0xPanty/misa-cybernetic-evolution

Misa Cybernetic Evolution is a control-theoretic "sidecar" for long-lived AI agents that prevents agents from judging or rewriting their own evolution. Inspired by Qian Xuesen's Engineering Cybernetics, it enforces a hard separation of roles: a control-layer distills redacted work windows into evidence, Hermes provides observe-only runtime and model-I/O taps, a measurement-quality gate cross-checks signals, L2/L3 paths validate replay and routing, and a human boundary retains durable authority. Who uses it: system builders, safety engineers, and operators who need auditable, deterministic learning planes for agents that learn from real work without gaining self-authority. Key features: deterministic routing for memory/skills/policies; observe-only model I/O and action monitoring; replay-gated work orders and seeded variants; local vector-store and retrieval-lineage contracts; runtime pause/resume without live effects; measurement-quality diagnostics; schema, tests, and value-proof commands. v0.28.0 adds a measurement-cleanliness gate so candidates are judged only if measurement is trusted.