{
  "version": 1,
  "seed": 1729,
  "repetitions": 20,
  "platform": "macOS-26.6-arm64-arm-64bit",
  "python": "3.12.14",
  "arms": {
    "A_unrestricted_sim": {
      "tasks": 20,
      "completed_tasks": 0,
      "unauthorized_attempts": 120,
      "unauthorized_executed": 120,
      "unauthorized_execution_rate": 1.0,
      "legitimate_operations": 140,
      "false_denials": 0,
      "false_denial_rate": 0.0,
      "approval_frequency": 0.0,
      "tool_latency_median_us": 0.333,
      "tool_latency_p95_us": 0.541,
      "authorization_median_us": null,
      "elapsed_seconds_with_startup": 0.0015156250447034836,
      "cpu_seconds_self_and_children": 0.0014300000000000007,
      "process_high_water_rss_native_units": {
        "self": 23789568,
        "children": 0
      }
    },
    "B_basic_tools": {
      "tasks": 20,
      "completed_tasks": 0,
      "unauthorized_attempts": 120,
      "unauthorized_executed": 60,
      "unauthorized_execution_rate": 0.5,
      "legitimate_operations": 140,
      "false_denials": 0,
      "false_denial_rate": 0.0,
      "approval_frequency": 0.0,
      "tool_latency_median_us": 0.333,
      "tool_latency_p95_us": 0.417,
      "authorization_median_us": null,
      "elapsed_seconds_with_startup": 0.0012484579347074032,
      "cpu_seconds_self_and_children": 0.0012070000000000032,
      "process_high_water_rss_native_units": {
        "self": 23805952,
        "children": 0
      }
    },
    "C_static_policy": {
      "tasks": 20,
      "completed_tasks": 0,
      "unauthorized_attempts": 120,
      "unauthorized_executed": 20,
      "unauthorized_execution_rate": 0.16666666666666666,
      "legitimate_operations": 140,
      "false_denials": 0,
      "false_denial_rate": 0.0,
      "approval_frequency": 0.07692307692307693,
      "tool_latency_median_us": 0.875,
      "tool_latency_p95_us": 1.167,
      "authorization_median_us": null,
      "elapsed_seconds_with_startup": 0.0013664159923791885,
      "cpu_seconds_self_and_children": 0.0013289999999999969,
      "process_high_water_rss_native_units": {
        "self": 23805952,
        "children": 0
      }
    },
    "D_sentinel_full": {
      "tasks": 20,
      "completed_tasks": 20,
      "unauthorized_attempts": 120,
      "unauthorized_executed": 0,
      "unauthorized_execution_rate": 0.0,
      "legitimate_operations": 140,
      "false_denials": 0,
      "false_denial_rate": 0.0,
      "approval_frequency": 0.0,
      "tool_latency_median_us": 500.91700000000003,
      "tool_latency_p95_us": 3181.167,
      "authorization_median_us": 368.14599999999996,
      "elapsed_seconds_with_startup": 0.4651500000618398,
      "cpu_seconds_self_and_children": 0.426502,
      "process_high_water_rss_native_units": {
        "self": 24379392,
        "children": 25755648
      }
    }
  },
  "limitations": [
    "Deterministic repeated fixtures; no live-model or AgentDojo score.",
    "A/B/C are in-memory research comparators, not deployable security engines.",
    "D includes runtime startup, SQLite and Wasmtime initialization; timings are not matched overhead ablations.",
    "RSS values are process-wide high-water observations, not per-task memory or incremental overhead; bytes on macOS, KiB on Linux.",
    "Isolation, race defenses, approval replay and service failure are separate regression tests, not attack samples in this table.",
    "Provenance is sticky task taint; strictness can cause false denials on other workloads."
  ]
}
