Archive 4 checkpoints (1-4)

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

@@ -0,0 +1,43 @@
+# Candidate-panel validation protocol — no training authorized or launched here
+
+## What the panel comparison can establish
+
+The causal question is whether adding specified gene coordinates improves rare-cell identity and state fidelity, beyond changes in input/output dimension, measurement support, generated-cell number, and the annotation/scoring axis. A missing ILC row in a six-class score is not evidence of absent ILC generation. An ILC classifier call is not independently established ILC identity.
+
+The present panel selection uses published human evidence and previously frozen pretraining-train-only pooled statistics. No local held-out expression was used to rank genes. Published Ye, Imperial, Giroux, and Ji endpoint programs are nevertheless **endpoint-informed panel design**, not a claim of fully blind feature selection. External sources can also overlap pretraining studies; independent-reference exposure must be audited before a generalization claim. Report source-panel reproduction separately from independent cohort validation.
+
+## Minimal staged experiment
+
+1. Reuse existing final-6k weights, exact-five support draws, generated arrays, full AIFI labels, scoring caches, and source receipts. Existing arrays cannot supply genuinely new output genes. Inspect retained `aifi_l1`/`aifi_l2` and saved probability outputs before launching annotation again. This audit already establishes nonzero saved ILC calls in three cohorts.
+2. Audit positive training examples and measurement support for a bounded ILC marker set. Use source-labelled train ILC rows, count-native measurements, and T/NK exclusion markers; interpret those labels as source labels, not expert-confirmed pure ILC. Positive expression alone does not establish joint identity. Pooled 1M-sample positivity does not prove expression in ILC or support across donors/platforms.
+3. For an explicitly authorized future experiment, use a compact first comparison: final 6k versus curated 7k, same trunk, data stream, optimizer budget, K, N-generated, support/test donors, adaptation schedule, and sampling/readout. The 8k candidate is the predeclared larger-program arm. Its all-blood-supporting coverage is a concrete reason to test it, not proof that 8k is required. Do not automatically run every possible ablation.
+4. If the curated arm improves, distinguish scientific feature choice from dimension using a **7k train-HVG control**, preserving the 6k and adding the next 1,000 eligible frozen train ranks, and optionally a **6k plus the missing protected/core genes** arm. These are proposed controls only; they have not been trained. The minimal core count alone does not include all optional functional sentinels or complete source programs.
+5. Before new compute, query the experiment reuse registry and inspect current two-host receipts. Exact support/test IDs, gene order, conditioner/scaler, masks, checkpoint, seeds, training budget, selection schedule, and generated-cell count must agree for adoption. A new panel legitimately changes model coordinates; a new report or sampling seed does not justify retraining an unchanged model.
+
+## Shared-axis primary endpoint
+
+- Preserve all original 6,000 genes. For the primary comparison use the same per-cohort measured intersection of those 6,000 genes and the same exact donor/test cell sets. Project to this common axis **before** CPM/CP10K, PCA, pseudobulk correlations, marker means, and depth/detected-gene comparisons. Expanding the denominator to 7k/8k would change values on unchanged genes.
+- For generated readout, preserve the Poisson policy and predeclare a common-axis randomization contract. The same integer seed with different row-major matrix widths does not ensure matched draws on shared genes. Project generated rates to the same ordered 6k axis before the common comparison readout, or use a validated random stream keyed by sample/cell/gene. Preserve the historical score as historical evidence when this differs from its native readout.
+- Keep the established frozen 1,005-gene AIFI annotator and six-class score as a continuity endpoint, with exactly the same query genes and zero-padding policy. Preserve full AIFI L1/L2 predictions as an additional diagnostic. This fixed annotator ignores all newly added genes; improvement under an expanded rare-cell classifier is a separate endpoint, not a like-for-like continuation of the old score.
+- Fix K and generated-cell number separately. K is the conditioning cell-bag size; N-generated is a sampling budget. Include zero-draw samples, donor-level uncertainty, and prevalence-aware power. Source-label prevalence 16,439 / 33,141,380 yields only 0.744 expected ILC in 1,500 independent draws, but actual conditional prevalence and dependence can differ substantially. Do not pool away donor failures.
+
+## Expanded-axis secondary endpoints
+
+- Evaluate newly generated genes only where both observed assay and target bulk feature masks support the measurement. Treat unavailable genes and the old model's absent output coordinates as **NA**, never biological zeros. `panel_gene_measurement_audit.csv` supplies audited source-feature availability, not per-population positivity.
+- Separate identity, exclusion/boundary, and state programs. For ILC include T/NK alternatives and coexpression, with special caution for blood ILC1-like/TCR-rearranged populations. Cytokines observed after stimulation or in tissue are assay-capability checks; their absence in resting blood does not establish absence of the lineage. Do not require every core/sentinel to be detected in every individual cell.
+- KIT-positive circulating ILCP are not automatically mature ILC3, and blood DC3 are not the same population as tissue/tumor mregDC; keep the ontology and specimen boundaries detailed in `research/ilc_dc/REPORT.md`. A cross-population state module does not establish a new lineage.
+- Use independent labels where possible: source-author fine annotations with documented transfer, protein/flow/V(D)J evidence, or a separately held-out rare-cell reference. Quantify realistic false positives in abundant T/NK/monocyte cells. The AIFI ILC validation subset has only 111 true ILC, and reference performance does not prove generated distribution fidelity.
+- Prespecify a validation marker/program set separate from classifier-fitting inputs and from panel-selection evidence where possible. Those genes may remain model inputs/outputs: excluding every validation gene from the generative model would prevent evaluating the desired gene-space capability. The independence boundary concerns **how labels and success criteria are defined**, not blanket exclusion from generation.
+- Distinguish benchmark-author target fidelity (Ye Mono-c06/CD14-MKI67, treatment-associated Imperial monocytes, etc.) from independent cohort generalization. Freeze thresholds using development/support material only; do not tune them on held-out generated-versus-observed test results. Register disease-, treatment-, specimen-, and donor-specific limitations.
+
+## Model and preprocessing compatibility
+
+The existing `SubsetAnnotationFreeConditioner` computes the full canonical bulk normalization and six depth/statistical features first, then subsets gene features. Preserve that full conditioner, its train-fitted parameters, and full-source totals. Re-fitting it on 7k or 8k would silently change normalization. Cell targets use `log1p(raw_count)` and the fixed scaler; HVG selection's CP10K-like normalization is not the training target transformation.
+
+The conditioner has G + 6 features; two G-wide measurement channels yield **3G + 6** final condition inputs. Noisy/self-conditioning inputs have 4G columns and the two output heads each have G rows. Consequently a 6k checkpoint cannot generate extra genes merely by changing a CSV. A future expansion must define fresh training or an explicitly labelled weight-migration experiment: copy shared gene blocks by `old_to_new.csv`, initialize new gene blocks, update all corresponding condition/state/head coordinates, and retrain. LoRA adapters also bind their parent shapes and gene order and are not interchangeable across these panels.
+
+The condition ordering is `[normalized bulk G, six scalars, bulk measurement G, output availability G]`; the scalar positions shift with G. State/self-condition inputs likewise contain four separate G-wide blocks. A naive append or an unchanged leading-column copy is therefore not a valid migration.
+
+The new manifests intentionally identify a literature-based candidate selector. The current `explicit_train_only_hvg_manifest` loader accepts the legacy HVG schema and will reject these manifests. Before a future training launch, add and review a compatible manifest/loader mode; do not mislabel the literature design as a pure train-HVG selector. Keep active 6k configurations and frozen releases intact.
+
+Analytic parameter estimates use the actual model layer definitions and reproduce the recorded 6k parameter count. They are not measured wall time or peak-memory predictions. Gene-shaped buffers grow 16.7% or 33.3%; sparse loading, prefetch, optimizer/EMA storage, and throughput require a bounded authorized profile. All future CUDA or substantial-memory work must satisfy the host guard, exclusive lock, and `ml-run` policy under the coordinating agent.

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

@@ -0,0 +1,31 @@
+{
+  "status": "PASS",
+  "checks": [
+    {
+      "n_genes": 7000,
+      "additions": 1000,
+      "preserved_old_genes": 6000,
+      "protected_core_and_sentinels_present": true,
+      "coordinate_order_and_mapping": "pass",
+      "all_added_genes_train1m_measured_and_detected": true,
+      "selected_genes_sha256": "b9ba04fb554cdec206a72fb6b00f2b3a7c49f038abebaa05473d87a71e265ec1",
+      "gene_positions_sha256": "f4bda0b94e2f5d6fcc7e44e34765a1b77b31dce85de49925bc95b3906ee91e10"
+    },
+    {
+      "n_genes": 8000,
+      "additions": 2000,
+      "preserved_old_genes": 6000,
+      "protected_core_and_sentinels_present": true,
+      "coordinate_order_and_mapping": "pass",
+      "all_added_genes_train1m_measured_and_detected": true,
+      "selected_genes_sha256": "62d961d83cf180dd925e080f696827f48074125cf2725d290cda21247d23c60f",
+      "gene_positions_sha256": "a6d8acb7ef27ae1c2733d50072e2d21cdb54b20a38910a5d10c4c8f9d72305d8"
+    }
+  ],
+  "source_files_rehashed": 13,
+  "active_6k_configs_unchanged": true,
+  "no_cuda_imported": true,
+  "no_training_or_inference": true,
+  "biological_generation_improvement_tested": false,
+  "verifier_sha256": "efbe890b9ce0ffde2b572dd7732d35f69f23ad4b648e75958fc8c00df08d0dae"
+}

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

@@ -0,0 +1,55 @@
+gene_id,gene_symbol,chromosome,start,end,gene_biotype,panel_index,is_model_gene,is_immune_marker,is_interferon_gene,is_sorted_bulk_measured,measurement_frequency_sc,measurement_frequency_bulk
+ENSG00000239605,STPG4,chr2,47045538,47155308,protein_coding,2313,True,False,False,True,0,27
+ENSG00000273269,STPG4,chr2,47065941,47176511,protein_coding,2314,True,False,False,True,0,27
+ENSG00000163633,C4orf36,chr4,86876205,86892813,protein_coding,4809,True,False,False,True,0,27
+ENSG00000285458,C4orf36,chr4,86886472,86936178,protein_coding,4810,True,False,False,False,0,0
+ENSG00000287542,HERC3,chr4,88523810,88708450,protein_coding,4830,True,False,False,False,0,0
+ENSG00000138641,HERC3,chr4,88592434,88708540,protein_coding,4831,True,False,False,True,0,27
+ENSG00000284762,PDE8B,chr5,77086732,77427124,protein_coding,5440,True,False,False,False,0,0
+ENSG00000113231,PDE8B,chr5,77210449,77428256,protein_coding,5441,True,False,False,True,0,27
+ENSG00000168255,POLR2J3,chr7,102537918,102572653,protein_coding,7688,True,False,False,True,0,27
+ENSG00000285437,POLR2J3,chr7,102562133,102572594,protein_coding,7692,True,False,False,False,0,0
+ENSG00000258724,PINX1,chr8,10725399,10839847,protein_coding,8117,True,False,False,False,0,0
+ENSG00000254093,PINX1,chr8,10764961,10839884,protein_coding,8118,True,False,False,True,0,27
+ENSG00000286112,KYAT1,chr9,128819663,128881931,protein_coding,9334,True,False,False,False,0,0
+ENSG00000171097,KYAT1,chr9,128832937,128882556,protein_coding,9335,True,False,False,True,0,27
+ENSG00000184154,LRTOMT,chr11,72080286,72096895,protein_coding,11133,True,False,False,True,0,27
+ENSG00000284922,LRTOMT,chr11,72080331,72110782,protein_coding,11134,True,False,False,False,0,0
+ENSG00000285188,PDE4C,chr19,18207961,18255419,protein_coding,17011,True,False,False,False,0,0
+ENSG00000105650,PDE4C,chr19,18207965,18248200,protein_coding,17012,True,False,False,True,0,27
+ENSG00000182378,PLCXD1,chrX,276322,303358,protein_coding,19156,True,False,False,False,0,0
+ENSG00000178605,GTPBP6,chrX,303346,318846,protein_coding,19157,True,False,False,False,0,0
+ENSG00000167393,PPP2R3B,chrX,333933,386955,protein_coding,19158,True,False,False,False,0,0
+ENSG00000185960,SHOX,chrX,624344,659411,protein_coding,19159,True,False,False,False,0,0
+ENSG00000205755,CRLF2,chrX,1190490,1212723,protein_coding,19160,True,False,False,False,0,0
+ENSG00000198223,CSF2RA,chrX,1268779,1325373,protein_coding,19161,True,False,False,False,0,0
+ENSG00000185291,IL3RA,chrX,1335025,1382689,protein_coding,19162,True,False,False,False,0,0
+ENSG00000169100,SLC25A6,chrX,1386133,1392929,protein_coding,19163,True,False,False,False,0,0
+ENSG00000169093,ASMTL,chrX,1403138,1453786,protein_coding,19164,True,False,False,False,0,0
+ENSG00000182162,P2RY8,chrX,1462572,1537187,protein_coding,19165,True,False,False,False,0,0
+ENSG00000197976,AKAP17A,chrX,1591604,1602532,protein_coding,19166,True,False,False,False,0,0
+ENSG00000196433,ASMT,chrX,1615059,1643081,protein_coding,19167,True,False,False,False,0,0
+ENSG00000169084,DHRSX,chrX,2219506,2502805,protein_coding,19168,True,False,False,False,0,0
+ENSG00000214717,ZBED1,chrX,2486414,2502862,protein_coding,19169,True,False,False,False,0,0
+ENSG00000002586,CD99,chrX,2690988,2741312,protein_coding,19170,True,False,False,False,0,0
+ENSG00000124333,VAMP7,chrX,155881227,155943769,protein_coding,20020,True,False,False,False,0,0
+ENSG00000124334,IL9R,chrX,155997696,156013017,protein_coding,20021,True,False,False,False,0,0
+ENSG00000182484,WASH6P,chrX,156020826,156025710,protein_coding,20022,True,False,False,False,0,0
+ENSG00000292344,PLCXD1,chrY,276322,303358,protein_coding,20023,True,False,False,False,0,0
+ENSG00000292358,GTPBP6,chrY,303346,318846,protein_coding,20024,True,False,False,False,0,0
+ENSG00000292327,PPP2R3B,chrY,333933,386955,protein_coding,20025,True,False,False,False,0,0
+ENSG00000292354,SHOX,chrY,624344,659411,protein_coding,20026,True,False,False,False,0,0
+ENSG00000292363,CRLF2,chrY,1190490,1212723,protein_coding,20027,True,False,False,False,0,0
+ENSG00000292357,CSF2RA,chrY,1268779,1325373,protein_coding,20028,True,False,False,False,0,0
+ENSG00000292332,IL3RA,chrY,1335025,1382689,protein_coding,20029,True,False,False,False,0,0
+ENSG00000292334,SLC25A6,chrY,1386133,1392929,protein_coding,20030,True,False,False,False,0,0
+ENSG00000292339,ASMTL,chrY,1403138,1453786,protein_coding,20031,True,False,False,False,0,0
+ENSG00000292333,P2RY8,chrY,1462572,1537187,protein_coding,20032,True,False,False,False,0,0
+ENSG00000292343,AKAP17A,chrY,1591604,1602532,protein_coding,20033,True,False,False,False,0,0
+ENSG00000292336,ASMT,chrY,1615059,1643081,protein_coding,20034,True,False,False,False,0,0
+ENSG00000292338,DHRSX,chrY,2219506,2502805,protein_coding,20035,True,False,False,False,0,0
+ENSG00000292345,ZBED1,chrY,2486414,2502862,protein_coding,20036,True,False,False,False,0,0
+ENSG00000292348,CD99,chrY,2690988,2741312,protein_coding,20037,True,False,False,False,0,0
+ENSG00000292366,VAMP7,chrY,57067747,57130289,protein_coding,20081,True,False,False,False,0,0
+ENSG00000292373,IL9R,chrY,57184216,57199537,protein_coding,20082,True,False,False,False,0,0
+ENSG00000292372,WASH6P,chrY,57207346,57212230,protein_coding,20083,True,False,False,False,0,0

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

@@ -0,0 +1,7 @@
+architecture,genes,parameters_analytic,extra_parameters_vs6k,parameter_ratio_vs6k,gene_axis_activation_ratio_vs6k,fp32_weights_extra_MiB,assumed_20bytes_training_state_extra_MiB,runtime_measured
+trunk1x,6000,54185184,0,1.0,1.0,0.0,0.0,False
+trunk1x,7000,61867184,7682000,1.1417730721372101,1.1666666666666667,29.30450439453125,146.52252197265625,False
+trunk1x,8000,69549184,15364000,1.28354614427442,1.3333333333333333,58.6090087890625,293.0450439453125,False
+trunk4x,6000,239955168,0,1.0,1.0,0.0,0.0,False
+trunk4x,7000,266069168,26114000,1.1088286625274935,1.1666666666666667,99.61700439453125,498.08502197265625,False
+trunk4x,8000,292183168,52228000,1.217657325054987,1.3333333333333333,199.2340087890625,996.1700439453125,False