fix: correct stale docs — disagreement threshold, reference set size, SEED-005 exemplar

docs/DISCOVERY_PREDICTION.md

@@ -545,7 +545,7 @@ in the 20-member pilot panel, P(≥1 from family) ≈ 1 − (1 − P_candidate)^
 | 3 | SEED-006 | INWKGIAAMAKKLL vars | 0.667 | 0.643 | ~14–22% | Balanced stability+novelty; mastoparan mechanism data |
 | 4 | SEED-009 | RRLPRPPYLPRP vars | 0.80+ | 0.467 | ~12–20% | Proline-rich intracellular (DnaK); orthogonal mechanism to all other seeds; 18 selected |
 | 5 | SEED-003 | RRWQWRMKKLG vars | 0.27 | 0.182 | ~8–16% | High ensemble (0.780-0.795) but poor serum stability; 16 selected |
-| 6 | SEED-005 | KRFFKKIGSALKFA | 0.52 | 0.467 | ~8–14% | FF motif aids membrane insertion; 4 selected |
+| 6 | SEED-005 | KRLFKKIGSALKFL (seed) | 0.52 | 0.467 | ~8–14% | Cecropin-magainin hybrid; best variant VAR_009 (KRFFKKIGSALKFA, FF motif) aids membrane insertion; 4 selected |
 
 > **Excluded from synthesis** (not in synthesis pool):
 > - SEED-004 (temporin, HIGH_CYTOTOX_RISK, GRAVY=+1.81): Correctly excluded by safety+cytotox gate
@@ -671,7 +671,7 @@ Required to strengthen confidence:
 
 ### 3. Novelty Verification
 
-The 45-sequence reference set used for novelty scoring does not represent the full landscape
+The 72-sequence reference set used for novelty scoring does not represent the full landscape
 of known AMPs. Real-world novelty check required against:
 - APD3 (Antimicrobial Peptide Database v3): > 3,000 natural AMPs
 - DRAMP v3.0 (Data Repository of Antimicrobial Peptides): > 19,000 entries

docs/METHODS.md

@@ -206,15 +206,15 @@ This score does not gate the ensemble — it is used only in the pilot panel ran
 ### 4.4 Novelty Score
 
 Novelty is computed as: `1 - max_similarity`, where max_similarity is the maximum normalized
-Levenshtein similarity between the candidate and any sequence in the 45-sequence reference set
+Levenshtein similarity between the candidate and any sequence in the 72-sequence reference set
 (`examples/known_reference/amp_curated_references.csv`).
 
 Normalized Levenshtein similarity: `1 - (edit_distance / max_length)`.
 
 A novelty of 0.0 means the sequence is identical to a known reference. A novelty of 1.0 means
 no sequence in the reference set is similar at all.
 
-**Limitation:** The reference set contains only 45 curated sequences. This is not a complete
+**Limitation:** The reference set contains only 72 curated sequences. This is not a complete
 survey of known AMPs. Real-world novelty should be checked against APD3, CAMP, DRAMP, and
 UniProt antimicrobial sequences.
 
@@ -375,7 +375,7 @@ power.
 |-------------|--------|--------|
 | Scoring not validated against lab data | Activity/safety scores may not correlate with assay results | Known; lab calibration needed |
 | Near-seed generation only | Misses genuinely novel AMP families | Known; future work |
-| Small reference set | Novelty may be overestimated | Partially mitigated; 45 refs used |
+| Small reference set | Novelty may be overestimated | Partially mitigated; 72 refs used |
 | Two scorers only | Disagreement is coarse uncertainty; not a calibrated posterior | Partially mitigated; Boman index added in v0.2 |
 | Helical amphipathicity assumption | Non-helical AMPs underscored | Known limitation |
 | No structural modeling | Cannot predict membrane interaction mode | Out of scope v0.1 |

docs/WET_LAB_HANDOFF.md

@@ -46,7 +46,7 @@ The pipeline uses four gates before a candidate is marked `selected: true`.
 |------|----------------|--------------|-----|
 | `min_novelty` | ≥ 0.20 | ≥ 0.05 | Near-duplicates waste budget on already-known peptides |
 | `max_safety_risk` → min safety | ≤ 0.70 → safety ≥ 0.30 | ≤ 0.40 → safety ≥ 0.60 | Phase 3 is the final synthesis batch — only clean candidates |
-| `max_disagreement` | ≤ 0.40 | ≤ 0.30 | When physchem and Boman scoring diverge > threshold, both models are guessing |
+| `max_disagreement` | ≤ 0.40 | ≤ 0.40 | When physchem and Boman scoring diverge > threshold, both models are guessing. Both configs use 0.40 — Trp-rich scaffolds (SEED-008) show disagreement ~0.37 due to mechanism divergence (interfacial insertion vs Boman electrostatic scale), not uncertainty |
 | Sequence validity | canonical AA only | same | Non-canonical residues are disqualified |
 
 A candidate must pass **all four gates** to be nominated. Failing any single gate → `selected: false`.
@@ -163,7 +163,7 @@ It also activates G-proteins and has calmodulin-binding properties.
 If fewer than 10 candidates are nominated from a batch:
 
 1. Check `disagreement` scores — if many candidates cluster near the configured `max_disagreement`
-   threshold (0.30 for phase3.yaml, 0.40 for pipeline.yaml), the two scorers may systematically
+   threshold (0.40 for both phase3.yaml and pipeline.yaml), the two scorers may systematically
    disagree for this peptide series. Consider whether to relax `max_disagreement` in a custom
    config (document the change and the scientific justification).
 2. Check `novelty` scores — if the reference set covers your template region well, novelty will