factory-tune

Constraints set once become either a permanent tax or a permanent hole. This command is the scheduled review that keeps them honest. Run it monthly, or after any escaped defect.

It proposes. It never edits CHARTER.md itself. A factory that can rewrite its own constraints has none.

Gather evidence

Look at the last 30 days (or since LAST_REVIEWED in the charter). Use immutable records under docs/factory/runs/ plus GitHub issue and PR timestamps:

1. Merged factory PRs. How many, and how many needed human fixes after merge?
2. Escaped defects. Anything that reached the default branch and later needed a fix. For each, trace which gate should have caught it. This is the single most valuable input here.
3. Rejected verifications. What did the verifier catch, and is there a pattern? A repeated catch is a candidate for a new deterministic gate, which is strictly better than catching it with a model every time.
4. Triage accuracy. Items marked ready-to-implement that turned out to need a spec. A high rate means the charter's AUTOMATABLE list is too generous.
5. Review queue depth over time. Was the human review queue the bottleneck?
6. Gate configuration. Which runs were MISCONFIGURED? Which optional gates reported SKIP, and should any become required?
7. Flow. Median queue age, implementation duration where records are complete, and time from awaiting-review to the human decision. State when records are incomplete; do not manufacture precision.

Propose

Tighten when

• An escaped defect traces to an automated gate you trusted. Say exactly which gate and what it missed. Tighten immediately; this is not a judgment call.
• A category keeps arriving in review needing real fixes.
• The verifier catches the same class of problem repeatedly. Propose the deterministic check that would catch it instead, because a gate cannot be talked out of its verdict and a reviewer can.

Loosen when