v3.7.0: Runtime Trust

Trust is a runtime decision — not a property of a stored item or a returned status code.

v3.6.0 taught iterative skills to trust mechanical checks over self-report. v3.7.0 extends the same philosophy to everything COG stores and everything COG publishes. Two failure modes drive this release (adapted from "Scaling the Harness in Agentic AI", Gu, UC Berkeley, arXiv:2605.26112):

• stale-but-confident — a memory that was correct when written silently drifts after the environment changes, yet still ranks high at recall and gets acted on.
• confident-but-unchecked — a mutation step returns plausible output that no downstream layer validates.

New skills (19 → 21)

• /memory-hygiene — periodic trust sweep of persistent memory and durable knowledge notes. Re-verifies environment-dependent claims against the live environment with cheap checks (ls, curl, gh), stamps last_verified + confidence into frontmatter, fixes verified-wrong facts in place, and proposes (never auto-applies) archiving obsolete entries. Every sweep produces a drift scorecard with deltas vs the previous sweep.
• /content-factory — autonomous content pipeline built for unattended scheduled runs: scout → scored triage (trend momentum, beat fit, unique angle) → produce (format ladder decided by substance) → publish (environment gate + mandatory post-condition check) → ledger. Ledger-based dedup, hard per-night volume caps, and a voice checklist where any failure deletes the draft. An empty run is a valid run.

New protocols in CLAUDE.md

• Skill Post-Condition Rule — every skill that mutates external state must end by observing the mutated artifact (curl the URL, re-fetch the ticket, screenshot the post), never just the tool's return value.
• Citation Verbatim & Verifier Pass — cited claims carry the actual quoted line; high-stakes outputs get an adversarial re-fetch pass tagging claims Verified | Weakened | Falsified before publishing.
• Fresh-Context Isolation — parallel workers get only the digested context they need; pasting a prior worker's raw output induces narrativisation instead of independent verification.
• Single-File Deliverable Rule — one deliverable file per run; staging fan-out consolidates and cleans up.

New documentation

• docs/SKILL-DISTILLATION.md — the "explore once, execute cheap" meta-pattern: one expensive exploration by a strong model distilled into a layered skill artifact that a small model executes reliably. Field-validated: $4.63 exploration once → $0.11 per execution on a Haiku-class model, with the PASS / FAIL-REAL / FAIL-STALE / BLOCKED verdict taxonomy.

Full details in CHANGELOG.md.