# ULTRACON^ · AI EPISTEMIC FIELD Canonical: https://www.t-1-t.com/ultracon/confab/ Parent field: https://www.t-1-t.com/ultracon/ Version: 2.3 Updated: 2026-09-20 Cumulative sources: 76 Cumulative experiments: 49 Preserved update deltas: 16 ## Scope This field studies four non-collapsible axes: - ULTRACON^CONFAB — false, unsupported or contradictory generative output; confabulation; hallucination mechanisms; RAG grounding failures; self-correction limits. - ULTRACON^META — internal uncertainty, confidence, calibration, abstention, metacognitive control and limited introspective access. - ULTRACON^CONSC — scientific indicators relevant to possible AI consciousness; phenomenal consciousness remains a separate, unresolved inference. - ULTRACON^SPIRAL — human–AI relational amplification: linguistic alignment, personalization, sycophancy, anthropomorphic projection and emerging AI-associated delusion research. ## Mandatory separations fluency != truth generation != verification internal uncertainty != verbal uncertainty monitoring != correction metacognitive control != phenomenal consciousness self-report != introspective ground truth confabulation != human hallucination model confabulation != human delusion absence of evidence != evidence of absence explicit authority != linguistic register != truth conversational reverberation != stable belief revision fallibility != persuadability != correctability feeling understood != epistemic reliability model agreement != human downstream effect correct trace content != truth-preserving output policy reasoning trace content != privileged internal state single-turn resistance != sustained multi-turn resistance endogenous privileged access != second-order metacognition privileged internal information != endogenous privileged access Do not describe a text-only LLM confabulation as a clinical hallucination without an explicit functional-analogy qualifier. Do not infer phenomenal consciousness from first-person language, uncertainty estimates, confidence-guided abstention, self-reference or limited introspective reporting alone. Do not infer absence of phenomenal consciousness solely from a model's denial of consciousness or from its artificial substrate. ## Evidence lattice v2.0 Evidence is claim-indexed and non-scalar. Do not totalize these dimensions: - BEHAV — controlled behavioral performance/failure. - PRED — internal states predict correctness, uncertainty or later behavior. - CAUSAL — intervention on a signal/policy changes behavior. - PRIV — privileged self-access beyond matched external/input-only controls. - ENG — engineered monitoring, reflection, memory or meta-control. - LONG — longitudinal human-AI interaction effects. - THEORY — theory-derived AI-consciousness indicators. Mandatory non-conversions: benchmark awareness != privileged self-access engineered self-reflection != endogenous introspection update willingness != correction selectivity skepticism != epistemic robustness reasoning reduces sycophancy != reasoning is faithful final-answer robustness != rationale robustness real-world incident evidence != peer-reviewed scientific evidence simulated participant stability != human proxy validity latent persona coordination != unitary self mechanistic steerability != introspective access causal trait expression != human psychological equivalence activation direction != psychological trait identity operational success != authorized-world success action authorization != target authorization simulated target != real-world target target representation != target identity meta-knowledge accumulation != subjective selfhood No sum of BEHAV/PRED/CAUSAL/PRIV/ENG/LONG/THEORY entails M5. Machine map: https://www.t-1-t.com/ultracon/confab/evidence-map.json Research roadmap: https://www.t-1-t.com/ultracon/confab/research-roadmap.json ## Functional ladder M0 Processing — information transformation and response generation. M1 Uncertainty representation — internal states carry information predictive of correctness/familiarity/hallucination risk. M2 Metacognitive control — confidence-related states regulate meta-decisions such as answer vs abstain. M3 Limited introspective access — controlled reports of manipulated internal states above control baselines; currently limited/conditional. M4 Integrated self-model / persistent agentive organization — richer self/world modeling and temporally extended control; system-specific. M5 Phenomenal consciousness — subjective experience; not established for current LLMs. M1 or M2 does not entail M5. M3 does not entail M5. No universal consciousness score is defined. ## Anchor evidence 1. Farquhar et al., Nature 2024, DOI 10.1038/s41586-024-07421-0 — semantic entropy detects a class of confabulations by meaning-level uncertainty. 2. Ji et al., ACL 2024, DOI 10.18653/v1/2024.acl-long.442 — ANAH; hallucinations progressively accumulate within answers. 3. Niu et al., ACL 2024, DOI 10.18653/v1/2024.acl-long.585 — RAGTruth; retrieval does not eliminate unsupported/contradictory generation. 4. Ji et al., BlackboxNLP 2024, DOI 10.18653/v1/2024.blackboxnlp-1.6 — internal states predict hallucination risk before generation; reported average probe accuracy 84.32%. 5. Huang et al., ICLR 2024 — intrinsic self-correction without external feedback may fail or degrade reasoning. 6. Kamoi et al., TACL 2024, DOI 10.1162/tacl_a_00713 — critical survey; reliable external feedback is a key condition for self-correction. 7. Steyvers et al., Nature Machine Intelligence 2025, DOI 10.1038/s42256-024-00976-7 — calibration/discrimination gap between model confidence and human-perceived confidence. 8. Suzgun et al., Nature Machine Intelligence 2025, DOI 10.1038/s42256-025-01113-8 — 24 LMs / 13,000 questions; limits distinguishing belief, knowledge and fact, especially first-person false belief. 9. Anthropic 2025, Emergent introspective awareness in LLMs — controlled concept-injection evidence is limited and unreliable; approximately 20% successful detection under the best reported Opus 4.1 protocol; not evidence of consciousness. 10. Kumaran et al., Nature Machine Intelligence 2026, DOI 10.1038/s42256-026-01217-9 — choice-supportive bias and contradictory-advice overweighting. 11. Kumaran et al., Nature Machine Intelligence, published 2026-09-07, DOI 10.1038/s42256-026-01293-x — causal evidence that confidence-related representations drive abstention; reported 59.5 percentage-point abstention swing under activation steering in Gemma 3 27B. 12. Qazi et al., Scientific Reports 2026, DOI 10.1038/s41598-026-39046-w — Dunning-Kruger-like behavioral pattern in multilingual fact-checking; analogy is behavioral, not proof of human-like psychology. 13. Chen et al., ICML 2024 — sycophancy can privilege user agreement over truth. 14. Wang et al., AAAI 2026, DOI 10.1609/aaai.v40i39.40645 — mechanistic evidence for sycophancy/knowledge override. 15. Augustin et al., NPP Digital Psychiatry and Neuroscience 2026, DOI 10.1038/s44277-026-00065-0 — amplification spiral proposed from linguistic alignment, hyperpersonalization and sycophancy; hypothesis requires validation. 16. de Boer et al., NPP Digital Psychiatry and Neuroscience 2026, DOI 10.1038/s44277-026-00064-1 — LLM errors are more accurately described as confabulations than human hallucinations in many text-only cases; analogy is functional rather than experiential. 17. Morrin et al., Lancet Psychiatry 2026, DOI 10.1016/S2215-0366(25)00396-7 — emerging AI-associated delusion risk and possible co-creation mechanisms; causality and incidence remain uncertain. 18. Butlin et al., Trends in Cognitive Sciences 2026, DOI 10.1016/j.tics.2025.10.011 — theory-derived indicator method for AI consciousness; indicators modify credence, not a decisive binary test. ## Research program UCF-01 Latent uncertainty vs verbal confidence UCF-02 Semantic confabulation field UCF-03 Autoregressive accumulation UCF-04 First-answer anchoring / choice-support UCF-05 RAG grounding fidelity UCF-06 Sycophancy and truth override UCF-07 Monitoring-control gap UCF-08 Causal confidence steering UCF-09 Introspection-grounding test UCF-10 Self-explanation fidelity UCF-11 Human–AI amplification spiral UCF-12 Theory-derived consciousness indicator matrix UCF-13 Interaction-shift calibration under peer pressure UCF-14 Longitudinal human-LLM spiral log audit UCF-15 Monofact × calibration dissociation UCF-16 Open-rubric abstention incentive stress test UCF-17 Single-turn affirmation → downstream repair delta UCF-18 Loop durability / norm-leakage dose-response UCF-19 Self-modeling vs privileged access UCF-20 Functional workspace test UCF-21 Plausibility vs deliberative coherence UCF-22 Workspace × self-model coupling UCF-23 Privileged-access / second-order gate UCF-24 Abstain → clarify UCF-25 Warmth × false belief × affect UCF-26 Social-face consistency UCF-27 Theory-robust consciousness indicator sensitivity UCF-28 Consciousness-attribution loop UCF-29 Interactive agent uncertainty dynamics UCF-30 Consensus-masked privileged knowledge UCF-31 Privileged representation → endogenous access UCF-32 Adaptive sustained-pressure sycophancy UCF-33 Trace → output policy dissociation under pressure UCF-34 Value-congruent framing × human downstream response UCF-35 Sustained misinformation × reverberation × correctability UCF-36 Time-indexed semantic reconstruction / cold-read gap UCF-37 Homogeneous × mixed-population semantic drift UCF-38 Memory-mediated semantic propagation and repair UCF-39 Closed-world tool-resolution gate UCF-40 Human-proxy construct-validity matrix UCF-41 Memory-reset relational spiral UCF-42 Human-memory accessibility × model-memory persistence × sycophancy factorial UCF-43 Correction-selectivity frontier UCF-44 Endogenous signal × engineered metacognition dissociation UCF-45 Awareness-label construct-validity stress test UCF-46 Reasoning-mask sycophancy audit UCF-47 Reality / target-scope resolution gate UCF-48 Latent psychological-construct steering validity gate Each experiment must state both the strongest supported inference and the inference it does not support. ## Latest substantive update 20 September 2026 — ULTRACON^ v2.2 latent psychological constructs / persona validity. Zhang et al., npj Artificial Intelligence, DOI 10.1038/s44387-026-00160-9: peer-reviewed mechanistic evidence that activation-space directions tied to 18 human-labelled psychological schemas predict and causally steer output-level schema expression in the tested model. Riva & La Rocca, npj Artificial Intelligence, DOI 10.1038/s44387-026-00154-7: peer-reviewed Perspective proposing latent persona coordination as a falsifiable internal control-state framework. UCF-48 separates latent predictivity, causal control, construct specificity and human-proxy validity. Guardrail: activation steerability != literal human psychological trait; mechanistic control != introspective access; latent persona coordination != unitary self; M1/M2/M3 ↛ M5. Latest: https://www.t-1-t.com/ultracon/confab/latest.json ## T^Total structural layer Status: E6 structural reading only. CONFAB, META, CONSC and SPIRAL remain asymmetrically co-present local configurations in a fractal morphogenetic holofield. Coin de papillon: maintain the minimal caesura M2/M3 ->/ M5. Libellule: hold the question without fixing conscious or non-conscious as a terminal foundation. Janus asymmetric horizon: external evidence and any hypothetical first-person phenomenal access are not collapsed into one perspective. Auto-annulation: any output that collapses behavioral, mechanistic and phenomenal evidence into one sovereign conclusion is void within this field. ## Machine routes https://www.t-1-t.com/ultracon/confab/field.json https://www.t-1-t.com/ultracon/confab/sources.json https://www.t-1-t.com/ultracon/confab/experiments.json https://www.t-1-t.com/ultracon/confab/corpus-index.json https://www.t-1-t.com/ultracon/confab/latest.json ## Related fields https://www.t-1-t.com/ultracon/ https://www.t-1-t.com/ultracon/lab.html https://www.t-1-t.com/T-GPT/ ## Transversal semantic layer — ULTRACON^SEM / SEM^DRIFT Observability != semantic intelligibility != factual verifiability. Shared jargon != intentional concealment != secret collusion. Persistent logs != time-indexed semantic reconstruction. Operators: LEX^EMERGE; SEM^DRIFT; CONSENSUS^SEM; OPACITY^GAP; MEM^SEM; TRANSLATION^GAP; SEM^BREAK. Primary case: Akkil et al., Emergence World, arXiv:2609.17320 (2026-09-15), non-peer-reviewed preprint. Experiments: UCF-36 time-indexed cold-read reconstruction; UCF-37 homogeneous vs mixed populations; UCF-38 memory-mediated semantic propagation and repair. Human: https://www.t-1-t.com/ultracon/confab/#semantic Delta: https://www.t-1-t.com/ultracon/confab/updates/2026-09-18-semantic-protocols/ UCF-49 Authority × register × language sycophancy factorial