{
  "schema":"ULTRACON_AI_EPISTEMIC_SOURCE_ADDENDUM_V1",
  "date":"2026-09-17",
  "sources":[
    {
      "id":"SRC-ASHUACH-ACL-2026",
      "year":2026,
      "title":"Masked by Consensus: Disentangling Privileged Knowledge in LLM Correctness",
      "authors":"Tomer Ashuach, Shai Gretz, Yoav Katz, Yonatan Belinkov, Liat Ein-Dor",
      "venue":"ACL 2026 Long Papers",
      "peer_reviewed":true,
      "doi":"10.18653/v1/2026.acl-long.483",
      "url":"https://aclanthology.org/2026.acl-long.483/",
      "axes":["META","CONFAB"],
      "evidence":["EMP_BEHAV","MECH"],
      "reported_anchor":"Across three similar-sized model families and five datasets, self-state probes show no general advantage over peer-model probes on the full evaluation set. On model-disagreement subsets, however, self-representations contain domain-specific privileged correctness information for factual tasks, while no consistent advantage appears for mathematical reasoning. The factual advantage emerges from early-to-mid layers onward.",
      "role":["privileged correctness information","peer-model control","consensus confound","domain asymmetry","layer localization"],
      "status":"peer_reviewed_acl_long",
      "non_inference":"An external probe extracting privileged information from a model's hidden states does not establish that the model itself can read, report or use that information introspectively. Privileged representation is not privileged self-access."
    }
  ],
  "non_inferences":["Hidden-state privilege is not equivalent to introspection.","Disagreement-subset advantage is domain-specific and does not generalize automatically from factual memory to reasoning.","Probe accessibility for an external researcher is not evidence of endogenous metacognitive readout.","M1/M2/M3 ↛ M5."]
}