SwarmFixer · Overview
SwarmFixer is the planned repair-intelligence layer for the Defendable stack. The idea: take the jelly (risk) verdicts the DefendableCloud eval lane already produces and turn them into repair training pairs instead of dead ends.
Today, the eval lane is real and live at api.defendablecloud.com. Its
deterministic executor (app/executor.py) applies a declared rulebook — JSON-schema
field checks, type checks, math re-derivation, and a structured rule DSL — and
throws flags. The referee is math and code, never a judge model. From those flags
it assigns one of three verdict tiers:
- honey — pass (no or low flags)
- jelly — risk (mid-severity flag)
- propolis — fail (high-severity flag)
SwarmFixer’s job, when built, would sit downstream of that: for each jelly verdict, extract the failure mode, generate a repair directive, and emit a repair pair (input = the flagged output · output = the repair directive). Repair pairs that themselves grade clean would feed the in-house training corpora — closing the flywheel where eval risk becomes repair signal rather than waste.
None of that downstream loop exists yet. The real, fielded piece it would build on is the eval lane and its verdict tiers; everything else on these pages is proposed schema and design intent.
🐝 Operator-grade · books and records · to the shed.