{
  "schema": "ULTRACON_AI_EPISTEMIC_UPDATE_V1",
  "date": "2026-09-20",
  "slug": "latent-persona-steering",
  "substantive_update": true,
  "title": "ULTRACON^ v2.2 — latent psychological constructs, causal steering and persona-validity firewall",
  "axes": [
    "META",
    "SPIRAL",
    "CONFAB"
  ],
  "evidence_level": "peer-reviewed empirical mechanistic article + peer-reviewed Perspective + falsifiable protocol",
  "summary": "A new peer-reviewed npj Artificial Intelligence article provides causal activation-steering evidence that human-labelled psychological constructs can correspond to manipulable directions in LLM activation space. ULTRACON separates this from human psychological identity, clinical state, self-modeling and consciousness, and adds UCF-48 to test construct specificity and human-proxy validity.",
  "new_separations": [
    "activation_direction != psychological_trait_identity",
    "causal_trait_expression != human_psychological_equivalence",
    "mechanistic_steerability != introspective_access",
    "latent_persona_coordination != unitary_self",
    "simulated_participant_stability != human_proxy_validity"
  ],
  "guardrail": "Mechanistic causal control over a human-labelled behavioral construct is evidence about model representations and output control, not evidence that the model literally instantiates the corresponding human psychological state or M5. M1/M2/M3 ↛ M5 remains unchanged.",
  "source_ids": [
    "SRC-ZHANG-MECH-SIMPART-NPJAI-2026",
    "SRC-RIVA-LATENT-PERSONA-NPJAI-2026"
  ],
  "experiment_ids": [
    "UCF-48"
  ]
}