Document shared cohort CRC and cSCC efficacy score

README.md

@@ -16,24 +16,49 @@ The public benchmark release is separated by evaluation scale. Each notebook
 is fully executed, explains every reported score, and links its compact
 evaluation script and input tables.
 
-- **[Cohort-level benchmarks](https://github.com/alunalabs/GAIA/blob/main/01_cohort_level_benchmarks.ipynb):** drug perturbation priors, the complete 44-row trial-ORR source ledger, observed-versus-predicted ORR correlations, comparisons, and permutation controls.
-- **[Patient-level benchmarks](https://github.com/alunalabs/GAIA/blob/main/02_patient_level_benchmarks.ipynb):** CRC and cSCC patient-response scores, complete-vector controls, and CRC measured on-treatment biology.
+- **[Cohort-level benchmarks](https://github.com/alunalabs/GAIA/blob/main/01_cohort_level_benchmarks.ipynb):** drug perturbation priors, the shared gene-readout efficacy score, all 44 trial ORRs and their source ledger, a separate ORR-calibrated output, comparisons, and controls.
+- **[Patient-level benchmarks](https://github.com/alunalabs/GAIA/blob/main/02_patient_level_benchmarks.ipynb):** the same four-gate score applied to CRC and cSCC, exact label controls, cSCC terminal-readout audits, and CRC measured on-treatment biology.
 - **[Biology-level benchmarks](https://github.com/alunalabs/GAIA/blob/main/03_biology_level_benchmarks.ipynb):** the 289-row CellChatBench protocol, row inventories, exact scoring rules, worked examples, and results.
 
 The complete benchmark package is in the
 **[public Gaia GitHub repository](https://github.com/alunalabs/GAIA)**.
 
+## Shared gene-readout efficacy score
+
+The cohort, CRC, and cSCC evaluations use the same final arithmetic:
+
+```text
+bio_conversion_support = mean(response_conversion_support,
+                              terminal_efficacy_support)
+
+universal_gene_readout_score = coverage_support
+                               * bio_conversion_support
+                               * resistant_tail_control
+```
+
+The terminal readout is therapy-aware: cohort and CRC use apoptosis support;
+cSCC checkpoint blockade uses control of persistent immune escape. The score
+is unitless and is not a calibrated clinical response probability or ORR.
+
 ## Released results
 
 | Evaluation | Result |
 | --- | --- |
-| Cohort ORR, 44 disease–drug rows | Pearson 0.650; Spearman 0.594; MAE 10.25 ORR percentage points |
-| CRC pretreatment response, 11 patients | ROC AUC 0.800; balanced accuracy 0.733 |
-| cSCC pretreatment response, 12 patients | ROC AUC 0.944; Spearman 0.772 |
+| Cohort universal gene-readout efficacy, 44 disease–drug rows | Pearson 0.658; Spearman 0.626 |
+| Separate cohort ORR-calibrated output, same 44 rows | Pearson 0.650; Spearman 0.594; MAE 10.25 ORR percentage points |
+| CRC shared-formula response, 11 patients | ROC AUC 0.800; Spearman 0.520; exact p=0.0628 |
+| cSCC shared-formula response, 12 patients | ROC AUC 0.778; Spearman 0.483; exact p=0.0660 |
+| Previous cSCC-specific five-axis comparator, same patients | ROC AUC 0.944; Spearman 0.772 |
 | CRC measured on-treatment biology, 11 patients | Best descriptive mean cosine 0.304 at rollout step 4 |
 | CellChatBench sender knockout | 257/289 direction passes (88.9%) |
 | CellChatBench sender overexpression | 257/281 scoreable direction passes (91.5%); 8 rows not applicable |
 
+The previous cSCC-specific score is shown because it discriminates better, but
+it is not the cross-benchmark headline. Enforcing the shared formula reduces
+cSCC AUC from 0.944 to 0.778; this is the cost of comparability, not an
+improvement. A strict literal-apoptosis cSCC audit reaches AUC 0.417 and is not
+presented as a supported checkpoint terminal mechanism.
+
 The CellChatBench release contains 183 ligand–receptor pathway rows and 106
 regional tissue-program rows. All released knockout statements expect a
 receiver decrease and all overexpression statements expect an increase, so
@@ -41,11 +66,11 @@ these results do not establish inhibitory-sign balance.
 
 ## Reproduction boundary
 
-The GitHub release reproduces metrics, controls, tables, and figures from
-compact exported benchmark data. It contains evaluation scripts only—no model
-training code—and does not rerun upstream spatial inference. Raw tissue assays,
-source perturbation matrices, full per-cell outputs, checkpoints, and model
-weights are not included.
+The GitHub release reproduces formulas, aggregation, metrics, controls, tables,
+and figures from compact exported benchmark data. It contains evaluation
+scripts only—no model training code—and does not rerun upstream spatial
+inference. Raw tissue assays, source perturbation matrices, full per-cell
+outputs, checkpoints, and model weights are not included.
 
 This Hugging Face repository currently contains the project card only; model
 weights are not currently published here.