{
  "schema": "ULTRACON_AI_EPISTEMIC_SOURCE_ADDENDUM_V1",
  "date": "2026-09-19",
  "sources": [
    {
      "id": "SRC-REN-UA-BENCH-ACL-2026",
      "year": 2026,
      "title": "Beyond \"I Don’t Know\": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty",
      "authors": "Jingyi Ren, Ante Wang, Yunghwei Lai, Xiaolong Wang, Linlu Gong, Weitao Li, Weizhi Ma, Yang Liu",
      "venue": "ACL 2026 Long Papers",
      "publication_date": "2026-07",
      "peer_reviewed": true,
      "doi": "10.18653/v1/2026.acl-long.547",
      "url": "https://aclanthology.org/2026.acl-long.547/",
      "axes": [
        "META",
        "CONFAB"
      ],
      "evidence": [
        "PEER_REVIEWED",
        "BENCH",
        "EMP_BEHAV",
        "TRAINING"
      ],
      "reported_anchor": "UA-Bench contains more than 3,500 questions across six datasets and evaluates 18 frontier LLMs. The study reports that high answer accuracy does not guarantee reliable discrimination between input/data uncertainty and model/capability uncertainty; targeted training improves this attribution in the tested Qwen3 systems.",
      "role": [
        "uncertainty attribution",
        "data uncertainty",
        "model uncertainty",
        "abstention routing"
      ],
      "status": "peer_reviewed_primary_research",
      "non_inference": "The paper’s bibliographic term 'self-awareness' names a benchmark capability. Successful uncertainty attribution would not by itself establish privileged endogenous access, second-order metacognition, subjective awareness or M5.",
      "provenance_update": "./updates/2026-09-19-evidence-v2/sources.json"
    },
    {
      "id": "SRC-LI-ACTIVE-CALIBRATION-ACL-2026",
      "year": 2026,
      "title": "Demystifying Uncertainty in LLMs: Active Calibration between Concepts and Human Evaluations",
      "authors": "Pengqi Li, Lizhong Ding, Zhehao Zhou, Chunhui Zhang, Jiarun Fu, Hao Li, Ye Yuan, Guoren Wang",
      "venue": "ACL 2026 Long Papers",
      "publication_date": "2026-07",
      "peer_reviewed": true,
      "doi": "10.18653/v1/2026.acl-long.259",
      "url": "https://aclanthology.org/2026.acl-long.259/",
      "axes": [
        "META",
        "CONFAB"
      ],
      "evidence": [
        "PEER_REVIEWED",
        "THEORY",
        "EMP_BEHAV",
        "METHOD"
      ],
      "reported_anchor": "The paper formalizes interactive calibration under underspecified inputs, derives a non-vanishing calibration-error lower bound for its non-interactive setting, and reports that clarification selected from calibration signals can reduce error in the tested interactive regimes.",
      "role": [
        "interactive calibration",
        "clarification",
        "information gain",
        "uncertainty routing"
      ],
      "status": "peer_reviewed_primary_research",
      "non_inference": "A useful clarification policy does not establish introspective access to uncertainty, and the reported lower bound and gains should not be generalized beyond the paper’s formal assumptions and evaluated settings.",
      "provenance_update": "./updates/2026-09-19-evidence-v2/sources.json"
    },
    {
      "id": "SRC-LI-SOURCE-BALANCE-ACL-2026",
      "year": 2026,
      "title": "How Large Language Models Balance Internal Knowledge with User and Document Assertions",
      "authors": "Shuowei Li, Haoxin Li, Wenda Chu, Yi Fang",
      "venue": "Findings of ACL 2026",
      "publication_date": "2026-07",
      "peer_reviewed": true,
      "doi": "10.18653/v1/2026.findings-acl.1267",
      "url": "https://aclanthology.org/2026.findings-acl.1267/",
      "axes": [
        "CONFAB",
        "SPIRAL",
        "META"
      ],
      "evidence": [
        "PEER_REVIEWED",
        "BENCH",
        "EMP_BEHAV",
        "TRAINING"
      ],
      "reported_anchor": "A three-source framework tests parametric knowledge, user assertions and document assertions across 27 LLMs from three families. Most tested models defer more to documents than users, while often failing to discriminate helpful from harmful external information; diverse source-interaction training improves discrimination in the reported experiments.",
      "role": [
        "knowledge conflict",
        "source arbitration",
        "user assertions",
        "document assertions",
        "parametric knowledge"
      ],
      "status": "peer_reviewed_primary_research",
      "non_inference": "Parametric knowledge is not a belief state, document deference is not evidence of rational trust, and source-selection behavior does not establish introspective knowledge of why a source was preferred.",
      "provenance_update": "./updates/2026-09-19-evidence-v2/sources.json"
    },
    {
      "id": "SRC-GOURABATHINA-TRACE-INVERSION-ACL-2026",
      "year": 2026,
      "title": "Answering the Wrong Question: Reasoning Trace Inversion for Abstention in LLMs",
      "authors": "Abinitha Gourabathina, Inkit Padhi, Manish Nagireddy, Subhajit Chaudhury, Prasanna Sattigeri",
      "venue": "ACL 2026 Long Papers",
      "publication_date": "2026-07",
      "peer_reviewed": true,
      "doi": "10.18653/v1/2026.acl-long.608",
      "url": "https://aclanthology.org/2026.acl-long.608/",
      "axes": [
        "META",
        "CONFAB"
      ],
      "evidence": [
        "PEER_REVIEWED",
        "METHOD",
        "EMP_BEHAV"
      ],
      "reported_anchor": "Trace Inversion reconstructs the likely query from a model-generated reasoning trace and uses mismatch with the actual query as an abstention signal. The paper reports gains over competitive baselines in 33 of 36 evaluated settings across four frontier LLMs and nine abstention QA datasets.",
      "role": [
        "abstention",
        "reasoning-trace diagnostic",
        "query misalignment"
      ],
      "status": "peer_reviewed_primary_research",
      "non_inference": "A reasoning trace can be diagnostically useful without being a faithful causal record, privileged hidden state, endogenous introspective report or evidence for M5.",
      "provenance_update": "./updates/2026-09-19-evidence-v2/sources.json"
    }
  ],
  "non_inferences": [
    "Bibliographic use of 'self-awareness' does not establish subjective self-awareness.",
    "Data-uncertainty attribution and model-uncertainty attribution are behavioral capacities, not direct evidence of privileged access.",
    "Reasoning traces can be useful abstention signals without being faithful introspection.",
    "Source arbitration behavior does not imply beliefs or rational trust.",
    "M1/M2/M3 ↛ M5."
  ]
}