/evolve — Goal-Driven Autonomous Loop

Measure what's wrong. Fix the worst thing. Measure again. Whether the fixes compound into a durable knowledge moat is a tracked hypothesis, not a promise — DEMOTED to unproven by ADR-0004 and ADR-0011. The proven product is the per-cycle verification (no verdict = not done), not the compounding. Do not market the flywheel ahead of the ruler.

Experimental tier. Autonomous long-loop; run attended or dispatched onto a substrate, never as an in-repo daemon (ADR-0009).

Cadence is pawl-gated, not per-tread (docs/contracts/pawls.md). Each cycle's heavy validation (pawl review, /validate --mixed, /pre-land-refuters) fires ONCE at the cycle's bead-acceptance / land pawl — not per slice or wave. The per-cycle regression gate (Step 5) is chaos: cheap and wrong-tolerant between pawls. Do NOT escalate every cycle to a cross-family panel "to be safe" — that re-creates the waterfall the ratchet avoids.

The loop runs as this skill. evolve selects work and invokes complete /rpi --auto cycles — that is the loop. Each cycle's post-mortem checkpoint is a re-plan point (re-scope / reorder / drop / add to the remaining queue from what the cycle taught), one altitude up from /rpi's agile re-plan loop — agile across cycles, not a fixed backlog. Substrates dispatch the whole loop as one unit through NTM, Agent Mail, or ao agent; the former RPI CLI wrappers are retired (ADR-0009).

Operator cadence: post-mortem finished work → measure repo state → select the next highest-value item → let rpi run research → plan → pre-mortem → implement → validate → harvest follow-ups → repeat until a kill switch, max-cycle cap, regression breaker, or real dormancy stops it.

Work selection ladder

Selection is a ladder re-read from the TOP after every productive cycle — never a one-shot check. Full per-rung procedure (ao loop next-work recommendation, scope filter, generator code, --quality cascade, dormancy hard-gate): references/work-selection-ladder.md.