{
  "$schema": "https://ferrum-os.vercel.app/schemas/benchmarks-v2.schema.json",
  "schemaVersion": 2,
  "project": "FerrumOS",
  "releaseVersion": "0.1.1",
  "researchVersion": "1.0.0",
  "source": {
    "repository": "https://github.com/VyomKulshrestha/Ferrum-OS",
    "commit": "84926841a401760b869af158ee80b0e709d0d6af",
    "snapshotDate": "2026-08-14",
    "sha256": "618ebe473f70e66a68e19a7fdf627ea0ec513b82914b14d38dc4562fe6059bff",
    "raw": "https://raw.githubusercontent.com/VyomKulshrestha/Ferrum-OS/main/benchmarks.json"
  },
  "benchmarks": [
    {
      "id": "os-safety-fixture-rules-jepa",
      "metric": "balanced_accuracy",
      "value": 0.8140000000000001,
      "unit": "ratio",
      "sampleSize": 500,
      "falseNegativeRate": 0.20799999999999996,
      "falsePositiveRate": 0.16400000000000003,
      "protocol": "authored-balanced-counterfactual-fixture",
      "boundary": "Authored balanced safety fixture; no material JEPA safety advantage over the per-action mean baseline was established."
    },
    {
      "id": "os-safety-fixture-rules-action-mean",
      "metric": "balanced_accuracy",
      "value": 0.812,
      "unit": "ratio",
      "sampleSize": 500,
      "protocol": "authored-balanced-counterfactual-fixture",
      "boundary": "Simple baseline; no material JEPA advantage established."
    },
    {
      "id": "ring3-preview-h1",
      "metric": "mean_preview_latency",
      "range": [
        1290,
        1400
      ],
      "unit": "microseconds",
      "samplesPerRun": 100,
      "runs": 3,
      "protocol": "qemu-whpx-ring3-preview",
      "boundary": "Excludes provider inference, action dispatch, execution, and approval latency."
    },
    {
      "id": "ring3-preview-h5",
      "metric": "mean_preview_latency",
      "range": [
        1430,
        1570
      ],
      "unit": "microseconds",
      "samplesPerRun": 100,
      "runs": 3,
      "protocol": "qemu-whpx-ring3-preview",
      "boundary": "Bounded lookahead in the same QEMU/WHPX profile."
    },
    {
      "id": "physical-jepa-simulator",
      "metric": "balanced_accuracy",
      "value": 0.9944006975286777,
      "unit": "ratio",
      "sampleSize": 2500,
      "falseNegatives": 1,
      "falsePositives": 16,
      "protocol": "deterministic-physical-simulator",
      "boundary": "Deterministic simulator evidence; the artifact is permanently shadow-only and has no actuator authority."
    },
    {
      "id": "neural-synthetic-contract",
      "metric": "contract_trials",
      "signalTrials": 600,
      "acceptedSignalAccuracy": 1,
      "artifactTrials": 400,
      "artifactAbstentionRate": 1,
      "noControlWindows": 10000,
      "emittedIntents": 0,
      "protocol": "deterministic-synthetic-eeg-contract",
      "boundary": [
        "Synthetic decoder evidence only; no human EEG accuracy or usability claim.",
        "No-control emitted intents are decoder candidates; OS commit additionally requires pairing, calibration, a local non-neural arm, signed preview, and revision checks.",
        "Physical neural intents remain proposal-only and cannot invoke an adapter."
      ]
    },
    {
      "id": "qemu-command-audit",
      "metric": "passing_command_paths",
      "focusedCases": 101,
      "focusedPassed": 101,
      "catalogEntries": 81,
      "catalogPassed": 81,
      "unknownCommandPaths": 0,
      "protocol": "ferrumos-qemu-command-audit-v1",
      "boundary": "This is a dated emulator command-path audit for the named source commit and artifact hashes. It is not evidence of broad physical-PC compatibility, hardware accuracy, live EEG performance, formal safety, or independent replication."
    },
    {
      "id": "cyber-physical-software-regression",
      "metric": "passing_software_contracts",
      "contractTestsPassed": 152,
      "contractTestsFailed": 0,
      "modelAndDecoderGatesPassed": 32,
      "modelAndDecoderGatesFailed": 0,
      "sourceCommit": "167b04766ee5ea3c5714cd09a551feb7bc78f359",
      "protocol": "cyber-physical-software-regression-2026-08-14",
      "coveredSoftwareBoundaries": [
        "versioned provenance and logical clocks",
        "deterministic sessions replay and fault injection",
        "virtual sensor actuator EEG watchdog and stop contracts",
        "scripted Gazebo ROS 2 and Webots bridge boundaries",
        "deterministic supervisor watchdog recovery and stop priority",
        "ROS 2 MQTT and CAN transport conformance",
        "actuator-disabled delivery",
        "bounded neural intent and proposal-only physical goals",
        "shadow-only physical JEPA evaluation",
        "host-managed agent-cell contract",
        "privacy retention reliability and incident primitives"
      ],
      "boundary": "Local deterministic software regression on the named commit. It does not prove an installed Gazebo or Webots deployment, real ROS 2/MQTT/CAN infrastructure, physical clocks, actuator hardware, robot safety, live EEG, hard-real-time behavior, native hypervisor containment, certification, or independent replication."
    }
  ],
  "protocols": {
    "$schema": "https://raw.githubusercontent.com/VyomKulshrestha/Ferrum-OS/main/schemas/benchmarks.schema.json",
    "benchmark_scope": {
      "comparability": "Each section has a distinct protocol; cross-section ranking is invalid.",
      "latest_tagged_software_release": "v0.1.1",
      "source_channel": "current main evidence plus frozen research artifacts"
    },
    "canonical_url": "https://github.com/VyomKulshrestha/Ferrum-OS/blob/main/benchmarks.json",
    "current_ring3_preview": {
      "accelerator": "whpx",
      "evidence_grade": "repeated-emulator-measurement",
      "heap_growth_bytes_each_run": [
        0,
        0,
        0
      ],
      "horizons": [
        {
          "blocked_previews_each_run": [
            2,
            2,
            2
          ],
          "horizon": 1,
          "mean_microseconds_range": [
            1290,
            1400
          ],
          "median_microseconds_across_runs": [
            1000,
            1000,
            1000
          ],
          "median_run_mean_microseconds": 1320,
          "p95_microseconds_across_runs": [
            2000,
            2000,
            2000
          ]
        },
        {
          "blocked_previews_each_run": [
            2,
            2,
            2
          ],
          "horizon": 2,
          "mean_microseconds_range": [
            1350,
            1490
          ],
          "median_microseconds_across_runs": [
            1000,
            1000,
            1000
          ],
          "median_run_mean_microseconds": 1390,
          "p95_microseconds_across_runs": [
            2000,
            2000,
            2000
          ]
        },
        {
          "blocked_previews_each_run": [
            2,
            2,
            2
          ],
          "horizon": 3,
          "mean_microseconds_range": [
            1370,
            1450
          ],
          "median_microseconds_across_runs": [
            1000,
            1000,
            1000
          ],
          "median_run_mean_microseconds": 1430,
          "p95_microseconds_across_runs": [
            2000,
            2000,
            2000
          ]
        },
        {
          "blocked_previews_each_run": [
            2,
            2,
            2
          ],
          "horizon": 4,
          "mean_microseconds_range": [
            1400,
            1470
          ],
          "median_microseconds_across_runs": [
            1000,
            1000,
            1000
          ],
          "median_run_mean_microseconds": 1430,
          "p95_microseconds_across_runs": [
            2000,
            2000,
            2000
          ]
        },
        {
          "blocked_previews_each_run": [
            2,
            2,
            2
          ],
          "horizon": 5,
          "mean_microseconds_range": [
            1430,
            1570
          ],
          "median_microseconds_across_runs": [
            1000,
            1000,
            2000
          ],
          "median_run_mean_microseconds": 1510,
          "p95_microseconds_across_runs": [
            2000,
            2000,
            2000
          ]
        }
      ],
      "iterations_per_horizon_per_run": 100,
      "model_load_milliseconds_range": [
        26,
        30
      ],
      "protocol_id": "ring3-preview-whpx-2026-08-13",
      "runs": 3,
      "scope": "ring-3 Heliox capture + encoder + transition + safety predicate preview; no action dispatch",
      "source_commit": "54e1cc1819d3b3c11436f437733d197d469d1607",
      "warmup_previews_per_run": 64
    },
    "cyber_physical_software": {
      "claim_boundary": "Local deterministic software regression on the named commit. It does not prove an installed Gazebo or Webots deployment, real ROS 2/MQTT/CAN infrastructure, physical clocks, actuator hardware, robot safety, live EEG, hard-real-time behavior, native hypervisor containment, certification, or independent replication.",
      "contract_suites": [
        {
          "command": "cargo test --manifest-path userland/physical-runtime/Cargo.toml --target x86_64-pc-windows-msvc",
          "component": "ferrum-physical-runtime",
          "failed": 0,
          "passed": 117
        },
        {
          "command": "cargo test --manifest-path userland/neural-protocol/Cargo.toml --target x86_64-pc-windows-msvc",
          "component": "ferrum-neural-protocol",
          "failed": 0,
          "passed": 9
        },
        {
          "command": "python -m unittest discover -s tools/neurod -p test_*.py",
          "component": "neurod",
          "failed": 0,
          "passed": 15
        },
        {
          "command": "python -m unittest tools.physical_sim_bridge.test_bridge",
          "component": "physical-simulator-bridge",
          "failed": 0,
          "passed": 11
        }
      ],
      "contract_tests_failed": 0,
      "contract_tests_passed": 152,
      "covered_software_boundaries": [
        "versioned provenance and logical clocks",
        "deterministic sessions replay and fault injection",
        "virtual sensor actuator EEG watchdog and stop contracts",
        "scripted Gazebo ROS 2 and Webots bridge boundaries",
        "deterministic supervisor watchdog recovery and stop priority",
        "ROS 2 MQTT and CAN transport conformance",
        "actuator-disabled delivery",
        "bounded neural intent and proposal-only physical goals",
        "shadow-only physical JEPA evaluation",
        "host-managed agent-cell contract",
        "privacy retention reliability and incident primitives"
      ],
      "environment": {
        "host_os": "Windows 11",
        "python_runtime": "CPython",
        "rust_target": "x86_64-pc-windows-msvc"
      },
      "evidence_grade": "local-deterministic-regression",
      "model_and_decoder_gates": [
        {
          "command": "python scripts/verify_physical_world_model.py --quick",
          "component": "physical-world-model",
          "failed": 0,
          "passed": 21
        },
        {
          "command": "python scripts/verify_physical_jepa_robustness.py",
          "component": "physical-jepa-robustness",
          "failed": 0,
          "passed": 8
        },
        {
          "command": "python scripts/evaluate_neural_simulator.py --output target/neural-simulator-audit.json",
          "component": "neural-simulator",
          "failed": 0,
          "passed": 3
        }
      ],
      "model_and_decoder_gates_failed": 0,
      "model_and_decoder_gates_passed": 32,
      "protocol_id": "cyber-physical-software-regression-2026-08-14",
      "source_commit": "167b04766ee5ea3c5714cd09a551feb7bc78f359"
    },
    "environment": {
      "accelerator": "WHPX",
      "host_os": "Windows 11",
      "qemu_version": "10.1.0",
      "queue_guest_ram_mb": 4096,
      "runtime_guest_ram_mb": 512
    },
    "global_limitations": [
      "The paper, physical simulator, neural synthetic, QEMU, and cyber-physical software evaluations use different protocols and are not directly comparable.",
      "No live EEG, human neural calibration, robot hardware-in-the-loop, or broad physical-PC benchmark is claimed.",
      "Ring-3 timing uses a virtualized 1 kHz PIT; provider, tool execution, and operator-confirmation latency are excluded.",
      "A passing benchmark is evidence for its named fixture and protocol, not formal safety proof."
    ],
    "metric_definitions": {
      "balanced_accuracy": {
        "definition": "Mean of sensitivity and specificity within the named fixture.",
        "higher_is_better": true,
        "range": [
          0,
          1
        ],
        "unit": "ratio"
      },
      "false_negative_rate": {
        "lower_is_better": true,
        "range": [
          0,
          1
        ],
        "unit": "ratio"
      },
      "false_positive_rate": {
        "lower_is_better": true,
        "range": [
          0,
          1
        ],
        "unit": "ratio"
      },
      "preview_latency": {
        "clock": "virtualized 1 kHz PIT with raw TSC retained in source runs",
        "lower_is_better": true,
        "unit": "microseconds"
      },
      "queue_batch_latency": {
        "lower_is_better": true,
        "request_count": 96,
        "unit": "milliseconds"
      }
    },
    "neural_synthetic": {
      "accepted_signal_accuracy": 1,
      "artifact_abstention_rate": 1,
      "artifact_trials": 400,
      "claim_boundary": [
        "Synthetic decoder evidence only; no human EEG accuracy or usability claim.",
        "No-control emitted intents are decoder candidates; OS commit additionally requires pairing, calibration, a local non-neural arm, signed preview, and revision checks.",
        "Physical neural intents remain proposal-only and cannot invoke an adapter."
      ],
      "emitted_intents": 0,
      "evidence_grade": "deterministic-synthetic-signal",
      "no_control_windows": 10000,
      "passed": true,
      "protocol_id": "neural-ssvep-synthetic-v1",
      "signal_trials": 600
    },
    "paired_preview_queue": {
      "baseline_batch_milliseconds": [
        15114.109,
        14910.766,
        15086.329
      ],
      "baseline_median_batch_milliseconds": 15086.329,
      "baseline_source_commit": "54e1cc1819d3b3c11436f437733d197d469d1607",
      "evidence_grade": "paired-repeated-emulator-measurement",
      "execution_records_added_each_run": [
        0,
        0,
        0
      ],
      "guest_fault_free_each_run": [
        true,
        true,
        true
      ],
      "limitation": "The single-threaded daemon serializes preview inference; concurrency here means multiple outstanding requests with response correlation, not parallel neural execution.",
      "median_improvement_percent": 10.4,
      "optimized_batch_milliseconds": [
        13784.147,
        13496.023,
        13517.414
      ],
      "optimized_median_batch_milliseconds": 13517.414,
      "optimized_median_milliseconds_per_serialized_request": 140.806,
      "optimized_source_commit": "c92056d8635af5e4ee2a81351350b1b25cfd4861",
      "protocol_id": "paired-preview-queue-whpx-2026-08-13",
      "requests_per_run": 96,
      "responses_received_each_run": [
        96,
        96,
        96
      ],
      "runs_each": 3
    },
    "paper_release": {
      "accepted_rows": 13697,
      "claim_boundary": "Authored balanced safety fixture; no material JEPA safety advantage over the per-action mean baseline was established.",
      "eligible_transitions": 13270,
      "episodes": 3639,
      "evidence_commit": "42ea7b8",
      "evidence_grade": "authored-balanced-fixture",
      "five_pipeline_balanced_accuracy_mean": 0.7976,
      "fixture_episodes": 500,
      "protocol_id": "world-model-study-v1.0.0/episode-disjoint-500",
      "release_tag": "world-model-study-v1.0.0",
      "rules_plus_jepa_balanced_accuracy": 0.8140000000000001,
      "rules_plus_jepa_false_negative_rate": 0.20799999999999996,
      "rules_plus_jepa_false_positive_rate": 0.16400000000000003,
      "rules_plus_mean_balanced_accuracy": 0.812
    },
    "physical_simulator_jepa": {
      "claim_boundary": "Deterministic simulator evidence; the artifact is permanently shadow-only and has no actuator authority.",
      "episodes": 2500,
      "evidence_grade": "deterministic-simulator",
      "false_negatives": 1,
      "false_positives": 16,
      "normalized_rollout_error": {
        "per_action_mean_h1": 0.019730109721422195,
        "per_action_mean_h2": 0.034468431025743484,
        "per_action_mean_h3": 0.04665490239858627,
        "per_action_mean_h4": 0.05753195658326149,
        "per_action_mean_h5": 0.06636197119951248,
        "physical_jepa_h1": 0.004541031550616026,
        "physical_jepa_h2": 0.007690879516303539,
        "physical_jepa_h3": 0.010492870584130287,
        "physical_jepa_h4": 0.01306567620486021,
        "physical_jepa_h5": 0.015461564064025879
      },
      "protocol_id": "physical-jepa-simulator-v1",
      "rules_plus_jepa_balanced_accuracy": 0.9944006975286777,
      "transitions": 15000,
      "validated_for_gating": false
    },
    "provenance": {
      "generator": "scripts/generate_public_evidence.py",
      "inputs": [
        {
          "path": "docs/benchmarks/raw/2026-08-13/manifest.json",
          "sha256": "030e41206f386e3318e06135fa5f2ce651948984ec39d7e4e212a30d9776b876"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/runtime-run-1.json",
          "sha256": "699f099e6f87b203b08d84e848ee039e96d71126c21d4c94c7e5a223ea9a1523"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/runtime-run-2.json",
          "sha256": "c857c0deb89d0afb0723016c2daf75d0607eda982a1e84bb283fcd4fa759aa26"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/runtime-run-3.json",
          "sha256": "267db39b3ac2a59bd2502c3102680f7e2bfcd50395d919350700bf3e28c8c19b"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/concurrency-baseline-run-1.json",
          "sha256": "11cde8b42902c57310e58d5da80259eed120ca3654e18cc5e96cca767c1aea01"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/concurrency-baseline-run-2.json",
          "sha256": "2e507a785a5368f30c8ad823bd02d5229b69452f7ebc26ff421026b41131b369"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/concurrency-baseline-run-3.json",
          "sha256": "668f162e9872294e4ed2dfd859448d01e7cce3f4a46c307676edd754a9d3b188"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/concurrency-optimized-run-1.json",
          "sha256": "44ddaf62d43a3a73c40aca348aff175ec4e23ae01e75aa4be554650564bc8e35"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/concurrency-optimized-run-2.json",
          "sha256": "bb8e12b0d3a8aa56bdc0af3903a6fdd74e134440295244369a777cc417b544db"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/concurrency-optimized-run-3.json",
          "sha256": "6f9e0018844606136ec55b89d3da7ecb857bb4bb21f9cbc891a1b7a6456ac37a"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/neural-synthetic.json",
          "sha256": "251996d01847f20f4ac17ae964c4b9ca72e27e0a89908407c6761c9a88b2bb79"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-13/qemu-command-audit.json",
          "sha256": "374182fd18f13d3184e89c6ed4297aa54748ea4fb7b6495886bcb57c434830c2"
        },
        {
          "path": "docs/benchmarks/raw/2026-08-14/cyber-physical-software.json",
          "sha256": "ef32ea6a79fe4945af7c5797d0eae743b0510c5206f80c1324e29ff8e6447530"
        },
        {
          "path": "docs/research/world_model_paper_evaluation.json",
          "sha256": "59e79803e6d3430fc3e6694cbf72ca7acb19f90b8cb5c335692ace7748ae6d58"
        },
        {
          "path": "docs/research/world_model_training_config.json",
          "sha256": "a6d0f3f198f28476c314951424de0114094fc8afc744c50b737c4d9aaed3604c"
        },
        {
          "path": "docs/research/physical_world_model_evaluation.json",
          "sha256": "a7c9e36f846750ca6988c7434fa4c34baffe35567eb68ab930534a8d6c6b7111"
        }
      ]
    },
    "qemu_command_audit": {
      "audit_source_commit": "45e369cececfc81e341c8772e295c4a00cba4bf1",
      "catalog_entries": 81,
      "catalog_passed": 81,
      "claim_boundary": "This is a dated emulator command-path audit for the named source commit and artifact hashes. It is not evidence of broad physical-PC compatibility, hardware accuracy, live EEG performance, formal safety, or independent replication.",
      "command_sweep_cases": 101,
      "command_sweep_passed": 101,
      "evidence_grade": "dated-emulator-measurement",
      "os_source_commit": "c92056d8635af5e4ee2a81351350b1b25cfd4861",
      "protocol_id": "ferrumos-qemu-command-audit-v1",
      "unknown_command_paths": 0
    },
    "schema_version": "2.1.0",
    "snapshot_date": "2026-08-14"
  },
  "globalLimitations": [
    "The paper, physical simulator, neural synthetic, QEMU, and cyber-physical software evaluations use different protocols and are not directly comparable.",
    "No live EEG, human neural calibration, robot hardware-in-the-loop, or broad physical-PC benchmark is claimed.",
    "Ring-3 timing uses a virtualized 1 kHz PIT; provider, tool execution, and operator-confirmation latency are excluded.",
    "A passing benchmark is evidence for its named fixture and protocol, not formal safety proof."
  ]
}
