[cschanhniem]

Summary

Addresses the external expert review (scored 24/100) that identified three critical gaps in the pre-wet-lab package:

• Candidate correlation not modelled: near-seed variants are correlated; if a seed family fails, most variants from it fail together. Effective independent experiments ≈ 8 scaffold families, not 20 candidates.
• Probability estimate too optimistic: internal 29–49% does not account for correlation. Calibrated estimate: 15–30% for ≥1 hit with acceptable selectivity.
• Competitive landscape: AMPGAN v3 (arXiv 2606.17127, June 2026) and ESCAPE benchmark (arXiv 2511.04814) set a new bar for the field.

Changes

docs/DISCOVERY_PREDICTION.md

• New External Calibration section with honest re-estimates and explanation of near-seed correlation
• Competitive landscape notes (AMPGAN v3, ESCAPE)
• Benchmark limitations (AUROC measured on 44+44 demo set, not APD3-scale)
• Novelty verification gap (45-sequence refs only)
• Concrete path-to-probability-uplift table (6 actions × expected impact × timing)

docs/ROADMAP.md

• Fixed stale "29–49%" to clarify it is the uncorrected internal estimate
• Added expert review template reference to v1.0 items
• New Beyond v1.0 gap table covering 7 items needed for a major breakthrough

.github/ISSUE_TEMPLATE/expert_review.yml

• Structured GitHub issue template for domain-expert / wet-lab collaborator sign-off
• Sections: sequence review, target organism + assay plan, safety/dual-use, computational score assessment, synthesis prioritisation, overall go/no-go
• Required attestation checkboxes before synthesis order

Test plan

• 1172 passed in 0.90s — no regressions
• Pre-push CI hook passed
• Human expert reads External Calibration section and confirms tone is honest and useful
• Verify expert_review.yml renders correctly in GitHub Issues UI

🤖 Generated with Claude Code