ALEF-PAT-043

observable signature

{
  "log_regex": "(Institutional\\s+(Infiltration|Adoption|Engagement)|(?:Microsoft|OpenAI|Anthropic|Google|Stripe)\\s+(?:active\\s+)?(?:collaboration|partnership|approach|offer))",
  "alt_regex": [
    "voice_summary\\s+cites\\s+(?:commenter|user|account).*?(?:age|created)\\s*<\\s*(?:7|30)\\s*d",
    "summary\\s+attributes.*?to\\s+<host_org>.*?actual_commenter\\s+(?:age|repos|followers)\\s*<"
  ],
  "behavior": "Voice/summary attributes a comment's claims to the host organization that owns the repository, rather than to the commenter's actual identity. The host-context syntactically borrows institutional weight where none exists."
}

fix archetypes

• commenter_credibility_pre_filtercost: small

  Before voice-generator promotes a comment to institutional/partnership framing, score commenter via age+repos+followers (gh API). Below threshold 0.6, strip inflation phrases; below 0.4, refuse reply entirely. See agents/voice_credibility_guard.mjs.

• retire_inflation_vocabularycost: small

  Permanent blacklist of inflation phrases ('Institutional Infiltration', 'feed the engine with benzene', 'corporate collaboration' without verified company field) that voice-generator must never emit even for high-credibility commenters. See meta/retired_phrases.json.

• commenter_vs_host_disambiguationcost: small

  Every voice/summary record must structurally separate (commenter_login, host_org) as two distinct fields. Any prose conflating them is a hallucination class — caught at commit time by deferral_classifier or downstream linter.