Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/archive/pre_rewrite_20261010/GIT_HEAD.txt

@@ -0,0 +1 @@
+fe5b8d98458cfc3ce6e0a48fbb63a32f1cfda686

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/editorial/rewrite_20261010.md

@@ -0,0 +1,118 @@
+# Manuscript rewrite of 2026-10-10 (from the paper-summary deck)
+
+**Request (user, 2026-10-10):** rewrite the paper from scratch on the basis of the current deck, with Figure 6 marked
+"to be continued".
+
+**Source of truth:** the paper-summary deck
+`immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/`, as rebuilt on 2026-10-10
+(76 slides).
+- Slide text, notes and full footnotes were the primary source.
+- The deck's `data/*.csv` tables were used to check numbers. Their rows carry source paths.
+- The figure split follows `data/figure_plan.json`.
+- The previous manuscript (a different story, older checkpoints) was not used as a content source. Only method facts were
+  reused, after re-checking them against the code and run records.
+
+**Previous version:** `archive/pre_rewrite_20261010/` holds sections/, main.tex, supplementary.tex, tables/,
+references.bib, both PDFs, and the git HEAD and status at the time of the rewrite. This includes the uncommitted
+2026-10-06 Figure 1 redesign edits. `archive/` is git-ignored, so this copy is local and in the paper archive delta.
+
+## Structure
+
+| Part | Files | Content |
+|---|---|---|
+| Main text | `sections/abstract.tex`, `introduction.tex`, `results_fig1..6.tex`, `discussion.tex` | About 5,500 words. Abstract 200 words. |
+| Methods | `sections/methods_model.tex`, `methods_paired.tex`, `methods_eval.tex` | Very descriptive, per the Nature-style guide. |
+| Figures | `sections/fig1..6.tex`, `figures/deck_20261010/` | Deck vector panels placed as-is. |
+| Supplementary | `supplementary.tex`; `sections/supp_notes.tex`, `supp_methods_model.tex`, `supp_methods_eval.tex`, `supp_table_cohorts.tex`, `supp_tables.tex`, `supp_figures.tex` | 5 Notes, Methods, 4 Tables, 17 Figures. |
+| Shared | `macros.tex`, `references.bib` | Figure macros, `\todo{}` and 55 cited references (every entry is cited). |
+
+Figures are typeset after the references, each starting on a new page, with the legend after the panels. A figure may
+run over more than one page because the deck panels are slide-sized. Provenance and sha256 of every copied panel are in
+`figures/deck_20261010/PROVENANCE.json`. The 180 mm journal layouts are still to be drawn.
+
+Figure 6 ("Discovery from public bulk-only data") is a placeholder. The text states only what was generated: LoRA-5 cells
+for 122 sepsis and 164 COVID-19 haemodialysis bulk-only samples, at K = 1,000 and K = 4,000. No findings are reported.
+
+## Title
+
+Chosen: "Five paired donors and minutes of fine-tuning turn blood bulk RNA-seq into single cells" (14 words). Other
+candidates: "Single blood cells generated from bulk RNA-seq carry disease signal beyond the bulk" and "A few paired samples
+and minutes of fine-tuning turn blood bulk RNA-seq into single cells".
+
+## Rules applied (user decisions, 2026-10-06 to 10-08)
+
+- The main comparison is LoRA-5 vs scSemiProfiler-5, with the same five donors and five draws per cohort. 5-fold CV is
+  reported separately and is never the head-to-head.
+- No bulk-broadcast baseline anywhere.
+  - The deck figure `sec04_metrics_overview` (Fig. 2d) still listed it as a comparator.
+  - It was redrawn on 2026-10-10 (`tools/sec04_metrics_overview.py`; old files in the deck's
+    `_archive/2026-10-10_metrics_overview_broadcast/`) and the deck was rebuilt.
+- Partial r, the cell-type signal beyond the bulk, is emphasised. So is the 5-fold disease logFC r of 0.82 (Imperial).
+- Time counts per-cohort GPU work only: no PBISC pretraining and no queue waits.
+- No PBISC-VAE, no refcal, no TPM input variants in the main text. Gene-wise calibration is a comparator.
+- "Paired-sc level" wording is avoided, while the paper keeps the claim that PBISC is the best of the tested methods.
+- BayesPrism and CIBERSORTx are always called cell-type deconvolution.
+- Modified-baseline runs appear in the Supplementary only (Note 4, Fig. 17). The from-scratch pilot is in the
+  Supplementary (Note 3, Fig. 16) and is framed positively in the Discussion.
+- Style: thesis first, bold lead-in keywords, short body, `[TODO]` for unfinished work, British spelling, no Korean.
+
+## Process
+
+1. Nine drafting agents wrote Results Figs 1–5, Methods (3 parts) and the Supplementary from a written brief. The
+   coordinator wrote the Fig. 6 placeholder.
+2. Test compile with tectonic.
+3. Eleven adversarial checkers (one per file group) traced every number and claim to the deck text, the deck data, the
+   plan READMEs and the code. Fixers applied the confirmed corrections (about 265).
+4. A separate agent wrote the abstract, introduction and discussion, which were then checked and fixed the same way.
+   The references were built and checked against Crossref and the reference PDFs.
+5. The coordinator made the cross-file fixes and wrote the Fig. 6 text.
+6. The main text was tightened from 6,364 to 5,543 words. An old-vs-new check restored 39 losses (definitions,
+   qualifiers, citations).
+7. A whole-paper review covered consistency, cross-references, rules and a referee read (three reviewers, 43 issues),
+   followed by per-file fixes, a few coordinator fixes and a final compile. Both documents compile with no undefined
+   reference, no overfull box and no bibliography warning (main 48 pages, Supplementary 46 pages).
+
+## Conflicts between the deck and its sources, resolved in the text
+
+- **SLE cohort (GSE135779).** The 42 scored donors are children and adults (25 cSLE, 5 aSLE, 8 cHD, 4 aHD), not a
+  childhood-only cohort. It also chose the width and gene panel, so the text calls it a development cohort.
+- **Corpus shares.** 88.5% healthy and 23.6% in vitro are shares of cells, not of donors.
+- **Top-up recipes.** They make every training cell a target at least once, not exactly once.
+- **Rare states (Fig. 5).** The analysis unit is the scored sample (several per donor in Imperial, Giroux and Ji).
+  "Donor ranking" therefore means Spearman over test samples. The deck axis label is kept.
+- **Bulk-only cells at K = 1,000.** The five draws shared one sampling seed, so they are not independent. Only the
+  K = 4,000 set has a separate seed for each draw.
+- **scSemiProfiler vs copying its representative.** It was never *significantly* better than the copy (Imperial 0.602 vs
+  0.608, P = 0.91), not "never better".
+- **CIBERSORTx top-50 DEG.** 6 of 17 scored draws. Three draws had no imputed contrast genes and are unscored.
+- **Composition wins vs scSemiProfiler-5.** 4 of 5 cohorts by the draw verdict; 5 of 5 by cohort means. The text names
+  the criterion each time.
+- **Behçet.** The bulk and single-cell donors are mostly different people. Three patients were matched afterwards by
+  expression; they are used only by the LoRA-3 to-do.
+- **Width-arm parameter counts.** Taken from the run records (102,106,848 and 312,649,440), not from the old manuscript.
+
+## Open items (`\todo{}` in the text)
+
+- **Fig. 5d.** The LoRA-3 Behçet arm.
+- **Fig. 5e.** Rare states in cohorts absent from the corpus (Kawasaki MS1, SLE ISG-high T).
+- **Fig. 6.** The whole figure and its findings.
+- **Release links.** Model, LoRA weights, code, generated cells, Source Data and a Zenodo DOI.
+- **Corpus datasets.** The list of the 22 corpus datasets with accessions; some CELLxGENE entries lack a collection ID.
+- **Author items.** Authors, funding, contributions, competing interests and ethics.
+- **Fig. 1 panels b and d.** These reuse manuscript images that print their own panel letters.
+
+## Deck changes made during the rewrite
+
+- `tools/sec04_metrics_overview.py` (Fig. 2d): the bulk-broadcast comparator was removed from the metrics table.
+- `tools/sec07_rank_sweep.py` (Supplementary Fig. 12a): the key "(used in the deck)" now reads "(used throughout)".
+- The deck was rebuilt after each change (76 slides). Old files are in the deck's `_archive/2026-10-10_*` folders.
+
+## Data oddity found during the rewrite (not used in any claim)
+
+scSemiProfiler's pooled coverage in the equal-pair-budget runs at n = 40 equals its 5-fold value to the last digit, in
+both ImMicro (617 of 16,050 real cells) and sepsis (1,181 of 12,300). Yet per-donor coverage differs for most donors, and
+composition L1 differs.
+- Sources: `plans/paired_finetune_intradataset_20260927/scores/<cohort>/scsemi_budget_20261002/pooled/` and
+  `.../scsemi_cv_20261002/pooled/`.
+- The claims (scSemiProfiler coverage 0.04 and 0.10 at every n) do not depend on it, but the scoring path is worth a
+  check.

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/editorial/rewrite_20261010_brief.md

@@ -0,0 +1,271 @@
+> Writing brief given to the drafting and checking agents of the 2026-10-10 rewrite (see rewrite_20261010.md).
+> The `/tmp/claude-1000/.../scratchpad/paper/` paths were a temporary staging area and no longer exist; the
+> staged files are now `main.tex`, `supplementary.tex`, `macros.tex`, `references.bib` and `sections/`.
+
+# BRIEF — rewrite of the PBISC-Diffusion manuscript from the current deck (2026-10-10)
+
+You are one of several agents rewriting the PBISC-Diffusion paper **from scratch**. The single source of truth for
+content is the **current paper-summary deck (76 slides, last built 2026-10-08)**. The old manuscript is NOT a content
+source (its story, checkpoints and numbers are superseded); you may reuse only verified *method facts* from it where
+noted below, after checking them against the current sources.
+
+Venue and audience: **Nature Communications Article**, read by immunologists and computational biologists. They care
+about results, not ML novelty: what the cells are good for, how few paired samples are needed, how fast it is, and
+how it compares with the methods they know.
+
+## 0. Paths
+
+| What | Path |
+|---|---|
+| Deck text, every slide incl. speaker notes and full footnotes (PRIMARY SOURCE) | `/tmp/claude-1000/-home-sj-server-1/cd3948fe-c9dc-46c9-874d-211078d8ba7a/scratchpad/paper/deck_text.txt` |
+| Deck folder | `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/` (`DECK/` below) |
+| Deck data behind every slide (CSV/JSON; most rows carry a `source` path) | `DECK/data/` |
+| Deck figure scripts | `DECK/tools/` |
+| Deck figures (vector PDF per slide figure) | `DECK/figures/<name>.pdf` |
+| Figure plan (single source of the figure split) | `DECK/data/figure_plan.json` |
+| Deck README (change log, decisions) | `DECK/README.md` |
+| Model 8 project root (`M8/` below) | `/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/` |
+| Paired fine-tuning plan (cohorts, splits, LoRA, scoring) | `M8/plans/paired_finetune_intradataset_20260927/` (README.md, data/<cohort>/QC_REPORT.md, tools/) |
+| Deconvolution comparison | `M8/plans/bayesprism_paired_20261006/` |
+| Rare states (Fig. 5) | `M8/plans/rare_states_paired_20261008/`, `M8/plans/rare_states_screen_20261004/` |
+| Bulk-only generation (Fig. 6, placeholder only) | `M8/plans/bulkonly_lora5_20261006/`, `M8/plans/bulkonly_lora5_k4000_20261006/` |
+| Final model run | `M8/runs/f01_6k_1x_seed20260627/` (effective_config.json, RUN_STATUS.json, ALLCELL_DATA_CONTRACT.json) |
+| Model source code | `M8/src/pbisc_sdd/` |
+| Old manuscript (method facts only) | `/home/sj_server_1/PBISC/paper-release-20260928/manuscript/` (`sections/methods.tex`, `sections/supplementary_methods.tex`, `source_data/corpus/`) |
+| Reference PDFs and index (citation metadata, DOIs) | `/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/INDEX.md` |
+| Staging folder for ALL outputs | `/tmp/claude-1000/-home-sj-server-1/cd3948fe-c9dc-46c9-874d-211078d8ba7a/scratchpad/paper/draft/` |
+
+Do not modify anything outside the staging folder. CPU only; never start GPU work. Never print credential files.
+
+## 1. Rules from the user (mandatory)
+
+1. **Main comparison = 5 paired donors.** LoRA-5 (LoRA fine-tuned on the pairs of 5 donors) vs **scSemiProfiler-5**
+   (scSemiProfiler given the same 5 donors' single cells as representatives). 5 draws per cohort; test = every other
+   usable sample, from its real bulk. This is the head-to-head claim.
+2. **5-fold CV is reported separately**, as an additional result, never mixed into the main 5-pair claim and never
+   called the head-to-head comparison.
+3. **No bulk-broadcast baseline** (bulk logFC copied to every cell type). Do not mention it anywhere, not even as a
+   control. Partial r stays (it uses the bulk only as a covariate).
+4. **Emphasise partial r** = "cell-type signal beyond the bulk": logFC r after both logFCs are residualised on the input
+   bulk's logFC. Present it as what single cells add over bulk.
+5. **Emphasise that LoRA reaches disease logFC r = 0.82** (Imperial, 5-fold). This is remarkable. Do not compare it with
+   the split-half ceiling as if it were a shortfall; do not mention the ceiling in a way that diminishes it.
+6. **Time:** count only per-cohort GPU time (scSemiProfiler: pretraining of each representative + one model per test
+   sample; LoRA: fine-tune + sampling). **Never add PBISC pretraining (15.9 h)** to LoRA's time and never mention GPU
+   queue waits. Neither method's provided assets are counted. Time is a main result: scSemiProfiler takes days, LoRA
+   minutes.
+7. **Wording:** do not claim "recovers paired single cells" / "paired-sc level". Do claim that PBISC-Diffusion with LoRA is
+   the best of the tested methods, with numbers.
+8. **No PBISC-VAE** anywhere. **No "refcal"/reference calibration/AIFI anchoring.** No TPM input variants in the main
+   text. "Gene-wise calibration" (alpha_g = mean pseudobulk CPM / mean bulk CPM over the paired training donors) IS a
+   comparator: show zero-shot (real bulk as measured) and zero-shot + calibration.
+9. Do not frame pretraining as an unfair advantage or a "fairness caveat". Pretraining is part of the method.
+10. A run that modifies a baseline's official behaviour (scSemiProfiler without its output threshold) is a
+    Supplementary-only sensitivity check, labelled "modified, not the official output". Main text uses official output.
+11. BayesPrism and CIBERSORTx are **cell-type deconvolution** (fractions and per-type expression, not single cells);
+    say so every time they appear.
+12. From-scratch (no pretraining) results are **Supplementary**, framed positively in the Discussion as a route to
+    other tissues (pilot: sepsis fold 0 only; folds 1–4 pending). Keep the pretraining advantages explicit (5-pair
+    donor composition r 0.70 vs 0.17; cycling T/NK closer 8 of 8; minutes vs ~1 h per cohort).
+13. **Figure 6 is "to be continued".** Only a placeholder subsection and a placeholder figure. State what exists (LoRA-5
+    cells generated for the bulk-only samples: sepsis 122, COVID-19 haemodialysis 164, at K = 1,000 and K = 4,000) and
+    mark the analysis `\todo{...}`. Report NO findings and do not mention earlier screens.
+14. Honest caveats belong in the text, briefly, where a reviewer would look: Giroux (5-donor test sets of 2 vs 2
+    subjects; scSemiProfiler ahead there), sepsis equal-budget logFC r where scSemiProfiler is higher, generated vs real
+    still separable by a classifier (AUC ~0.9), most rare states generated below their observed fraction, zero-shot
+    closer on some rare-state fractions. State them as facts with numbers, without hedging the main claims.
+
+## 2. Style (user's standing rules for this manuscript)
+
+- **Thesis first.** Every paragraph opens with its claim; evidence follows. Results subsection titles are claims.
+- **Bold lead-in keywords**: `\textbf{Five pairs are enough.} ...` style, at most one per paragraph.
+- **Short body.** Off-thread experiments get one sentence and a pointer to the Supplementary.
+- **Plain English.** Short sentences with verbs, common words. No symbol strings ("→", "≈", "·") in prose; no lists of
+  fragments. Define every technical term at first use in each part (e.g. "coverage, the share of real cells with a
+  generated neighbour").
+- Every number comes with what it is, which data and the comparator: "coverage 0.79 vs 0.04 for scSemiProfiler-5
+  (ImMicro, 5 draws)". Use counts of wins as "in 13 of 20 draws".
+- `\todo{...}` for anything not done yet (macro defined in main.tex; renders red). Never invent a result.
+- **No Korean anywhere.** British spelling (normalised, modelling, colour, haemodialysis). Use "pseudobulk", "single-cell"
+  (adj.), "scRNA-seq", "logFC", "top-50 DEG", "5-fold cross-validation (CV)".
+- Nature Communications limits (targets): title ≤ 15 words; abstract ≤ 200 words, no references; main text
+  (Intro + Results + Discussion) about 5,000 words; legends ≤ 350 words each; Discussion without subheadings.
+
+## 3. Terms (use exactly)
+
+- **\method{}** = PBISC-Diffusion (macro). The pretrained model = "the pretrained model" / "\method{}".
+- **zero-shot** = pretrained model fed the real bulk as measured. **zero-shot + calibration** = fed the bulk after
+  gene-wise calibration fitted on the same training pairs.
+- **LoRA-5** = LoRA (rank 16) fine-tuned on all pairs of 5 donors, fixed 3,000 steps; **scSemiProfiler-5**.
+- **5-fold CV**: donor-level; every sample generated by the fold model that never trained on it; one pooled scoring
+  call per method; scSemiProfiler's representatives = exactly LoRA's training (+ inner-validation) pairs.
+- **pairs-needed curves** ("learning curve"): inside the 5-fold design, n = 5, 10, 20, 40 or all training pairs per fold,
+  early-stopped on 7–8 inner-validation pairs. These "5 pairs" are NOT the 5-donor draws; say which one you mean.
+- Cell metrics: **precision** (share of generated cells inside the k-NN ball of a real cell; Kynkäänniemi 2019),
+  **coverage** (share of real cells with a generated cell inside their k-NN ball; Naeem 2020), k = 10, PCA-50 fitted on
+  real cells; **composition L1** over six scANVI cell types (lower is better); **classifier AUC** (kNN classifier
+  two-sample test, 0.5 = indistinguishable); **MMD²**; **E_N** (scSemiProfiler's own normalised error, Wang 2024).
+- Disease metrics: **logFC r** (Pearson r of case-vs-control logFC per cell type, generated vs observed, mean over T,
+  B, monocyte, NK); **partial r**; **top-50 DEG** overlap (Welch + BH, summed over types); **CAT** (correspondence at
+  the top, Irizarry 2005).
+- Cohorts: ImMicro (107 healthy donors; Babdor 2026); Ye 2026 sepsis ("sepsis"); Imperial COVID-19 haemodialysis
+  ("Imperial"; bulk Gisby 2022, CITE-seq Stephenson 2025); Giroux 2022 COVID-19 outpatients ("Giroux"); Ji 2025
+  checkpoint-inhibitor liver injury ("Ji", whole-blood bulk); childhood SLE GSE135779 (Nehar-Belaid 2020); Behçet
+  disease (bulk GSE198533, sc GSE198616; Zheng 2022).
+
+## 4. Manuscript structure, owners and word targets
+
+Main text (`main.tex` inputs, in order): title; abstract; Introduction (650 words); Results = six subsections
+(Fig.1 550, Fig.2 800, Fig.3 600, Fig.4 950, Fig.5 550, Fig.6 placeholder ~120); Discussion (750); Methods (very
+descriptive; ~4,500 words over three files, overflow to Supplementary Methods); declarations; references; then the
+figures with legends (each figure starts on a new page).
+
+Results subsections map to deck slides (the deck text has "===== SLIDE n" markers):
+
+| Results / Figure | Deck slides (main panels) | Details slides usable as Supplementary |
+|---|---|---|
+| Fig. 1 Design and performance | 6, 7, 8, 9, 10, 11, 12 | 13, 14 |
+| Fig. 2 Fine-tuning with LoRA (healthy ImMicro only; NO disease metrics) | 17–27 | — |
+| Fig. 3 vs scSemiProfiler on healthy donors + time | 30–34 | 62, 63 |
+| Fig. 4 Disease signal, paired cohorts | 37–43 | 44, 45, 46 |
+| Fig. 5 Rare cell states | 49–52 (+ planned 5e) | 72 |
+| Fig. 6 Bulk-only discovery | placeholder | — |
+| Supplementary | 58–72, 74 | |
+| Discussion | 76 (+ caveats) | |
+
+## 5. Figures: panel → deck PDF (copied later to `figures/deck_20261010/<name>.pdf`)
+
+Main figures (write `\includegraphics` paths exactly as `figures/deck_20261010/<name>.pdf`):
+
+- **Fig. 1** a `sec02_backbone_compare`; b `fig01_architecture_nature` (manuscript figure, Nature-style overview);
+  c `sec02_bulk_conditioning`; d `fig01_corpus_nature` (manuscript figure); e `sec03_width_loss` + `sec03_width_bench`;
+  f `sec03_panel`.
+- **Fig. 2** a `sec05_condition_space` + `sec05_shift_cohorts`; b `sec05_cells_umap` + `sec05_cells_metrics`;
+  c `sec06_ft_method` + `sec06_eval_design`; d `sec04_metrics_overview`; e `fig2_healthy_cells`;
+  f `fig2_healthy_realism`; g `fig2_healthy_modes`; h `fig2_healthy_pairs`; i `fig2_healthy_rank`.
+- **Fig. 3** a `fig3_healthy_fewpair`; b `fig3_healthy_pairs`; c `fig3_healthy_copy`; d `fig3_time_days` +
+  `sec09_time_to_apply`.
+- **Fig. 4** a `sec09_cohorts`; b `fig4_disease_fewpair`; c `fig4_disease_fivefold`; d `sec09_cat_curves`;
+  e `secS_deconv_disease`; f `sec04_sle_benchmark` + `sec04_sle_per_celltype`.
+- **Fig. 5** a `fig5_rare_fraction`; b `fig5_rare_donors`; c `fig5_rare_fivefold`; d `sec05_behcet_direction`
+  (status Adapt: the LoRA-3 Behçet result is to be added → `\todo`); e planned panel (Kawasaki MS1, SLE ISG-high T;
+  status Build → `\todo`, no image).
+- **Fig. 6** placeholder box only (macro `\tbcpanel{...}`), no deck image.
+
+Supplementary Figures (fixed numbering; refer to them as "Supplementary Fig.~N" in plain text):
+
+| # | Content | Deck PDF(s) | Deck slide |
+|---|---|---|---|
+| 1 | Pretraining corpus, detailed | manuscript `figS_corpus` | 9 |
+| 2 | Model details: training pairs and denoiser | `sec01_slide_a`, `sec01_fig1a_denoiser` | 13, 14 |
+| 3 | 5 donors, all five cohorts: cell metrics | `sec09_fewpair_cells` | 58 |
+| 4 | 5 donors: realism beyond precision/coverage | `sec09_extra_fewpair` | 59 |
+| 5 | 5 donors: precision/coverage at k = 3, 5, 10 | `sec09_extra_ksens` | 60 |
+| 6 | 5-fold CV: realism metrics, all cohorts | `sec09_extra_fivefold` | 61 |
+| 7 | Same training pairs (5-fold): LoRA vs scSemiProfiler | `sec09_same_pairs` | 62 |
+| 8 | scSemiProfiler copies its representative, all cohorts | `sec09_scsemi_representative` | 63 |
+| 9 | Full fine-tuning vs LoRA vs condition-encoder-only | `sec07_ft_modes` | 64 |
+| 10 | Pairs needed: composition, coverage, training steps | `sec08_pairs_realism` | 65 |
+| 11 | Pairs needed: donor and disease signal | `sec08_pairs_signal` | 45 |
+| 12 | LoRA rank and K diversification | `sec07_rank_sweep`, `sec07_k_aug` | 66, 67 |
+| 13 | Sepsis: fine-tuning variants | `fig4_disease_variants` | 44 |
+| 14 | Cell-type deconvolution: composition | `secS_deconv_composition` | 46 |
+| 15 | Pairs from other cohorts do not transfer | `sec08_cross_cohort`, `sec08_cross_cohort_steps` | 68, 69 |
+| 16 | Training without pretraining (same cohort, other cohorts, rare states) | `sec07_no_pretraining`, `sec08_cross_scratch`, `secS_scratch_rare_states` | 70, 71, 72 |
+| 17 | scSemiProfiler without its output threshold (modified, sensitivity only) | `secA_scsemi_readout` | 74 |
+
+Supplementary Notes (fixed numbering): 1 Pretraining corpus; 2 One LoRA per cohort: pairs from other cohorts;
+3 Training without pretraining; 4 scSemiProfiler read-out sensitivity; 5 scSemiProfiler representative choice.
+Supplementary Tables: 1 Paired cohorts and contrasts; 2 Per-cohort values of the 5-donor comparison; 3 Per-cohort
+values of the 5-fold comparison; 4 GPU time per cohort. (Tables may be built by the supplementary agent from DECK/data.)
+
+LaTeX macros available (defined in main.tex / supplementary.tex; do not redefine):
+```
+\method{}                       PBISC-Diffusion
+\todo{text}                     red [TODO: text]
+\mainfigure{label}{Title sentence}   starts a new page, numbers the figure, prints "Figure N | Title"
+\figpanel{a}{figures/deck_20261010/name.pdf}          one panel image, full width (max 0.31\textheight)
+\figpanelh{a}{file}{0.25}       same with an explicit max height as a fraction of \textheight
+\figpanelpair{a}{fileA}{fileB}{0.48}{0.48}            two images side by side at the given widths (fraction of \textwidth)
+\tbcpanel{text}                 framed "To be continued" box with text
+\figlegend{...}                 the legend paragraph (bold title sentence first, then **a**, ... panel descriptions)
+\suppfigure{label}{Title}       like \mainfigure for Supplementary Fig. N
+```
+Example of one main figure block (the figure file holds only this):
+```
+\mainfigure{fig:fig3}{LoRA cells are closer to real cells than scSemiProfiler's, in minutes instead of days}
+\figpanel{a}{figures/deck_20261010/fig3_healthy_fewpair.pdf}
+...
+\figlegend{\textbf{LoRA-5 against scSemiProfiler-5 on healthy donors.} \textbf{a}, ... \textbf{b}, ...}
+```
+Refer to panels in text as `Fig.~\ref{fig:fig3}a` (labels: `fig:fig1` ... `fig:fig6`). Citations: `\cite{key}` placed
+before the period with no space (`... bulk\cite{wang2024scsemiprofiler}.`).
+
+## 6. Citation keys (use these; if you need another, use a key `firstauthorYEARword` and list it in your return value
+with the full reference so the references agent can build it)
+
+Existing in references.bib: `ostheimer2025sparse` (SDD backbone), `luo2024scdiffusion`, `ho2020ddpm`, `song2021ddim`,
+`wang2024scsemiprofiler`, `newman2019cibersortx`, `chu2022bayesprism`, `avilacobos2020benchmarking`, `lopez2018scvi`,
+`cui2024scgpt`, `theodoris2023geneformer`, `nehar2020mapping` (GSE135779 childhood SLE), `benjamini1995fdr`,
+`robinson2010edger`, `mcinnes2018umap`, `babdor2026immicro` (ImMicro), `randolph2021genetic`, `zheng2022c1qbehcet`.
+New (to be added): `hu2022lora`, `kynkaanniemi2019improved`, `naeem2020reliable`, `lopezpaz2017revisiting`,
+`gretton2012kernel`, `irizarry2005multiple`, `webber2010similarity`, `squair2021confronting`, `vinastorne2026systema`,
+`roohani2024gears`, `li2026systematic`, `khayatkhoei2023emergent`, `law2014voom`, `xu2021scanvi`, `gong2025multiomic`
+(AIFI immune health atlas), `ye2026sepsis`, `gisby2022multiomics`, `stephenson2025temporal`, `giroux2022differential`,
+`ji2025chili`, `reyes2020immune` (MS1), `tukiainen2017landscape` (X escape genes), `arunachalam2020systems`,
+`liu2021timeresolved`, `wang2021kawasaki`, `nichol2021improved` (cosine schedule), `chen2023analog` (self-conditioning),
+`loshchilov2019decoupled` (AdamW).
+
+## 7. Numbers discipline
+
+- Take every number from the deck text (slide keywords, notes, full footnotes) or from `DECK/data/*.csv`. If two sources
+  disagree, use the CSV and report the conflict in your return value.
+- Round as the deck does (r to 2 decimals; L1 to 2–3 decimals; coverage to 2 decimals). Keep the n and the unit.
+- Never compute new statistics unless you state exactly how and from which file (and the verifier can redo it).
+
+## 7b. Resolved decisions after the first drafts (2026-10-10; these OVERRIDE anything above)
+
+1. **SLE cohort GSE135779 is children AND adults**: the 42 scored donors are 25 childhood SLE, 5 adult SLE, 8 childhood
+   and 4 adult controls (30 SLE, 12 controls). Never write "childhood SLE" or "paediatric" for it. Write "an SLE cohort
+   of children and adults (GSE135779)". It is not in the pretraining corpus, but SLE itself is (Perez 2022 adult SLE;
+   GSE163121 SLE B cells), so say "cohort absent from the corpus", never "disease absent from the corpus".
+2. **The SLE cohort was used to choose the design** (width, gene panel; Fig. 1e,f), so it is a development cohort, not
+   an untouched test set. Fig. 4f text and the Discussion say so once. The final model was fixed before any paired
+   cohort was used.
+3. **Corpus shares are shares of cells**: 88.5% of cells from healthy donors, 23.6% of cells cultured in vitro.
+4. **Top-up recipes make every training cell a target at least once** (not "exactly once").
+5. **Rare states (Fig. 5)**: the unit in the analysis code is the scored *sample* (Imperial, Giroux and Ji have several
+   samples per donor). Inclusion = at least 30 observed state cells (native depth) in at least 3 samples; ranking =
+   Spearman over test samples. Use "samples" in text and legends. Plasmablast definition: cite `mathew2020deep`
+   (not arunachalam2020systems).
+6. **Bulk-only K = 1,000 draws shared one sampling seed** (cell slot noise shared across draws); only the K = 4,000
+   set has per-draw seeds. Never call the K = 1,000 draws "independent".
+7. **Width-arm parameter counts**: use the run records (102,106,848 and 312,649,440), not the old manuscript values.
+8. **Annotator**: final model and paired cohorts use the frozen scANVI annotator `aifi_scanvi_f01common_v1` (1,005
+   genes; L2 accuracy 0.869, macro-F1 0.850; six-type gate 0.992 / 0.981); the width/proportion-hint ablations used the
+   2,980-gene annotator (0.889 / 0.870). Do not quote 0.880 / 0.860 (an annotator not used here).
+9. **Metric definitions differ between benchmarks**: paired cohorts: logFC r = mean over T, B, monocyte, NK (mean CP10K);
+   top-50 DEG = Welch + BH (FDR < 0.05) overlap summed over T, B, monocyte, NK, DC (max 250); CAT = mean over T, B,
+   monocyte, NK. SLE benchmark: logFC r and partial r averaged over six types (incl. DC and Other) on log(1+CPM); top-50
+   summed over six types. Say which one applies.
+10. **Batches in LoRA training are pair-balanced** (a sample pair drawn uniformly, then a cell), not donor-balanced.
+11. **Inner validation**: 7 pairs (ImMicro), 7–8 (sepsis), ~10% of donors elsewhere (Imperial 5 patients, Giroux 1–2
+    subjects, Ji 2 patients).
+12. **Identity checks** are sex-gene and expression-match checks; there is no genotype check.
+13. **Behçet**: no designed pairs; the bulk (9 BD, 10 HC) and single-cell donors (4 BD, 4 HC) are mostly different people
+    (three BD donors were matched afterwards by expression; used only by the LoRA-3 to-do). Panel 5d uses a looser C1Q
+    rule (any C1Q gene) than Fig. 5a–c (at least 2 of 3).
+14. **scSemiProfiler vs copying its representative**: "never significantly better than the copy" (Imperial 0.602 vs
+    0.608, P = 0.91), significantly worse in three of five cohorts.
+15. **CIBERSORTx top-50 DEG**: 6 of 17 scored draws (3 draws had no imputed contrast genes and are unscored).
+16. **Composition wins of LoRA-5 vs scSemiProfiler-5 across cohorts**: 4 of 5 cohorts by the deck's draw verdict
+    (>= 80% of draws; Giroux 3 of 5 draws); 5 of 5 by cohort means. Always name the criterion.
+17. **Deck figure for Fig. 2d was redrawn** without the bulk-broadcast comparator (logFC r row now reads "zero-shot;
+    partial r (vs bulk)"). Remove any LaTeX comment or to-do about that redraw.
+18. **Rare-state averaging**: a draw in which a method generated no cell of a state is skipped in the mean log2 ratio;
+    a state never generated in any draw counts as that method's miss. State this in Methods.
+
+## 8. Return value
+
+Return a short JSON-like summary: files written (absolute paths), word count per file, every `\todo` you left, every
+citation key used, any new keys requested (with full reference), and any conflicts or doubts about numbers.