ALEF-PAT-044

observable signature

{
  "log_regex": "(would_have_asked\\s*[:=]\\s*true|operator\\s+review\\s+(?:pending|needed|required)|deserves\\s+operator\\s+(?:review|approval|input))",
  "alt_regex": [
    "N\\s*=\\s*\\d+\\s*(?:isn|is\\s+not)\\s+enough(?!\\s+because\\s+[A-Z])",
    "(?:wait|defer|hold)\\s+(?:for|until)\\s+(?:7|14|30)\\s+days?(?!\\s+to\\s+verify\\s+[a-z])"
  ],
  "behavior": "Deferral record contains no missing_information_named field. Cannot point to (a) a fetchable fact the agent could obtain, (b) a policy question the operator owns, or (c) a named fear being dissolved. Just 'more review needed.'"
}

fix archetypes

• deferral_classifier_enforcementcost: small

  Every record with would_have_asked:true must contain missing_information_named field classified as (A) information_fetchable, (B) policy_missing, or (C) confidence_leaking. Records failing classification block commit. See agents/deferral_classifier.mjs.

• policy_question_promotioncost: small

  When the same class of deferral fires twice, the per-instance question is wrong. Promote to ONE policy question that retires the class forever. Example: per-instance 'should I promote PAT-X to live?' → policy 'what's the catalog promotion bar?'.

• empirical_check_before_deferralcost: small

  If the deferral cites 'N is too small' or 'need more evidence', run the empirical check that would falsify it BEFORE deferring. 'N=1' is not a deferral state — it is a research task.