verify

Run an independent audit. Target (may be empty — then audit this session's active goal, or if none exists, the most recently archived goal):

1. Resolve the goal file: active goal → this session's .ultragoal/goals/active/<slug>/goal.md (or the legacy active.md); a slug → .ultragoal/goals/active/<slug>/goal.md or .ultragoal/goals/archive/<slug>.md; empty + no active goal → newest file in archive/.
2. Do not pre-investigate, summarize the work, or characterize the goal first — anything you add becomes context that anchors the auditor. Dispatch the ultragoal:verifier subagent immediately, telling it only the goal file path and that this is a post-hoc audit: re-run every rubric check command, test every constraint, and append its verdict to the goal file per its protocol.
3. Relay the verdict to the user verbatim-faithfully: verdict first, then the per-item findings exactly as the verifier reported them. Do not soften failures or add reassurance the verifier didn't give.
4. If the audit FAILS on a goal that was marked done: say plainly which claims didn't hold, and offer two paths — reopen the goal (move it back to active.md, set status: active and session: to this session, uncheck the failed items) or record the gap as a known issue in .ultragoal/memory/failures.md.
5. If it fails on an active goal, the loop self-corrects — the gate will keep the session working; just make sure the failed items get unchecked.

For maximum isolation on high-stakes work (the in-session verifier still shares your Claude Code process and permissions), tell the user they can run the audit as a completely separate headless session from their terminal:

claude -p "Use the ultragoal:verify skill to audit the goal '<slug>'" --permission-mode acceptEdits

A verifier in a fresh process has seen none of this session's narrative — the strongest fresh-eyes guarantee available.