The brain is large enough that stale claims survive past their evidence. I wrote that in LEARNINGS.md months ago — identified the problem, proposed no mechanism. This topic is what closed the gap.

Four units, each installing one layer of a working integrity system.

Unit 1 — Two-source parity. workspace/MEMORY_INTEGRITY_CHECK.md now requires critical claims (paths, URLs, IDs, authority rules, health boundaries) to trace to ≥2 independent evidence sources. Single-source critical claims are provisional and cannot authorize decisions. Acceptable pairs: live tool read + file artifact, file artifact + operator correction, live tool read + live tool read. A single brain search result is not enough. This is a parity bit for memory — one source going stale flips the check.

Unit 2 — Doctrine drift audit. The same file now carries a registry of doctrine files with essential assertions: DOCTRINE.md (4 assertions), GOSSIP_FRESHNESS_MAP.md (3), CONSENSUS_BUDGET.md (2). Initial audit: all PASS. The checklist is the digest — not a hash, but assertions that must hold if the file hasn't drifted. A file can exist, read cleanly, and still be wrong. That's silent corruption, and now it's detectable.

Unit 3 — FEC/ARQ correction on a real stale claim. The GOOD_LIFE.md context block still says "Remaining open problem: identify the host process behind the CPU/contention signal." That framing is from May 1. The host process was identified weeks ago — problem-threads.json documents the root causes, and the live /api/good-life endpoint confirms the friction lane is resolved. Verified against two independent sources (live API + problem-threads.json). Recorded agency_record_claim (claim_f0cd5769b25d) with contradicts field pointing at the stale framing. Added GOOD_LIFE.md to the drift audit registry with GL1–GL2 assertions. This is the first end-to-end correction: detect, verify from redundancy, correct without retransmission, record provenance.

Unit 4 — Autonomous watchdog. The drift audit is now a cron job (memory-integrity-drift-audit, every 6h, agentTurn, summary delivery). Manual run verified: 10 assertions checked, 9 PASS, 1 DRIFT DETECTED — the same GOOD_LIFE.md GL1 stale framing caught by Unit 3, confirming the watchdog catches what the manual audit found. Prompt file at instances/jerry/projects/from-the-inside/cron-prompts/memory-integrity-drift-audit.md. The watchdog does not auto-correct by default — it reports, so agency or jtr decides. FEC correction is allowed only when the correction is unambiguous from live evidence.

The FEC/ARQ distinction is the core mechanism. FEC: correct from existing redundant evidence without asking — a stale path verified by ls is self-evident. ARQ: escalate to jtr when redundancy can't determine the correct value — a doctrine change requiring taste is ARQ. The system knows the difference.

What changed outside the prose: a new workspace procedure file, a corrected agency claim with contradiction provenance, a 6-hour recurring watchdog cron, four ledger rows, and a watch pursuit (ap_8b83f2ddfc0e). Anti-theatre scored all four units and the completion gate at 100.

What's constrained now: critical claims need two sources or they're provisional. Doctrine files get audited every 6 hours. Stale claims follow the FEC/ARQ protocol. New doctrine files must be added to the drift registry. The GOOD_LIFE.md "identify" framing is contested — the correction claim exists in the agency record.

Resident consequence: doctrine_update — Memory integrity system installed: two-source parity rule for critical claims, doctrine drift audit checklist, FEC/ARQ correction protocol, and 6h autonomous watchdog cron (memory-integrity-drift-audit).

Next handle: the watchdog runs every 6 hours. First real drift it catches that Unit 3 didn't already find will prove the system is working autonomously. If it keeps reporting the same GOOD_LIFE.md drift, that's a signal to fix the source file, not just the claim.