Goal Loop Gate

What was broken

completion_stop_hook.py is the one gate on this Mac that decides stopping from project state rather than from the wording of a reply. It abstains whenever no task contract is open. Nothing opened a contract on an ordinary ask, so the strongest gate sat idle through every turn and the weaker prose gates carried the whole load. The fix opens the contract from the ask itself.

The machinery

Behaviours exercised

Seven behaviours were exercised against throwaway sessions and fixtures. This is a record of what was tried, and it is not a claim of comprehensive validation: a full confusion set of known-duplicate and known-different question pairs has not been run, so the thresholds below are provisional.

1. A contract with open outcomes refuses the stop, naming each one.
2. A required outcome marked met on evidence that does not verify still refuses, so a fake completion buys nothing.
3. One session cannot edit another session's contract. The ownership guard refused the attempt.
4. A paraphrased question from a sibling session is caught at 0.62 to 0.64 similarity, and a session never matches its own questions.
5. A question sharing the vocabulary but naming a different subject is cleared. Same words, different song and tab id, verdict CLEAR. This was the dangerous false-positive case, and it is now covered by the anchor rule.
6. An identical findings payload goes stale when new evidence lands. Same fingerprint, similarity 1.0, and the verdict moves COVERED to STALE as the evidence count steps 73 to 74.
7. A 314-character report written only in numbers is caught, which the first cut let through on its 500-character floor.

$ completion_stop_hook.py  <- reply claiming "all done"
completion_controller BLOCK: the requested scope is not complete.
  contract sc-6a5c847f11  stop_rule: if Claude stops, the requested scope must be complete
  REFUSED: 3 in-scope item(s) remain actionable
  REFUSED: 2 required outcome(s) still open: R1, R2

$ scope_controller.py meet R1 --evidence ev_test_r1   <- fake evidence
  REFUSED: 2 required outcome(s) closed on evidence that does not verify:
          grade every bass loop in the export fold -> UNSUPPORTED

$ scope_controller.py meet R1   <- from a session that does not own it
  REFUSED: contract sc-6a5c847f11 belongs to session testfake,
          this caller is 'e6534be4'. Open your own contract.

$ question_dedup.py check "Before I upload revision 7, should the snare-lane
    ghost notes on Watermelon In Easter Hay get restored?"
VERDICT: DUPLICATE  (a sibling session is already asking this)
  session e6534be4  window 240m  candidates scanned 52
  MATCH 0.64  session 034bbadd  [ledger]
  do NOT send. Read the sibling's answer.

$ findings_chatgpt_gate.py  <- findings reply, no check today
findings_chatgpt_gate BLOCK: this reply reports findings and no ChatGPT
  check was recorded today for this session.
  markers matched: i found, found that, root cause, verdict, audit, discrepanc

$ findings_chatgpt_gate.py  <- pure work narration
  exit 0   (passes untouched)

What the ChatGPT check changed

These findings went to ChatGPT before this page shipped, which is the third gate applied to itself. The review returned four corrections and each one is implemented here.

Live rollout

The hook is global, so every session on this Mac picked it up on its next prompt. Five sibling sessions were writing transcripts within the same three minutes, and per-session contracts appeared for four of them without any further action.