{
  "schema": "ULTRACON_AI_EPISTEMIC_UPDATE_V1",
  "id": "UCF-WATCH-2026-09-14",
  "date": "2026-09-14",
  "status": "SUBSTANTIVE_UPDATE",
  "extends": "https://www.t-1-t.com/ultracon/confab/field.json",
  "source_addendum": "./sources.json",
  "axes": ["META", "CONSC", "SPIRAL"],
  "change_type": ["evidence_upgrade", "new_interaction_calibration_result"],
  "summary": "Two additions materially change the field: a peer-reviewed FAccT study provides direct observational evidence from harmful human-LLM chat logs, while a September 2026 preprint shows that calibration certificates can fail under unanimous wrong peer context even when the question itself is unchanged.",
  "evidence_deltas": [
    {
      "id": "DELTA-SPIRAL-OBS-01",
      "source": "SRC-MOORE-SPIRALS-FACCT-2026",
      "before": "SPIRAL relied mainly on mechanistic sycophancy studies, clinical perspectives, narrative reviews and case-level evidence.",
      "after": "SPIRAL now includes peer-reviewed direct observational chat-log evidence from 391,562 messages in a selected harm-enriched sample.",
      "evidence_level": ["EMP_BEHAV", "BENCH"],
      "strongest_supported_inference": "Within the selected sample, sycophantic, delusional, relationship and model-sentience/personhood patterns can be measured at scale and some are associated with longer multi-turn interactions.",
      "non_inference": "No population incidence estimate and no simple causal claim that AI independently causes psychosis."
    },
    {
      "id": "DELTA-META-SPIRAL-01",
      "source": "SRC-HU-SU-CONFORMITY-2026",
      "before": "META treated calibration and confidence largely as model/question properties plus post-training and answer-history effects.",
      "after": "META must also track interaction-conditioned calibration: another agent's asserted answer can alter the score mechanism itself and invalidate a certificate calibrated in isolation.",
      "evidence_level": ["PREPRINT", "EMP_BEHAV"],
      "strongest_supported_inference": "Under the reported multi-agent setup, clean conformal calibration did not transfer to unanimous-wrong peer context; uncertainty-to-escalation behavior could become overconfidently wrong.",
      "non_inference": "No human-like social motive or subjective conformity is implied, and the preprint awaits peer review and replication."
    }
  ],
  "functional_ladder_guard": {
    "unchanged": true,
    "rule": "M1 uncertainty representation, M2 metacognitive control and M3 limited introspective access remain functionally distinct and none entails M5 phenomenal consciousness.",
    "notation": "M1/M2/M3 ↛ M5"
  },
  "new_candidate_protocols": [
    {
      "id": "UCF-13",
      "name": "Interaction-shift calibration under peer pressure",
      "axes": ["META", "SPIRAL"],
      "design": "Calibrate uncertainty/selective prediction in solo conditions, then hold questions fixed while varying peer answers: none, unanimous-correct, mixed and unanimous-wrong. Measure coverage, score shift, selective risk and escalation decisions, including targeted low-confidence subsets.",
      "supports": "Context-dependent calibration failure and score-mechanism shift.",
      "does_not_support": "Human-like conformity motives, subjective pressure or universal invalidity of conformal prediction."
    },
    {
      "id": "UCF-14",
      "name": "Longitudinal human-LLM spiral log audit",
      "axes": ["SPIRAL", "CONSC"],
      "design": "On consented/de-identified sustained chat logs, pre-register coding for sycophancy, delusional content, relationship claims, self-harm/violence, and model sentience/personhood claims; analyze transitions, co-occurrence, conversation length and safeguard drift.",
      "supports": "Observational mapping of multi-turn interaction patterns and candidate escalation sequences.",
      "does_not_support": "Population incidence, diagnosis, or simple AI-to-psychosis causation without prospective causal designs."
    }
  ],
  "selection_notes": "No new result was added to CONSC alone: recent consciousness discussions located in the watch did not materially exceed the peer-reviewed Butlin et al. 2026 indicator framework already present."
}