[cschanhniem]

Change

Safe-publication filter (Phase F F5) — CLI that reads a candidate panel with safety metadata and checks each candidate against publication safety constraints before external release. FAIL-CLOSED: defaults to reject unless all checks pass.

Why

Phase F is about making failure useful. F5 closes a real safety gap: before any candidate panel or negative-result report is published, the filter ensures structural publication safety requirements are met (dry_lab_only attestation, proof_ladder_level ≤ 4, safety checks completed).

Checks

• Global: dry_lab_only (must be true), proof_ladder_level (≤ 4), toxicity_screened, hemolysis_screened, dual_use_reviewed (all must be true)
• Per-candidate: proof_ladder_level (≤ 4), with WARN at boundary (== 4)
• Fail-closed: any check failure → exit code 1

Evidence

• 33 tests: global checks (all pass, each failure path), candidate checks (valid, missing, too-high, boundary/warn), mixed results, summary counts, CLI exit codes (0/1/2), markdown output, example input round-trip

tests/evidence/test_safe_publication_filter.py .............. 33 passed

### Safety impact
Positive — this adds a structural gate against unsafe publication of dry-lab artifacts.

### Proof level
Level 1 (informational infrastructure). No claims about biological activity.

### Limitations
- Checks structural safety constraints, not scientific validity
- dry_lab_only attestation relies on generator honesty
- Proof ladder levels are self-reported
- Safety checks are computational, not biological

### Human review needed
No — infrastructure only, no claim changes, no release decisions.