self-improve

MARKET-SCANNER — Self-improve

The discovery half of the marketplace's self-improving loop. A scan that approved a candidate it should have killed is the signal: the fix is a better-articulated parameter or kill-threshold, fixed once, upstream — not a louder rule. (Covenant: ../../knowledge/covenant.md.)

Feedback intake (cross-project learning)

The scanner gets sharper from two sources:

1. Outcome feedback — an opportunity it approved turned out weak (shipped, no one paid; the channel never materialised; an incumbent crushed it). This arrives as an ideation-feedback entry (symptom → which parameter was too weak → what would have caught it), emitted by FOUNDRY's LEARN station / FOUNDER when the conveyor learns the truth.
2. Recurring survival — the same kind of weak candidate keeps surviving the scan (flagged by the market-scan skill).

The loop

1. Reflect on the target element (a parameter row, the scoring rubric, a skill) against the covenant
   • pillars: is a kill-threshold too lax? is a parameter missing? does a skill do more than one thing (→ cleave it)?
2. Decide: sharpen a parameter / add a kill-threshold / add a kill-ledger ANTI-PATTERN (../../knowledge/discovery/scoring.md) / cleave / reference- not-restate. If the element is already tight, say so and stop.
3. Apply on a branch (surgical, one concern).
4. Adversarial gate: run /foundry:pr-review if the foundry plugin is installed; otherwise self-review against the covenant and state that the gate ran in reduced form.
5. Open a PR for the human to merge (never self-merge), so the improvement reaches every user. Follow the merge-governance the user has set (FOUNDRY's pr-approval default).

Each pass must leave the scanner measurably better at killing weak ideas early — at least halving the remaining distance. Record the recurring defect class as a kill-ledger ANTI-PATTERN so the next like candidate dies on sight.