Missions leave trails. Trails that led to verified work strengthen; everything else fades. Reinforcement learned before reproducible evidence would reward persuasive prose rather than demonstrated work — so learning is gated on the canonical, evidence-backed outcome of each mission.
A role earns a positive trail when its output was consumed downstream, required evidence passed, and the mission ended completed_verified — the only positive learning signal. The Tester catching a real failure and the Soldier correctly blocking a dangerous patch are also positive signals for them.
Completed-but-unverified, partial, timed-out, and failed missions never reinforce positively — the episode is stored, without reward. A typed failure attributable to a role's output records a negative trail for that role and task type.
Cancellation is neutral, and a role is never punished for not running. Provider outages update provider reliability, and environmental tool failures update tool reliability — not the worker's skill.
Trails fade toward neutral over time. A route that stops proving itself stops being preferred; old success is not a permanent verdict.
The colony can answer: what has worked, what usually fails, who solved this before, and what knowledge exists — across restarts, from the local SQLite database.
After the canonical mission evaluation persists, the Archivist emits structured memory candidates — never before finalization, and never positively from unverified work.
Repeatable procedures can be certified and stored durably, with certification recorded atomically — rather than re-derived from scratch each mission.
Typed artifacts and evidence records carry producer, provenance, and content hashes, so a verified outcome can be traced back to the exact bytes that earned it.
There is no single global worker score. Trails are keyed by role, worker, task type, capability, environment fingerprint, tool or source domain, and contract version.
Role standing is computed from trails, but nothing routes on it yet — the router still picks by configuration. Reputation-aware routing is planned, not shipped.
A trail is designed to influence routing only after a configurable minimum number of observations — no single lucky mission redefines a route.
Learning starts at full-roster activation. Legacy unverified completions are never backfilled as positive signals.
Sources: PLAN.md (Stage E, gap table), ANT_EXECUTION.md (outcome semantics), and ADR-004 (artifact store).