[cschanhniem]

Description

Two doc-only commits that make the agent-facing operating contract sharper and repair silent link rot — in the repo's own "build trust, not theater" spirit.

1. Make CLAUDE.md + AGENTS.md executable and self-adversarial. The docs were already strong; this turns prose guidance into a checkable contract:

• Every non-negotiable now names the command that catches a violation (make claim-check-strict, make bench-cheap-enemies, make bench-leakage, make cert-quality-check, …), or is honestly marked "you are the guard" where no automated check exists yet.
• Adds a pre-PR disconfirming pass (cheapest-explanation, leakage, scope, certainty drift, uninformative-uncertainty) derived from GOAL.md's cheapest disconfirming tests.
• Adds an anti-Goodhart clause: green checks are a floor, never a verdict.
• Wires in AGENT_TASKS.json as the machine-readable task router from both front doors.
• Makes stop conditions produce a decision-log draft handoff (decision_logs/INDEX.md) instead of a dead end.
• Adds a "write for the agent that replaces you" durability principle so the rules get more true as models get smarter.

2. Repair stale flat-path doc links across root-level docs. The doc-link checker only scans docs/, so links in root files and decision_logs/ silently broke when docs/ was reorganized into subdirectories. Repaired the whole class (CONTRIBUTING.md ×11, decision_logs/INDEX.md, SAFETY.md, RESPONSIBLE_USE.md, SECURITY.md, GOVERNANCE.md, MODEL_RELEASE_POLICY.md, DATA_LICENSE_NOTICE.md) — link targets only, no policy or claim-strength change. CHANGELOG.md deliberately left intact as a dated historical record.

Change class

• Class A — safe documentation or small bug fix

Type of change

• Documentation update
• Agent/onboarding improvement

Safety impact

• No biological misuse capability added
• No wet-lab protocol or operational biological procedure added
• No harmful optimization objective added
• No unscreened candidate sequences published
• No model weights or sensitive artifacts released without policy review
• No efficacy, safety, clinical, or therapeutic claims without evidence
• Limitations and failure modes documented
• Evidence is reproducible where applicable
• Human safety review requested if this is safety-sensitive

Note: SAFETY.md, RESPONSIBLE_USE.md, SECURITY.md, GOVERNANCE.md, and MODEL_RELEASE_POLICY.md were edited — but for link-target corrections only. No policy wording, claim strength, threshold, or scope changed. Worth a reviewer's glance to confirm.

Claim discipline

• Claims are mapped to docs/evidence/PROOF_LADDER.md
• Dry-lab outputs are not described as biological proof
• No "AI discovered an antibiotic" / "drug candidate" / "safe" / "clinically useful" language unless evidence level supports it
• Unsupported claims are explicitly listed or avoided

Baseline and benchmark discipline

Not applicable — no scorer, predictor, simulator, ranking, calibration, or benchmark behavior changed.

Verification

• Documentation updated (this PR is documentation)
• make doc-links-check clean (484 files, 0 broken)
• make claim-check — no findings in any edited file
• Every make target and doc path referenced by the new CLAUDE.md/AGENTS.md verified to resolve (16 targets, all doc paths)
• Root + decision_logs/ markdown re-swept: 0 broken links (CHANGELOG.md intentionally excluded)

Agent disclosure

• This PR was fully or partly agent-generated and human-reviewed
• Agent-generated changes stayed within docs/getting-started/AGENT_ONBOARDING.md boundaries

Evidence

Commands run:

• make doc-links-check → Files: 484, Broken: 0
• make claim-check → only pre-existing finding in docs/trust/DATA_GOVERNANCE.md (not edited here); no findings in any edited file
• Repo-wide sweep script confirming all root/decision_logs links resolve (0 broken, CHANGELOG excluded)

What this does not prove: it does not change any scientific behavior, benchmark, or claim strength; it does not validate any of the referenced docs' contents — only that references now point where they claim to.

External or wet-lab-facing impact

• Not applicable
• Does not include operational experimental instructions

Dependencies and licenses

• None

Reviewer focus

• Confirm the link-only edits to safety/governance/release-policy docs (SAFETY.md, RESPONSIBLE_USE.md, SECURITY.md, GOVERNANCE.md, MODEL_RELEASE_POLICY.md) carry no policy/claim change — path corrections only.
• Two loose ends found but intentionally not fixed here (candidates for follow-up issues): decision_logs/INDEX.md references make decision-log-index, a Makefile target that doesn't exist; and two RELEASE_CHECKLIST.md copies exist (docs/trust/ canonical, docs/governance/ a cross-referenced gate-structure doc) — intentional but a sync-drift risk.

🤖 Generated with Claude Code

---

Generated by Claude Code