docs: update ROADMAP, add AUROC reproducibility, update RISK_REVIEW

docs/METHODS.md

@@ -320,6 +320,10 @@ Phase 2 benchmarks verified that the pipeline:
 benchmark set using the full ensemble scorer). Benchmarked after PRs #48–54; the scoring
 improvements in those PRs moved AUROC from 0.7926 → 0.8138 → 0.8164.
 
+**To reproduce:** `make validate-scoring` (runs `cli validate-scoring` on
+`examples/validation/known_amps.csv` vs `examples/validation/random_background.csv`)
+or `make validate-scoring-strict` for the composition-matched (scrambled-decoy) variant.
+
 **Key scoring changes in v0.2.x (PRs #48–54):**
 
 | PR | Change | AUROC impact |

docs/RISK_REVIEW.md

@@ -35,14 +35,18 @@ OpenAMP Foundry pipeline. It covers:
 
 | Risk | Status | Mitigation |
 |------|--------|-----------|
-| Extreme hydrophobicity (hemolysis proxy) | **FILTERED** | max_safety_risk=0.40 in phase3.yaml excludes sequences with hydrophobic_fraction > 0.65 |
-| Extreme charge density | **FILTERED** | Poly-cationic sequences excluded by safety scorer |
+| Extreme hydrophobicity (hemolysis proxy) | **FILTERED** | max_safety_risk=0.40 in phase3.yaml excludes sequences with hydrophobic_fraction > 0.65 or strong amphipathicity (μH > 0.55) |
+| Extreme charge density | **FILTERED** | Poly-cationic sequences excluded by safety scorer (charge_density_ph74 > 0.55 triggers risk increment); uses pH-7.4 charge for accuracy |
+| Predicted mammalian cytotoxicity | **FILTERED + DOWN-RANKED** | Selectivity proxy (charge/GRAVY ratio) flags HIGH_CYTOTOX_RISK candidates (proxy < 0.5); cytotox_penalty doubly demotes them in pilot panel selection |
 | Long repeat runs (degenerate composition) | **FILTERED** | Sequences with repeat_run ≥ 6 penalised; few reach selection threshold |
+| High proline content | **REPORTED** | proline_fraction > 15% incurs synthesis penalty (−0.10); flagged in expert review for pseudoproline dipeptide use |
 | Cysteines (disulfide bridges) | **REPORTED** | Flagged in synthesis feasibility report; reviewer must assess |
 | Sequences resembling known toxins | **NOT ASSESSED** | Computational toxin screening is out of scope for this pipeline version |
 
 **Action required:** A qualified reviewer must inspect the synthesis feasibility report
 for candidates with high cysteine fraction or proline content before ordering synthesis.
+Candidates with HIGH_CYTOTOX_RISK flag (selectivity_proxy < 0.5) require extra hemolysis
+assay scrutiny before any mammalian cell work.
 
 ---
 

docs/ROADMAP.md

@@ -1,47 +1,64 @@
 # Roadmap
 
-## v0.1 — Safe starter
+## v0.1 — Safe starter ✓
 
 - toy data;
 - transparent heuristic scoring;
 - JSON evidence certificates;
 - report generation;
 - tests and CI.
 
-## v0.2 — Dataset layer
+## v0.2 — Dataset layer ✓
 
 - license-aware dataset loaders;
 - dataset cards;
 - deduplication;
-- cluster splitting;
-- leakage checks.
+- cluster splitting (`benchmark/splits.py`);
+- leakage checks (`bench leakage` CLI + `test_cluster_split.py`).
 
-## v0.3 — Predictor adapters
+## v0.3 — Predictor adapters ✓
 
-- adapters for open AMP predictors;
-- versioned model manifests;
-- ensemble scoring;
-- uncertainty score.
+- Boman index adapter (independent second scorer, `scoring/boman.py`);
+- ensemble scoring with configurable weights (`scoring/ensemble.py`);
+- model disagreement uncertainty signal (disagreement = |activity − boman|).
 
-## v0.4 — Benchmark suite
+## v0.4 — Benchmark suite ✓
 
-- hidden-positive recovery;
-- negative control rejection;
-- novelty stress test;
-- toxicity penalty benchmark.
+- hidden-positive recovery (`bench baseline`, `test_hidden_active_recovery.py`);
+- negative control rejection (`test_negative_penalization.py`, `test_negative_robustness.py`);
+- novelty stress test (`test_novelty_pressure.py`);
+- toxicity penalty benchmark (`test_toxicity_penalty.py`);
+- retrospective AUROC benchmark (`validate-scoring` CLI, AUROC = 0.8164 on demo set).
 
-## v0.5 — Lab-batch package
+## v0.5 — Lab-batch package ✓
 
-- pre-registration template;
-- batch manifest;
-- candidate certificates;
-- expert-review checklist;
-- safe publication template.
+- pre-registration template (`docs/SELECTION_RULE.md`);
+- batch manifest (`outputs/run_manifest.json` with SHA-256 input hashes);
+- candidate certificates (JSON + `schemas/candidate.schema.json`);
+- expert-review checklist (`docs/EXPERT_REVIEW_PACK.md`);
+- wet-lab handoff guide (`docs/WET_LAB_HANDOFF.md`);
+- safe publication template (`docs/NOMINATION_REPORT.md`).
+
+## v0.5.x — Scoring enhancements (shipped alongside v0.5)
+
+Implemented during the pre-wet-lab improvement loop (PRs #31–#54):
+
+- Serum stability scoring — HNE elastase + trypsin + chymotrypsin site density
+- Selectivity proxy — charge/GRAVY-based mammalian cytotoxicity risk detector
+- Pilot panel selection — `pilot_priority` formula with serum stability, novelty, selectivity, cytotox_penalty
+- Aggregation propensity scoring — consecutive hydrophobic run + β-branched density
+- Aggregation-safe mutation generation — prevents creating VILMFW runs ≥ 4
+- Proline synthesis penalty — DKP/coupling-difficulty penalty at > 15% proline
+- Helix propensity bonus — Chou-Fasman Pα weight increased 0.01 → 0.03 in activity scorer
+- Safety score uses pH-7.4 charge consistently with activity scorer
+- C-terminal amidation and N-terminal acetylation recommendations in QC
+- Wave 2 D-amino guidance in synthesis order
+- DISCOVERY_PREDICTION.md — quantified discovery probability model (~29–49%)
 
 ## v1.0 — Validated dry-lab-to-wet-lab loop
 
 - independently reviewed assay batch;
 - lab results ingested;
 - negative results archived where safe;
-- active-learning batch 2;
+- active-learning batch 2 (D-amino variants, MDR strain panel);
 - public reproducibility report.