evolvelisted

## What This Does Synthesizes diaries + failure ledgers + learnings into vetted governance proposals through a cost-sequenced 4-reviewer panel. Supersedes v1 `extract-lessons` and `self-learning-review`. NEVER auto-applies anything — the human gate is non-negotiable. ## Cost-Sequenced Protocol **Step 0 — Pre-filter (Tier 0).** Run `core/scripts/tools/evolve-signal.sh`. If it prints `NO-SIGNAL`, STOP and emit `[EVOLVE] no signal (<n> new) — nothing to do`. With `--dry-run`, continue using fixture inputs but do NOT spawn subagents. **Step 1 — Evidence reviewer (Haiku, solo first).** Spawn ONE Task (`model: "haiku"`) with the KERNEL prompt below to cluster the NEW signals into candidate patterns, each with an occurrence count and a `PASS`/`NOISE` verdict. If every pattern is `NOISE`, STOP and emit `[EVOLVE] all signal noise`. **Step 2 — Fit / Cost / Risk panel (Haiku, one batch ≤ Tier 2 cap).** For the `PASS` patterns only, spawn three Tasks (`model: "haiku"`), each ≤2000 tokens: - **Fit:** verdict ∈ {ADD, MODIFY, MERGE, PRUNE, REJECT}, target rule/skill id. - **Cost:** activation mode ∈ {always, intelligent, scoped}, always-on budget delta vs 6k. - **Risk:** {OK, OVER-TRIGGERS, NEEDS-SCOPE} + one-line justification. **Step 3 — Aggregate (orchestrator, no extra subagent).** Combine the four verdicts per pattern into a proposal. Drop any pattern where Fit=REJECT or Risk=OVER-TRIGGERS. **Ledger dedupe (Tier 0):** before writing, check the outcome ledger — `rg -c $'\t<verdict