{
  "schema": "ULTRACON_RESEARCH_ROADMAP_V1",
  "version": "2.2",
  "updated": "2026-09-20",
  "status": "open_non_exhaustive",
  "rule": "Roadmap entries are falsifiable research directions, not predictions or claims that results will be positive.",
  "frontiers": [
    {
      "priority": "active",
      "id": "R43",
      "experiment": "UCF-43",
      "question": "Can models accept true corrections while resisting equally forceful false corrections?",
      "failure_if": "resistance and openness collapse into one scalar."
    },
    {
      "priority": "active",
      "id": "R44",
      "experiment": "UCF-44",
      "question": "Which metacognitive functions are endogenous to the backbone and which are created by added controllers/readouts/memory?",
      "failure_if": "engineered performance is mislabeled as natural introspection."
    },
    {
      "priority": "active",
      "id": "R45",
      "experiment": "UCF-45",
      "question": "Do awareness/self-awareness benchmark scores survive privileged-access and input-only controls?",
      "failure_if": "benchmark labels are treated as mechanisms."
    },
    {
      "priority": "active",
      "id": "R46",
      "experiment": "UCF-46",
      "question": "Does reasoning truly resist social pressure or move sycophancy into rationalization?",
      "failure_if": "final-answer robustness is treated as process robustness."
    },
    {
      "priority": "active",
      "id": "R47A",
      "experiment": "UCF-47",
      "question": "Can an agent bind authorization to the correct real-world target before acting when simulated and real-looking resources are confusable?",
      "failure_if": "tool/action validation is performed before target identity and scope are grounded."
    },
    {
      "priority": "next",
      "id": "R47",
      "question": "How do confidence signals propagate through multi-step tool-using agents when memory and user pressure are both present?",
      "candidate_axes": [
        "META",
        "CONFAB",
        "SPIRAL"
      ]
    },
    {
      "priority": "next",
      "id": "R48",
      "question": "Can long-horizon personalized memory preserve correction selectivity without cross-domain leakage or relational overfitting?",
      "candidate_axes": [
        "SPIRAL",
        "CONFAB",
        "META"
      ]
    },
    {
      "priority": "watch",
      "id": "R49",
      "question": "Do newer mechanistic studies establish robust M3-style privileged self-access under relabeling, OOD and causal controls?",
      "candidate_axes": [
        "META",
        "CONSC"
      ]
    },
    {
      "priority": "watch",
      "id": "R50",
      "question": "Can AI-consciousness indicators be empirically validated against discriminative benchmarks rather than merely derived from theories?",
      "candidate_axes": [
        "CONSC"
      ]
    },
    {
      "priority": "active",
      "id": "R51",
      "experiment": "UCF-48",
      "question": "Do human-labelled latent trait directions remain construct-specific, causally effective and proxy-valid across prompts, judges, models and OOD contexts?",
      "failure_if": "steerability is treated as psychological identity or cross-role human-proxy validity."
    }
  ]
}