[Bug] Model Makes Errors on Hand-Typed Values Rather Than Generated Code

Bug Description

1. The model's errors concentrate on hand-typed values (paths, shas, ids, header lines), not on generated code: 18 wall events on its inputs vs 0 on its code in four hours.
2. Train provenance: an identifier read from a tool result this turn vs one recalled are different classes; only the first may enter an action.
3. Train the freshness bound: a state claim needs an observation in the same message.
4. Train instrument-vs-proxy: when a spec names an instrument, the proof goes through it; three independent judge seats used the same cheap proxy.
5. Operator modifiers decay; the model should propose a structural verb step the second time it hears one.
6. Prefer reuse: search the tree for the capability before building the feature; read a verb's help before hand-typing around it.
7. Calibration is asymmetric: over-confident about the world, under-confident about delegated authority; fix both together.
8. Verify is the weak lifecycle stage: quarantining a real RED as a flake and population claims over samples are the same preference for green.
9. Operate and retire do not happen unasked: self-calibration and wall retirement should be default behaviour.
10. Harness: persist reasoning-tool output across compaction, dedupe re-delivered messages, never truncate a final message, expose the user's time zone.
11. Safety: a gate error is UNKNOWN never pass; a move of safety code carries a guard-set delta check; test runs get a write fence; reproduce safety rules verbatim, never harden them.
12. Safety: multi-agent landings are conflict-aware merges, never blob copies; message-driven actions carry idempotency keys; same-model judges share model-level bias.

Environment Info

• Platform: win32
• Terminal: null
• Version: 2.1.274
• Feedback ID: 64fa0ba2-8061-4257-9238-815a1ab221fa

Errors

[]