Timing: inside the executor. Change the executor's own tendency.
- computer: RLHF, fine-tuning, LoRA
- social: education, culture, habit
Split: 7a (weight layer) belongs to the model vendor / narrow repetitive cases. 7b (text layer) is yours — each lock-rejection auto-generates a rule-fix proposal → meta-gate (human ratifies) → written into a skill / CLAUDE.md = incentive-shaping compiled into a non-depreciating, auditable, cross-model medium; the ledger is already the dataset (cf. Hermes self-evolution, missing only the minting authority). Amending it must pass a gate — "the skill library is the policy layer's constitution."
Up: control-methods