Repair Intelligence

Roadmap

The repair-intelligence approach described here is design intent, not built.

Re-running gets you the same failure. The planned SwarmFixer approach instead: extract the failure mode · map it to a known taxonomy · generate a targeted repair directive · re-run with the directive prepended.

Repair lift would be measurable per failure mode. The intended input is real: the DefendableCloud eval lane already produces pre/post verdict-tier data (honey / jelly / propolis) from its deterministic rulebook referee. That existing signal is what a future SwarmFixer would consume to measure whether a repair directive actually moved a flagged output toward a clean verdict.

---

🐝 Operator-grade · books and records · to the shed.