-led (13–33% of the covariance) and mostly between lineages (within-lineage r 0.01–0.09).
+- **Why qualified.**
+  - The severity and fatal agreement comes from the 29 donors present in both sets: the generator copies each donor's own deviations (same-donor r 0.17–0.20 vs about 0 for other donors) (verify/IMP-PROG-2_recompute/v07_fingerprint.csv).
+  - "A broadcast has zero structure by construction" is false: through the same log-CPM pipeline a broadcast gives r 0.03–0.15.
+  - Results depend on normalisation.
+  - Some "type-specific" genes are contamination-like: NKG7 in CD14 monocytes, HLA-DP in naive CD4.
+- **New?** Methodological. It replicates SEP-PROG-4 for disease vs control, at about half the sepsis strength (bulk-only r 0.10–0.16 vs 0.17–0.35).
+
+#### G. Artefacts
+
+**IMP-CLIN-5: T cells fall and CD14 monocytes rise with severity — a mapping artefact.** Lens imp_clinical; verdicts qualified / qualified.
+
+- **What it is.** In bulk-only samples the generated T share falls and the CD14 share rises with peak severity, WHO grade and fatality in 5/5 draws (T −0.49 to −1.07; CD14 +0.66 to +1.17 logit).
+  - Observed cells show neither for peak severity: T −0.06 (p 0.88), CD14 −0.07 (p 0.87).
+  - On the same paired test samples, the generator over-calls CD14 with severity in wave 1 in all five draws (+1.05 to +1.69) (verify/IMP-CLIN-5_recompute/v02_discordance.csv).
+- **Mechanism.** These are not neutrophil transcriptomes mislabelled as monocytes: generated cells almost never express neutrophil genes (v05_ldg_cell_labels.csv).
+  - The generator reads the bulk S100A8/9/12 alarmin rise, which comes from LDGs and activated monocytes, as extra classical monocytes, and dilutes the T cells.
+  - Monocyte-restricted bulk genes (CD14, VCAN, FCN1, CSF1R, MAFB) do not rise with severity.
+  - Adjusting for the bulk alarmin score removes nearly all of the CD14 effect (verify/IMP-CLIN-5_calib/v02_B_adjusted.csv).
+- **Draws.** No cellular finding. CD14 artefact in 0–4; T artefact cleanly in 2–3.
+- **Failed sub-claims.**
+  - "The generator makes no severity effect on paired tests" is wrong.
+  - "Observed shows neither" overreaches for fatal outcome, and for within-donor WHO grade in wave 1, where observed CD14 rises (+0.66 per step, p 0.011) (verify/IMP-CLIN-5_calib/v05_within_donor_loo.csv).
+- **Consequence.** Adjust every monocyte, T-cell and platelet association for bulk alarmin, LDG and T-cell scores.
+- **New?** The ingredients are documented in the cohort papers: Gisby 2022 p.5 (LDG signature in the bulk), p.18 (fresh-PBMC bulk; LDGs lost on freeze–thaw); Stephenson 2025 p.20 (thaw protocols by wave). The generator artefact itself is new.
+
+**IMP-LONG-5: platelet decline over days — a generator artefact.** Lens imp_longitudinal; verdicts qualified / qualified.
+
+- **What it is.** The generated platelet share in wave-1 bulk-only samples falls by −0.112, −0.031, −0.063, −0.059, −0.035 logit/day (q < 0.03 in draws 0, 2, 3). Observed platelets are flat (+0.006/day in wave 1).
+- **Calibration.** No draw passes a wave-1 time gate: levels are 2.4–19x too high.
+- **Mechanism.** The generator ties platelets to the bulk ISG and granulocyte scores, which both fall over time. Adjusting for them removes the decline in all draws (verify/IMP-LONG-5_calib/v05_gate_summary.csv; verify/IMP-LONG-5_recompute/v05_coupling_diff.csv).
+- **Limits of the evidence.**
+  - A direct excess on identical paired samples is shown only for draw 0 (−0.080/day, p 0.0035).
+  - Observed wave-1 platelets are near zero (median 2 cells per sample), so "flat" rules out only the steep draws (0, 2, 3).
+  - The early-severity direction is calibrated and is not an artefact.
+
+**IMP-POP-5: a generated-only lineage-mixed "T cell" cluster.** Lens imp_populations; verdicts qualified / qualified.
+
+- **What it is.** Cluster T r1.0 c1:
+  - HLA-DR/DP/DQ+, CD74+, CD79A+, low library;
+  - 75% of its cells lack CD3E/CD3D/CD2;
+  - over-produced 11–19x;
+  - fails calibration in all 20 draw × contrast rows;
+  - recurs at the same noise slots across draws (odds ratio 112 vs 1.7) (verify/IMP-POP-5_calib/tables/v06_per_draw_summary.csv).
+- **Clinical rise.** It is mostly the generated T-cell denominator. After adjusting for the T share, fatal holds in 0/5 draws and severity in 1/5.
+- **Recommendation.** Exclude these cells.
+- **Two smaller mixed clusters differ** (verify/IMP-POP-5_recompute/tables/v11_side_tests.csv). NK c15 (NK cells with myeloid genes) and Mono c4 (monocytes with NK genes):
+  - have observed counterparts that rise with COVID (p 0.0045, 0.0008);
+  - carry platelet genes, which suggests cell aggregates;
+  - Mono c4 passes the gate in draw 3.
+
+## 3. Candidate new populations
+
+None is new. Sources: imp_populations/tables/s05_candidates.csv, s06_stability.csv, s04_screen_r1.0.csv, s10_dexmono_summary.csv, s11_artefact_coexpression.csv, s12_shared_noise.csv, s13_rule_markers.csv, and the verifier tables in section 2.
+
+| Candidate | Markers | Observed support | Artefact risk | Verdict |
+|---|---|---|---|---|
+| ISG-high, HLA-DR-low CD14 monocytes (IMP-POP-1; Mono r2.0 c12, part of r1.0 c0) | IFITM1/2/3, MT2A, PLAC8, XAF1, EPSTI1, RNASE2, APOBEC3A; HLA-DP/DR, CD74 and AP-1 genes down | severity +0.180 (p 0.0011), but the HLA-low shift is global | medium: HLA axis uncalibrated; cluster unstable (Jaccard 0.25) | overlap of two known axes |
+| CD163+ CD14 monocytes (IMP-POP-2) | CD163; FKBP5, RNASE2, PLBD1 (generated); IL1R2, THBS1, MCEMP1 (observed only) | severity +0.217 (p 2.6e-4) | medium–high: glucocorticoid confound; a level shift, not a state | known (Stephenson 2025; COMBAT 2022) |
+| HLA-II-high CD14 monocytes (IMP-POP-3; Mono r1.0 c6, Jaccard 0.59–0.63) | HLA-DP/DQ/DM, CD74, CPVL up; S100A8/9/12 down | WHO −0.047 per grade (p 3e-5) | high in bulk-only cells: dilution by CD14 over-production | known |
+| ISG-high states in all lineages (IMP-POP-4; Jaccard 0.55–0.76) | 10-gene ISG rule | strong in every lineage | low | positive control |
+| Generated-only lineage-mixed "T" (IMP-POP-5; T r1.0 c1) | HLA-II, CD74, CD79A, S100A8/9, LYZ; no CD3 | none (8 CD3-negative mixed cells, 6 COVID+ donors) | artefact | exclude |
+| NK r1.0 c15 / Mono r1.0 c4 | myeloid genes in NK / NK genes in monocytes; platelet genes | observed COVID rise (p 0.0045 / 0.0008) | over-produced 1.2–9.5x; probably aggregates | lead only (Mono c4 gated in draw 3) |
+| IFN-negative HLA-II-low (MS1-like) CD14 monocytes (Mono r2.0 c17) | S100A, HLA-II low, no ISG | observed falls in COVID (−0.078, p 0.0023) | over-produced 1.66–2.12x; unstable | negative: not expanded in this cohort |
+| Dex-monocytes (CD163 + ADAMTS2/VSIG4/SLC1A3/MFGE8/MRC1) | per Stephenson 2025 | 11.9% of wave-2 severe monocytes vs 1.1% COVID− | generated at 0.02–0.09x | not reproduced |
+| C1Q+ CD16 monocytes | C1Q genes | - | C1Q-high state ratio 0.06–0.60, Spearman about 0 | untestable |
+| Activated HLA-DR+ NKG7+ GZMH+ CD8 T (T r2.0 c32) | HLA-DR, NKG7, GZMH | no association | unstable (Jaccard 0.16–0.25), 2.6x over | negative |
+| Effector/atypical B (B r1.0 c13) | FCRL5, FGR, FCRL3 | no association | Spearman about 0 | negative |
+| S100A8/9+ DC3-like cDC2 (DC r1.0 c3) | S100A8/9 | fatal +0.081 (p 0.006) | about 5 DCs per sample; no embedding-free check possible | lead only |
+| Platelet sub-cluster (Other r1.0 c0) | TSC22D1, HIST1H2AC, TUBB1 | fatal +0.28 (p 0.013) | about 3 parent cells per sample | untestable |
+| IEG-high NK (NK r1.0 c7) | FOS, JUN, DUSP1 | lower in COVID+ (p 0.0013) | sample handling | not a population |
+| Re-found L2 labels: CD16 monocyte loss, memory CD4/CD8 subsets, plasma cells | - | yes | - | not new |
+
+The analyses also corrected the identity of three labels:
+- "Proliferating NK" is mostly ISG-high, non-dividing NK cells (IMP-LONG-3).
+- "CD8aa" marks CD8B+ IL7R+ GZMK+ memory-like CD8 cells (IMP-CLIN-4).
+- AIFI "naive B" absorbs ISG-high switched memory B cells (IMP-LONG-2).
+
+The two cell sets separate in a plain joint embedding, and generated monocytes detect fewer genes than observed ones (226–315 vs 387 per cell). Every reported population was therefore re-checked with a marker rule that does not use the embedding.
+
+## 4. Refuted and negative results
+
+### 4.1 Refuted
+
+**IMP-CLIN-4: CD8αα T cells fall with severity.** Verdicts: recompute qualified, calib refuted.
+- The bulk-only drop of CD8aa as a share of all cells (−0.38, −0.59, −0.41 in draws 1, 2, 4) is a T-cell-wide fall. CD8aa per T cell does not fall in any draw (+0.20, −0.16, +0.03, +0.24, −0.08).
+- Observed cells show the opposite structure: the T share is unchanged (−0.06, p 0.88), and the CD8aa share within T cells falls (−0.73, p 3e-4) (verify/IMP-CLIN-4_calib/v04_models.csv, v05_tsubset_controls.csv).
+- Within-T calibration fails in draws 1 and 3 and is direction-only (0.23–0.30x) in draws 0, 2, 4 (v07_withinT_calibration.csv).
+- A bulk T-cell score explains 37–100% of the generated effect.
+- In this cohort the label marks CD8B+ memory-like cells, not innate-like CD8αα cells (v08_label_biology.txt).
+- **What survives:** an observed single-cell result (CD8aa share of T falls with peak severity, 35 donors), not a bulk-only finding.
+
+**IMP-LONG-2: the memory share of B cells rises over the illness.** Verdicts: refuted, refuted.
+- The labeller reads acute interferon as "naive". Adding acute-phase ISG counts to late B cells flips 41–64% of memory labels to naive; a matched non-ISG swap flips 1–14% (verify/IMP-LONG-2_recompute/v05_flips_obs_bo.csv).
+- 45% of isotype-validated IgG/IgA switched memory cells in the top ISG quartile are called naive (verify/IMP-LONG-2_calib/v03_crosstab.csv).
+- With the cohort authors' CITE-seq/BCR labels, the wave-1 memory share does not rise (+0.015/day, p 0.15). The generated wave-1 slope is 2.8–4.2x that truth in every draw (v03_truth_slopes.csv, v03_gen_vs_author.csv).
+- **What survives:** AIFI memory/naive B labels track the bulk interferon clock. A real rise exists only in wave 2 (+0.075/day, 15/15 donors).
+
+### 4.2 Sub-claims that failed inside qualified findings
+
+- **IMP-PROG-4:** the cross-type ordering of the ISG response; "the severity ISG difference does not carry over".
+- **IMP-LONG-1:** "falls steadily over 3 weeks"; "ISG-high T in every draw".
+- **IMP-POP-4:** "severity reproduced only in monocytes"; "every lineage" (cDC2 holds in 4/5 draws).
+- **IMP-PROG-1:**
+  - "B cells unchanged" (observed B HLA-II rises in wave-1 COVID+ and falls in fatal patients);
+  - cDC2 involvement for COVID;
+  - "progressive loss" for severity.
+- **IMP-POP-1:** "a coupled IFN+ HLA-II-low sub-state"; "ESKD differs from non-ESKD".
+- **IMP-PROG-3:**
+  - "all draws, all three monocyte subsets";
+  - the MS1 genes ALOX5AP, CYP1B1 and RETN;
+  - "S100 adds to bulk";
+  - fatal as a separate contrast.
+- **IMP-POP-2:** "mostly HLA-II-low"; "FKBP5+ RNASE2+ PLBD1+ state"; "not dexamethasone-induced".
+- **IMP-PROG-6:** TSC22D3 and VSIG4 rise; a glucocorticoid signal in wave-1 fatal patients; "within-severe cannot be calibrated" (it can, and it fails).
+- **IMP-CLIN-1:** "the bulk adds attribution to naive B"; the lens's "total-B" bulk score is a plasma-cell score.
+- **IMP-CLIN-2:** "draws 1–4 calibrated" (size only in 3–4); the beyond-bulk WHO residual in draws 0, 1, 4.
+- **IMP-CLIN-3:** "does not track severity within a patient"; "platelets are most of 'Other'".
+- **IMP-PROG-2:** "shared across COVID, severity and fatal"; "a quarter of the reliable signal"; "zero by construction".
+- **IMP-PROG-5:** "all composition-driven" (29/30).
+- **IMP-LONG-3:** "proliferating NK cells fall"; "peak in days 0–7" in wave 1.
+- **IMP-LONG-4:** "rises from week 1 to weeks 2–3"; "the overshoot is beyond observed cells"; "recovery not reported".
+- **IMP-LONG-5:** "the early-severity effect is an artefact".
+- **IMP-CLIN-5:** "the generator makes no severity effect on paired tests"; memory CD8 and gdT do not fit the same pattern.
+- **IMP-POP-5:** "the companion clusters behave the same way".
+
+### 4.3 Negative results from the lenses
+
+1. **Mortality.**
+   - No generated feature is linked to fatal outcome once WHO grade at sampling is adjusted for (0 of 50 at q < 0.05 in every draw), and none within severe/critical patients.
+   - The observed fatal-beyond-WHO signals (Treg −2.51, platelets +1.96, memory CD4 −2.40) come from wave-2 donors, and the generator does not reproduce them (imp_clinical/s08_extra_models.parquet, s13_extra_summary.csv).
+2. **Within-patient severity changes** are not detected for any feature; all bulk-only WHO-grade associations are between patients.
+3. **No predictive gain over the bulk.** Adding generated features to a model with 10 bulk principal components plus age and sex changes leave-one-donor-out AUC from 0.743 to 0.723–0.806 (severity), with no consistent gain (imp_clinical/s12_fullbulk_cv.csv).
+4. **T-cell subset losses in COVID+** (T, CD8aa, memory CD4, MAIT, gdT) do not hold in wave-1 observed cells; memory CD4 even reverses (+0.42, p 0.003).
+5. **TNF loss in monocytes**, Stephenson's top severity gene (observed −0.86 to −1.23), is not reproduced, and its truth comes from wave 2 (imp_programs/programs_summary.csv).
+6. **Severity ISG in NK, B and T cells** is attenuated in bulk-only cells because severe bulk-only samples were taken later. Generated monocytes also fail to show the monocyte recovery seen in wave-1 observed cells (+0.08 to +0.11 logit/day) (imp_longitudinal tables s15–s16).
+7. **Granulocyte (LDG) genes** cannot be assigned to a generated cell type (LDG state ratio 0.01–0.03). Reading them as monocyte expression is the IMP-CLIN-5 artefact.
+8. **Attribution of severity and fatal bulk genes** to cell types does not beat a reference atlas or a constant guess (imp_programs/attribution_calib.csv).
+9. **Negative control.** The "platelet signal in monocytes" state is associated with severity and WHO grade in bulk-only cells (severity q ≤ 0.007 in 4/5 draws; WHO q ≤ 0.002 in 4/5). Yet every gate is NULL_genfx or FAIL: observed cells show no WHO or severity effect (WHO +0.02, p 0.91) (imp_clinical/s05_summary.csv; synthesis_imperial_eskd/i05_spotcheck.txt). Observed fatal patients do show more of it (+1.11, p 0.007), which fits platelet–monocyte aggregates, but the generator fails that check in 5/5 draws. Strong but uncalibrated generated associations are not findings.
+10. **Sex.** CD8aa is lower in males in generated and observed cells. This is treated as a labeller artefact: chrY genes are labeller inputs (SEP-CLIN-6).
+11. **Other trajectories the generator imposes but observed wave-1 cells lack:** six-type NK decline, proliferating T decline, memory CD4 rise in draws 0 and 4.
+12. **Rare states that fail calibration** carry no claim: C1Q-high monocytes, LDG, cycling monocytes, DN2/ABC B, HSPC, cycling activated T.
+13. **Plasmablast, T-activation, T-exhaustion and ILC-family programmes** cannot be reported: observed contrasts are null or gates fail.
+14. **Gene-length axis.** Generated within-type COVID logFC correlates with gene length in draws 0–2 (Spearman 0.07–0.21). So do observed fatal logFCs (0.14–0.24). The long "bulk-discordant" fatal genes are suspect (imp_programs/tech_length.csv).
+
+### 4.4 Dendritic cells and NK cells
+
+Source: synthesis_imperial_eskd/i03_dc_nk_clinical_screen.csv, from imp_clinical/s05_summary.csv.
+
+- **Six-type DC:** no calibrated association in any contrast.
+  - Draw 2 alone rises with COVID, severity and fatal outcome (+0.56 to +0.82, q 0.03–0.04). Observed cells do not show it; it is an artefact of draw 2.
+- **cDC2:** observed cDC2 falls with COVID (−0.35, p 0.038), severity (−0.71, p 0.012) and WHO grade (−0.32, p 0.0098).
+  - Only draws 0 and 3 (COVID) and draws 3 and 4 (severity, WHO) pass the gate, and none is significant.
+  - Draw 2 goes the opposite way (+0.56 to +0.60).
+  - The observed cDC2 loss is not reproduced in bulk-only cells. Its over-time recovery is covered in IMP-LONG-4.
+- **pDC, cDC1, ASDC:** nothing.
+- **NK cells:**
+  - Generated six-type NK and CD56dim NK fall with severity, WHO grade and fatal outcome in draws 0, 2 and 4 (CD56dim WHO −0.23 to −0.32, q ≤ 0.009).
+  - Observed cells show no such fall (CD56dim WHO +0.15, p 0.096). The gates FAIL or are NULL: this is lymphoid dilution, as in IMP-CLIN-5.
+- **CD56bright NK:**
+  - Observed CD56bright NK falls with severity (−0.63, p 0.005; wave 1 −0.79).
+  - The generated fraction does not track it (Spearman −0.13 to 0.25); only draw 1 passes the gate.
+- **Proliferating NK:**
+  - Rises in COVID+ (PASS in 4/5 draws; observed +1.03, p 2e-11).
+  - But the label is mostly ISG-high NK (IMP-LONG-3), and the bulk carries it.
+
+## 5. What would turn each qualified finding into a confirmed one
+
+For most findings, "confirmed" can only mean a confirmed reproduction of a known, bulk-visible effect.
+
+### 5.1 Steps that help every finding
+
+1. **A corrected, wave-matched trust layer.** Rebuild calibration.csv with:
+   - severity among COVID+ only, which fixes the recovery bug;
+   - wave-1 rows;
+   - donor-aware contrasts with confidence intervals;
+   - WHO-grade, time, severity × time and fatal-within-severe rows;
+   - the programmes and states the lenses used: HLA-II, S100/MS1, ISG per lineage, CD163.
+
+   synthesis_imperial_eskd/i01_calibration_summary.py is a template.
+2. **Wave-1 single-cell truth.** 88% of the bulk-only samples are wave 1, but the severe and fatal truth is mostly wave 2, and there is only 1 fatal paired donor in wave 1. Single cells from frozen wave-1 PBMC of bulk-only donors, if such samples exist, would test the predictions directly.
+3. **Handle the fresh-bulk vs frozen-cell mismatch.**
+   - Report every monocyte, T-cell and platelet association adjusted for bulk alarmin, LDG and T-cell scores, and for the generated parent-cell share.
+   - Longer term, give the generator a target that knows about granulocytes.
+4. **Interferon-robust labels.**
+   - Use the authors' CITE-seq/BCR labels for B subsets, as verify/IMP-LONG-2_calib did, or a reference that includes ISG-high memory B cells.
+   - Use marker rules for "Proliferating NK" and "CD8aa".
+   - Remove chrY and XIST inputs before any sex analysis.
+5. **Clinical metadata.** Steroid exposure per sample, reconciled fatal labels, and the processing site of each sample.
+6. **More, designed draws.** These need GPU time, which is the coordinator's call.
+   - At least 2 sampling seeds per draw, to separate shared noise.
+   - Draws trained only on wave-1 donors.
+   - Draws without the steroid-era CD163-rich donors (C33, C108).
+7. **Fixed bulk baselines for every claim.** An own-gene score, a lineage-normalised score, a ridge model trained on paired samples, deconvolution, and a constant prior, with per-draw partial correlations.
+
+### 5.2 Per finding
+
+- **IMP-PROG-4, IMP-POP-4, IMP-LONG-1:** keep as positive controls.
+  - Report lymphoid states and ISG-high T only in the calibrated draws (0–2).
+  - State severity within a fixed illness window (≤ 14 days) or adjusted for days since swab.
+  - Test any cell-type ordering on a common sample set without IFI27.
+- **IMP-LONG-3:** rename the feature "IFN-stimulated NK (label)", use marker-defined cycling NK for proliferation, and match generated depth over time.
+- **IMP-PROG-1:**
+  - Report severity, with intermediate monocytes and cDC2 as the robust carriers and a CD14-fraction-adjusted size.
+  - Retest B-cell sparing with a normalisation-robust score.
+  - Obtain monocyte or B-cell HLA-DR flow data for bulk-only patients, if they exist.
+- **IMP-POP-1:** report the ISG and HLA-II axes separately and calibrate HLA-II per cell before any joint-state claim.
+- **IMP-POP-3:** report absolute counts and share-adjusted values; calibrate on wave 1 with several thresholds.
+- **IMP-PROG-3:**
+  - Report CD14 only, with a count-adjusted size and the observed-supported genes (S100A8/9/12, MCEMP1, PLAC8, CD163, IL1R2, CLU).
+  - Fix IL1R2 and MS1-state under-generation.
+- **IMP-POP-2, IMP-PROG-6:** get steroid exposure and reconciled fatal labels, use a depth-robust CD163 level, and adjust for WHO grade at sampling. Draws trained without steroid-era donors would show whether the effect survives.
+- **IMP-CLIN-1:** retest the naive attribution with author B labels, and report the ridge-adjusted generated residual per draw.
+- **IMP-CLIN-2:** restrict to wave 1 and report draws 3–4. Fix plasma under-generation in draw 0. Show a residual beyond the bulk plasma score that the truth also has.
+- **IMP-PROG-5:** fix the non-dividing plasma-cell output, use an objective cell-cycle gene set, and score against a constant prior and deconvolution.
+- **IMP-CLIN-3, IMP-LONG-5:** obtain site-adjusted wave-1 platelet truth. Require the effect beyond bulk platelet and granulocyte scores on paired truth.
+- **IMP-LONG-4:** calibrate cDC2 relative to monocytes and within donors, and separate the generator's ISG and glucocorticoid dependence.
+- **IMP-PROG-2:**
+  - Make the donor-disjoint, wave-matched double-centred r with a pipeline-broadcast null the standard beyond-bulk metric.
+  - Use a pseudocount-free estimator on well-expressed genes.
+  - Severity and fatal versions need paired donors who are not in the bulk-only set.
+- **IMP-CLIN-5, IMP-POP-5:** these are artefacts to fix, not findings to confirm.
+  - Condition on, or remove, the alarmin/LDG bulk signal.
+  - Filter generated cells that lack the lineage genes of their label (for example, no CD3 in "T").
+- **IMP-CLIN-4, IMP-LONG-2 (refuted):** revisit only with a within-T calibration row and marker-confirmed CD8aa identity, and with author B labels.
+
+## Provenance
+
+- **Scripts** (synthesis_imperial_eskd/):
+  - i00_counts.py: generation counts.
+  - i01_calibration_summary.py: trust-layer summary; reproduces calibration.csv to within 5e-6 and adds the COVID+-only and wave-1 restrictions.
+  - i02_extract_graded.py: verbatim findings and verdicts from workflow wf_d188d252-beb. The journal's sha256 (7929331f…b2) is in i02_graded_input.json. All 1,500 cited files exist (i02_cited_files.csv).
+  - i03_checks.py: the scale='X2' search, DC/NK screen rows, shared noise.
+  - i03b_recovery_and_training_samples.py: recovery-sample severity labels and training-donor samples.
+  - i04_build_findings_json.py: writes findings_imperial_eskd.json and asserts that every cited source exists.
+  - i05_spotcheck.py: checks headline verifier numbers against the verifiers' own tables. All matched.
+- **Logs and tables** sit next to the scripts. Every other number comes from the lens and verifier files cited above.
+- **Compute.** CPU only, run with systemd-run MemoryMax=6G and no swap. No GPU work.
+- **Writes.** Only synthesis_imperial_eskd/ and findings_imperial_eskd.json. REPORT_imperial_eskd.md is saved by the coordinator; its content is this text.
+- **Archive.** No PBISC-Diffusion archive delta from this agent. Registration under paper-release-20260928/maintenance/POLICY.md is left to the coordinator.
+
+
+## Completeness review
+
+The completeness critic wrote this for imperial_eskd on 2026-10-06. `REPORT_imperial_eskd.md` did not exist yet, so I took the report text from the workflow journal and saved a copy as `critic_imperial_eskd/c00_report_text_from_journal.txt`. All work was CPU only (systemd-run, MemoryMax=6G, no swap). The scripts and tables (c00–c06) are in `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/critic_imperial_eskd/`. This agent made no archive change.
+
+### A. Number spot-check: 58 statements against their sources (`c01_spotcheck_verdicts.csv`)
+
+55 match, 2 are wrong and 1 is only partly right. Where I could, I recomputed values from meta.csv, composition.parquet and calibration.csv rather than reading them from the synthesis tables.
+
+**Wrong or partly right:**
+- **"Each test set holds about 13–14 severe/critical donors."** It is 12–13. The part about 7 fatal donors, 6 of them from wave 2, is right.
+- **Negative result 3, "AUC from 0.743 to 0.723–0.806 (severity)".** The severity models give 0.726–0.781. The 0.806 comes from a fatal model, whose own baseline is 0.770 (`imp_clinical/s12_fullbulk_cv.csv`).
+- **IMP-POP-5, "over-produced 11–19x".** This holds only in the discovery embedding. The verifier gives 2.1x (kNN-mapped) to 16x depending on the definition, 3–60x per draw, and a calibration ratio of 1.9–31.
+
+**Right but incomplete:**
+- The transfer shifts for severe/critical samples (+0.69 granulocyte, −0.82 ISG) have BH q 0.21. The report gives only p.
+- IMP-PROG-2's "p 0.005 in 5/5 draws" is the smallest p a 200-permutation test can give (1/201). With the other two gene filters, p is 0.01–0.12, and 4 of those 10 values are above 0.05.
+- IMP-PROG-6's "Spearman 0.87–0.97, n = 5" rests on 4 fatal training donors, because C101 trained both draws 1 and 2.
+- Section 4.4 says only draw 1 passes the gate for CD56bright NK. That is true under the lens gate. Under the synthesis's own wave-1 gate, draws 1 and 4 both carry it (bulk-only q 3e-4 and 2e-4).
+- "Plasmablast (label)" is the same feature as "l2 | Plasma cell": their rows in calibration.csv are identical. So the 50 trust-layer features are 49 distinct ones, and the pooled tallies count plasma cells twice.
+
+**Everything else matched**, including:
+- all counts in section 1.1;
+- all 45 numbers in the per-draw trust-layer table;
+- the pooled counts: 24 of 50 pass the level gate, 8 have Spearman ≥ 0.5, 12 are below 0.2, and 31 and 24 keep their sign in 5/5 draws;
+- the per-draw notes:
+  - draw 0: plasma 0.23x, platelets 3.66x, MS1 0.045x;
+  - draw 1: NK 0.40x;
+  - draw 2: cDC2 Spearman 0.12, with wrong signs;
+  - draw 3: 0.29x and 0.37x;
+  - draw 4: MS1 0.37x, and wave-1 Spearman −0.15 and −0.09;
+- training-donor plasma 0.28% vs 1.16%;
+- C141 at 7.6% platelets and C65 at 8.5% plasma cells (both are means of per-sample fractions);
+- shared noise of 22% and 4%, and the odds ratio of 39–86 vs 5–14;
+- the per-draw headline values of IMP-PROG-1/3/4/5/6, IMP-LONG-1/2/3/4/5, IMP-POP-1/2/4 and IMP-CLIN-1/2/3/4/5, and IMP-PROG-2's r and share.
+
+### B. Gated associations the report never examines (`c02_gated_unreported.csv`)
+
+I merged two inputs:
+- the clinical lens's per-draw bulk-only models (`imp_clinical/s02_bulkonly_assoc.parquet`);
+- the synthesis's own corrected gates (`i01_codes_by_feature.csv`, R0/R1/R2).
+
+A draw "carries" a feature when its gate is PASS or DIR, its bulk-only estimate has the observed sign, and BH q < 0.05. Five such results appear nowhere in the findings, in "What is reproduced" or in the negative results.
+
+1. **Proliferating T cells rise in COVID+ samples.** This is the cleanest reproduction in the cohort outside interferon.
+   - **Draws.** Draws 0–4 carry it under the official gate and under the wave-1 gate: bulk-only +0.62 to +1.05 logit, q ≤ 0.026. It survives dropping the sex-flag sample and the training donors.
+   - **Within T cells.** It also holds as a share of T cells (+0.85 to +1.55), including in wave 1 (`c04_within_T_checks.txt`).
+   - **Observed.** +0.65 (p 4e-4) over all samples and +1.39 (p 3e-8) in wave 1. On wave-1 test samples the generated effect is +0.88 to +1.35 against an observed +1.31 to +1.49.
+   - **New?** No. Stephenson 2025 reports transient activated T cells (Fig. 4I).
+   - **Beyond bulk (`c06_prolifT_beyond_bulk.txt`).**
+     - A bulk cell-cycle score tracks the observed fraction better than the generated cells do (Spearman 0.47–0.52 vs 0.35–0.49).
+     - After adjusting for bulk cell-cycle, plasma and T scores, observed wave-1 cells keep +0.67 (p 0.033).
+     - Generated cells keep it only in draw 3 (+0.64, p 0.018). Draws 0, 1, 2 and 4 fall to +0.02 to +0.26, not significant.
+   - It belongs in the report as a positive control that the bulk carries.
+2. **MAIT cells fall with severity.**
+   - **Draws.** The COVID+-only gate is DIR in draws 0–3. Bulk-only, the MAIT share of all cells falls by −0.37 to −0.88, with q ≤ 0.002 in draws 1–3.
+   - **Observed.** −1.22 (p 0.0015) in wave 1, and the fall holds within T cells (−1.17, p 0.005).
+   - **The IMP-CLIN-4 lesson was not applied.** Within T cells, the bulk-only effect survives only in draw 1 (−0.66, p 0.001). Draws 0, 2, 3 and 4 give −0.01 to −0.27, not significant. So most of it is the same T-cell dilution that refuted CD8aa.
+   - On wave-1 test samples, draw 1 does reproduce the within-T loss: −1.14 against an observed −1.23.
+3. **Effector B cells rise in COVID+ samples.**
+   - **Draws.** Draws 0, 2 and 4 carry it (+0.83, +1.10, +0.95; q ≤ 2e-4). The gates are PASS or DIR in 5/5 draws, also in wave 1.
+   - **Observed.** +0.47 (p 0.004) over all samples and +0.62 (p 0.012) in wave 1.
+   - **Not reconciled.** Section 3 calls the effector/atypical B cluster "no association, negative", but this label-level result says otherwise.
+4. **DN T cells rise in COVID+ samples in wave 1.**
+   - **Draws.** The wave-1 gate is PASS or DIR in 5/5 draws, and draws 0–2 carry it (q ≤ 0.018).
+   - **Observed.** +0.84 (p 0.006) in wave 1.
+   - **How the report treats it.** It lists DN T only as "over-produced", which is a level problem, not a contrast.
+   - **Possible label artefact.** The IMP-POP-5 verifier found that the observed CD3-negative mixed cells are mostly JCHAIN-high cells labelled "DN T cell". The label may therefore follow the plasma-cell rise. This was not checked.
+5. **Plasma cells are higher in men (among COVID+ samples).**
+   - **Not reported.** The lens screened sex for every feature, but the report mentions only CD8aa.
+   - **Observed.** 2.19% in men vs 0.47% in women, +0.94 logit (p 0.018).
+   - **Not a severity proxy.** Men are more often severe (79% vs 52% of paired donors), but the effect survives adjustment for WHO grade and peak severity: +0.85, p 0.026 (`c03_trustlayer_checks.txt`).
+   - **Draws.** Draws 3 and 4 reproduce it after the same adjustment (+0.66 and +0.58; p 0.026 and 0.002). Draws 0–2 do not.
+   - **Not simple imprinting.** Draw 1 was trained on a plasma-rich man (C65, 8.5% plasma cells) and a plasma-poor woman (C31), yet it does not carry the effect (`c05_sex_plasma_training_donors.txt`).
+   - **Status.** It is a lead only. Before it counts, it needs the report's own rule for sex analyses: remove the chrY and XIST genes from the labeller's inputs.
+
+### C. Other analyses not run
+
+- **Age.** The clinical lens screened age for all 50 features, and the report never mentions the results.
+  - Only memory B cells (draw 3) give a gated, significant result.
+  - Observed B cells fall with age (−0.25 per decade, p 0.004), and the gate passes in 5/5 draws, yet bulk-only cells show no fall (q ≥ 0.3).
+  - The observed effect comes from wave 2 (wave 1: −0.05, p 0.66). The report should state age as a negative result.
+- **Comorbidities.** `data/imperial_eskd/bulkonly_meta.csv` holds cause of ESKD, ischaemic heart disease, COPD and smoking. None of them was used, not even as a confounder in the mortality analysis. Diabetes and ethnicity were used only as covariates.
+- **Pre-infection "COVID−" samples.**
+  - The bulk-only COVID− group contains pre-infection samples of four donors who later tested positive.
+  - C105_negative was taken 7 days before the first positive swab and has the highest bulk ISG of any COVID− sample.
+  - The other three, C101, C141 and C33, are training donors of draws 1–2, 0 and 4.
+  - Only the IMP-POP-4 verifier dropped C105_negative. Every other COVID+ vs COVID− contrast keeps all four.
+- **WHO-grade and fatal gates in wave 1.** The wave-1 recalibration (R2) covers only the COVID and peak-severity contrasts. The WHO-grade claims (IMP-CLIN-2, IMP-CLIN-3, IMP-POP-3) still rest on gates over both waves, and the truth behind those gates is mostly wave 2.
+- **The 17 wave-2 bulk-only COVID+ samples** were never compared with the wave-2 truth. That is the only place where dexamethasone-era and CD163 truth exists (IMP-PROG-6, IMP-POP-2).
+- **CD16 monocytes over time.** A lens negative result is missing from the report. Observed CD16 monocytes rise by +0.114/day in wave 1 and fall by −0.072/day in wave 2, and the generator reproduces neither course.
+
+### D. Trust-layer problems the report understates
+
+1. **The recovery-sample bug is worse than described.**
+   - The severity test in calibration.csv compares donors. In every draw, 6–7 donors sit in both arms: their mild/moderate COVID+ visits in the mild arm and their recovery visit in the severe arm. The donors are C140, C168, C169, C170, C187, C190 and C58.
+   - The "severe" arm has 19–20 donors, but only 12–13 of them are severe COVID+ donors (`c03_double_arm_donors.csv`).
+2. **The two trust-layer implementations disagree, and the report quotes both.**
+   - Section 1.3 says the generator invents the six-type T-cell COVID contrast (NULL_genfx in 5/5 draws). Section 4.3 item 4 follows the lens, which says these T-cell losses "pass calibration on all paired test samples", and treats them as real but wave-2-only.
+   - **The cause is visit weighting.** Wave-2 donors have 4.1 COVID+ visits each, against 2.6 in wave 1. They make up 52–57% of the test samples but only 41–44% of the donors.
+     - Averaged per donor, the observed difference is −0.03 (Mann-Whitney p 0.52–0.63).
+     - Averaged per sample, it is −0.08 (`c03_tcell_weighting.csv`).
+   - **The IMP-CLIN-5 verifier tables show the same.**
+     - Adjusting for wave turns the observed T-cell COVID effect from −0.54 into +0.50.
+     - On wave-1 test samples, generated T cells fall (−0.20 to −0.79) where observed T cells rise (+0.43 to +0.56); the difference has p ≤ 0.001.
+   - The report should state one weighting rule and say which features change gate under the other.
+
+### E. Overstatements
+
+1. **The summary bullets claim more than the findings support.**
+   - "An interferon response in every lineage ... carried by draws 0–4": lymphoid ISG-high states are carried by draws 0–2 only, and MAIT and pDC by no draw (IMP-POP-4, IMP-LONG-1, IMP-PROG-4).
+   - "HLA class II loss ... in CD14 and intermediate monocytes and cDC2 (draws 0–4)": cDC2 severity holds in 4/5 draws, and CD14 holds in 4/5 once the CD14 share is adjusted for (IMP-PROG-1).
+2. **"All of these are known from Stephenson 2025 and Gisby 2022, and the bulk already carries each one"** contradicts the report's own entries.
+   - Neither paper reports platelets (IMP-CLIN-3).
+   - IMP-CLIN-1 and IMP-CLIN-3 are rated "partly" beyond the bulk.
+3. **"More platelets with severity" is listed as reproduced**, but:
+   - no draw passes the wave-1 level gate (generated platelets are 2.4–19x too many);
+   - observed platelets are about 10x higher in wave-2 samples;
+   - the support comes from all-wave gates whose truth is mostly wave 2.
+4. **"Interferon-driven labelling errors for naive B, Proliferating NK and CD8aa."** By IMP-CLIN-4's own text, the CD8aa label marks CD8B+ IL7R+ GZMK+ memory-like cells. That is not an interferon error.
+5. **"Wave 2 ... had dexamethasone treatment."** Only 7 of 16 (Stephenson 2025) or 8 of 17 (Gisby 2022) wave-2 patients received glucocorticoids.
+6. **IMP-PROG-2 as the one result that "survives a donor-disjoint, wave-1 test".** This holds for one gene filter, at the permutation floor (section A). The bottom line should say so.
+
+### F. Lessons from the sepsis run
+
+| Lesson | Done? |
+|---|---|
+| Shared noise | Yes. It is quantified (`imp_clinical/s14`, `imp_populations/s12`) and stated. But each finding still lists "5/5" under replication without its feature's noise share, for example MAIT 19%, pDC 25%, cDC2 11%. |
+| Recalibrate on the analysed population | Partly. COVID+-only (R1), wave 1 (R2) and the time windows were recalibrated. WHO grade, fatal outcome, age and sex were not, and the corrected gates were never applied to the features in section B. |
+| Calibration transfer | Yes for every finding (each verifier, and `imp_programs/transfer_set_shift.csv`). Not for the features in section B. |
+| statsmodels `scale='X2'` | Yes, none is used. I grepped the scripts: every binomial GLM uses clustered standard errors or GEE. |
+| Gated per-draw headline numbers | Mostly. Several "What survives" lines also quote draws that fail the gate, for example IMP-CLIN-2 draw 0 (0.21x) and IMP-CLIN-3 draw 0 (4.2x). |
+| Population ratios and z-scoring | Partly. Embedding-free marker rules and an independent re-clustering (s06) exist. But the main embedding is still z-scored per domain, and IMP-POP-5's 11–19x is an embedding figure. |
+| DC/NK and bulk baselines | Partly. DC and NK composition was screened (section 4.4), but no bulk baseline exists for the associations in section B. |
+
+### G. Cheap next steps (CPU)
+
+- **Report contents.**
+  - Add proliferating T cells as a positive control that the bulk carries, and add the age null.
+  - Add MAIT, effector B and DN T with within-parent and label checks; `c04` is a template.
+  - Add the sex–plasma lead once the chrY and XIST labeller inputs are removed.
+- **Trust layer.**
+  - Rebuild the severity row on COVID+ samples, with no donor in both arms.
+  - Choose donor or sample weighting.
+  - Add wave-1 rows for WHO grade and fatal outcome.
+  - Remove the duplicated plasma feature.
+- **Data hygiene.**
+  - Exclude or relabel C105_negative in every COVID contrast.
+  - Add the comorbidities as confounders in the mortality analysis.
+- **Wording.** Fix the two wrong numbers and the overstatements in section E.

paper-releases/pbisc-diffusion/updates/20261006T033406Z-bulkonly-lora5-k1000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_20261006/analysis/wf/REPORT_ye2026_sepsis.md

@@ -0,0 +1,757 @@
+# Bulk-only LoRA-5 discovery: synthesis for the Ye 2026 sepsis cohort
+
+Cohort ye2026_sepsis, 2026-10-06. This report combines three lenses (clinical, programs, populations), their 18 candidate findings and 36 adversarial verdicts.
+
+- Paths are relative to /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/ unless they start with /.
+- Synthesis scripts and tables are in synthesis_ye2026_sepsis/.
+- Every finding, both verbatim verdicts and my notes are in findings_ye2026_sepsis.json.
+- Grading rule: refuted if any verifier refuted; confirmed only if both confirmed; otherwise qualified.
+
+## Bottom line
+
+- **No finding is confirmed.** 17 are qualified and 1 is refuted (synthesis_ye2026_sepsis/y02_grades.csv).
+- **Direction reproduced.** In the 121 bulk-only adults the generated cells reproduce the direction of the known sepsis-vs-healthy shifts:
+  - monocytes lose HLA class II and gain an MS1/S100-high state;
+  - MAIT, GZMK+ memory T and NK cells fall;
+  - activated CCL4+ and NR4A2+ lymphocyte states rise, and so do plasma cells.
+- **None of this is new.** Every one of these is in Ye 2026 or earlier work, and every one is visible in the bulk.
+- **Sizes depend on the draw,** because each draw imprints its five training donors.
+  - Draw 0 (MS1-rich sepsis donors) is the only draw that gets MS1 levels right.
+  - Draw 1 fails most lymphocyte states and makes too many monocytes.
+- **Sepsis vs ICU controls cannot be separated reliably.** The trust layer has 5-6 adult ICU controls per draw, and this is the weakest calibrated contrast. ISG-high monocytes are the only lead, and observed cells do not confirm it.
+- **Little beyond the bulk.** The only clear piece is a modest cell-type-specific pattern (SEP-PROG-4). Simple bulk scores match or beat most population and program findings.
+- **No new population survives.** The screen found these artefacts:
+  - chrY labeller inputs, which break sex contrasts;
+  - an immediate-early/handling program;
+  - depth-dependent marker rules;
+  - chrY leakage from the generator;
+  - draw-specific clusters;
+  - over-produced doublet-like clusters.
+- **Refuted:** "ICU controls carry the same MAIT and memory CD8 depletion as sepsis" (SEP-CLIN-1).
+
+## 1. What was generated and how much to trust it
+
+### 1.1 Generated data
+
+- **Bulk-only samples:** 122 (68 adult sepsis, 36 adult HC, 17 adult ICU controls, 1 paediatric ICU control).
+  - The lenses analysed the 121 adults.
+  - Three samples have a sex flag: HC23, HC30, Res-S34 (synthesis_ye2026_sepsis/y00_samples_by_set_group.csv, y00_counts.txt).
+- **Draws:** five LoRA-5 models, each fine-tuned on 5 donors.
+  - Each generated 1,500 cells per bulk-only sample: 183,000 per draw, 915,000 in total (y00_cells_by_set_draw.csv).
+- **Calibration cells:** 1,500 cells for each of 77 paired test samples per draw, 577,500 in total.
+  - Each test set has 42-43 adult sepsis, 6-7 adult HC, 5-6 adult ICU controls and 21-23 children (y00_paired_test_groups.csv).
+- **Observed truth:** 82 paired samples, 589,478 cells (2,107-15,576 per sample).
+
+Training donors per draw (y00_training_donors.csv):
+
+| Draw | Training donors | Control among them |
+|---|---|---|
+| 0 | HC07, Abd-S66, Res-S150, Skin-S74, Res-S77 | adult HC |
+| 1 | HC12, Uri-S134, Uri-S147, Res-S164, Res-S166 | adult HC |
+| 2 | PIC01, Res-S177, Skin-S74, Res-S136, Res-S77 | paediatric ICU control |
+| 3 | PHC05, Res-S139, Res-S174, Skin-S142, Abd-S57 | paediatric HC |
+| 4 | IC23, Skin-S142, Res-PS22, Abd-S102, Skin-S46 | adult ICU control (Res-PS22 is paediatric sepsis) |
+
+Only draws 0 and 1 saw an adult HC in training.
+
+### 1.2 How the draws differ
+
+Generated monocyte share in bulk-only adult sepsis is 0.23, 0.50, 0.09, 0.12 and 0.22 in draws 0-4. Observed paired adult sepsis is 0.22. Observed adult ICU controls are monocyte-rich (0.39) and NK-poor (0.065), and no draw reproduces that (y00_bulkonly_l1_composition_by_draw_group.csv, y00_observed_l1_composition_by_group.csv).
+
+Per draw (y01_codes_S_vs_HC.csv, y01_calibration_long.csv):
+
+- **Draw 0:** most sepsis-vs-HC passes (13 of 50). The only draw with the right MS1 level (ratio 0.98); it overshoots MS1-like states.
+- **Draw 1:** makes too many monocytes (paired sepsis−HC +0.28 vs +0.06 observed). Intermediate monocytes have ratio 7.55, Spearman 0.06 and the wrong sign. It makes almost none of the immediate-early lymphocyte states. Fewest passes (5).
+- **Draw 2:** weakest for monocyte HLA class II and MS1 (MS1 ratio 0.05).
+- **Draw 3:** wrong sign for the sepsis-vs-HC monocyte share. Overshoots CCL4+ cytotoxic T (1.84x).
+- **Draw 4:** wrong sign for the CD14 monocyte share and the six-type NK share. 11 passes.
+
+### 1.3 The trust layer (calibration.csv)
+
+I summarised analysis/features/ye2026_sepsis/calibration.csv: 50 features (35 composition, 15 states) and 3 contrasts. Gates were fixed in advance (synthesis_ye2026_sepsis/y01_calibration_summary.py):
+
+- **Level:** ratio 0.5-2 and Spearman ≥ 0.3.
+- **PASS:** level, same sign, and a generated/observed contrast of 0.5-2.
+- **Direction only:** ratio 1/3-3, Spearman ≥ 0.2 and same sign.
+- **FAIL:** anything else.
+
+Per draw (y01_draw_summary.csv):
+
+| Draw | Level pass (of 50) | Sepsis vs HC: same sign / PASS / direction only / FAIL | Sepsis vs ICU: same sign / PASS | Sepsis vs all controls: same sign / PASS |
+|---|---|---|---|---|
+| 0 | 21 | 42 / 13 / 12 / 25 | 30 / 4 | 33 / 9 |
+| 1 | 18 | 38 / 5 / 14 / 31 | 37 / 4 | 32 / 4 |
+| 2 | 16 | 39 / 6 / 18 / 26 | 32 / 8 | 35 / 7 |
+| 3 | 22 | 38 / 6 / 19 / 25 | 33 / 6 | 34 / 8 |
+| 4 | 23 | 36 / 11 / 14 / 25 | 38 / 10 | 31 / 8 |
+
+Pooled (y01_pooled.csv): 24 of 50 features meet the level gate. The sign holds in 5/5 draws for 24 features (sepsis vs HC), 16 (sepsis vs ICU) and 17 (sepsis vs all controls).
+
+**Sepsis vs HC: what is reproduced** (y01_codes_S_vs_HC.csv)
+
+- **Size and direction in at least three draws:**
+  - memory CD8 T: PASS in 0, 2, 3, 4; draw 1 overshoots 2.7x;
+  - MAIT: PASS in 0-2; draws 3-4 at about half size;
+  - gdT: PASS in 1, 2, 4; draw 0 has the wrong sign;
+  - six-type "Other".
+- **Direction in all five draws, size in one or two:**
+  - six-type B, memory B, naive B;
+  - CD56dim NK: size only in draw 0; about zero in draw 4;
+  - ISG-high T: ratio 0.35-0.69.
+- **Depends on the draw:**
+  - monocyte share: PASS only in draw 0; wrong sign in 3-4;
+  - CD14 monocytes: PASS in 0-1; wrong sign in 4;
+  - intermediate monocytes: PASS in 3-4; FAIL in 0-1;
+  - ISG-high monocytes: PASS in 0 and 4; FAIL in 1;
+  - MS1: PASS only in draw 0; ratio 0.03-0.14 elsewhere.
+
+**Sepsis vs HC: what is not reproduced**
+
+- **FAIL in all five draws:**
+  - ASDC, cDC1, CD56bright NK, CD8aa, ILC, DN2/ABC B;
+  - transitional B (ratio 0.08-0.31);
+  - cycling monocytes (0.05-0.11);
+  - LDG (0.01-0.09);
+  - HSPC (Spearman about 0);
+  - plasmablast markers (0.09-0.24).
+- **FAIL in at least four draws:** CD16 monocytes, DN T, effector B, erythrocytes, progenitors, Treg, C1Q-high monocytes (0.05-0.52), cycling activated T.
+
+**Sepsis vs ICU** (y01_codes_S_vs_ICU.csv):
+
+- Monocyte share passes in draws 2-4; intermediate monocytes in 0, 2, 3, 4.
+- MS1 and CD56dim NK keep the observed sign in 0/5 draws.
+- Memory CD8 has the wrong sign in draws 2-4.
+- MAIT keeps the sign in 5/5, but at only 1-37% of the observed size.
+
+**Limits of the trust layer**
+
+1. **Few, reused controls.** Each draw's observed contrast rests on 6-7 HC and 5-6 ICU controls, the same people in every draw. Five passing draws are not five independent confirmations on the control side.
+2. **Sign alone is weak.**
+   - The negative control "platelet signal in monocytes" keeps the sepsis-vs-HC sign in 5/5 draws with Spearman 0.15-0.23.
+   - HSPC and progenitors keep the sign on observed differences of 0.0000-0.0004.
+3. **Missing contrasts.** There are no ICU-vs-HC, age, sex or source contrasts and no confidence intervals. The verifiers had to compute these themselves.
+4. **Raw means can disagree with adjusted contrasts.** For ISG-high monocytes, sepsis vs HC keeps the raw sign in 4/5 draws, but the adjusted sign is wrong in 5/5 (verify/SEP-CLIN-2_calib/v01_paired_contrasts.csv).
+
+## 2. Findings
+
+### 2.1 Confirmed
+
+None. SEP-CLIN-6 had one "confirmed" verdict, but it confirms an artefact.
+
+### 2.2 Qualified findings at a glance
+
+"Draws" means the draws that reproduce the feature's paired contrast and so carry the result.
+
+| ID | What survives | Draws | Observed paired cells | Beyond bulk | New | Strength |
+|---|---|---|---|---|---|---|
+| SEP-PROG-1 | Lower HLA-II in CD14 and intermediate monocytes, sepsis vs HC | 0-4 (draw 2 weakest) | yes, strong | location only, weak | no | moderate |
+| SEP-POP-2 | Fewer HLA-DR-high monocytes | 0-4 cluster-free; cluster c4 only draw 1 | yes | no | no | moderate |
+| SEP-PROG-2 | MS1 and S100 genes up in monocytes | direction 0-4; size 0-1 | yes | no | no | moderate (direction) |
+| SEP-POP-1 | More MS1-like HLA-II-low CD14 monocytes | 2, 3, 4 (0 overshoots) | yes vs HC, no vs ICU | no | no | moderate |
+| SEP-POP-6 | Low-library S100A12+ HLA-DRA-low cluster is the same MS1-like shift | 0, 2, 3, 4 (rule) | yes; not granulocytes | no | no | weak as a population |
+| SEP-PROG-3 | CD14 monocytes carry most bulk-up genes | 0, 2, 3, 4 (descriptive) | yes | not shown | no | weak |
+| SEP-CLIN-5 | More MS1 in ICU controls than HC | 0 only | yes, 2 samples | no | no | weak |
+| SEP-CLIN-2 | More ISG-high monocytes in sepsis than ICU controls | direction 0-4; size 1-4 | not significant | partly | not established | weak lead |
+| SEP-CLIN-4 | Fewer ISG-high T cells in abdominal sepsis | none under full controls | marginal | no | no | weak |
+| SEP-POP-5 | Fewer KLRB1+ GZMK+ memory T / MAIT-like cells | 0-4 (cluster) | yes | no | no | moderate, not sepsis-specific |
+| SEP-POP-4 | More CCL4+ activated cytotoxic T cells, fewer quiet TEMRA | 0, 2, 3, 4 | yes, mostly lymphocyte-wide activation | no | no | moderate up, weak down |
+| SEP-POP-3 | More NR4A2+ GPR183+ AREG+ memory CD4 cells | 0, 2, 3, 4 | yes | overstated | no | weak-moderate; handling suspect |
+| SEP-PROG-6 | Lymphocyte CCL/TNF and NR4A programs up | CCL 0, 4 (3); NR4A 0, 2-4 | yes, one sample-wide factor | claim wrong | no | artefact-suspect |
+| SEP-PROG-5 | Plasmablast genes up in B lineage | direction 0-4 | yes, marginal | no | no | weak |
+| SEP-CLIN-3 | B cells decline with age | 1, 3, 4 | same direction, p 0.05-0.12 | no | not in Ye; not textbook | weak |
+| SEP-PROG-4 | Cell-type-specific response a bulk broadcast cannot give | 0-4 | yes (modest share) | yes, modestly | methodological | moderate |
+| SEP-CLIN-6 | Male excess of "Proliferating NK" labels is a labeller artefact | none (not biology) | artefact shown | - | - | artefact |
+
+### 2.3 Details
+
+For each finding: what survives and the draws that carry it; the four honesty items (1 calibration, 2 replication across draws, 3 observed paired cells, 4 beyond bulk); the verdict reasons; novelty; sources.
+
+#### A. Monocyte HLA class II
+
+**SEP-PROG-1: HLA-II loss in CD14 and intermediate monocytes.** Lens sep_programs; verdicts qualified / qualified.
+
+- **What survives.** Generated CD14 and intermediate monocytes of bulk-only sepsis express less HLA class II than HC in all five draws. CD14: −0.93, −1.06, −0.71, −1.37, −1.13 log2; Welch p 5e-8 to 2e-27 (verify/SEP-PROG-1_recompute/v01_bulkonly_effect.csv).
+- **Draws.** CD14 and intermediate: 0-4, draw 2 weakest. CD16: 1, 3, 4.
+- **(1) Calibration.**
+  - The program is not in calibration.csv.
+  - The paired direction passes 5/5. The size is 0.48, 0.55, 0.38, 0.77, 0.59 of observed, and the CI excludes 1 in 4/5 draws.
+  - Per-sample Spearman is 0.51-0.64, and 0.30-0.46 within adult sepsis (v02_calibration_program.csv).
+  - Bulk-only effects are lower bounds; rescaled by each draw's ratio they are about −1.8 to −1.9.
+- **(2) Replication.** 5/5 draws, q ≤ 4.8e-7. It survives age, sex, site, sex flag, matching, cell number, depth, label confidence and renormalisation.
+- **(3) Observed.** CD14 −1.56 (43 vs 7, p 4.5e-12). Children replicate it: −1.38 (13 vs 4) (verify/SEP-PROG-1_calib/v01_observed.csv).
+- **(4) Beyond bulk.** Weak.
+  - The bulk falls too: −0.76, or −1.71 relative to monocyte markers.
+  - The bulk ranks patients better (Spearman 0.68-0.78 vs 0.51-0.64).
+  - The monocyte residual after bulk adjustment is reproduced at 0.33-0.51x, and is significant only in draws 1 and 4 (v05_bulk_adjusted.csv).
+  - Across all cell types the cross-type r is only 0.14-0.45.
+- **Why qualified.**
+  - Not sepsis-specific: ICU controls carry 55-75% of the drop (bulk-only ICU−HC −0.41 to −0.91; observed −0.97).
+  - Sepsis vs ICU is not calibrated: generated is about 0 in draws 2-4 vs −0.49 to −0.59 observed (v07_icu.csv).
+  - In draws 0-1, whole-sample HLA-DRA moves against the bulk.
+- **New?** No. Ye 2026 Mono_c01-CD14-HLA-DRA is reduced (PDF p.6, p.14). Also Reyes 2020 and Winkler 2017.
+
+**SEP-POP-2: fewer HLA-II-high CD14 monocytes.** Lens sep_populations; qualified / qualified.
+
+- **What survives.** A cluster-free HLA-DR-high share falls in bulk-only sepsis in all five draws. The rule is HLA-DRA+DRB1 above the observed 75th percentile, with observed cells thinned to generated depth. Draws 0-4: −0.185, −0.195, −0.085, −0.200, −0.229; q ≤ 1.3e-5 (verify/SEP-POP-2_calib/tables/v06_per_draw_summary.csv).
+- **Draws.** Cluster-free: 0-4. Cluster c4: draw 1, and draw 3 partly.
+- **(1) Calibration.**
+  - Cluster-free: the gate passes 5/5 (Spearman 0.40-0.62; contrast 0.38-0.81x).
+  - c4: 0.29, 1.10, 0.04, 0.45, 0.22x of observed. Draw 2 Spearman 0.13; draw 0 has only 3 generated HC (verify/SEP-POP-2_recompute/tables/v09_summary.csv).
+- **(2) Replication.** 5/5 negative under both definitions; robust to the usual confounders.
+- **(3) Observed.** HLA-DR-high share −0.268 (p 6.8e-5); c4 −0.243 (p 0.0038).
+- **(4) Beyond bulk.** None. The bulk HLA index gives AUC 0.088. Raw bulk HLA-DRA tracks the observed share better in 5/5 draws (rho 0.71-0.74 vs 0.40-0.62) (v04_tracking.csv).
+- **Why qualified.**
+  - c4 mixes HLA-II with an uncalibrated AP-1 axis that has no observed sepsis association (p 0.58).
+  - Sample sizes were smaller than reported (for example 26 of 68 sepsis in draw 2).
+  - In draw 0 the drop comes from the monocyte denominator.
+  - Not sepsis-specific.
+- **New?** No. Ye 2026 Fig. 4d, Fig. 7e and ED Fig. 9e. ED Fig. 9a deconvolves the same in-house bulk.
+
+#### B. MS1 / S100 monocyte program
+
+**SEP-PROG-2: MS1 and S100 genes up in monocytes.** Lens sep_programs; qualified / qualified.
+
+- **What survives.** Generated sepsis CD14 monocytes express more MS1 genes (+2.25, +1.31, +0.40, +1.09, +0.60 log2) and more S100A8/A9/A12 (+1.37, +1.00, +0.14, +0.84, +0.19) than HC, in all five draws (verify/SEP-PROG-2_recompute/v99_headline.csv).
+- **Draws.** MS1 size: 0-1, with 3 borderline. MS1 direction: 0-4. S100 in CD14: 0, 1, 3. CD16: none at gene level.
+- **(1) Calibration.**
+  - Paired MS1 ratio 1.06, 0.58, 0.19, 0.45, 0.26; Spearman 0.62-0.73 (verify/SEP-PROG-2_calib/v05_per_draw_verdict.csv).
+  - The trust-layer MS1 state passes only in draw 0 (0.98; 0.03-0.14 elsewhere).
+  - The sepsis-vs-ICU sign is wrong in 5/5 draws.
+- **(2) Replication.** 5/5 draws, q < 1e-5; robust to confounders and to granulocyte leakage.
+- **(3) Observed.** CD14 MS1 +1.93 (p 6e-15); S100 +1.55 (p 2e-11). RETN and CYP1B1 do not rise in observed cells.
+- **(4) Beyond bulk.** None.
+  - Bulk MS1 +2.17 and S100 +2.59 (t about 14).
+  - The bulk ranks patients better (0.83-0.88 vs 0.62-0.73).
+  - In draws 2 and 4 the bulk explains the whole contrast.
+- **Why qualified.**
+  - A lenient gate hid the attenuation.
+  - Draws 1-4 shift MS1 genes without making MS1 cells.
+  - CLU is partly a platelet confound.
+  - PLAC8 falls in draw 4; some named genes are flat in some draws.
+  - The CD16 gene pattern r is −0.33 to 0.42 (v03_gene_pattern_concordance.csv).
+  - Not sepsis-specific.
+- **New?** No. Reyes 2020; Ye 2026 Mono_c03-CD14-RETN (PDF p.6, 8, 14); Yao 2023.
+
+**SEP-POP-1: more MS1-like HLA-II-low CD14 monocytes.** Lens sep_populations; qualified / qualified.
+
+- **What survives.** A state with S100A12, MCEMP1, RETN, ALOX5AP, PLBD1 and CLU, and low HLA-II, takes a larger share of monocytes in sepsis than in HC (verify/SEP-POP-1_recompute/tables/v10_summary.csv):
+  - draw 0: 0.228 vs 0.010;
+  - draw 2: 0.175 vs 0.018;
+  - draw 3: 0.195 vs 0.013;
+  - draw 4: 0.134 vs 0.019.
+- **Draws.** Clean in 2, 3, 4 (matched contrast 1.44, 0.74, 0.95x). Draw 0 overshoots (2.0-3.3x). Draw 1 does not carry it (0.15x).
+- **(1) Calibration.**
+  - Ratio 2.00, 0.46, 1.75, 1.05, 1.16; Spearman 0.37, 0.36, 0.72, 0.55, 0.45 (verify/SEP-POP-1_calib/tables/v07_per_draw_summary.csv).
+  - In draws 1-4, IL1R2 is in 2-6% of state cells vs 42-45% of observed cells at matched depth.
+- **(2) Replication.** q ≤ 3e-4 in 5/5 draws, but draw 1's effect is about 1% (pooled Fisher p 0.17).
+- **(3) Observed.** +0.070 (30 vs 7, p 0.0014).
+  - Observed ICU controls are at least as high as sepsis (0.107 vs 0.070; p 0.93).
+  - The generated cells put ICU below sepsis.
+- **(4) Beyond bulk.** None. Bulk MS1 genes give AUC 0.96-0.97 and track patients better in 5/5 draws (v04_bulk_auc.csv).
+- **Why qualified.**
+  - Draw 1 and missing IL1R2.
+  - The observed counterpart holds only about 16% of observed MS1-rule cells (v05_obs_ms1_by_cluster.csv).
+  - Draw 0 overshoots with only 3 generated HC.
+  - The ICU direction is wrong.
+- **New?** No. Ye 2026 Mono_c03 (printed p.155; Fig. 4d also in ICU controls); Reyes 2020; Leite 2024.
+
+**SEP-POP-6: low-library S100A12+ HLA-DRA-low cluster.** Lens sep_populations; qualified / qualified.
+
+- **What survives, reinterpreted.** These cells are not granulocyte-like. At matched depth, observed members lack neutrophil genes:
+  - FCGR3B 0.6% vs 9.1% in other monocytes;
+  - ≥2 of 14 neutrophil genes in 6.1% vs 15.2%;
+  - they lean MS1 (32.8% vs 18.2%).
+  - The sepsis > HC signal survives library adjustment; it is the same shift as SEP-POP-1.
+- **Draws.** The c0 cluster is significant in 0 and 3. A HLA-DR-zero S100A12+ rule passes in 0, 2, 3, 4 (Spearman 0.54-0.59; −0.01 in draw 1).
+- **(1) Calibration.** c0 ratio 0.64, 0.50, 0.81, 1.05, 0.72; Spearman 0.11-0.33; contrast 0.14-0.71x.
+- **(2) Replication.** 5/5 positive; q < 0.05 in 2/5 on the subsample and 5/5 on all cells.
+- **(3) Observed.** +0.117 (p 1.4e-4), with all 7 HC at zero. Sepsis vs ICU +0.044 (p 0.24).
+- **(4) Beyond bulk.** None. The bulk S100A12-minus-HLA-DRA index gives AUC 0.98.
+- **Why qualified.**
+  - Library size sets membership, not the disease contrast.
+  - The granulocyte reading is wrong.
+  - CSF3R is lower in c0, not higher.
+  - Reclustering recovers c0 poorly (Jaccard ≤ 0.35).
+  - The low-library part of the generated cells is uncalibrated.
+- **New?** No.
+- **Sources.** verify/SEP-POP-6_recompute/tables/v09_key_numbers.csv; verify/SEP-POP-6_calib/tables/v08_key_numbers.csv; v02_gene_detection.csv.
+
+**SEP-CLIN-5: more MS1 in ICU controls than HC.** Lens sep_clinical; qualified / qualified.
+
+- **What survives.** In draw 0 only, 12.7% vs 0.33% of monocytes (quasi-binomial p 1e-5) (verify/SEP-CLIN-5_recompute/v02_perdraw.csv).
+- **Draws.** 0 only. Draws 1-4 make a median of 0-1 MS1 cells per ICU sample.
+- **(1) Calibration.**
+  - There is no ICU-vs-HC row.
+  - The paired check keeps the sign in 5/5, at 0.32x (draw 0) and 0.02-0.04x (draws 1-4).
+  - At fixed bulk IL1R2 the generator under-calls ICU MS1 in 5/5 draws, so its sepsis > ICU ordering is a model bias (verify/SEP-CLIN-5_calib/v06_icu_residuals.csv).
+- **(2) Replication.** Count and rank tests are positive in 5/5 draws.
+- **(3) Observed.** 24.4% vs 0.67% (6 vs 7; exact permutation p 0.017, MWU 0.051). It is driven by IC16 (87%) and IC19 (53%); IC19 is missed by every draw.
+- **(4) Beyond bulk.** None. Bulk IL1R2 explains R² 0.71-0.78; the generated value adds 0.01-0.03 (v04_bulk_vs_gen_paired.csv).
+- **New?** No. Ye 2026 Fig. 4d, ED Fig. 9a, ED Fig. 9e; Reyes 2020; Leite 2024.
+
+**SEP-PROG-3: CD14 monocytes carry most bulk-up genes.** Lens sep_programs; qualified / qualified.
+
+- **What survives.** Draws 0, 2, 3, 4 put most top bulk-up genes in CD14 monocytes, as observed cells do (agreement 72-92%).
+- **Draws.** Descriptive in 0, 2, 3, 4. A gain over the healthy reference in paired calibration only in draw 0. Draw 1 fails.
+- **(1) Calibration.**
+  - Paired top-30 agreement 0.87, 0.37, 0.80, 0.77, 0.73. The reference scores 0.67-0.70; a constant "CD14" call scores 0.80-0.86.
+  - At top 100, draws 2-4 are worse than the constant (q 0.005-0.016) (verify/SEP-PROG-3_calib/v02_summary.csv; verify/SEP-PROG-3_recompute/v07_calib_baselines.csv).
+- **(2) Replication.** 4/5 draws.
+- **(3) Observed.** CD14 monocytes carry 24 of 30 top genes; 27/30 are expression-driven.
+- **(4) Beyond bulk.** Not shown. A constant CD14 prior (80-84%), NNLS deconvolution and each draw's own training donors match or beat every draw (v03_baselines.csv).
+- **Why qualified.**
+  - The reference was weak.
+  - Library equalisation removes the gains in draws 2-4.
+  - "Within-cell induction" is false for draws 0 and 2.
+  - The IL1R2 within-cell share is about 51%, not 96%.
+  - Bulk-down genes: 1-11 of 30 correct vs 29 for a constant NK call.
+- **New?** No. Reyes 2020/2021; Ye 2026. VSIG4, ADAMTS2, TPST1 and FAM20A as monocyte glucocorticoid genes: Molendijk & Koenderman 2026; Hofer 2008.
+
+#### C. Interferon
+
+**SEP-CLIN-2: more ISG-high monocytes in sepsis than ICU controls.** Lens sep_clinical; qualified / qualified.
+
+- **What survives.** Higher in bulk-only sepsis than ICU controls in all five draws: mean draw logit +0.83 to +0.90, p 5e-5 to 3e-4. It is the only one of 50 features with q < 0.05 for this contrast (verify/SEP-CLIN-2_recompute/v02_summary.txt, v10_screen.txt).
+- **Draws.**
+  - Direction: 0-4.
+  - Size: draws 1-4 at 0.49, 0.48, 0.96 and 0.38 of the paired difference; draw 0 at only 0.09.
+  - A strict gate leaves draw 3 alone (+1.08 to +1.34; p 0.001 and 0.0002) (v11_gated.txt).
+- **(1) Calibration.**
+  - Ratio 0.43-0.93, Spearman 0.48-0.71; sepsis-vs-ICU sign 5/5.
+  - Sepsis vs HC: the adjusted sign is wrong in 5/5 draws.
+  - At equal truth the generator puts sepsis +0.70 logit above HC (p 3.5e-6).
+- **(2) Replication.** Positive in 5/5 draws. p < 0.05 in 4-5/5 depending on the monocyte-count rule. Beta-binomial 4/5; quasi-binomial 0/5 (p 0.07-0.15).
+- **(3) Observed.** +0.71 to +0.94 logit (46 vs 6, p 0.20-0.35) with equal medians (4.8% vs 4.4%).
+  - The gap comes from 6 of 46 very-high-ISG patients.
+  - The generator instead shifts the typical patient (AUC 0.74) (verify/SEP-CLIN-2_calib/v03_group_quantiles.csv).
+- **(4) Beyond bulk.** Partly.
+  - No bulk interferon program separates sepsis from ICU (Hallmark IFN-alpha p 0.33).
+  - After bulk adjustment, observed truth keeps +0.38 (p 0.026) and generated keeps +0.68 (v09_conditional.txt).
+  - But real paired cells separate them no better than bulk IFI27.
+  - Generated monocytes carry IFI27 that real monocytes lack (8-100 CPM vs about 1).
+- **Failed sub-claims.**
+  - "Sepsis = HC": observed sepsis is below HC (−0.65 to −0.94), and Ye 2026 reports it (PDF p.6).
+  - "ICU < HC" is real in observed cells (−1.72 to −1.88), but the generator reproduces it at ≤ 0.34x, it is fragile, and the bulk explains it.
+- **New?** Not established.
+  - ICU < HC is in Ye Fig. 4d and the MOFA interferon factor (PDF p.14).
+  - Sepsis > ICU is not significant in Ye's own abundance data, and Ye Fig. 4i has ICU controls lowest.
+  - If confirmed, it would be the only sepsis-specific result here.
+
+**SEP-CLIN-4: lower interferon in abdominal sepsis.** Lens sep_clinical; qualified / qualified.
+
+- **What survives.** Direction only: fewer ISG-high T cells in abdominal sepsis than other sources (−0.71, p 0.0027, q 0.067, so not BH-significant). The monocyte half fails.
+- **Draws.**
+  - The gate picks draws 1 and 3, but draw 3 is depth-confounded (abdominal T cells shallower, p 0.0007) and draw 1 depends on the 10 Emergency-Center samples.
+  - With site, flexible age and depth together: −0.46, p 0.053, and no single draw is significant (verify/SEP-CLIN-4_recompute/v10_summary.csv).
+- **(1) Calibration.**
+  - Ratio 0.35-0.69, Spearman 0.28-0.66.
+  - The source contrast keeps its sign 5/5 on paired tests, but the observed test-set p is 0.14-0.30 and the samples are shared.
+  - The generator compresses respiratory and inflates skin (verify/SEP-CLIN-4_calib/v04_source_profile.csv).
+- **(2) Replication.** 5/5 negative. Logit OLS 4/5; beta-binomial 1/5; quasi-binomial 0/5.
+- **(3) Observed.** −0.84 (p 0.10); rank-ANCOVA p 0.044. It vanishes after bulk-ISG adjustment.
+- **(4) Beyond bulk.** None. The bulk 10-ISG score is −0.65 to −0.72 z (p 4e-4) and removes the effect (−0.18, p 0.33).
+- **New?** Largely reported: Ye 2026 Fig. 4i; Scicluna 2025.
+
+#### D. Lymphocytes
+
+**SEP-POP-5: fewer KLRB1+ GZMK+ memory T (MAIT-like) cells.** Lens sep_populations; qualified / qualified.
+
+- **What survives.** The share of T cells falls in sepsis in all five draws: −0.074, −0.060, −0.061, −0.038, ���0.050; AUC 0.05-0.16, q ≤ 2.2e-6 (verify/SEP-POP-5_recompute/tables/v11_summary_per_draw.csv).
+- **Draws.** At cluster level 0-4 (paired contrast 0.67-1.17x; CIs exclude 0). On embedding-free definitions only draws 0-1 get the size right; draws 2-4 have the sign only (0.06-0.40x).
+- **(1) Calibration.** Ratio 0.86-1.08. Spearman 0.28-0.42, and 0.07-0.28 within sepsis. calibration.csv l2 MAIT: sign 5/5, ratio 0.52-0.55.
+- **(2) Replication.** 5/5 draws; robust to confounders including depth GEE and T-cell count.
+- **(3) Observed.** −0.074 (p 3e-4), from MAIT (−0.041) and memory CD4 (−0.019). Observed ICU controls sit at the HC level.
+- **(4) Beyond bulk.** No.
+  - The raw bulk score separates groups through T-cell content (pan-T AUC 0.052).
+  - A MAIT-core bulk score tracks patients better (Spearman 0.44-0.59 vs 0.28-0.42).
+- **Why qualified.**
+  - The generator also depletes ICU controls (bulk-only ICU−HC −0.030 to −0.060, p ≤ 0.002 in 5/5). Observed cells and Ye Fig. 3d show no such depletion (v11_icu_vs_hc.csv).
+  - Sepsis vs ICU is only 0.17-0.41x.
+  - Errors cancel inside the cluster: MAIT loss too small and memory CD8 loss too large in draws 2-4 (verify/SEP-POP-5_calib/tables/v07_c9_components_calibration.csv).
+  - MAIT transcription-factor genes are barely emitted.
+- **New?** No. Ye 2026 Fig. 3d and 2b; COMBAT 2022.
+
+**SEP-POP-4: CCL4+ activated cytotoxic T up, quiet TEMRA down.** Lens sep_populations; qualified / qualified.
+
+- **What survives.** CCL4+ GADD45B+ NR4A2+ cytotoxic T cells (c10) rise in draws 0, 2, 3, 4 (AUC 0.89-0.93), and GNLY+ GZMB+ FGFBP2+ TEMRA (c13) fall (verify/SEP-POP-4_recompute/tables/v09_per_draw_verdict.csv). This is mostly the cytotoxic view of a lymphocyte-wide activation program, and it is not sepsis-specific.
+- **Draws.** c10: 0, 2, 3, 4 (1.14, 0.74, 1.84, 1.24x); draw 1 0.12x. c13: sign in 0, 2, 3, 4; strict gate 2, 3, 4.
+- **(1) Calibration.** c10 Spearman 0.23-0.51 (0.04-0.38 within sepsis). For c13 the observed CI includes 0 in every test set.
+- **(2) Replication.** c10 5/5, but draw 1 is tiny. c13 4/5.
+- **(3) Observed.** c10 +0.070 (p 0.0047); c13 −0.034 (p 0.047; permutation 0.24).
+  - The B-cell activation share explains 92% of the observed c10 effect (v08_mediation_bcell.csv).
+  - ICU controls: 0.055 vs sepsis 0.073 (p 0.56).
+- **(4) Beyond bulk.** No at group level (a 10-gene bulk index gives AUC 0.895). Per patient only draw 4 adds information (verify/SEP-POP-4_calib/tables/v05_bulk_beyond_paired.csv).
+- **Why qualified.**
+  - Draw 1 carries nothing.
+  - Urinary sepsis is over-called (generated 0.04-0.10 vs observed 0.003) (v02_source_calibration.csv).
+  - TNF, IFNG and CCL3 are under-expressed 3-4x.
+- **New?** No. Ye 2026 T_CD8_c04 / T_CD8_c05 (PDF p.4; Fig. 3; ED Fig. 9a).
+
+**SEP-POP-3: NR4A2+ GPR183+ AREG+ memory CD4 state up.** Lens sep_populations; qualified / qualified.
+
+- **What survives.** Increases of +0.108, +0.171, +0.140, +0.111 in draws 0, 2, 3, 4 (draw 1 +0.001) (verify/SEP-POP-3_recompute/tables/v10_per_draw_summary.csv).
+- **Draws.** 0, 2, 3, 4. Draw 4 is most robust; draw 2 depends on the definition (rule Spearman 0.09).
+- **(1) Calibration.** Ratio 1.52, 0.13, 0.98, 0.67, 0.98; Spearman 0.28, 0.20, 0.30, 0.39, 0.36 (0.09-0.38 within sepsis).
+- **(2) Replication.** 4/5 draws (q ≤ 6.6e-11); robust to confounders.
+- **(3) Observed.** +0.142 (p 0.0042), with all 7 HC at zero. A clustering-free rule gives p 0.013-0.085.
+- **(4) Beyond bulk.** Overstated. A bulk score of the claim's own genes, or bulk CCL5 alone, gives AUC 0.92, the same as the generated cells.
+- **Why qualified.**
+  - ICU controls get 2-5x too much: generated ICU−HC +0.06 to +0.14 vs observed +0.03.
+  - It is one sample-level program across lineages (Spearman 0.60-0.94).
+  - Single-cell immediate-early levels are invisible in the same sample's bulk (Spearman 0.03-0.11) (verify/SEP-POP-3_calib/tables/v03_summary.csv).
+  - Site differs within sepsis (CEMC 0.24 vs First Affiliated 0.115, p 0.03).
+  - A handling component is likely.
+- **New?** No. Ye 2026 T_CD4_c02 (abstract; PDF p.2; Fig. 2b, 7f; ED Fig. 9a).
+
+**SEP-PROG-6: CCL3/CCL4/TNF/IFNG and NR4A programs in lymphocytes.** Lens sep_programs; qualified / qualified.
+
+- **What survives.** The generated cells reproduce a pan-cell-type immediate-early factor (verify/SEP-PROG-6_calib/v01_bulkonly.csv):
+  - CCL program in memory CD8: +1.68, +0.06, +0.69, +0.94, +1.48;
+  - NR4A program in memory CD4: +1.70, +0.22, +1.43, +1.85, +1.87.
+- **Draws.** CCL: 0 and 4 strict, plus 3 on a lenient gate. NR4A: 0, 2, 3, 4.
+- **(1) Calibration.**
+  - CCL ratio 0.77, 0.00, 0.24, 0.39, 0.61; within-sepsis Spearman −0.06 to 0.33.
+  - Per-draw size ranks exactly with the training donors' levels (Spearman 1.0) (v03_draw_summary.csv).
+- **(2) Replication.** Memory CD8 5/5. In gdT, draw 1 is significantly opposite.
+- **(3) Observed.** +1.6 to +1.7. It is one factor across all cell types; a generic immediate-early score removes about 60%, and a CD8/NK remainder of about 0.6 log2 is left (v05_ieg_adjustment.csv).
+- **(4) Beyond bulk.** The claim that the bulk lacks the signal is wrong.
+  - The 7 paired HC bulks sit in a high-immediate-early batch.
+  - Against the other HC, sepsis bulk rises (CCL3 +1.59, NR4A2 +2.29).
+  - Bulk-only NR4A +0.74 (p 5e-7) (verify/SEP-PROG-6_recompute/v06_bulk_genes_vs_clean_HC.csv).
+- **New?** Ye 2026 reports these as biology (PDF p.12, 14), but its validation used PMA/ionomycin-stimulated cells.
+- **Strength.** Artefact-suspect.
+
+**SEP-PROG-5: plasmablast genes up in B lineage.** Lens sep_programs; qualified / qualified.
+
+- **What survives.** Direction only: +3.13, +1.86, +1.55, +0.40, +2.62 log2, through more plasma cells.
+- **Draws.** Direction 0-4. Draw 0 passes only marginally (2.4x overshoot). Draw 3 is best calibrated but has the weakest effect.
+- **(1) Calibration.**
+  - Spearman 0.03-0.28 (bulk 0.67-0.71).
+  - The log ratios are inflated by a near-zero HC floor: in draws 0, 1, 2 and 4, 25-53% of HC have no generated plasma cell.
+  - On a linear scale it is 0.13-0.73x; l2 Plasma cell ratio 0.15-0.46.
+  - Sources: verify/SEP-PROG-5_recompute/v04_zero_plasma_floor.csv; verify/SEP-PROG-5_calib/v01_calibration.csv.
+- **(2) Replication.** 5/5, q < 0.05 in 4/5.
+- **(3) Observed.** +1.02 (p 0.022; MWU 0.056).
+- **(4) Beyond bulk.** None (bulk +0.94; +1.38 relative to B markers).
+- **New?** No. Ye 2026 PDF p.6, p.19; ED Fig. 6e.
+
+**SEP-CLIN-3: B cells decline with age.** Lens sep_clinical; qualified / qualified.
+
+- **What survives.** All B subsets decline: naive B −0.0136 logit/yr (about 13% per decade), all B −0.0121 (verify/SEP-CLIN-3_recompute/v01_targets.txt). This is a bulk-level observation.
+- **Draws.** Naive B: 1, 3, 4, or 0, 1, 4 with adult-only level calibration. All B: 1, 3, 4. A strict gate halves the B slope (p 0.055) (v05_gated_estimates.csv).
+- **(1) Calibration.** There is no age row. Paired slopes keep the sign 5/5, but the observed slope is never significant, the checks reuse the same ~54 adults, and draw 4 overstates it 2.5x.
+- **(2) Replication.** Naive B 5/5 negative, 4/5 at p < 0.05.
+- **(3) Observed.** −0.0153/yr (p 0.12) and −0.0148/yr (p 0.052). Without the Emergency-Center samples p is 0.27-0.49.
+- **(4) Beyond bulk.** None. Bulk B scores fall with age (p 0.002-0.003, also in HC alone). After bulk adjustment, p is 0.15-0.21.
+- **Why qualified.**
+  - The decline is pan-B, not naive-specific.
+  - A negative control is flagged at the same q.
+  - The generator weakens the true naive CD8 age effect (−0.049/yr observed) to 27-55%.
+- **New?** Not in Ye 2026. Not textbook either: Caraux 2010 and Blanco 2018 report stable naive B; Gong 2025 reports CD27− B expanding while the generator lowers effector B.
+
+#### E. Beyond the bulk
+
+**SEP-PROG-4: a cell-type-specific sepsis response a bulk broadcast cannot give.** Lens sep_programs; qualified / qualified.
+
+- **What survives.** After removing gene and cell-type means (double-centring), the generated within-type sepsis-vs-HC logFC pattern still agrees with observed cells in all five draws (verify/SEP-PROG-4_recompute/v15_summary.csv, v13_double_centred_calibration.csv):
+  - bulk-only r 0.35, 0.21, 0.17, 0.21, 0.26;
+  - paired r 0.21, 0.18, 0.13, 0.18, 0.15;
+  - permutation null ≤ 0.018;
+  - a bulk broadcast gives −0.09 to 0.05.
+- **Draws.** 0-4.
+- **(1) Calibration.** It passes 5/5 and holds with log2(CPM+1) (0.08-0.14), without immediate-early genes, with test-only truth and at equal depth.
+- **(2) Replication.** 5/5.
+- **(3) Observed.** The observed interaction is reliable (Spearman-Brown 0.65-0.70), but the generated cells capture only about 5-16% of it (verify/SEP-PROG-4_calib/v04_reliability.csv).
+  - It is concentrated in well-expressed genes (r 0.42-0.64).
+  - The top 50 genes, mostly CD14-monocyte genes, carry 32-42% of the covariance.
+  - It is mostly between lineages (lineage-centred r 0.09-0.22).
+- **(4) Beyond bulk.** Yes, modestly; this is the one clear beyond-bulk content.
+  - The claim's single-centred r (0.16-0.32) is matched by a bulk broadcast (0.22-0.30) and by simple bulk predictors (0.37-0.43).
+  - "Bulk carries none of it by construction" is false.
+- **Why qualified.** Gene-level gains hold only for ensembles; monocyte-normalised bulk beats every draw in monocytes; no sepsis-vs-ICU signal (r 0.01-0.04).
+- **New?** Methodological.
+
+#### F. Artefact
+
+**SEP-CLIN-6: "Proliferating NK" labels more frequent in men.** Lens sep_clinical; qualified / confirmed (the "confirmed" verdict confirms the artefact).
+
+- **What it is.** Not biology. Generated 1.09% vs 0.62%; observed 1.72% vs 0.88%.
+  - The labeller uses RPS4Y1, USP9Y and UTY as inputs.
+  - Zeroing them removes the observed per-NK male excess (+0.451 to −0.007, p 0.97). Giving women chrY counts creates it. An autosomal control leaves it unchanged (verify/SEP-CLIN-6_calib/v09_counterfactual_models.csv).
+- **Generated cells.**
+  - chrY explains 20-82% of the excess per draw.
+  - The rest is a generator excess of NK cells in men (+0.31, p 1.5e-5 after adjusting for bulk NK), which the bulk does not show.
+  - Draws 0 and 2 leak chrY into 31-35% of women's cells.
+- **Calibration.** Ratio 0.41-0.72, Spearman 0.23-0.51; sepsis-vs-HC sign 3/5 (the claim said 4/5).
+- **Consequence.** Any sex contrast of L2 labels is suspect. Transitional B, CD8aa, naive B, memory CD4 and effector B also shift when the chrY inputs are zeroed (verify/SEP-CLIN-6_recompute/v11_other_labels.txt).
+
+## 3. Candidate new populations
+
+None is new. Sources: sep_populations/tables/s07_candidate_detail.csv, s04_markers_r2.0.csv, s14_artefact_clusters.csv, s08_obs_only_c18.csv, and the verifier tables above.
+
+| Candidate | Markers | Observed support | Artefact risk | Verdict |
+|---|---|---|---|---|
+| MS1-like HLA-II-low CD14 mono (Mono r2.0 c8; SEP-POP-1) | S100A12, MCEMP1, RETN, ALOX5AP, PLBD1, CLU; IL1R2 only in draw 0 | +0.070 vs HC (p 0.0014); ICU ≥ sepsis | medium: IL1R2 missing in draws 1-4; covers ~16% of observed MS1 cells | known (MS1; Ye Mono_c03) |
+| Low-library S100A12+ HLA-DRA-low mono (Mono r2.0 c0; SEP-POP-6) | S100A12, low HLA-DRA, low library | +0.117 (p 1.4e-4); not neutrophil-like | high as a separate population (Jaccard ≤ 0.35) | same MS1-like state |
+| HLA-II-high CD14 mono (Mono r1.0 c4; SEP-POP-2) | HLA-DRA/DRB1/DPB1, CD74, LGALS2, MS4A6A, plus AP-1 | −0.24 to −0.27 | medium: uncalibrated AP-1 axis | known (Ye Mono_c01) |
+| NR4A2+ GPR183+ AREG+ memory CD4 (T r2.0 c4; SEP-POP-3) | NR4A2, GPR183, AREG, FOS, JUN, CD69 | +0.142 (p 0.0042) | high: shared immediate-early program, imprint, ICU over-production | known (Ye T_CD4_c02) |
+| CCL4+ GADD45B+ cytotoxic T (T r2.0 c10; SEP-POP-4) | CCL4, GADD45B, NR4A2, GNLY, GZMH | +0.070 (p 0.0047) | high: activation program; TNF/IFNG/CCL3 weak | known (Ye T_CD8_c05) |
+| KLRB1+ GZMK+ IL7R+ memory T / MAIT-like (T r2.0 c9; SEP-POP-5) | KLRB1, GZMK, IL7R, CCL5 | −0.074 (p 3e-4) | low-medium: MAIT TF genes barely emitted | known (Ye MAIT, CD4 GZMK TEM) |
+| Observed-only granulocyte-like (Mono r1.0 c18) | MMP9, IL1R2, CSF3R, ALPL, CYP4F3 | 65 observed cells, 0 generated; 3/46 sepsis vs 0/7 HC (p 0.51) | generator blind spot (LDG ratio 0.01-0.09) | untestable in generated cells |
+| Draw-specific monocyte clusters (r1.0 c9, c13; r2.0 c10, c20-c22) | TMEM176B/FCN1/PLBD1; FCGR3A/MS4A7 | - | training-donor imprint (up to 94% of cells from one draw) | not reportable |
+| 71 generated-only micro-clusters (≤ 30 cells; 592 cells) | RPRD2, RICTOR, PDE4D, NFKB1 | none | noise | not tested |
+| Doublet-like mixed-lineage clusters (Mono r1.0 c5, T r1.0 c10, NK r1.0 c22) | cross-lineage genes | lower in observed | over-produced (e.g. 13% vs 7.5% of monocytes) | not populations |
+| Library-size clusters (Mono r1.0 c2, T r1.0 c8, NK r1.0 c15) | few markers | - | depth | not populations |
+| C1Q+ FCGR3A+ mono (Mono r2.0 c16) | C1Q genes, HLA-II, FCGR3A | no effect (p 0.60) | C1Q-high under-generated (0.05-0.52) | negative |
+
+## 4. Refuted and negative results
+
+### 4.1 Refuted: SEP-CLIN-1
+
+**Claim.** ICU controls (17) show the same lymphoid depletion as sepsis vs HC (36):
+- MAIT 0.47% vs 1.35% (sepsis 0.40%);
+- NK 3.6% vs 9.6%;
+- memory CD8 11.5% vs 20.5%;
+- none of the three differs from sepsis.
+
+**Recompute verdict: qualified.**
+- The numbers reproduce and are robust.
+- The contrast is bulk-derivable: a single-marker bulk baseline gives MAIT −1.06 to −1.09 (generated −1.28) and NK −1.28 (generated −1.26) (verify/SEP-CLIN-1_recompute/v05_bulk_baseline.csv).
+- The MAIT/T loss is not supported: bulk −0.11 (p 0.65); observed −0.31 (p 0.60).
+- On paired samples the generator places ICU controls nearer sepsis than the observed cells do in 5/5 draws (v03_icu_position.csv).
+- The memory CD8 sepsis-vs-ICU direction is reversed.
+
+**Calibration verdict: refuted.** One generator bias produces the inflated ICU < HC effect, the "ICU = sepsis" null and the beyond-bulk residual.
+- ICU MAIT is under-generated (generated/observed 0.28-0.41).
+- At fixed bulk SLC4A10 the generator adds an ICU offset of −0.55 to −1.35; the observed offset is +0.03 to +0.15 (verify/SEP-CLIN-1_calib/v06_beyond_bulk_offset.csv).
+- Observed MAIT runs HC 2.84%, ICU 1.74%, sepsis 0.89%.
+- Sepsis is below ICU for MAIT/T (p 0.042).
+- ICU controls are more lymphopenic than sepsis (57% vs 73%, p 0.003) (verify/SEP-CLIN-1_calib/v02_observed.csv).
+- Grimaldi 2014 agrees: MAIT is low in non-infected ICU patients and lower still in severe infection.
+
+**What survives** is already reported and visible in the bulk: ICU controls are lymphopenic vs HC, with NK clearest (observed −1.24, p 0.017; Ye ED Fig. 1e).
+
+### 4.2 Sub-claims that failed inside qualified findings
+
+- **SEP-CLIN-2:** "sepsis = HC" for ISG-high monocytes is contradicted.
+- **SEP-CLIN-5:** "ICU ≥ sepsis" for MS1 is a mean driven by two samples.
+- **SEP-POP-6:**
+  - the granulocyte/LDG identity is contradicted;
+  - "generated cells lack IL1R2/HP" is false for draw 0 (IL1R2) and draws 3-4 (HP).
+- **SEP-PROG-3:** "within-cell induction" and "IL1R2 96% within-cell" are wrong.
+- **SEP-POP-1 and SEP-POP-4:** "5/5 draws all calibrated" is wrong; draw 1 fails.
+- **SEP-PROG-6:** "the bulk lacks the signal" comes from an HC bulk batch.
+- **SEP-PROG-4:** "a bulk broadcast carries none of it" is false for the claimed statistic.
+- **SEP-CLIN-3:** "naive-specific immunosenescence" is wrong; the decline is pan-B and not textbook.
+- **SEP-PROG-2:** the CD16 gene pattern is wrong.
+- **SEP-CLIN-1:** the memory CD8 sepsis-vs-ICU direction is reversed.
+- **SEP-PROG-5:** the MKI67 attribution is misstated for draw 4.
+
+### 4.3 Negative results from the lenses
+
+1. **Infection source.** Nothing passes BH (minimum q 0.37). The observed urinary pattern (MS1 −2.02, intermediate monocytes +2.57, q 0.021) is not reproduced (sep_clinical/s02_summary.csv, s04_observed_assoc.csv).
+2. **Monocyte interferon program.** Observed CD14 ISG falls in sepsis (−1.38, p 0.001). Generated cells do not reproduce it (FAIL 5/5) (sep_programs/isg_check_observed.csv, programs_calib.csv).
+3. **T-cell programs.** Exhaustion has the opposite sign. Activation rises in every generated T subset, but observed shows a rise only in memory CD8.
+4. **Within-type sepsis-vs-ICU DE.** 51 FAIL, 9 WEAK, 0 PASS; only the six-type T pseudobulk passes (sep_programs/calib_types.csv).
+5. **Bulk-down gene attribution.** 1-11 of 30 correct vs 29 for a constant NK call (sep_programs/attribution_mcnemar.csv).
+6. **Ageing.** The naive CD8 age decline is reproduced at only 0.27-0.55x, and the MAIT age decline is missed. Control-only age effects do not survive.
+7. **Sex.** Observed sex effects do not replicate. The generated NK-in-men excess is in neither the observed cells nor the bulk.
+8. **Monocyte share in controls.** Unusable: the draws trained on an adult HC (0, 1) and the others (2-4) give opposite ICU-HC signs (sep_clinical/s11_monocyte_share.txt).
+9. **Strong hits in poorly calibrated features** (ILC, progenitor, HSPC, erythrocyte, plasmablast markers, platelet-in-monocyte negative control) are artefact-suspect. Generated HSPC go up while bulk HSPC goes down.
+10. **Domain gap.** A plain joint embedding separates observed from generated cells. About 270-410 genes per lineage are almost never emitted (sep_populations/tables/s02_domain_gap_genes.csv).
+11. **States with no observed support:** IL1B+ inflammatory, ISG-high cluster, VCAN/CSF3R and platelet-complexed monocytes, and C1Q+ FCGR3A+ monocytes.
+12. **No adult cycling MKI67+ monocytes.** This agrees with Ye, where the subset is paediatric-only.
+
+## 5. What would turn each qualified finding into a confirmed one
+
+For most findings "confirmed" can only mean a confirmed reproduction of a known, bulk-visible effect.
+
+### 5.1 Steps that help every finding
+
+1. **More paired controls.** The binding limit is 6-7 HC and 5-6 ICU controls per draw, the same people each time. Controls with both bulk and single-cell data, from more samples or an external cohort, would make magnitudes testable and sepsis vs ICU testable at all.
+2. **A richer trust layer.** Add these to calibration.csv:
+   - ICU-vs-HC, age, sex and source contrasts;
+   - adjusted contrasts;
+   - contrast ratios with bootstrap CIs;
+   - the programs and states the lenses used: HLA-II, MS1, double-centred interaction.
+3. **Fixed bulk baselines.** Each finding must beat four baselines at predicting per-sample observed values on paired tests, with per-draw partial correlations:
+   - an own-gene bulk score;
+   - a lineage-normalised bulk score;
+   - deconvolution;
+   - a constant prior.
+4. **More, designed draws** (10-20, some with MS1-rich controls, some without extreme-state sepsis donors). Regress each draw's effect on its training donors' observed level, and require a non-zero bulk-driven part. This needs GPU time, which only the coordinator can schedule.
+5. **Depth-matched, embedding-free state definitions** (thinning plus marker rules) for every state claim.
+6. **Handling metadata and a dissociation/immediate-early covariate** (processing time, cryopreservation, site) before any IEG/CCL/NR4A claim.
+7. **Relabel without chrY (and XIST) inputs** before any sex analysis.
+
+### 5.2 Per finding
+
+- **SEP-PROG-1, SEP-POP-2:** Get per-patient monocyte HLA-DR truth.
+  - Check whether Ye's flow cohort (ED Fig. 9e) includes bulk-only patients. If so, test generated CD14 HLA-II against measured HLA-DR beyond bulk HLA-DRA.
+  - Report the cluster-free share, not c4.
+  - Sepsis-specific claims need more paired ICU controls.
+- **SEP-PROG-2, SEP-POP-1, SEP-POP-6:**
+  - Fix IL1R2 under-generation in draws 1-4.
+  - Use a depth-matched HLA-DR-zero S100A12+ rule.
+  - Benchmark against bulk MS1 deconvolution (Reyes 2020; Leite 2024).
+  - Use only the genes that hold in all draws (IL1R2, CLU, CD163, S100A8).
+  - Get more paired ICU controls.
+- **SEP-PROG-3:** Rerun with a library-equalised decomposition against constant-prior and deconvolution baselines. Check steroid exposure: FKBP5 and PER1 are up, and there is no steroid column.
+- **SEP-CLIN-5:** Needs an MS1-calibrated draw besides draw 0, and more paired ICU controls, both MS1-rich and MS1-poor. Compare per sample with bulk IL1R2.
+- **SEP-CLIN-2:** Needs more paired ICU controls or an external bulk plus single-cell cohort with ICU controls. Test the median and the high tail separately. The IFI27-free, depth-equalised rule already leaves the bulk-only result unchanged.
+- **SEP-CLIN-4:** Pre-specify the omnibus test (currently p 0.022, q 0.37). Use a depth-matched rule and flexible site/age adjustment, and replicate with more urinary and skin cases.
+- **SEP-POP-5:** Fix MAIT TF emission (SLC4A10, RORC, ZBTB16) and recalibrate on L2 MAIT. Test the ICU over-depletion once there are more paired ICU controls.
+- **SEP-POP-4:** Test the CD8-specific residual after adjusting for B/NK activation. Calibrate the TNF/IFNG/CCL3 module and the urinary contrast.
+- **SEP-POP-3, SEP-PROG-6:** Collect handling metadata and dissociation scores, and replicate in fresh blood or an independent cohort. Until then, do not report these as in vivo biology.
+- **SEP-PROG-5:** Fix plasma-cell under-generation, use linear-scale contrasts, and require per-sample tracking beyond the bulk.
+- **SEP-CLIN-3:** Add an age contrast to the trust layer, get more paired adults across ages and flow-measured B subsets, and resolve the effector-B sign error.
+- **SEP-PROG-4:**
+  - Make the double-centred statistic with a bulk-broadcast null the standard beyond-bulk metric.
+  - Identify which genes and cell types carry it.
+  - Repeat sepsis vs ICU when more controls exist.
+  - It becomes a confirmed method result if it also holds in the Imperial cohort.
+- **SEP-CLIN-6:** Already an established artefact. To close it, relabel without chrY and check whether the generator's NK-by-sex excess remains.
+- **SEP-CLIN-1 (refuted):** Revisit only with an ICU-vs-HC calibration row and more paired ICU controls.
+
+## Provenance
+
+- **Scripts** (synthesis_ye2026_sepsis/):
+  - y00_counts.py: generation counts;
+  - y01_calibration_summary.py: trust-layer summary;
+  - y02_extract_graded.py: verbatim findings and verdicts from workflow wf_e137f359-578, with the journal's sha256 in y02_graded_input.json;
+  - y03_build_findings_json.py: writes findings_ye2026_sepsis.json.
+- **Logs and tables** sit next to the scripts.
+- **Other numbers** come from the lens and verifier files cited above. I spot-checked the key per-draw values against those tables.
+- **Compute.** CPU only, run with systemd-run MemoryMax=6G and no swap. No GPU work.
+- **Writes.** Only synthesis_ye2026_sepsis/ and findings_ye2026_sepsis.json. REPORT_ye2026_sepsis.md was not written because the harness blocks subagents from writing report files; its content is this text.
+- **Archive.** No PBISC-Diffusion archive delta from this agent. Registration under paper-release-20260928/maintenance/POLICY.md is left to the coordinator.
+
+
+## Completeness review
+
+The completeness critic wrote this for ye2026_sepsis on 2026-10-06. Note: `REPORT_ye2026_sepsis.md` does not exist on disk. The synthesis agent's report text exists only in the workflow journal; I saved a copy as `critic_ye2026_sepsis/c00_report_text_from_journal.txt`. So I could not append this section to the report, and I did not create the file myself. The coordinator should append the text below when it writes the report.
+
+All work was CPU only, run with systemd-run (MemoryMax=6G, no swap). Scripts and tables are in `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/analysis/wf/critic_ye2026_sepsis/` (c00–c07). There is no archive delta from this agent; registering it under POLICY.md is left to the coordinator.
+
+### A. Number spot-check: 41 report statements checked against their cited sources (`c01_spotcheck.csv`)
+
+35 match, 3 are partly wrong, 2 are wrong (both minor), and 1 cites the wrong file:
+- **Children per test set:** the report says 21–23; the source gives 22–23 (`y00_paired_test_groups.csv`).
+- **MAIT, sepsis vs ICU:** the report says 1–37% of observed size; `calibration.csv` gives 0.8–34%.
+- **"No draw reproduces" the monocyte-rich, NK-poor ICU controls:** the report's own table contradicts this in part. Draw 1 gives ICU monocytes 0.400 (observed 0.389), draw 0 gives ICU NK 0.065 (observed 0.065), and draw 4 puts ICU above sepsis. What fails is the ICU-minus-sepsis monocyte excess: observed +0.17, generated −0.10 to +0.04.
+- **"Sepsis vs ICU is the weakest contrast":** true only by PASS count (32 vs 36 vs 41). By FAIL count (138 vs 145) and same-sign count (170 vs 165), sepsis vs all controls is weaker.
+- **"Only the six-type T pseudobulk passes" sepsis vs ICU:** six-type B cell also passes, in draw 0 (`sep_programs/calib_types.csv`).
+- **"270–410 never-emitted genes":** the numbers are right (271–406), but they come from `sep_populations/logs/s02_domain_gap.log`. The cited table contains none of them.
+- **Small inconsistencies:**
+  - SEP-POP-2 contrast is 0.38–0.81x in the report but 0.35–0.89x in the JSON.
+  - SEP-POP-3 draw 3 ratio is 0.665, which rounds to 0.66, not 0.67.
+
+Everything else matched, including:
+- all 45 numbers of the trust-layer table and the pooled counts (24/50, and 24/16/17);
+- the draw-1 monocyte (+0.28 vs +0.06) and intermediate-monocyte (7.55 / 0.06) values, and the MS1 ratios;
+- the ratios in the FAIL list;
+- the headline per-draw values of SEP-PROG-1/2/4/5/6, SEP-POP-1 to 5 and SEP-CLIN-1 to 6;
+- source omnibus q 0.37, urinary q 0.021, attribution 1–11/30.
+
+### B. Cross-cutting problems the report does not address
+
+1. **The five draws share their per-cell noise, so "k/5 draws" is not five independent replications.**
+   - All draws used seed 20260927, batch size 256 and the same Poisson seed. `run_sampling.sh` says each cell keeps its seed.
+   - At the same cell slot, draws agree on the six-type label 58–80% of the time, against 38–56% expected if independent. L2 labels agree 35–53% vs 9–11%. Per-cell library size correlates 0.75–0.90 (Spearman) vs ~0 at shifted slots (`c03`).
+   - For the per-sample fractions the analyses use, shared noise explains 23–33% of cross-draw agreement for rare types (DC, cDC2, pDC, plasma cells, progenitors, CD56bright NK), about 15% for MAIT and Treg, and under 5% for major types (`c03b`).
+   - Draws 0 and 2 also share 2 of their 5 training donors (Skin-S74, Res-S77), and draws 3 and 4 share Skin-S142.
+2. **The trust layer is computed on a different population from the bulk-only analysis.**
+   - Its ratio and Spearman use all 77 test samples, 22–23 of them children; every bulk-only analysis is adults-only.
+   - My re-implementation reproduces `calibration.csv` exactly (difference ≤ 5e-6). Restricting it to adults changes the level gate in 38 of 250 feature-draw rows (22 lose it, 16 gain it). Per-draw level passes go from 21/18/16/22/23 to 23/13/13/20/25 (`c02`, `c02b`).
+   - Effects on findings:
+     - ISG-high T loses draw 1. SEP-CLIN-4's gate then leaves only draw 3, which is depth-confounded, so no draw carries it.
+     - MAIT loses draws 1–3; memory CD8 loses draw 3; CD14 monocytes lose draw 0.
+     - Naive B carriers change from {1,3,4} to {0,1,4}.
+     - Plasma cells gain draws 0 and 4.
+   - The "sepsis vs all controls" truth includes 15 paediatric sepsis and 8 paediatric controls. Its sign agreement changes in 58/250 rows when restricted to adults.
+3. **Nobody checked whether calibration transfers from paired to bulk-only samples.** In bulk RNA, the paired (trust-layer) samples differ from the bulk-only samples within every group (`c05`):
+   - **HC:**
+     - immediate-early genes +2.36 log2 (q 1e-4), CCL +1.32, erythrocyte −2.17, cDC2 +0.64;
+     - a classifier separates the 7 paired from the 36 bulk-only HC with AUC 0.93.
+   - **ICU controls:** immediate-early +2.25 (q 0.03).
+   - **Sepsis:**
+     - erythrocyte −1.25, granulocyte −0.59, platelet −0.82, MS1 −0.49, immediate-early +0.81 (q ≤ 0.04);
+     - classifier AUC 0.64.
+   - Some bulk contrasts change sign between the two sets:
+     - immediate-early, sepsis − HC: +0.67 bulk-only vs −0.89 paired;
+     - immediate-early, sepsis − ICU: +0.51 vs −0.93;
+     - CCL, sepsis − ICU: +0.47 vs −0.30;
+     - B, ICU − HC: +0.33 vs −0.11;
+     - T, sepsis − ICU: −0.53 vs +0.08.
+   - So calibration does not transfer for the immediate-early/CCL states (SEP-POP-3, SEP-POP-4, SEP-PROG-6), for contaminant features, or for B-cell ICU contrasts. It plausibly does transfer for ISG, MAIT, NK, HLA-II and MS1, whose contrasts agree between the sets. The report mentions the paired-HC batch only inside SEP-PROG-6.
+4. **A statsmodels bug inflates some quasi-binomial p-values.** In statsmodels 0.14.6, a two-column count response fitted with `scale='X2'` gets the dispersion wrong. I confirmed it on synthetic data: scale 0.04 instead of 13.5, standard errors 18x too small (`c06`).
+   - Nine verifier scripts use this form:
+     - SEP-POP-1 calib v01 and recompute v04;
+     - SEP-POP-2 recompute v02 and v04;
+     - SEP-POP-3 recompute v02;
+     - SEP-POP-6 recompute v03 and v05b;
+     - SEP-CLIN-3 calib v02;
+     - SEP-CLIN-4 recompute v03 (v03b fixes it).
+   - I traced SEP-POP-1, -2 and -3. Their robustness conclusions rest on rank models and survive (SEP-POP-1 with all covariates: rank-INT p 0.034 to 8e-7). The quasi-binomial p-values in those tables (as low as 1e-167) are invalid and must not be quoted. SEP-POP-6 and SEP-CLIN-3 were not traced.
+5. **SEP-CLIN headline numbers come from the five-draw mean of logits,** which DATA_GUIDE rule 3 forbids as primary evidence. For example, SEP-CLIN-5's +1.33 is driven by draw 0 (+2.73); the complete-case estimate is +0.86. The lens's own gate table (`sep_clinical/s12_gated.csv`) is never used in the report.
+6. **Population-level ratios near 1 are partly built in.** Each domain is z-scored separately before co-clustering, which pushes cluster-level ratios toward 1. Embedding-free MAIT-like definitions give 0.42–0.77 instead (SEP-POP-5 verifier). The report uses cluster ratios as calibration evidence for SEP-POP-1, -3, -4 and -5 without this caveat.
+
+### C. Analyses not run
+
+- **Outcome and severity.** Survival, SOFA, shock, steroids and pathogen are not in `meta.csv` or the GEO characteristics. But Ye 2026 Methods ran Kaplan–Meier survival using clinical data and bulk RNA-seq of 174 adults, and reports survival links for plasma cells, T_CD4_c02-NR4A2 and NK_c03-CCL4hi. Those are the same states as SEP-PROG-5, SEP-POP-3 and SEP-POP-4. Testing "generated state predicts survival beyond the bulk signature" is the clearest beyond-bulk test available, and needs the per-patient outcomes.
+- **Sepsis endotypes (SRS/MARS)** can be computed from bulk; Kwok 2023 (SRS1) is in the reference folder. Not run.
+- **Infection source** was tested only on the 50 trust-layer features. The programs and populations lenses never used `infection_source`, although source-specific biology is Ye 2026's headline.
+- **ICU vs HC calibration.** I computed it on adult paired tests (`c02`):
+  - Same sign in 154/245 rows, and within 0.5–2x of observed size in only 66.
+  - MS1: 0.02–0.04x in draws 1–4.
+  - ISG-high monocytes: wrong sign in draw 0, at most 0.34x elsewhere.
+  - Memory CD8 overshoots 1.7–3.6x in draws 1–4.
+- **DC (a missing negative result; no lens analysed DCs).**
+  - Observed DC share falls in sepsis: 1.42% vs 0.34% (p 4e-4). Bulk cDC2 markers fall −1.61 log2, pDC −1.09.
+  - Generated cells show the loss only in draws 0 and 3. Draw 1 goes significantly the wrong way (+0.0018, p 0.034); draws 2 and 4 show about zero (`c04`).
+- **NK.** The bottom line says "NK cells fall", but no finding or verifier covers it. Generated NK falls in draws 0–3 and rises in draw 4 (+0.005, p 0.87). The lens gate carries only draws 0 and 3.
+- **The observed truth itself is unstable for one key feature.** For ISG-high monocytes, sepsis − HC changes sign across the five test sets (−0.005 to +0.031), depending on which donors are held out (`c07`).
+
+### D. Thin verification and omitted verifier caveats
+
+- **SEP-CLIN-2, the only sepsis-specific lead.** The report omits three verifier results:
+  - At equal truth, the generator itself adds about +0.34 logit to sepsis vs ICU (95% CI −0.24 to 0.92), about 40% of the +0.83–0.90 effect.
+  - The observed sepsis-vs-ICU difference vanishes at equal depth (monocytes downsampled to 1,000 UMI): +0.49, p 0.62, AUC 0.47.
+  - Draw 2 loses significance without the MIN_PARENT/nanmean rule.
+- **SEP-POP-5.** The report omits:
+  - cluster c9 is unstable across resolutions (Jaccard 0.09 and 0.18);
+  - marker concordance is 0.68, not the claimed 0.90;
+  - within-sepsis tracking of L2 MAIT is −0.03 to 0.13.
+  
+  Despite this, section 3 rates its artefact risk "low-medium".
+- **SEP-PROG-4.** For a double-centred statistic, a bulk broadcast is about 0 by construction, so it is a trivial comparison. The informative baseline is the csSAM deconvolution (`verify/SEP-PROG-4_recompute/v08_cssam.csv`): single-centred r 0.05–0.07, and partial r of generated given csSAM 0.17–0.31. The report does not cite it, and it was never run on the double-centred statistic.
+- **All population findings** were verified on the discovery agent's own Leiden labels. Independent re-clustering recovered SEP-POP-6's cluster with Jaccard ≤ 0.35 and SEP-POP-1's with 27–47% overlap.
+- **Lens table error.** In `sep_populations/tables/s11_training_donor_imprint.csv`, the train_groups column is misaligned with train_donors (PHC05 is listed as adult_sepsis). A verifier flagged it; the report does not mention it.
+- **Draw 0's "MS1-rich sepsis donors".** Skin-S74 has 0 MS1 cells among 8 monocytes, so only 3 of draw 0's 4 sepsis training donors are MS1-rich.
+
+### E. Overstatements
+
+1. **"The generated cells reproduce, in direction, the known sepsis-vs-healthy shifts" is selective.**
+   - It leaves out known shifts that fail: DC loss (2/5 draws), the drop in monocyte ISG (fails in 5/5), exhaustion (opposite sign), cycling cells, plasmablasts and low-density granulocytes.
+   - It counts as reproduced:
+     - NK loss (draw 4 reversed);
+     - plasma cells (no draw passes the lens gate; trust-layer ratio 0.15–0.46);
+     - CCL4+/NR4A2+ states (artefact-suspect, and their calibration does not transfer, see B3).
+2. **The "Draws" column lists direction-only draws that fail the DATA_GUIDE gate:**
+   - SEP-PROG-2: draws 2 and 4 at 0.19x and 0.26x;
+   - SEP-PROG-5: Spearman 0.03–0.28;
+   - SEP-CLIN-2: draw 0 at 0.09x.
+3. **"MS1 ... the cell state in draws 2–4" overstates.** The c8 cluster expresses IL1R2 in only 2–6% of cells (observed 42–44%). It covers 16% of observed MS1-rule cells, and the trust-layer MS1 state is at 0.05–0.14x.
+4. **"The only clear piece is SEP-PROG-4"** should say "modest". The generated cells capture 5–16% of the observed pattern, the comparison is against a trivial null, and there is no sepsis-vs-ICU signal (double-centred r 0.02–0.06).
+5. **The report presents honesty items 1 (calibration) and 3 (observed cells) as separate evidence,** but they share about 94% of their samples.
+
+### F. Cheap next steps (CPU)
+
+- **Trust layer:** rebuild it adult-only, add ICU-vs-HC/age/sex/source rows (`c02` is a template), and flag features whose bulk contrasts differ between paired and bulk-only samples (`c05`).
+- **Statistics:**
+  - Refit the affected quasi-binomial models with proportions and `var_weights`.
+  - Headline gated per-draw estimates, not the five-draw mean.
+- **Report contents:**
+  - Add DC and NK to the negative results, and cite `s12_gated.csv`.
+  - Request per-patient survival and SOFA data (Ye 2026 supplement or the authors).
+- **GPU (coordinator's call):** generate at least 2 seeds per draw, so draw agreement can be separated from shared noise.

paper-releases/pbisc-diffusion/updates/20261006T033406Z-bulkonly-lora5-k1000/diffusion/dataset/server_1/model8/plans/bulkonly_lora5_20261006/share/README.md

@@ -0,0 +1,125 @@
+# Generated single cells for bulk-only samples (LoRA-5, K = 1,000)
+
+Two cohorts, 2026-10-06:
+
+- **Ye 2026 sepsis** (GSE279448): 122 bulk samples that have no single-cell partner.
+  - 68 adult sepsis, 36 adult healthy controls, 17 adult ICU controls, 1 paediatric ICU control.
+- **Imperial ESKD COVID** (Gisby 2022 bulk; Stephenson 2025 single cell): 164 bulk samples without a single-cell
+  partner, from 76 donors.
+  - 143 COVID+, 19 COVID−, 2 recovery samples.
+
+**These cells are model predictions, not measurements.** Read the caveats below before using them.
+
+## How the cells were made
+
+- **Model.** PBISC-Diffusion (f01, 6k panel, 1x, epoch 16), fine-tuned with LoRA r16 on the paired bulk + single-cell
+  data of only 5 donors of the same cohort, for 3,000 steps. This is the "LoRA-5" setting of the paper's paired
+  benchmark.
+- **Five draws.** There are five such models per cohort (draws 0–4). Each was trained on a different set of 5 donors,
+  so every bulk sample has five independent model versions.
+  - The training donors of each draw are in each file's `uns["provenance"]`.
+  - `obs["donor_trained_draw<j>"]` marks Imperial samples whose donor trained draw j (their other visits were paired).
+- **Input.** Each sample's real bulk RNA-seq, as a fixed-depth condition at recipe size K = 1,000. This is the same
+  arithmetic as the paired benchmark.
+- **Sampling.** EMA weights, DDIM 50 steps, Bernoulli gene-detection read-out, 1,500 cells per sample.
+  - Sepsis: 183,000 cells per draw. Imperial: 246,000 cells per draw.
+- **Counts.** The model outputs expected counts. The h5ad files hold one Poisson draw of them (seed 20260916), as
+  integer UMIs.
+- **Labels.** The frozen f01common scANVI annotator gives a six-type gate (`celltype_l1_pred`), AIFI L1 (`aifi_l1`),
+  AIFI L2 (`aifi_l2`, 29 types) and its confidence.
+
+## Files
+
+| path | what |
+|---|---|
+| `<cohort>/lora5_bulkonly_<cohort>_draw<j>.h5ad` | One file per draw (AnnData; details below). |
+| `<cohort>/samples.csv` | Bulk-only sample metadata (clinical columns, sex flags, training-donor flags). |
+| `raw_npz/<cohort>/draw<j>/*.npz` + `*.log` | The sampler's own outputs and their JSON receipts. Imperial is in two row parts per draw. |
+| `analysis/REPORT_<cohort>.md` | The analysis report for each cohort. |
+| `analysis/DATA_GUIDE.md` | How the analysis files are organised. |
+| `analysis/features/<cohort>/calibration.csv` | Paired-test calibration (the trust layer). |
+| `analysis/findings_<cohort>.json` | Every candidate finding with its two adversarial verdicts. |
+| `analysis/analysis_bundle.tar.gz` | The full analysis tree: features, scripts and tables of every analysis agent. |
+| `MANIFEST.json` | Byte size and sha256 of every file under `<cohort>/`. |
+
+**h5ad contents:**
+
+- `X`: integer UMI counts (int32 CSR), cells × 6,000 panel genes.
+- `obs`:
+  - `sample_id`;
+  - the sample's clinical metadata;
+  - labels: `celltype_l1_pred`, `aifi_l1`, `aifi_l2`, `aifi_l2_confidence`;
+  - `lib_measured`: UMIs on the genes the cohort's single-cell assay measured.
+- `var`: `gene_id` (Ensembl), `gene_symbol`, `measured_in_cohort`.
+  - Genes the cohort's assay did not measure are generated as 0 by design. Drop them for cohort comparisons.
+- `uns["provenance"]`: a JSON string with the model, training donors, checkpoint sha256, seeds, generated-NPZ sha256
+  and label sha256.
+
+```python
+import json, anndata as ad
+a = ad.read_h5ad("lora5_bulkonly_ye2026_sepsis_draw0.h5ad")
+prov = json.loads(a.uns["provenance"])
+a = a[:, a.var.measured_in_cohort.to_numpy()]
+```
+
+**Raw NPZ contents:** `generated_counts` (expected counts), `generated_log1p`, `active_logits`, `active_mask`,
+`support_mask`, `bulk_index` (cell → sample row), `sample_ids`, gene ids/symbols and `metadata_json`. Each `.log` is
+the JSON receipt: seed, checkpoint and condition sha256, and output sha256.
+
+## Caveats
+
+1. **Predictions.** Each cell is what the model predicts for that bulk profile. Nothing here was measured at the
+   single-cell level.
+2. **Draws differ a lot.** Each draw imprints its five training donors.
+   - Example: in sepsis only draw 0 reproduces the MS1 monocyte level of the paired truth; draw 1 makes too many
+     monocytes.
+   - Analyse each draw separately and check that draw's calibration before trusting a feature.
+   - Do not average absolute levels over draws.
+3. **The draws share noise.** All five draws were sampled with seed 20260927, and the Poisson read-out uses one seed.
+   So cell slot c of sample i starts from the same noise in every draw.
+   - "4 of 5 draws agree" is therefore weaker evidence than five independent replications.
+   - The K = 4,000 version (shared separately) gives every draw its own seed.
+4. **Calibration.** `calibration.csv` compares generated and observed cells on each draw's paired test samples:
+   - mean levels and their ratio;
+   - the Spearman correlation over samples;
+   - whether each clinical contrast keeps its sign.
+
+   A feature or contrast that fails there should not be interpreted in the bulk-only cells.
+5. **Labels are predictions too.** They are scANVI calls on generated cells. Rare types (DC subsets, plasmablasts,
+   progenitors) are the least reliable.
+6. **Sex flags.** The bulk sex call differs from the metadata for HC23, HC30 and Res-S34 (sepsis) and C137_positive_2
+   (Imperial). These samples were generated, but they are flagged in `obs["sex_flag"]`.
+
+## Main results so far
+
+- **Sepsis** (`analysis/REPORT_ye2026_sepsis.md`; 18 candidate findings, each checked by two adversarial verifiers):
+  - **No finding is confirmed.** 17 are qualified and 1 is refuted.
+  - **Direction reproduced.** The generated cells reproduce the direction of known sepsis-vs-healthy shifts:
+    - monocyte HLA class II loss and an MS1/S100-high state;
+    - fewer MAIT, GZMK+ memory T and NK cells;
+    - more activated CCL4+ and NR4A2+ lymphocytes and more plasma cells.
+  - **Nothing new.** All of these are already reported and visible in the bulk.
+  - **Sepsis vs ICU controls** cannot be separated reliably.
+  - **No new cell population** survived verification.
+  - **Report caveats.** The report ends with a completeness review that lists its weak points, including the shared
+    draw noise above and a statsmodels dispersion bug in some verifier p-values.
+- **Imperial ESKD COVID** (`analysis/REPORT_imperial_eskd.md`; 21 candidate findings, each checked by two
+  adversarial verifiers):
+  - **No finding is confirmed.** 19 are qualified and 2 are refuted.
+  - **Direction reproduced.** The cells reproduce the direction of known COVID-19 shifts:
+    - an interferon response in every lineage that fades over the first weeks;
+    - monocyte HLA class II loss and an S100/CD163 monocyte programme with severity;
+    - more B-lineage cells, and more plasma cells at higher WHO grade.
+  - **Nothing new.** All of these are already reported (Stephenson 2025, Gisby 2022) and visible in the bulk.
+  - **Traps for anyone re-analysing this cohort:**
+    - Wave: 88% of bulk-only COVID+ samples are wave 1, but about half of the paired truth is wave 2, which had
+      another processing site and dexamethasone.
+    - Fresh bulk vs frozen cells: the bulk carries granulocyte RNA, which the model turns into extra CD14 monocytes.
+    - Interferon-linked label errors: naive B, "Proliferating NK" and CD8aa.
+  - **Fatal outcome** adds nothing beyond peak severity.
+  - **No new cell population** survived verification.
+
+## Contact and provenance
+
+- Plan directory on the lab server: `model_8_sparse_data_diffusion/plans/bulkonly_lora5_20261006/`.
+- The models are the few-pair draws of `plans/paired_finetune_intradataset_20260927/`.