Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261015T141849Z-rare-cell-7k-8k-panels/diffusion/dataset/server_1/model8/ara/trace/sessions/session_index.yaml

@@ -0,0 +1,27 @@
+sessions:
+- id: 2026-10-13_001
+  date: '2026-10-13'
+  summary: Expanded142 to154slides with corpus/split audit, saved adaptation and disease
+    metrics, and original-paper rare-target observed diagnostics.
+  turn_count: 3
+  events_count: 6
+  claims_touched: []
+  open_threads: 5
+- id: 2026-10-14_001
+  date: '2026-10-14'
+  summary: Redrew Figure1a and inserted exact training-corpus Figure1b in155-slide
+    master.
+  turn_count: 1
+  events_count: 2
+  claims_touched: []
+  open_threads: 1
+- id: 2026-10-15_001
+  date: '2026-10-15'
+  summary: Implemented Figure1 suggestions in156-slide master; all numerical source
+    counts preserved. Verified five-marker co-detection in selected training candidates.
+    Prepared verified 7k/8k rare-cell gene panels and audited native ILC signal; no
+    new training.
+  turn_count: 3
+  events_count: 3
+  claims_touched: []
+  open_threads: 2

paper-releases/pbisc-diffusion/updates/20261015T141849Z-rare-cell-7k-8k-panels/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/DETERMINISM.json

@@ -0,0 +1,21 @@
+{
+  "status": "PASS",
+  "rebuilds_compared": 2,
+  "identical_files": {
+    "panels/7k/additions.csv": "948d00ea714362e44eec144e2be999c9d462765f66db1402d73f13cdedf2699e",
+    "panels/7k/gene_positions.npy": "f4bda0b94e2f5d6fcc7e44e34765a1b77b31dce85de49925bc95b3906ee91e10",
+    "panels/7k/old_to_new.csv": "05e14a36058b0bc47cd4f750d64d59e19edaac47b78f4dece7e9289e6ced5ec9",
+    "panels/7k/selected_genes.csv": "b9ba04fb554cdec206a72fb6b00f2b3a7c49f038abebaa05473d87a71e265ec1",
+    "panels/7k/selection_manifest.json": "c31f4adfcb2aa1692f2a54e967de1338276c4c98c7f5744fd97373cb476648d1",
+    "panels/8k/additions.csv": "bd4f723f4750afb38133e1150423bcf33b52207faddfa09998c1a521f675db13",
+    "panels/8k/gene_positions.npy": "a6d8acb7ef27ae1c2733d50072e2d21cdb54b20a38910a5d10c4c8f9d72305d8",
+    "panels/8k/old_to_new.csv": "b303cbe00d1f7c90ff65b50492135def80ce92d4587b6577490db1aab504e1be",
+    "panels/8k/selected_genes.csv": "62d961d83cf180dd925e080f696827f48074125cf2725d290cda21247d23c60f",
+    "panels/8k/selection_manifest.json": "d9e46d3d6a0fbd7d7496c3f075a6c7c236eb5062926fd81f34bd071f2d79d8cb",
+    "selection_manifest.json": "2c8922df98464c99b83a1402bb7cb5a2fc83122be4fc1c7a1bd1924b673b67f6",
+    "evidence_resolved.csv": "efca78457641bc5cb87da891c24dba6a6d8ad71fd71e75ff6a8bbc0f0c3a7b60",
+    "coverage.csv": "2b933b7c1e2d85871f7bbaac4e6352bdfe2fa277897270719c2b88bc27807a2e",
+    "coverage_by_evidence_class.csv": "c668d80dc450ed843c63af9cd6b7df0382b8104ecd0e5d54e169f9a5b2f9f91e"
+  },
+  "same_exact_input_hashes": true
+}

paper-releases/pbisc-diffusion/updates/20261015T141849Z-rare-cell-7k-8k-panels/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/README.md

@@ -0,0 +1,11 @@
+# Rare-cell context candidates, 2026-10-15
+
+Read [REPORT.md](REPORT.md) for the recommendation, causal audit and limitations. The broader candidate is 8k, with nested 7k as a compact comparison. Both preserve all original 6k genes; neither is an active training configuration.
+
+- [7k panel](panels/7k/selected_genes.csv) and [8k panel](panels/8k/selected_genes.csv): exact gene lists, with positions, mappings, additions and manifests alongside them.
+- [Population coverage](coverage.csv), [evidence-class coverage](coverage_by_evidence_class.csv), [98 source targets/programs](rare_cohort_targets.csv), and [resolved per-gene evidence](evidence_resolved.csv).
+- [ILC/DC source research](research/ilc_dc/REPORT.md) and [disease source research](research/disease/REPORT.md), including primary-paper supplements and provenance.
+- [Saved generated ILC classifier calls](existing_generated_ilc_calls.csv) and [matched-sample observed classifier calls](existing_observed_ilc_calls.csv): diagnostics, not validated ILC generation.
+- [Validation protocol](VALIDATION_PROTOCOL.md), [independent artifact verification](VERIFIED.json), and [deterministic rebuild evidence](DETERMINISM.json).
+
+`audit_local.py`, `audit_measurement_and_labels.py`, `build_panels.py`, and `verify_panels.py` are CPU-only artifact tools. They do not launch training or inference. `build_panels.py` consumes the frozen research CSVs and existing train-only statistics; updating those inputs deliberately changes the candidate version and requires renewed review.

paper-releases/pbisc-diffusion/updates/20261015T141849Z-rare-cell-7k-8k-panels/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/REPORT.md

@@ -0,0 +1,144 @@
+# Rare-cell gene-context expansion: verified 7k and 8k candidates
+
+**Recommendation:** retain the 8k panel as the broader research candidate, with the nested 7k panel as the compact comparison. Both preserve every current 6k gene and all available core/sentinel genes. The additional 1,000 coordinates in 8k complete the reviewed blood-supporting and stimulation-program coverage; this is a reason to test 8k, not evidence that it improves generation. No training, inference, active configuration change, or manuscript claim was made in this task.
+
+The claim that ILC are not generated *because their markers are missing* is not established. Some important markers are missing, but most common ILC identity genes are present, and existing saved outputs contain ILC classifier calls. The published broad scorer deliberately merges those calls into **Other**.
+
+## Concrete deliverables and exact budget
+
+| Candidate | Total genes | Original genes preserved | Literature-supported additions | HVG padding | Missing available core/sentinel genes |
+|---|---:|---:|---:|---:|---:|
+| 7k | 7,000 | 6,000 | 1,000 | 0 | 0 |
+| 8k | 8,000 | 6,000 | 2,000 | 0 | 0 |
+
+Files: [7k selected genes](panels/7k/selected_genes.csv), [7k raw positions](panels/7k/gene_positions.npy), [8k selected genes](panels/8k/selected_genes.csv), [8k raw positions](panels/8k/gene_positions.npy). Each directory also contains `additions.csv`, `old_to_new.csv`, and a hashed `selection_manifest.json`.
+
+The input evidence comprises 8,774 gene–population–source rows, 4,821 distinct requested symbols, and 4,163 resolved raw-axis Ensembl genes. Of those, 2,091 are absent from final 6k. Every one of these 2,091 has positive measured and detected counts in the frozen 1M training sample. Neither panel invents genes that are absent from the canonical axis. There are 218 available distinct core genes; 37 were missing from 6k. A minimal core-only expansion would therefore be 6,037 genes; adding the protected functional sentinels gives 6,044. The full available literature union would be 8,091 genes, but broad optional programs are not all necessary identity features.
+
+Selection protects core identity/boundary and named functional-sentinel genes first. It then alternates fairly among populations and source programs, using author rank or within-program effect-size order, before optional stimulation and tissue programs. Broad generic DC contrasts therefore cannot consume the entire budget ahead of smaller disease programs. These are deterministic design choices, not an optimized model-performance ranking. `evidence_resolved.csv` records every source row, role, direction, tissue/assay, alias decision, selection rank, training-sample signal, and 6k/7k/8k membership.
+
+## Actual ILC coverage and why the global selector missed genes
+
+The final panel is `artifacts/panel_ladder_floor01_1m_20260923/tiers/6k`, with positions SHA-256 `653ffa45ac519000baa41086c52408009391d6e68cfb218bc6a50d34163fcb24`. Both the original final-6k configuration and the newer matching-panel trunk×4 configuration bind that exact axis. The historical legacy 6k shares only 4,258 genes with it; an old “6k” result is not automatically a comparison on this panel.
+
+The final selector used 1,048,576 sampled occurrences from 8,192 **train-only** recipes, with a minimum of 1,048 detected and 104,857 measured occurrences. Its 0.1% detection floor can remove sparse but meaningful rare-cell genes. The full training corpus has 16,439 source-labelled ILC among 33,141,380 cells (0.0496%), but the recipe sample is not a uniform prevalence survey; that arithmetic explains a risk, not an exact proof of why every marker was omitted.
+
+The final generative configuration is annotation-free: ILC labels used here identify source rows for audit, rather than supervised class targets or an ILC-specific generation head.
+
+| Gene / raw locus | Final 6k | 7k | 8k | Detected in frozen 1M sample | Interpretation |
+|---|---|---|---|---:|---|
+| PTGDR2 / CRTH2 | missing | included | included | 3,562 | ILC2 receptor/identity support; below top-6k rank, not below detection floor |
+| NCR2 / NKp44 | missing | included | included | 582 | tissue/state-sensitive ILC3 receptor; protected capability sentinel |
+| IL5 | missing | included | included | 672 | functional cytokine; no requirement for resting-blood expression |
+| IL13 | missing | included | included | 524 | functional cytokine |
+| IL22 | missing | included | included | 195 | functional cytokine |
+| IL17A | missing | included | included | 45 | functional cytokine |
+| IL17F | missing | included | included | 103 | primary human stimulated tissue evidence includes protein readout; not validated resting-blood RNA identity |
+| IL12RB1 | missing | missing | included | 97,664 | broader receptor/program coverage supplied by 8k |
+| CRLF2, X primary Ensembl locus | missing | missing | included | 1,451 | Y-associated duplicate symbol already present in 6k; locus availability must remain explicit |
+| IL9R, X primary Ensembl locus | missing | included | included | 1,469 | another locus with a Y-associated symbol copy already in 6k |
+| TRAC | absent from raw axis | unavailable | unavailable | NA | requires upstream feature-axis/preprocessing work |
+
+IL7R, KLRB1, KIT, GATA3, RORA, RORC, IL1RL1, IL17RB, IL23R, IL1R1, ID2, TCF7, AHR, ZBTB16, TBX21, and EOMES are already present in final 6k. CD3D/CD3E and multiple NK boundary genes are also present. See [the exact audit](ilc_marker_audit.csv) and [the 48-locus targeted-audit list](ilc_targeted_audit_genes.csv). Some TCR constant features are outside the protein-coding raw axis; marker expansion cannot substitute for V(D)J or independent protein evidence.
+
+The 17 protected functional sentinels are source-supported assay-capability choices. Their priority does not turn tissue/stimulated expression into a universal blood identity requirement. KIT-positive circulating ILCP must remain distinct from mature ILC3, and blood DC3 from tissue/tumor mregDC. The [ILC/DC research report](research/ilc_dc/REPORT.md) explains those boundaries.
+
+## Direct support in source-labelled training ILC
+
+After the candidate panels were frozen, a separately authorized bounded CPU audit read native counts for the **16,439 registered train ILC rows**, covering 12 source datasets, 125 shards, 937 corrected dataset-scoped donor identities and 2,417 sample keys. It inspected 48 raw loci without changing labels or using the result to retune gene selection. The cached cell counts, masks, donor/dataset summaries, chunk receipts and verification are under [coordination/ilc_training_support](coordination/ilc_training_support/).
+
+| Missing-6k locus | Detected / measured source-labelled train ILC | Corrected donors with detection | Datasets with detection |
+|---|---:|---:|---:|
+| PTGDR2 | 518 / 16,439 | 201 | 11 |
+| IL12RB1 | 1,063 / 16,439 | 323 | 12 |
+| CSF2 | 113 / 16,439 | 82 | 12 |
+| IL5 | 19 / 16,354 | 10 | 5 |
+| IL13 | 46 / 16,354 | 17 | 5 |
+| NCR2 | 9 / 16,354 | 8 | 4 |
+| IL17A | 0 / 16,354 | 0 | 0 |
+| IL17F | 2 / 16,354 | 2 | 2 |
+| IL22 | 1 / 16,354 | 1 | 1 |
+| CRLF2, X locus | 57 / 8,954 | 46 | 6 |
+
+Nine of the ten audited missing-6k loci have at least one positive source-labelled train ILC. This supports available training signal for PTGDR2 and IL12RB1 across donors, while the functional cytokines are much more weakly supported. **IL17A has no detected count in this measured selection**, even though it has 45 detections in the pooled 1M train sample. Adding its coordinate does not create positive ILC training examples. This is a precise statement about the audited source-labelled rows, not proof that human blood ILC can never express IL17A.
+
+Existing core-gene signal is substantial: IL7R is detected in 9,318/16,439 selected cells and KIT in 4,289/16,439. Source-label quality and state composition still matter. Parse cytokine-perturbation data contribute 7,485 of the 16,439 ILC-labelled cells, so this selection is not solely resting natural blood. The existing CRLF2 Y-associated raw coordinate is measured in those 7,485 Parse-labelled cells and detected in 655; the added X coordinate is measured in the other 8,954 cells and detected in 57. Collapsing this to one symbol-level present/absent bit would conceal an assay/source coverage difference.
+
+The same perturbation cohort supplies 1,872 of 1,880 IL1RL1-positive cells, 35 of 46 IL13-positive cells, 13 of 19 IL5-positive cells, and the sole IL22-positive cell in this selection. Thus a positive training count must retain its stimulation/source context. Descriptive co-detection also shows 6,936 cells with IL7R and KLRB1, 2,188 with CD3D or CD3E, and 1,401 with NKG7 and GNLY. Of the IL7R/KLRB1 double-positive cells, 1,381 also have CD3D/CD3E and 1,106 also have NKG7/GNLY. These overlapping patterns are **not a reclassification or a purity estimate**: source labels are not biological truth, and transcript detection alone cannot settle ILC/T/NK identity.
+
+The canonical extraction passed count/mask checks and a separate direct-CSR replay of 29 rows, including five chunk-boundary rows. Eight original-H5AD spot rows also matched counts, masks and canonical totals. Historical original H5AD paths for four source datasets (300b, 301, 302, 304) are absent locally; their canonical CSR replay passed, but original-file replay is explicitly unavailable. These checks do not establish expert-confirmed ILC identity or generated-cell fidelity. Extraction alone took 189.5 seconds; the complete process including grouped summaries took 245.4 seconds, with a full-process maximum RSS of 418 MiB. The cgroup peak was approximately 2 GiB including file cache, with zero swap. This audit was CPU-only and did not launch model inference.
+
+## The scorer hides ILC within Other
+
+The paired scorer uses `aifi_scanvi_f01common_v1` with **1,005 input genes**, not the entire 6k. Its fine ontology contains ILC, but `predict_scanvi.py` collapses ILC, platelet, erythrocyte, and progenitor labels to `Other` in `celltype_l1_pred`. Full `aifi_l1` and `aifi_l2` remain in saved label files. A reference validation table contains 105 correct ILC calls among 111 true ILC (94.59% recall) and 123 total ILC predictions (85.37% precision); those small reference counts do not validate generated cells under distribution shift.
+
+The following counts were read from completed **existing** exact-five draw-0 outputs, without annotation or generation. The recorded source-array hashes, label-file hashes, and corresponding score-summary hashes are retained in `existing_generated_ilc_calls_sources.json`.
+
+| Cohort | LoRA-5 generated ILC classifier calls / cells | Generated fraction | Observed ILC classifier calls / cells on same scored sample IDs | Observed fraction |
+|---|---:|---:|---:|---:|
+| ImMicro | 944 / 153,000 | 0.6170% | 1,109 / 371,922 | 0.2982% |
+| Ye sepsis | 208 / 115,500 | 0.1801% | 2,381 / 556,369 | 0.4280% |
+| Imperial | 637 / 213,000 | 0.2991% | 664 / 451,540 | 0.1471% |
+
+These are pooled classifier diagnostics, **not author-gold ILC counts, proportion accuracy, or validated biological recovery**. Observed cells are not equal-weighted per donor, whereas the generated outputs have 1,500 cells per scored sample. All listed generated ILC calls are included in broad Other. The same audit also finds ILC calls in zero-shot raw outputs. New marker genes are ignored by the fixed 1,005-gene annotator, so a richer panel needs a separately validated rare-cell endpoint alongside the unchanged broad score.
+
+## What 8k adds over 7k
+
+Counts below are overlapping gene sets and must not be summed. “Available” means a resolved raw-axis gene, not proven expression in that population.
+
+| Evidence class | Available raw genes | 6k covered | 7k covered | 8k covered |
+|---|---:|---:|---:|---:|
+| Core identity / boundary | 160 | 131 | 160 | 160 |
+| Core state / program | 45 | 36 | 45 | 45 |
+| Blood-supporting programs | 2,603 | 1,406 | 2,225 | 2,603 |
+| Same-benchmark author targets | 674 | 431 | 674 | 674 |
+| Optional stimulation states | 1,288 | 673 | 838 | 1,288 |
+| Optional tissue programs | 1,399 | 770 | 962 | 1,308 |
+
+Both panels cover all 128 available ILC2-program genes, 39 ILC1/T-boundary genes, 108 NK/ILC1-boundary genes, and 81 ASDC genes. Broader coverage increases from 390 to 484 available ILCP genes, 498 to 667 DC3 genes, and 293 to 331 pDC genes. The 91 genes still omitted by 8k belong only to optional tissue programs. These are formal feature-coverage gains; whole-transcriptome DEG lists contain shared genes and do not imply lineage specificity.
+
+All 1,081 resolved genes in the disease agent's catalog, including its 140 available core genes, are already covered by **both** 7k and 8k. None of its 63 disease-target rows gains a gene in 8k over 7k. The incremental 8k breadth therefore comes from the ILC/DC supporting, stimulation and tissue programs. In particular, the eight highlighted Ye Mono-c06 genes and twelve Imperial dexamethasone-associated marker/boundary genes were already present in 6k; absence of these published states cannot be explained simply by those headline genes being excluded.
+
+The target inventory spans ILC/ILCP and DC boundaries; Ye source-defined monocyte, lymphoid, plasma, megakaryocyte and HSC programs; Imperial treatment-associated monocytes; SLE DN2/LDG; RA IFN states; JDM inflammatory/proliferative states; circulating CD34-enriched progenitors; DRESS; and selected bulk/immune-state findings. [rare_cohort_targets.csv](rare_cohort_targets.csv) contains 98 source-defined populations or programs with source DOI, specimen, context, and coverage. Those 98 rows are not 98 experimentally validated rare cell types. Influenza infection source cells do exist in the training corpus: COMBAT contributes 18,243 train and 990 test cells. The separate vaccine B-cell publication is external corroboration, not infection-source identity evidence.
+
+The [disease research report](research/disease/REPORT.md) preserves important qualifications: the Imperial state is steroid-associated, JDM inflammatory CD16 monocytes do not have uniformly positive CD163, and proliferative/IFN programs are not independent lineages. Ye fine-cluster identity cannot be established by generic marker co-detection alone.
+
+## Raw-axis, assay, and leakage audit
+
+The raw axis contains 20,097 unique Ensembl genes. Approved symbols, aliases, and previous symbols were resolved using the [official HGNC download](https://hgnc.genenames.org/download/) retrieved on 2026-10-15, last modified 2026-10-13; its payload and hash are frozen under `reference/`. Duplicate symbols such as CRLF2/IL9R were resolved to the primary HGNC Ensembl locus while preserving every original raw locus. Final uniqueness is by Ensembl ID, not symbol.
+
+There are 86 resolved alias/renaming queries. CAST has a raw-versus-current-HGNC Ensembl conflict and AGPAT9 has a multi-gene alias ambiguity; these evidence rows are flagged rather than silently assigned. Any CAST coordinate already present in 6k is still preserved. There are also 264 known HGNC genes absent from the raw axis and 355 unresolved source symbols absent from that axis. These include noncoding and immune-receptor features. Three requested core symbols are unavailable: IGHD, AC026369.3, and MEG3. All unavailable/conflicting rows remain in `unavailable_or_ambiguous_evidence.csv`; they are not made into zero-valued coordinates. A model with 8k genes is therefore not a claim of complete human rare-cell transcriptome coverage.
+
+The train-HVG statistics, ranked positions, ordered recipes, and sampled-reference payload hashes match their original selection manifest, including a streamed hash check of the 565 MB sampled-reference CSV. All recipe rows are train and non-heldout. The selection sample covers 22 datasets, 1,399 donor labels and 2,853 sample keys across 10x, Parse, ScaleBio and Seq-Well. These are sampler metadata, not gene-specific donor replication or population-specific signal. The 1M counts are sampled occurrences and are not asserted to be 1M distinct cells.
+
+Existing source-feature masks were audited without expression access: 133 Ye sc library masks, 41 Giroux, 46 Ji, the shared Imperial sc mask, and each cohort's bulk mask. The number of genes measured in both modalities is:
+
+| Cohort | 6k | 7k | 8k |
+|---|---:|---:|---:|
+| Ye sepsis | 5,975 | 6,975 | 7,975 |
+| Imperial | 5,927 | 6,924 | 7,923 |
+| Giroux | 5,563 | 6,534 | 7,508 |
+| Ji | 5,752 | 6,749 | 7,745 |
+
+Masks were used for audit, not ranking. They establish assay feature availability, not expression or cell identity. Missing expanded-axis observations must be masked/NA in evaluation. No local held-out expression informed panel ranking, but benchmark-author marker tables are endpoint-informed evidence and must not be described as blind generalization. Independent cohort/label validation remains required.
+
+## Compatibility, cost, and the next scientifically distinct work
+
+`panel_position` remains the original 20,097-axis coordinate; CSV row order is the new canonical tensor order. The old 6k is a preserved **set**, not necessarily the first 6,000 tensor rows. `old_to_new.csv` supplies exact remapping. Both candidate manifests intentionally require a reviewed loader extension; the present legacy HVG schema rejects this literature-driven selector. Active configurations are unchanged.
+
+The full canonical bulk normalization and six scalar features are computed before subsetting, so their normalization must be retained. With G genes the model condition is **3G + 6**, its noisy/self-conditioning input is **4G**, and each output head has G rows. A 6k checkpoint cannot sample new genes by pointing to a larger CSV; future fresh training or an explicitly labelled, tested parameter-migration/retraining procedure is necessary. Existing LoRA adapters are also shape- and gene-order-specific.
+
+| Trunk | 6k parameters | 7k parameters | 8k parameters |
+|---|---:|---:|---:|
+| ×1 | 54,185,184 | 61,867,184 | 69,549,184 |
+| ×4 | 239,955,168 | 266,069,168 | 292,183,168 |
+
+For ×4, 7k/8k add 26.114M/52.228M parameters (+10.88%/+21.77%). At an illustrative 20 bytes per parameter for FP32 weights, gradients, Adam states and EMA, those additions imply about 498/996 MiB of persistent training state, before activations and data loading. Gene-shaped arrays grow 16.7%/33.3%. These are analytic estimates, not measured wall time or host-memory peaks; a future authorized profile must follow the host guard and exclusive `ml-run` policy.
+
+[VALIDATION_PROTOCOL.md](VALIDATION_PROTOCOL.md) specifies paired shared-6k scoring, expanded-marker secondary endpoints, fixed support/K/N-generated/seed/budget, normalization/readout fairness, a proposed matched 7k-HVG control, the compact 6k-plus-core control, and independent rare-label/program validation. Existing 6k outputs and full labels should be reused. No control training or bulk-only resampling was started.
+
+## Verification and evidence status
+
+[VERIFIED.json](VERIFIED.json) independently checks exact size, Ensembl uniqueness, raw-coordinate order, nested sets, all-6k retention, old-to-new remapping, train-sample measurement/detection, protected marker coverage, input hashes and unchanged active configurations. [DETERMINISM.json](DETERMINISM.json) records two bit-identical builds across 14 output files. The candidate gene lists are complete and reproducible; biological improvement is untested. The independent targeted source-labelled ILC audit is maintained separately under `coordination/ilc_training_support/` and was not used to retune these frozen panels.
+
+Archive publication is handled by the coordinating agent under the PBISC-Diffusion maintenance policy. A completed local artifact or verifier is not itself an archive receipt or an experimental-performance result.