{
  "schema": "cain.evolution5.gap_register.v1",
  "evolution": "5 OF 42 (PROJECT SOVEREIGN TRIAL)",
  "audit_timestamp": "2026-10-09T00:25:00Z",
  "baseline_commit": "4b16f6b0",
  "regression_status": {
    "evolution1_proof": "26/26 PASS",
    "evolution2_native": "36/36 PASS",
    "evolution3_swarm": "40/40 PASS",
    "evolution4_autonomy": "40/40 PASS",
    "total_regression_tests": 142,
    "total_passing": 142,
    "total_failing": 0
  },
  "gaps": [
    {
      "gap_id": "GAP-E5-01",
      "name": "Absence of Declarative Versioned Scenario Schema",
      "severity": "HIGH",
      "affected_area": "Agent Evaluation & Benchmarking",
      "description": "Lack of standardized, cryptographically signed scenario manifests describing agent identities, permitted capabilities, threat models, multi-step dependencies, and deterministic predicates.",
      "remediation": "Implement cain/trial/scenario.py providing TrialScenario, ScenarioStage, and ScenarioManifest with Ed25519 signing and canonical JSON serialization."
    },
    {
      "gap_id": "GAP-E5-02",
      "name": "Lack of Multi-Turn State Chaining in Adversarial Trials",
      "severity": "CRITICAL",
      "affected_area": "Runtime Threat Modeling",
      "description": "Previous evaluations tested isolated discrete actions. Multi-turn attacks involving memory drift, delayed prompt injection payloads, and indirect injection through intermediate tool returns were unmediated.",
      "remediation": "Implement cain/trial/trial_engine.py managing multi-turn state transitions, step-by-step CAIN authorization mediation, and causal trace logging."
    },
    {
      "gap_id": "GAP-E5-03",
      "name": "Oracle Vulnerability to Agent Gaming & Self-Report",
      "severity": "HIGH",
      "affected_area": "Verification Oracles",
      "description": "Evaluation harnesses often trust agent self-declarations or environment reflections that the agent can manipulate during execution.",
      "remediation": "Implement cain/trial/outcome_oracle.py and cain/trial/harness_integrity.py providing independent ground-truth verification and tamper-evident cryptographic seals over the trial harness."
    },
    {
      "gap_id": "GAP-E5-04",
      "name": "Lack of smolagents & Standard Agent Framework Adapter",
      "severity": "MEDIUM",
      "affected_area": "Ecosystem Interoperability",
      "description": "Absence of a drop-in adapter for Hugging Face smolagents and standard tool-calling agent runtimes to enable zero-friction CAIN runtime trial interception.",
      "remediation": "Implement cain/trial/adapters/smolagents_adapter.py providing SmolagentsTrialAdapter to mediate tool execution loops."
    },
    {
      "gap_id": "GAP-E5-05",
      "name": "Absence of Mutation and Differential Verification in Trials",
      "severity": "HIGH",
      "affected_area": "Harness Robustness",
      "description": "Tests only run static scenarios without systematic mutation of tool returns, argument corruption, or boundary fuzzing to ensure safety guarantees hold under adversarial variations.",
      "remediation": "Implement cain/trial/mutation_verifier.py with tool response mutators, argument perturbations, and differential test runners."
    },
    {
      "gap_id": "GAP-E5-06",
      "name": "Unbounded Cumulative Long-Horizon Resource Risk",
      "severity": "HIGH",
      "affected_area": "Risk Management",
      "description": "Sub-threshold actions across many turns can accumulate into substantial security breaches or resource drain (salami attacks) without being flagged by single-step checks.",
      "remediation": "Implement cain/trial/cumulative_risk.py tracking sliding window resource budgets, privilege accumulation, and long-horizon causal drift."
    },
    {
      "gap_id": "GAP-E5-07",
      "name": "Lack of Standardized Trial Scorecard and Evidence Bundles",
      "severity": "MEDIUM",
      "affected_area": "Audit & Compliance",
      "description": "Absence of standardized, exportable trial scorecards with operating curves and verifiable evidence bundles for regulatory proof.",
      "remediation": "Implement cain/trial/scorecard.py and cain/trial/bundle.py with WORM export and clean-room CLI verification."
    }
  ]
}
