Training Data

Roadmap

This training-data design is intent for the planned SwarmFixer pipeline. No SwarmFixer corpus exists today.

Every successful repair would generate two training examples:

1. Repair-model pair — failure → repair directive.
2. Agent pair — original input + repair directive → correct output.

Both would feed downstream training corpora. This is the honest extension of the real flywheel doctrine — eval verdicts feeding training corpora — but a SwarmFixer corpus of these pairs does not exist yet; only the upstream eval verdicts are real and live.

---

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