Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261015T143826Z-gene-context-selection-rationale/diffusion/dataset/server_1/model8/AGENTS.md

@@ -0,0 +1,111 @@
+# 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 (see `presentations/2026-09-30--paper-summary--r00/CURRENT_MASTER.json` for the live page count; all 156 Figure1-focused parent pages and original76 IDs are preserved):
+`presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261015_figure1_focused.pptx`.
+Use the build and page map named by `CURRENT_MASTER.json`; the frozen156-page parent remains in `data/x1_priority_20261015/base/`. New x1 core and full CPU evaluations are explicitly distinguished.
+Earlier 155/154/142/140/131-slide decks remain preserved source versions, not the current master.
+The 34-slide review is a historical subset, not the master. The approved method palette is
+`presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_color_palette_20261013.md`:
+PBISC LoRA #168FE5, calibrated #59A8DF, raw #C3E2F4; scSemiProfiler #83C6A8;
+BayesPrism #D9B23F and CIBERSORTx #F0D77D. Preserve biological/cohort colours and quantitative heatmap scales.
+Older gold/blue source figures are preserved historical assets; use the final semantic palette derivatives in current decks.
+
+## Current manuscript figure scopes
+
+The user's 2026-10-13 definitions supersede old figure numbers inside the deck and earlier plans:
+Figure 1 overview/corpus statistics; Figure 2 architecture/gene-context ablations on SLE pseudobulk;
+Figure 3 healthy measured bulk–sc paired mismatch, full FT, LoRA ranks and support budgets;
+Figure 4 disease measured-paired quantitative integrity comparisons;
+Figure 5 paired rare-population qualitative recovery; Figure 6 paired adaptation followed by
+bulk-only published-finding reproduction (TODO).
+The authoritative scope and proposed 131-page mapping are in
+`presentations/2026-09-30--paper-summary--r00/data/figure_assignment_20261013/`.
+Read `AUTHOR_SCOPE.md` and `ASSIGNMENT.md` before figure work. Preserve the distinction
+between total-five pairs and five training plus extra validation; rank16 is not universally
+best in existing metrics, and disease comparisons have cohort/endpoint exceptions.
+Existing rare-state quantitative slides do not substitute for paired qualitative panels.
+The scope is now applied in the actual reordered PPTX/PDF, not only in planning documents.
+Use `tools/build_x1_priority_master_20261015.py --if-changed` in the presentation directory for current x1 result refreshes. It consumes the immutable156-page base and verified evaluation receipts. `build_figure1_focused_master_20261015.py` is the preserved historical Figure1 branch; rerunning it would omit the new x1 result additions.
+Figure 2 now includes corpus-validation curves and cost metadata alongside SLE pseudobulk
+benchmarks; these are not paired real-bulk performance panels. Figure 1's recorded PCA
+uses an existing ctx2048 E12 illustration, with endpoint colours revealed by denoising progress,
+not per-step confidence. Display real R2 as a real-cell repeat from the same donors, not new donors.
+Healthy pair-budget panels cover all six types with both Pearson correlation and MAE, including adverse results.
+They use historical n+7 validation access, not total-n. Do not omit NK/DC/Other to make a five-pair claim appear stronger.
+
+The latest strengthening proposal is `data/figure_assignment_20261013/STRENGTHENING_PROPOSAL.md` under the deck.
+Corpus 41,395,184 vs train 33,141,380 is a split difference, not K exclusion; all training cells are covered after top-up.
+Healthy 88.51% is a source-cell share. The remainder includes ACPA+ at-risk and is labelled Disease / at-risk.
+Historical LoRA all means 78–79 training +7 validation, keeping21–22 test/fold. Nested random training subsets and
+outer/inner splits replay exactly. The legacy total-five102-test/draw design differs from strict healthy outer-fold tests.
+Figure5 original-paper targets require author identity: Ye Mono_c06 pediatric pulmonary and Imperial wave2
+steroid-associated monocytes. Ye observed-marker co-detection is a diagnostic, not reproduced cluster identity;
+fine labels were not found in133 audited canonical obs schemas. Existing generic rare-state rules do not substitute
+for those source-defined populations. Giroux/Ji program results are not automatically rare-population evidence.
+
+Figure 1 main corpus panel now uses exact train metadata: 33,141,380 cells, 5,232 samples, 2,572 corrected donor identities; Healthy 88.4241%. The 88.51% healthy share applies to the full source pool only. Training cell-type proportions were recomputed across 432 metadata shards, not scaled from full-source proportions. Use `data/figure1_revision_20261014/train_corpus/result.json`. Figure 1a uses organic vector glyphs; cells/model nodes are schematic, and K is distinct from generated-cell count.
+
+Figure 1a now explicitly shows noisy gene-space state -> iterative PBISC denoising -> sparse expression with exact zeros; those profiles are schematic, not recorded data. Figure 1b main proportions use grouped small categories with source membership in `data/figure1_focus_20261015/corpus/display_aggregations.csv`; Unlabelled and ACPA+ at-risk remain separate. Complete training counts are preserved on support page9, full-source composition page10. Do not revert the denoiser to a generic molecular-network icon or drop the detail panel.
+
+For C1QA/CD163/CD68/WARS1/LGALS3 training-cell questions, reuse `plans/corpus_macrophage_markers_20261015/REPORT.md` and verified cell-level cache. All five are measured in the four audited cohorts; WARS1 is stored as WARS. Raw-count co-detection found 1,319 of33,219 unflagged C1Q-source-labelled training candidates, plus109 flagged. This is an existence check of a selected subset, not full-corpus prevalence or confirmed macrophage identity. Do not rerun the sparse extraction for the same question or relabel source monocytes solely from this count.
+
+## Gene-context selection rationale
+
+User instruction, 2026-10-15: preserve how and why 2k/3k/6k/7k/8k were selected for paper writing.
+Read `plans/rare_gene_context_expansion_20261015/SELECTION_RATIONALE.md` before describing or changing gene-context panels.
+Historical 2k/3k/legacy6k share a 32,768-occurrence, 1%-detection-floor ranking; final floor01 6k uses 1,048,576 occurrences and 0.1%.
+Historical 2k/3k are not exact subsets of final6k. The curated 7k/8k preserve final6k and use literature priority plus deterministic program/population round-robin.
+7k/8k are user-requested budgets, not biological minima or performance optima; core+sentinel coverage requires 6,044 available genes, while the full available literature union is 8,091.
+Separate within-selector budget ablation from literature-guided feature expansion. Benchmark-paper marker use is endpoint-informed; the later native ILC audit did not select genes.
+Retain exact versions, source manifests and per-gene reasons; do not invent an original numeric-budget rationale or rewrite it after seeing results. Current candidate lists are frozen and no 7k/8k training has been launched by these tasks.

paper-releases/pbisc-diffusion/updates/20261015T143826Z-gene-context-selection-rationale/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/README.md

@@ -0,0 +1,13 @@
+# 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.
+
+- [Selection rationale and paper reporting contract](SELECTION_RATIONALE.md): historical panel versions, exact curated selector, fixed-budget rationale, and limits on causal interpretation.

paper-releases/pbisc-diffusion/updates/20261015T143826Z-gene-context-selection-rationale/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/SELECTION_RATIONALE.md

@@ -0,0 +1,86 @@
+# Gene-context selection rationale and paper reporting contract
+
+Date: 2026-10-15. User instruction: preserve how and why the 2k/3k/6k/7k/8k panels were selected so the paper can explain the experiment design. This document records verified implementation/provenance separately from present scientific interpretation. It does not amend the manuscript, change panels, establish an optimal size, or authorize a new training campaign.
+
+## 1. Scientific questions and the meaning of the budgets
+
+**Two questions must remain distinct.**
+
+1. Within a fixed selection recipe, how does a larger gene budget change reconstruction, disease-program fidelity, cost and validation behaviour? This supports a gene-budget comparison only when selector, data, normalization and evaluation are matched.
+2. Does adding literature-grounded rare-cell identity, boundary and state information to a frozen broad-expression panel improve rare-program fidelity? This is a targeted feature-expansion question. It changes both feature content and dimension relative to 6k.
+
+The historical 2k/3k/6k budgets can be discussed as progressively larger feature budgets, but that is a current interpretation. The bounded audit located the numerical defaults and exact algorithms, **not a contemporaneous record explaining why those precise three numbers were chosen**. Do not retrospectively call them optimized, biologically necessary, or selected by rare-cell recovery. Record any later recovered original decision evidence as a separately dated addition.
+
+The 7k and 8k totals were explicitly requested by the user before the expansion audit. They are **fixed comparison budgets**, not numbers inferred by optimizing coverage or held-out performance. The deterministic selection filled +1,000 and +2,000 positions; coverage explains the trade-off after construction. Selection was frozen before any 7k/8k training; it was not globally prespecified before all earlier PBISC experiments.
+
+## 2. Versioned lineage: a size label alone is insufficient
+
+| Family | Selection sample and eligibility | Relationship |
+|---|---|---|
+| Historical 2k / 3k / legacy 6k | 256 train recipes × 128 sampled cells = 32,768 occurrences; seed 20260627; measured ≥3,276 (10%); detected ≥327 (1%) | The v3 ladder uses prefixes of one eligible ranking: 2k ⊂ 3k ⊂ 6k ⊂ broad 11,589. Original 2k/3k arrays replay exactly. |
+| Final floor01 6k | 8,192 train recipes × 128 = 1,048,576 occurrences; same seed; measured ≥104,857 (10%); detected ≥1,048 (0.1%) | New eligible universe of 15,114; its own 2k/6k/broad ladder is nested. This is the frozen baseline for the new candidates. |
+| Curated 7k / 8k | Preserve final floor01 6k and add source-grounded genes using the selector below | final 6k ⊂ curated 7k ⊂ curated 8k; a new selection strategy, not continuation of pure HVG rank. |
+
+Historical selection uses library normalization target 10,000, measurement-aware log1p mean/variance, log1p(variance/mean) standardized within 20 mean-expression bins, eligible immune/interferon marker priority, dispersion ranking, and canonical-index tie-breaking. These are sampled occurrences, not asserted distinct cells.
+
+**Direct cross-family overlaps with final 6k:** historical 2k: 1,998/2,000; historical 3k: 2,975/3,000; legacy 6k: 4,258/6,000. Therefore even the historical 2k/3k are not exact subsets of final 6k. Never draw one unqualified nested 2k→3k→6k→7k→8k ladder across these versions. Identify each plotted result by selector family, exact gene list/hash, checkpoint and experiment receipt. The recorded Figure1 ctx2048 illustration is not automatically this exact 2,000-gene panel.
+
+Sources:
+- `artifacts/hvg_2000_train256x128_seed20260627/manifest.json`
+- `artifacts/hvg_3000_train256x128_seed20260627/manifest.json`
+- `artifacts/panel_ladder_train256x128_seed20260627_v3/manifest.json` and `FINAL_VERIFICATION.json`
+- `artifacts/panel_ladder_floor01_1m_20260923/manifest.json`
+- `src/pbisc_sdd/gene_selection.py` and `src/pbisc_sdd/panel_ladder.py`
+
+## 3. Why expand beyond final 6k?
+
+A pooled dispersion/detection selector may omit low-frequency lineage or state information. Some ILC-related features are already present, while PTGDR2, NCR2 and several cytokine/receptor loci are absent. A gene can be excluded by rank or by a prevalence threshold; do not attribute every omission to the 0.1% floor. PTGDR2, for example, is detected above that floor but is outside the selected 6k rank.
+
+The expanded design aims to retain broad reconstruction coverage while adding complementary identity, lineage-boundary and state programs. It includes ILC/ILCP and DC programs plus source-defined disease targets such as SLE DN2/LDG, sepsis monocyte states, steroid-associated monocytes, JDM states and circulating progenitors. These include shared genes and state contrasts: the additions are **not 1,000/2,000 independently specific rare-cell markers**, and the source target rows are not all separate rare lineages.
+
+## 4. Exact frozen 7k/8k selection rule
+
+1. Preserve every final-6k Ensembl ID. Resolve primary-literature symbols through the frozen HGNC reference onto the 20,097-gene raw axis; unresolved/conflicting mappings are not invented as zero-valued features. Preserve distinct existing raw loci when symbols repeat.
+2. New candidates must be raw-axis-resolved and have `measured_cells > 0` in the frozen train sample. **Positive detection and the old 0.1% detection floor are not curated-candidate filters or ranking keys.** All selected additions happen to have positive pooled detection; this is an audit result, not their selection rule.
+3. Assign tiers: **tier 0 = core genes and the 17 protected functional sentinels together**; tier 1 = supporting blood and bounded disease programs; tier 2 = optional stimulation states; tier 3 = optional tissue programs. Protection does not turn stimulated/tissue cytokines into universal resting-blood identity requirements.
+4. Within each population/source-program queue, order by explicit source_rank, otherwise source-statistic rank (ascending); otherwise absolute log2FC/avg_logFC/avg_diff (descending); otherwise input-evidence row order. Ensembl ID resolves equal rank/effect ties.
+5. Traverse source-program queues lexicographically in round-robin within each population, then population queues likewise within each tier. Skip duplicates; earlier tiers win. This limits domination by long gene lists, but does not guarantee equal coverage of biological families or optimize a biological fairness objective.
+6. Take the first 1,000 or 2,000 unique additions and union with the frozen 6k. HVG backfill is implemented as a fallback but contributed **zero** genes. Serialize in ascending canonical raw-coordinate order and retain exact old-to-new maps; the old 6k is a preserved set, not necessarily the first 6,000 tensor rows.
+
+Authoritative implementation: `build_panels.py`; inputs and hashes: `selection_manifest.json`; gene-level provenance: `evidence_resolved.csv`; output checks: `VERIFIED.json` and `DETERMINISM.json`. Selection code SHA-256: `ec36872c10d3a560fa60ae8d42760cf70b596806453eedba7a69612a826302f7`.
+
+## 5. What the coverage supports, and what it does not
+
+Evidence-derived counts differ from the requested budgets:
+- Preserved 6k + all available missing core genes: **6,037**.
+- Also protect available functional sentinels: **6,044**.
+- Preserved 6k + the entire available literature union: **8,091**.
+
+| Audited available gene set | 7k | 8k |
+|---|---:|---:|
+| Disease evidence catalog | 1,081/1,081 | 1,081/1,081 |
+| Blood-supporting programs | 2,225/2,603 | 2,603/2,603 |
+| ILCP program | 390/484 | 484/484 |
+| DC3 program | 498/667 | 667/667 |
+| pDC program | 293/331 | 331/331 |
+| Same-benchmark author targets | 674/674 | 674/674 |
+
+Sets overlap and must not be summed. Both candidates retain all available core/protected genes. The 91 available genes omitted by 8k belong only to optional tissue programs. Raw-unavailable genes remain explicit gaps; these counts do not represent all human rare-cell biology.
+
+Thus **7k is the compact targeted-expansion candidate; 8k is the broader ILC/DC program candidate**. The recommendation for 8k concerns feature coverage and remains a proposal, not an adopted optimal size or demonstrated gain. The first experiment's execution order is a separate decision; the staged validation protocol permits a compact 6k-vs-7k start with 8k as the larger-program arm. Not every possible control must be launched.
+
+## 6. Paper comparison and evaluation contract
+
+- A within-family 2k/3k/6k comparison can address feature budget if the training/evaluation contracts match. Cross-family historical results need explicit selector labels; do not attribute their differences only to gene count.
+- Final6k versus curated7k/8k is **literature-guided rare-program expansion**. To isolate curated gene content from larger dimensionality, a same-size train-HVG extension is a proposed control. It has not been executed. The 6,044 compact protected expansion is another optional diagnostic, not mandatory new work.
+- Curated7k versus curated8k shares a selector and nested ranking, but larger G also changes input/output parameter count and compute. Report parameters, time/memory when measured, training budget and shared-axis performance. Do not claim constant model capacity solely because trunk width is fixed.
+- Gene selection did not use local held-out expression. Nevertheless, marker tables from benchmark publications informed the panel: call this **endpoint-informed design**, not completely blind or wholly train-only feature selection. Distinguish source-target reproduction from independent validation.
+- Native source-labelled ILC expression was audited **after panel freezing** and did not rank or retune the candidates. It is evidence of training signal, not the original reason a particular gene was selected and not confirmed ILC identity.
+- The fixed 1,005-gene annotator ignores new genes and maps ILC into broad Other. Retain that score for continuity and add separately justified rare-cell/program endpoints. Existing saved predictions must be reused before new compute.
+- Freeze panel/threshold/readout choices before expanded-model results. Report all tested arms and failures; never relabel whichever size wins as a prospectively optimized size. A feature-coverage rationale is not a generation-success claim.
+
+## 7. Durable writing rule
+
+Methods should report the **source, sample, eligibility, ranking, priority, fixed budget and exact panel version**. Results should report the **measured performance/cost and limits**. Keep the design rationale separate from the outcome. Gene context denotes the selected modeled gene space here; internal conditioner width is a different architecture variable.
+
+No manuscript text, gene list, checkpoint or active configuration was modified by this documentation task. For the numerical evidence and proposed evaluation details, read `REPORT.md` and `VALIDATION_PROTOCOL.md`. Future edits must preserve this provenance; new selection rules or evidence require a new panel version rather than rewriting the original selection history.

paper-releases/pbisc-diffusion/updates/20261015T143826Z-gene-context-selection-rationale/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/coordination/RATIONALE_PROVENANCE.json

@@ -0,0 +1,43 @@
+{
+  "schema": "pbisc.gene_context_rationale.v1",
+  "status": "documented_from_existing_evidence",
+  "date": "2026-10-15",
+  "provenance": "User requested durable rationale; independent read-only historical and curated-selector audits",
+  "historical_exact_numeric_budget_reason_found": false,
+  "panels_changed": false,
+  "training_launched": false,
+  "sources": [
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/artifacts/hvg_2000_train256x128_seed20260627/manifest.json",
+      "sha256": "30ce7c67505b97c2e225a13c5d50f5ee6705db8587b8c30c48c351d35627ba74"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/artifacts/hvg_3000_train256x128_seed20260627/manifest.json",
+      "sha256": "912ba92f1e314c12ef9010c99586b48d505249c32b8f19a36c847c09f52282c1"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/artifacts/panel_ladder_train256x128_seed20260627_v3/manifest.json",
+      "sha256": "bcd80d9555f74775048f32ddca1dfbcaf9883909ae70fd2bca4e11c66efe81b5"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/artifacts/panel_ladder_train256x128_seed20260627_v3/FINAL_VERIFICATION.json",
+      "sha256": "fa078766d49b38ff7236511036c019520244a22be82232d4e1c74a6a64256bc6"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/artifacts/panel_ladder_floor01_1m_20260923/manifest.json",
+      "sha256": "acbc8bc2cca5ecf3d3002e711b1d4a55b4069f17ae9666f0d64c197ca5cbc775"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/rare_gene_context_expansion_20261015/build_panels.py",
+      "sha256": "ec36872c10d3a560fa60ae8d42760cf70b596806453eedba7a69612a826302f7"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/rare_gene_context_expansion_20261015/selection_manifest.json",
+      "sha256": "2c8922df98464c99b83a1402bb7cb5a2fc83122be4fc1c7a1bd1924b673b67f6"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/rare_gene_context_expansion_20261015/SELECTION_RATIONALE.md",
+      "sha256": "3286937aa3fedd8ef388dd0176b4850f2db56cf3d3b1544b9dc4751a5fa17cfe"
+    }
+  ]
+}