binoc

AI "fuzzing" for binoc. A naive idea-generation subagent invents dataset-change scenarios from a bare description of what binoc is for — deliberately blind to binoc's actual formats, rules, and test vectors — then the orchestrator builds each scenario, runs binoc, and judges whether the output is good or reveals a bug or missing feature. Use to find things that ought to work but nobody designed for.

Review checklist for binoc correspondence rule packs, renderers, and core changes. Verifies the invariants mechanical lints cannot: descriptor honesty, deterministic proposals, security, settled behavior, layering, projection/extract honesty, and the recurring model-to-surface failure modes that produce layering regressions.