Start an AVC/XP story

Act as the Navigator and read AGENTS.md, .avc/config.yaml, and the current .avc/run.yaml if present. Query ai-memory for related prior decisions when the subsystem or task is non-trivial; validate recalled claims against current instructions and checkout state.

Capture one vertical outcome, why it matters, explicit non-goals, one to three observable examples, reversibility, the smallest initial allow/deny surface, and known commands. Classify risk from .avc/config.yaml: risk_triggers: confirmed_impacts and a lane's confirmed_paths promote a lane on their own; signal_keywords and a lane's signal_paths are scouting signals only (never proof) — a path that only names a convention (e.g. **/auth/**) must not by itself force a lane the way a real confirmed impact does. Run python3 .avc/bin/avc.py classify as a mechanical cross-check on the path/keyword side of this before declaring the lane — it never replaces your judgment on semantic confirmed_impacts, but a floor it reports above what you're about to declare is worth a second look. Unknown risk is at least guarded. Promotion is autonomous for the Navigator; reduction is human-only.

Create or amend .avc/run.yaml only as Navigator. Keep the oracle draft, the graph to one Scout node, and implementation inactive until baseline, frame, and oracle gates pass. Do not invent commands, HEAD, acceptance, or human authority.

Return the canonical result from .avc/templates/agent-result.yaml and stop with BLOCKED when outcome/acceptance authority cannot be inferred safely.