Archive 4 checkpoints (1-4)

.gitattributes

@@ -241,3 +241,5 @@ paper-releases/pbisc-diffusion/updates/20261013T022035Z-pbisc-red-scsemi-blue/di
 paper-releases/pbisc-diffusion/updates/20261013T022035Z-pbisc-red-scsemi-blue/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_results_review_20261013.pptx filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20261013T023822Z-pbisc-yellow-scsemi-blue/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_results_review_20261013.pdf filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20261013T023822Z-pbisc-yellow-scsemi-blue/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_results_review_20261013.pptx filter=lfs diff=lfs merge=lfs -text
+paper-releases/pbisc-diffusion/updates/20261013T034842Z-expanded-131-slide-master/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261013.pdf filter=lfs diff=lfs merge=lfs -text
+paper-releases/pbisc-diffusion/updates/20261013T034842Z-expanded-131-slide-master/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261013.pptx filter=lfs diff=lfs merge=lfs -text

paper-releases/pbisc-diffusion/updates/20261013T034842Z-expanded-131-slide-master/diffusion/dataset/server_1/model8/AGENTS.md

@@ -0,0 +1,57 @@
+# Model 8 PBISC-Diffusion paper maintenance
+
+This active Model 8 source and its experiments are part of the
+`pbisc-diffusion` paper. At task start, read
+`../../paper-release-20260928/maintenance/POLICY.md` along with this
+directory's README and applicable parent instructions.
+
+At the end of each related task, follow that policy's versioned update
+workflow for the exact source, experiment, checkpoint, or evidence assets that
+changed. From the release root, use
+`maintenance/update.py prepare --request FILE --id ID`, inspect the prepared
+update, then run `maintenance/update.py publish ID --execute` and
+`maintenance/update.py status` under the policy. Record a no-op or pending
+status when no
+asset changed or publication failed; never describe that as an upload. Do not
+regenerate or alter the frozen 2026-09-28 release, its root manifests, or the
+HF `paper-20260928` revision. Inference assets are location plus receipt by
+default unless the task explicitly needs their payload.
+
+The private dataset must stay private. Include datalake, VAE, or Model 7 assets
+only with exact evidence that they are dependencies of this diffusion paper.
+The parent host GPU safety rules still apply.
+
+## Reuse existing experiments before starting compute
+
+User instruction, 2026-10-13: avoid repeating completed Model 8 experiments.
+Start with `EXPERIMENT_INDEX.md` and query `EXPERIMENT_REUSE_INDEX.jsonl` using
+`scripts/audit_experiment_reuse.py`. A missing output in a new plan directory
+does not establish that an experiment has never been done.
+
+- Before training, generation, or deconvolution, compare existing source
+  receipts: base checkpoint, exact support/test sample and donor IDs, data and
+  condition content, measured genes/masks, validation access, optimization and
+  selection schedule, seed, and implementation/configuration provenance.
+- Registry matches are candidates, not automatic equivalence. Verify payloads
+  and write truthful adoption evidence. Preserve original execution provenance;
+  do not relabel an old artifact as newly executed strict work.
+- Changing a sampling seed or receipt format does not justify retraining an
+  existing model. Reuse verified checkpoints and run only the genuinely changed
+  operation. Missing native receipts call for adoption, not automatic reruns.
+- Record the candidate sources and the exact scientific reason for any new
+  operation before launching it. Unknown provenance requires investigation,
+  not a conclusion that new compute is necessary. Consult newer live receipts
+  when the inventory predates an ongoing campaign.
+- Distinguish existing results, CPU-only statistics/plotting, a new protocol
+  update, and missing data in plans and decks. Never label all four as an empty
+  TODO. Archive/backup copies and tensor-identical models are not independent
+  experimental evidence.
+
+The 2026-10-13 audit and confirmed duplicate-training proof are in
+`plans/paper_figures_strictpairs_20261012/audit_reuse_20261013/`.
+Sepsis deconvolution uses verified historical draw-specific outputs through
+`tools/adopt_deconv_legacy.py` in that plan; its methods must not be rerun as a
+fallback when adoption validation fails. Current expanded evidence master (all original 76 slides retained, plus additions):
+`presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261013.pptx`.
+Use its `data/expanded_master_20261013/original_76_coverage.csv` and source manifest.
+The 34-slide review is a historical subset, not the master. PBISC uses muted yellow/gold and scSemiProfiler blue.

paper-releases/pbisc-diffusion/updates/20261013T034842Z-expanded-131-slide-master/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261013.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d63a97a37008aae9fcbd7388ae58accc1aab68378d75c082d6edc41b51b955e5
+size 16957455

paper-releases/pbisc-diffusion/updates/20261013T034842Z-expanded-131-slide-master/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261013.pptx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f2d1d38cf73e0d0482a6e9255941337f8d13dba365ded0b95af37efb5dd9bcc
+size 19526237

paper-releases/pbisc-diffusion/updates/20261013T034842Z-expanded-131-slide-master/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/README.md

@@ -0,0 +1,338 @@
+# PBISC-Diffusion expanded evidence master
+
+**Current working master: 131 slides = all original 76 slides + 55 additions.**
+
+[Editable PPTX](PBISC-Diffusion_expanded_master_20261013.pptx) · [PDF](PBISC-Diffusion_expanded_master_20261013.pdf)
+
+This is the comprehensive manuscript-working deck requested by the user, not a shortened talk. Every original slide is retained in its original relative order; new evidence is inserted next to the relevant topic. Original protocol wording is clarified where necessary.
+
+- [Original 76 → expanded page mapping](data/expanded_master_20261013/original_76_coverage.csv)
+- [All slide titles and provenance](data/expanded_master_20261013/slide_map.csv)
+- [Expansion report and evidence boundaries](data/expanded_master_20261013/REPORT.md)
+- [Verification](data/expanded_master_20261013/qa/VERIFICATION.json)
+- [Sources and immutable input hashes](data/expanded_master_20261013/manifest.json)
+
+The 55 additions comprise 20 healthy/input-control views, 25 disease/rare-state/external or exploratory views, six reconstructed review figures and four evidence/protocol guides. Existing saved results and coordinates are reused; no new training, inference or statistical fitting was launched. Historical scoring, diagnostic-only reference access, completed strict subsets and Figure 6 awaiting new data are labelled explicitly.
+
+**Palette:** PBISC muted yellow/gold #C5AA4B; scSemiProfiler blue #2166AC. Method variants use related shades, while biological/cohort colors retain their own meaning. The complete policy is `data/expanded_master_20261013/palette.json`.
+
+Builder: `tools/build_expanded_master_20261013.py`. CPU export/verification entry: `data/expanded_master_20261013/render.py` (use the host's bounded-memory execution policy). Plot caches are reused only after module, input and output hashes match.
+
+The previous [34-slide review](PBISC-Diffusion_results_review_20261013.pptx) is a historical subset; it is not the working master. The original 76-slide source deck and frozen archives remain unchanged.
+
+## Original 2026-09-30 r00 deck and historical change log
+
+76-slide English deck, ordered by the six main figures of the Nature-style manuscript plan (user request 2026-10-06,
+following `paper-release-20260928/reference-papers/Nature Style Manuscript Writing.pdf`):
+- Title and outline: 6 figures, Supplementary, Appendix, Discussion.
+- "Manuscript plan: six main figures": a table of each figure's claim and how many panels are ready.
+- For each figure:
+  - a divider;
+  - a plan slide, a table of panel / content / evidence slide / status (Ready, Adapt, Draft, Build, Run, Plan);
+  - the existing slides that serve its panels, each labelled "Fig. 2c" etc. in the top-right corner.
+- Supplementary: a plan slide plus the from-scratch results. Appendix: the modified scSemiProfiler run. Discussion: the
+  other-tissues outlook.
+
+User decisions of 2026-10-06:
+- No PBISC-VAE: scSemiProfiler, a VAE-based method, is the comparison (Fig. 3).
+- Fig. 2 shows LoRA first on the healthy cohort (ImMicro) with the realism metrics (precision, coverage, classifier
+  AUROC, MMD, E_N); the disease-cohort comparison comes later.
+- CYBERSORTx and other bulk deconvolution methods go to the Supplementary only (PBISC generates single cells).
+- Figs 4 and 6 stay unplanned. Their existing slides are kept in the section as "Fig. 4 candidate" / "Fig. 6 candidate".
+
+Only order, section labels, the agenda and new plan or draft slides changed. The new slides are "Overview draft (Nature
+style)", the manuscript's candidate Fig. 1, and "Pretraining corpus", the Fig. 1d strip. No new result figures were
+made.
+- The plan lives in `data/figure_plan.json`.
+- `tools/figure_plan.py` draws the plan tables and the slide text, and checks that every slide is placed exactly once.
+- `tools/restructure_figure_order_20261006.py` rebuilds the order from the plan. It is rerun after every plan change and
+  keeps all slide fields.
+- The earlier section order (10 sections, 45 slides) is in `_archive/2026-10-06_before_figure_order/`.
+
+Short keyword lines only; the figures carry the results. Every slide that uses a published method, metric or
+baseline has a one-line "References:" footer; full citations are in the speaker notes.
+
+- `PBISC-Diffusion_paper_summary_20260930.pptx`: the deck
+- `PBISC-Diffusion_paper_summary_20260930.pdf`: LibreOffice PDF of the deck
+- `data/slides_spec.json`: slide order, titles, keywords, footnotes, references, speaker-note text
+- `generate_deck.py`: builds the pptx from `data/slides_spec.json` only (python-pptx)
+- `figures/`: exactly the figures the slides show (PNG + PDF), made by `tools/sec*.py`. Intermediate panels go to a
+  temporary folder; superseded figures are deleted, not kept (2026-10-02, user request; every deleted file was
+  checked to be in the paper archive by sha256 first).
+- `data/sec*.csv` and `data/*_slide_text.json`: every plotted or quoted number with its source path
+- `render/`: PDF and 70-dpi slide JPGs for checking
+
+## Slide order
+
+| section | slides |
+|---|---|
+| 1 Model overview | bulk in, cells out; denoiser |
+| 2 Backbone | SDD → PBISC-Diffusion; bulk conditioning |
+| 3 Architecture ablation | width & context (loss, benchmark); gene panel |
+| 4 Metrics + SLE benchmark | **how we score generated cells** (metric table with references); SLE benchmark; per cell type; bulk broadcast baseline |
+| 5 Real bulk vs pseudobulk | input shift; same shift in three cohorts; cells off-manifold; disease direction flips |
+| 6 Paired fine-tuning | method; evaluation design |
+| 7 LoRA ablations | full vs LoRA vs cond.-encoder-only; LoRA rank r = 4 / 16 / 64; K diversification |
+| 8 Pairs needed | composition / coverage / steps (with scSemiProfiler at the same n); donor and disease signal; pairs from other cohorts; longer training on them (with training minutes) |
+| 9 Paired validation | cohorts; 5 donors: cells, realism beyond P/C, P/C at k = 3 / 5 / 10, scSemiProfiler copies its representative, disease signal, CAT curves, time to apply; 5-fold CV (with scSemiProfiler on the same pairs); same training pairs: LoRA vs scSemiProfiler; 5-fold realism metrics |
+| 10 Next steps | paired-subset cohorts (planned); other tissues from a few dozen pairs (Discussion framing of the from-scratch results) |
+| Supplementary | no pretraining: same cohort, from scratch (pilot); no pretraining: other cohorts, from scratch (pilot); rare states: from scratch vs pretrained; deconvolution with the same pairs: composition; deconvolution with the same pairs: disease signal |
+| Appendix | scSemiProfiler without its output threshold (sensitivity; modified read-out, not the official output) |
+
+## Rebuild
+
+Everything is CPU only; no GPU or checkpoint is touched.
+
+```bash
+# prerequisites (plan folder plans/paired_finetune_intradataset_20260927/, CPU):
+#   scsemi/tools/representative_dependence.py      -> scsemi/analysis/representative_dependence_20261002/
+#   tools/contrast_extra.py (+ tools/contrast_extra_tables.py) -> scores_contrast/   (Spearman, ceiling, limma, CAT, RBO)
+#   tools/build_tpm_zeroshot.py + tools/score_tpm_zeroshot.sh -> scores/<ds>/zeroshot_tpm_20261002/pooled/
+#     (the length-corrected zero-shot arm needs one GPU sampling step: tools/sample_lockwait.sh, coordinator only)
+cd /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00
+# shared helpers, not run directly: style.py, sec03_common.py, scsemi_results.py, fold_sizes.py
+for f in sec01_overview sec02_backbone sec03_width_loss sec03_width_bench sec03_panel \
+         sec04_metrics_overview sec04_sle_benchmark sec05_condition_space sec05_cells_offmanifold \
+         sec05_behcet_direction sec05_shift_cohorts sec06_finetune_method sec07_ft_modes \
+         sec07_k_aug sec07_rank_sweep sec08_pairs_needed sec08_cross_cohort sec09_paired_validation sec09_extra_metrics \
+         sec09_time_to_apply sec09_scsemi_representative sec09_cat_curves sec10_paired_subset_cohorts \
+         secA_scsemi_readout sec07_no_pretraining sec08_cross_scratch sec09_same_pairs \
+         secS_scratch_rare_states secS_deconvolution fig3_time_days fig4_disease fig23_healthy fig5_rare_states sec10_other_tissues figure_plan; do   # keep this order (sec10 reads secS; figure_plan last)
+  python3 "tools/$f.py" || break
+done
+python3 tools/restructure_figure_order_20261006.py   # order the slides by data/figure_plan.json
+python3 tools/sync_sec09_text.py      # copy every generated slide text (sec09_index / text_source) into the spec
+python3 generate_deck.py              # -> PBISC-Diffusion_paper_summary_20260930.pptx
+soffice --headless --convert-to pdf --outdir render PBISC-Diffusion_paper_summary_20260930.pptx
+pdftoppm -jpeg -r 70 render/PBISC-Diffusion_paper_summary_20260930.pdf render/slide
+cp render/PBISC-Diffusion_paper_summary_20260930.pdf .
+```
+
+`generate_deck.py` stops with an error if a figure is missing, a title does not fit on one line, a footnote needs
+more than 2 lines (1 line when the slide has references), the reference line needs more than 1 line, or any text
+contains Korean or placeholder words. `tools/restructure_20261002.py` is the one-off script that produced the current
+slide order and the references (it refuses to run twice).
+
+## Notes
+
+- **No "all" label (2026-10-02).** The 5-fold arm trains each fold model on the other four folds minus 7–8
+  inner-validation donors; its test fold is never trained on (RUN.json `test_rows_never_loaded_for_training`). Figures
+  label it "5-fold CV (other 4 folds)" and learning-curve ticks show the training-pair count per fold (78–79 ImMicro,
+  58–59 sepsis), read by `tools/fold_sizes.py`.
+- **Learning curve x = 0 uses no pairs (2026-10-02).** It shows the zero-shot model fed the real bulk as measured
+  and after exon-length correction (counts / Ensembl union-exon kb, then the same fixed-depth arithmetic;
+  `plans/.../tools/build_tpm_zeroshot.py`, scored in `scores/<ds>/zeroshot_tpm_20261002/pooled`). The old x = 0
+  calibration point, whose alpha used every training pair of the fold, is gone; calibration appears only fitted on the
+  same n pairs as the LoRA point (squares; `cal_budget_20261001`).
+- **Disease metrics (2026-10-02).** Next to Pearson logFC r and the Welch top-50 DEG overlap: Spearman r, the
+  split-half ceiling (observed vs observed, Spearman–Brown), limma-voom top-50 DEG, each arm's chance overlap, and
+  CAT curves with the k / 2G chance line (`plans/.../scores_contrast/`, tools/contrast_extra.py; every Pearson r and
+  Welch count reproduced exactly). Repeated samples of one patient (Imperial, Giroux, Ji) count as separate samples in
+  both DE tests, as in the scorer (footnote).
+- Section 9 text is generated: `tools/sec09_paired_validation.py`, `sec09_extra_metrics.py`, `sec07_rank_sweep.py`,
+  `sec09_time_to_apply.py`, `sec09_scsemi_representative.py`, `sec09_cat_curves.py` write slide text JSON; sync copies
+  it. Hand-written: the keywords of the two "Pairs needed" slides (section 8) and of every slide in sections 1–7 and 10.
+- Giroux's disease metrics exclude 196 sex-linked genes (severity is confounded with sex), from
+  `plans/.../scores_nosex/`. The calibration comparator is "Zero-shot + calibration" (user decision 2026-10-01).
+- Story rule (user decision 2026-09-30): no PBISC-VAE and no reference calibration (refcal). The exon-length-corrected
+  input is used only as the no-pair x = 0 point of the learning curve (user decision 2026-10-02).
+  `tools/sec04_immicro.py` (refcal vs scSemiProfiler on ImMicro) is kept but is not part of the deck or the rebuild.
+- scSemiProfiler 5-fold and equal-budget results are picked up automatically when they appear (`tools/scsemi_results.py`;
+  arm `scsemi__raw` in `scores/<cohort>/scsemi_cv_20261002/pooled*/`, `scsemi_n{5,10,20,40}__raw` in
+  `scores/<cohort>/scsemi_budget_20261002/pooled*/`; same donors and reference cells as the main pooled call only).
+  Then rerun sections 8 and 9, sync, recheck the section 8 keywords and rebuild.
+- Earlier versions: `_archive/2026-09-30_before_no_vae_refcal/`, `_archive/2026-10-01_before_extra_metrics/`,
+  `_archive/2026-10-02_before_time_to_apply/`, `_archive/2026-10-02_before_restructure/` (pptx, PDF, spec; their
+  figures are in the paper archive). `DRAFT_1150_*`: the copy shown at the 2026-09-30 12:00 meeting.
+
+- Added 2026-10-03 (user decision: the cross-cohort experiment is reported at K = 1000 only and is the figure that explains
+  why PBISC fits LoRA on a few pairs of each cohort): slides 25-26 in section 8, `tools/sec08_cross_cohort.py`.
+  - Slide 25 "Pairs from other cohorts do not transfer": per held-out cohort, composition L1 / coverage / disease logFC r of
+    zero-shot, calibration fitted on the other cohorts, LoRA r16 / r64 / full fine-tune trained on the other cohorts
+    (+ Liu 2021) at step saturation, and LoRA on the cohort's own pairs (5-fold pooled).
+  - Slide 26 "Longer training on other cohorts" (": gap stays" since 2026-10-06): the same metrics at every 20k-step
+    milestone.
+  - Source plan: `plans/paired_crosscohort_20261002/` (README; same-reference and same-sample checks in the script).
+    Agenda item 8 now reads "How many pairs, and do other cohorts' pairs transfer". The deck has 37 slides.
+    Pre-update copy: `_archive/2026-10-03_before_cross_cohort/`.
+
+## Sources
+
+Final model: `runs/f01_6k_1x_seed20260627` (epoch 16, EMA). Scores and receipts come from
+`plans/utility_studies_final6k_20260925/`, `plans/paired_finetune_intradataset_20260927/`,
+`plans/immicro_scsemi_comparison_20260927/`, `plans/floor01_campaign_20260923/`,
+`analysis/scsemi_benchmark_diffusion_20260913/` and the run folders named in each `data/sec*.csv`. Reference PDFs:
+`/home/sj_server_1/PBISC/paper-release-20260928/reference-papers/` (INDEX.md). Manuscript:
+`/home/sj_server_1/PBISC/paper-release-20260928/manuscript/`.
+- 2026-10-04 appendix (user decision: supplementary, appendix only, because it changes scSemiProfiler's official
+  output). Slide 38 is the divider "Appendix" and slide 39 is "Appendix: scSemiProfiler without output threshold"
+  (`tools/secA_scsemi_readout.py`, `data/secA_scsemi_readout{.csv,_slide_text.json}`).
+  - Inputs: `plans/paired_finetune_intradataset_20260927/scsemi/analysis/output_variants_20261003/`. The script checks
+    that its LoRA-5 and official scSemiProfiler values equal the section 9 slide data draw by draw.
+  - Result: without upstream's output threshold, scSemiProfiler's coverage rises in 5/5 cohorts, and 23/23 LoRA-5
+    comparisons (coverage, precision, composition, logFC r, DEG) keep their winner.
+  - The outline keeps its 10 sections (the builder requires 10). The pre-change copy is in
+    `_archive/2026-10-04_before_appendix/`. The deck has 39 slides.
+- 2026-10-05 update (user request: put the results so far into the deck). Pre-change copy: `_archive/2026-10-05_before_scratch_cv/`.
+  - New slide 23, "No pretraining: same cohort, from scratch" (`tools/sec07_no_pretraining.py`). Pilot on sepsis fold 0
+    (17 donors). Reads `fold0_scratch_20261005/` and the runs' RUN.json. The donor composition r is recomputed from
+    per_donor.csv and reproduces ablation_extra_donor.csv.
+  - New slide 28, "No pretraining: other cohorts, from scratch" (`tools/sec08_cross_scratch.py`). Held-out sepsis pilot.
+    Every arm is read from `paired_crosscohort_20261002/scores_masked/`: genes no training cohort measured are zeroed
+    in all arms, because the unmasked scratch arm puts 91–94 % of its UMI there (read from the generated_masked logs
+    since 2026-10-06).
+  - New slide 38, "Same training pairs: LoRA vs scSemiProfiler" (`tools/sec09_same_pairs.py`). scSemiProfiler 5-fold
+    CV with LoRA's own training pairs, plus GPU hours per fold from the receipts on either host. Cohorts that are still
+    running show "in progress"; rerun the script as they finish.
+  - These two slides now draw the scSemiProfiler CV / budget runs automatically:
+    - Slide 24, "Pairs needed" (diamonds; ImMicro so far): keyword 2 reworded, footnote extended.
+    - Slide 37, "5-fold CV" (yellow bar; ImMicro and Ji so far).
+  - `tools/sec09_paired_validation.py` reads the scSemiProfiler CV arm's Spearman / limma values from
+    `scores_contrast/<cohort>/scsemi_cv_20261002/pooled/` (contrast_extra.py --design scored). Pearson r and the Welch
+    DEG must match the scorer; the values stay blank until computed. Cohorts not yet run say "in progress".
+  - Agenda item 7 now reads "LoRA: compute, rank, K; training from scratch". The deck has 42 slides.
+  - New helper: `tools/grouped_bars.py`.
+- 2026-10-06 update (user request: check that the cross-cohort LoRA / full FT, cross-cohort from-scratch and
+  same-cohort from-scratch (+ recipes) results, with benchmarks and training times, and the scratch-vs-pretrained rare
+  states are in the deck; user decision: the from-scratch results go to the paper's Supplementary and the Discussion
+  presents them as a route to other tissues). Pre-change copy: `_archive/2026-10-06_before_supplementary/`.
+  - **Supplementary block (new divider, slides 40-43).**
+    - "No pretraining: same cohort, from scratch" and "No pretraining: other cohorts, from scratch" moved here from
+      sections 7 and 8. Their content is unchanged; slide 42's notes now name the cross-cohort slides by title.
+    - New "Rare states: from scratch vs pretrained" (`tools/secS_scratch_rare_states.py`).
+      - It reads the rare-state screen (`plans/rare_states_screen_20261004/out/sepsis_fold0_*` and
+        `sepsis_loco_masked_*`).
+      - The cross-cohort arms use the masked scoring (46 never-measured genes zeroed in every arm).
+      - The donor bootstrap reproduces the verified analysis in `out/scratch_vs_lora_20261005/REPORT.md`.
+  - **Section 10 "Next steps".**
+    - New "Next: other tissues from a few dozen pairs" (`tools/sec10_other_tissues.py`): a comparison table, pretrained
+      + LoRA vs from scratch.
+    - Every value is read from the deck's own source tables. The say text is the user's Discussion framing.
+  - **Slide 26 "Longer training on other cohorts".** The first keyword is now training minutes to saturation (RUN.json
+    wall seconds), with per-cohort minutes and steps in the notes (`tools/sec08_cross_cohort.py`).
+  - **Slide 34 "5 donors: time to apply".**
+    - The pretraining hours are read from `data/sec06_ft_method.csv`, no longer hard-coded.
+    - The note names the 'Fine-tuning with paired bulk + sc' slide by title. Its old "slide 15" pointed at the wrong
+      slide.
+  - Agenda items 7 and 10 updated. Spec change: `tools/restructure_20261006.py` (refuses to run twice). The deck has 45
+    slides; no Korean text.
+  - **Review pass (same day).** An independent 5-agent review of the changed slides confirmed 39 issues; all are
+    fixed. Pre-review copies of the touched scripts and the spec: `_archive/2026-10-06_before_supplementary/*.pre_review`.
+    - Slide 39: pretraining now reads 33.1 M train cells (`train_split_cells`; the 41.4 M corpus includes val/test).
+      The rare-state row follows the bootstrap verdicts on both sides (closer on ISG-high T / on DN2 B; labels and
+      states both arms miss by > 4-fold are not counted). "No corpus" became "come close": the scratch runs still took
+      the panel, value scaler and z-score constants from the PBMC checkpoint (RUN.json `kept_from_checkpoint`, now in
+      the notes). The say text carries the one-fold pilot qualifier; the footnote names both pilots.
+    - Slide 43: "no significant difference on average" (not "no difference"). Keyword 2 is the cycling T/NK count
+      (8 of 8 comparisons). The notes add the ISG-high T split by panel, the scratch-closer states (DN2 B, MS1), the
+      52-call tally, the fold-0 caveat (fold 0 is the own-pair LoRA's best fold for C1Q-high and cycling T/NK,
+      computed from `folds.csv`), label states unthinned, and "8 of 13 states shown".
+    - Slide 26: title "...: gap stays" and "drifts in Sepsis and Ji" are computed from the milestones (gap = no milestone
+      reaches the own-cohort coverage; drift = full fine-tune L1 up by >= 0.05). Minutes are in
+      `data/sec08_cross_cohort.csv` (wall_seconds, minutes, run_receipt) and in the numbers list.
+    - Slide 42: UMI share (94 %, 91–94 % across milestones), the observed share (0.014 %, from `cells_panel6k.npz`)
+      and both arms' stopping settings are read from the receipts; the "none" label sits inside the axes.
+    - Slide 34: keywords use the cohort means of the figure (17–59×, 57 min–4.1 h); the per-draw range 11–63× is in the
+      notes. The right panel shares the bars' rows and has a title.
+    - Talk text and pointers: slides 23, 25, 26, 45 (say), 22 (recipes description), 32 and 36 (slide pointers by
+      title), slide 1 subtitle "updated 2026-10-06".
+- 2026-10-08 update (user request: put the BayesPrism, CIBERSORTx and scSemiProfiler 5-fold results into the deck).
+  Pre-change copy: `_archive/2026-10-08_before_deconvolution/`.
+  - **Two new Supplementary slides** (`tools/secS_deconvolution.py`, `data/secS_deconvolution{.csv,_slide_text.json}`):
+    "Deconvolution with the same pairs: composition" and "...: disease signal".
+    - Source: `plans/bayesprism_paired_20261006/` (README, `results/RESULTS_full.md`, `RESULTS_cibersortx.md`).
+    - BayesPrism (recommended genes) and official CIBERSORTx (Fractions S-mode + HiRes) use the same paired donors
+      as the reference and the test samples' real bulk as the mixture, with the benchmark's own truth and scorer.
+    - Keywords and notes are computed from the tables.
+  - **Supplementary plan.** The planned "Bulk deconvolution methods" row is replaced by the two Ready slides.
+    `tools/figure_plan.py` now takes an optional per-slide note (`supplementary.notes`) and drops "Planned:" when
+    nothing is planned.
+  - **New scSemiProfiler results picked up (server_1):**
+    - Sepsis and Giroux 5-fold CV. Giroux was finalized on server_1 from sha-checked copies of its fold NPZs, and
+      rescored without sex-linked genes in `scores_nosex/giroux2022/scsemi_cv_20261002/pooled`.
+    - Sepsis equal-budget n = 5/10/20/40.
+    - Imperial 5-fold: finished on server_2 at 13:45 and was finalized on server_1 from sha-checked fold copies.
+      scSemiProfiler 5-fold now covers all five cohorts.
+    - Spearman/limma extras for sepsis, Giroux and Imperial; realism extras for all five cohorts.
+  - **Text changes from the new results:**
+    - "Same training pairs": keywords are data-driven. LoRA leads on coverage and precision in 5/5 cohorts;
+      composition 4/5 and logFC r 3/4, where scSemiProfiler leads in Giroux. The GPU-time ratio is 19–62×.
+    - sec09_same_pairs.py now reads the sepsis fold-0 pilot queue tag (`sep_f0`). For Giroux it quotes DEG from
+      scores_nosex, and for older pooled LoRA calls from pooled_deg.
+    - "5-fold CV" keyword: logFC r 3/4.
+    - "Pairs needed" (hand-written): scSemiProfiler coverage 0.04 / 0.10. On sepsis disease logFC r,
+      scSemiProfiler at 5–40 pairs (0.52–0.58) is above LoRA (0.29–0.48), while its top-50 DEG stays at 2–6.
+- 2026-10-08, new figure split (user decision; supersedes the 2026-10-06 split; pre-change copy
+  `_archive/2026-10-08_before_new_figure_plan/`).
+  - Fig. 2: LoRA fine-tuning with no disease metrics.
+  - Fig. 3: PBISC vs scSemiProfiler on healthy donors, with time as a main result.
+  - Fig. 4: disease benchmarks on the paired cohorts (PBISC versions, scSemiProfiler, and BayesPrism / CIBERSORTx
+    labelled as cell-type deconvolution; the deconvolution slides moved here from the Supplementary).
+  - Fig. 5: rare-cell fractions and detection.
+  - Fig. 6: new findings from the bulk-only sepsis and COVID samples (K = 1,000 and 4,000).
+  - The cross-cohort transfer slides moved to the Supplementary; their placement is open.
+  - New slide "Time: days for scSemiProfiler, minutes for LoRA" (`tools/fig3_time_days.py`, Fig. 3d).
+    - Inputs: GPU hours per 5-fold experiment from the receipts (`data/sec09_same_pairs.csv`, checked against
+      the scSemiProfiler event logs on both hosts), the 5-donor draws (`data/sec09_time_to_apply.csv`) and the
+      one-time pretraining (`data/sec06_ft_method.csv`).
+    - Results: ImMicro 24 h vs 1.2 h, sepsis 38 h vs 37 min. All runs: scSemiProfiler 184 h (7.7 days) vs PBISC
+      22 h including pretraining.
+    - The equal-budget scSemiProfiler runs are excluded: they reuse the 5-fold run's representative models.
+  - Panel status: Build / Run marks new slides or analyses still to do; Adapt marks slides to re-scope (healthy-only
+    Fig. 2-3, disease columns into Fig. 4).
+- 2026-10-08, comparison rules (user decision; memory feedback_paper_comparison_rules):
+  - **Time.** Only per-cohort GPU work counts: no PBISC pretraining, no GPU queue waits. The time slide now shows the
+    5-donor draws (main) next to the 5-fold runs (separate). All runs: 7.7 GPU-days vs 5.7 h.
+  - **No bulk-broadcast baseline.** It is removed from the SLE benchmark, CAT, deconvolution and "Pairs needed"
+    figures.
+    - The "Bulk broadcast baseline" slide is gone. Its partial r panel is now on the SLE benchmark slide
+      (0.21 vs 0.17).
+    - Still to rebuild in their own phases: Fig. 2e (ImMicro only) and Fig. 6a.
+  - **Main comparison = 5 pairs** (LoRA-5 vs scSemiProfiler-5); 5-fold is reported separately. Partial r is
+    emphasised (bold panel titles).
+  - **New `tools/fig4_disease.py`**, which checks all 200 logFC r / DEG values against the section-9 tables:
+    - "Disease signal with 5 pairs: LoRA-5 vs others" (4b);
+    - "5-fold CV: disease signal" (4c);
+    - "Sepsis: fine-tuning variants" (details).
+  - The old "5 donors: LoRA vs scSemiProfiler — disease signal" slide is retired. `tools/sync_sec09_text.py` lists it
+    in RETIRED.
+  - The deconvolution slides are retitled "Cell-type deconvolution, same pairs: ...".
+- 2026-10-08, healthy-only Figs 2-3 (`tools/fig23_healthy.py`).
+  - Eight ImMicro-only slides, re-plotted from the deck's own data tables; nothing is re-scored.
+  - Fig. 2 (no disease metrics):
+    - cells, realism, full / LoRA / condition encoder, pairs needed, rank and K;
+    - 5 donors (main) shown next to 5-fold (separate).
+  - Fig. 3 (vs scSemiProfiler):
+    - LoRA-5 vs scSemiProfiler-5 on 6 cell metrics: LoRA ahead on 6 of 6;
+    - the same pairs, 5 to all;
+    - scSemiProfiler copies its representative;
+    - time.
+  - The earlier all-cohort slides moved to the Supplementary (note "all five cohorts").
+  - "5-fold CV: train on the other 4 folds" is retired (it carried the broadcast; its content is now in Fig. 4c and
+    the healthy slides). It is listed in RETIRED of `tools/sync_sec09_text.py`.
+- 2026-10-08, Fig. 5 rare cells (`tools/fig5_rare_states.py`).
+  - Source: new plan `plans/rare_states_paired_20261008` (README). It covers 4 paired disease cohorts, the 5-donor
+    draws and 5-fold, and the arms zero-shot, calibration, scSemiProfiler and LoRA. Each NPZ is sha-checked and the
+    benchmark's Poisson read-out and labels are reproduced cell by cell.
+  - Three slides:
+    - fraction (5a);
+    - donors and disease direction (5b);
+    - 5-fold, separate (5c).
+  - Results over 56 state × cohort pairs, 5 donors:
+    - LoRA-5 closer to the observed fraction than scSemiProfiler-5 in 51/56;
+    - within 2-fold: 19 vs 5;
+    - better donor ranking: 43/56;
+    - disease direction kept: 73% vs 61%.
+  - Fig. 6a no longer mentions the broadcast. No slide text mentions it now.
+- 2026-10-10, last broadcast mention removed from a figure (found while rewriting the manuscript from this deck).
+  - Fig. 2d "How we score generated cells" (`tools/sec04_metrics_overview.py`) still listed "broadcast; split-half
+    ceiling" as the comparator of logFC r inside the table image. It now reads "zero-shot; partial r (vs bulk)".
+  - Old tool, CSV and figure: `_archive/2026-10-10_metrics_overview_broadcast/`. The deck was rebuilt (76 slides); a
+    text search of every slide, the notes and the PDF twin of every figure a slide uses finds no "broadcast". Three
+    figure files of retired slides still contain it and are not used: `sec04_sle_broadcast`, `sec09_fewpair_disease`,
+    `sec09_fivefold`.
+  - The manuscript was rewritten from this deck on 2026-10-10 (`paper-release-20260928/manuscript/`,
+    `editorial/rewrite_20261010.md`); its figures are copies of this folder's figure PDFs.
+- 2026-10-10, LoRA rank figure key (`tools/sec07_rank_sweep.py`): "LoRA rank r = 16 (used in the deck)" now reads
+  "(used throughout)", because the same panel is Supplementary Fig. 12a of the manuscript. Old files:
+  `_archive/2026-10-10_rank_sweep_label/`. Deck rebuilt (76 slides).