Communicator Training

Roadmap — not yet built

The Communicator and SwarmFixer training loops described here are designed, not yet built in the audited codebase. This page documents the intended tiering. (See the overview for full status.)

Designed routing

StreetChat is designed to route each graded street pair by its verdict tier:

• Honey / Royal Jelly → feeds the next Communicator training cycle. These are the clean, client-ready pairs.
• Jelly → routed to SwarmFixer as repair-training data — risk-flagged but recoverable.
• Propolis → held for root-cause review. These do not enter training; they are studied for the failure pattern first.

This tiering maps directly to the verdict tiers that ARE real in DefendableCloud (defendable-cloud-v2): honey = pass, jelly = risk, propolis = fail. The classification a pair carries here is the same rulebook vocabulary the Cloud referee applies — math and code, not a judge model.

The training step itself is designed to be operator-extensible to any HuggingFace-compatible training stack.

---

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