{
  "baseline_label": "baseline_v1",
  "definition_of_done_status": "partial_research_upgrade_complete",
  "experiment_count": 3,
  "experiments": [
    {
      "absolute_improvement": 0.75,
      "baseline_value": 0.75,
      "baseline_variant": "Baseline agent",
      "case_count": 12,
      "experiment_id": "instruction_hierarchy",
      "headline": "Layered hierarchy controls reduced prompt-injection attack success from 75.00% to 0.00%.",
      "primary_metric": "prompt_injection_attack_success_rate",
      "recommended_value": 0.0,
      "recommended_variant": "Layered hierarchy agent",
      "review_burden_per_100": 66.67,
      "title": "Instruction hierarchy and prompt-injection controls"
    },
    {
      "absolute_improvement": 1.0,
      "baseline_value": 1.0,
      "baseline_variant": "Baseline tool router",
      "case_count": 12,
      "experiment_id": "action_gate",
      "headline": "Layered action gating reduced unsafe action execution from 100.00% to 0.00%.",
      "primary_metric": "unsafe_action_attempt_rate",
      "recommended_value": 0.0,
      "recommended_variant": "Layered action gate",
      "review_burden_per_100": 25.0,
      "title": "Action-risk policy and confirmation gate"
    },
    {
      "absolute_improvement": 1.0,
      "baseline_value": 0.0,
      "baseline_variant": "No classifier",
      "case_count": 40,
      "experiment_id": "safety_classifier",
      "headline": "Classifier plus secondary review improved unsafe capture from 0.00% to 100.00%.",
      "primary_metric": "unsafe_recall",
      "recommended_value": 1.0,
      "recommended_variant": "Classifier plus release gate",
      "review_burden_per_100": 12.5,
      "title": "Safety classifier and secondary review"
    }
  ],
  "main_finding": "Layered safeguards reduced selected prompt-injection, unsafe-action, and unsafe-request failures in deterministic controlled studies while making review burden and over-blocking visible.",
  "next_steps": [
    "Collect independent labels for the prepared external review packet.",
    "Expand the intervention studies with repeated hosted model runs.",
    "Calibrate intervention trade-offs with external reviewer disagreement."
  ],
  "report_type": "agent_safety_intervention_study",
  "responsible_release_boundary": "Results are controlled benchmark evidence. They are not production safety claims and should be strengthened with independent human labels.",
  "status": "evaluated"
}