{
  "schema_version": 2,
  "method": "gavagai",
  "manifest_ref": "foundation-models:gavagai",
  "dataset": {
    "id": "synthetic:nov04-gavagai-v1",
    "revision": "sha256:ad8389f6d94d104bb9f76740381ec40dadb3005cde6d437c232e0345ab8e132f",
    "recipe": "small seeded mechanism/state/index contracts; no public capability benchmark",
    "capability_benchmark": false
  },
  "seed": [
    42,
    43,
    44
  ],
  "seeds": [
    42,
    43,
    44
  ],
  "per_seed": [
    {
      "seed": 42,
      "diagnostic_only": true,
      "evaluation_tier": "l1_mechanism",
      "metrics": {
        "scalar_dual_max_error": 1.1102230246251565e-16,
        "inactive_cap_softmax_error": 1.1102230246251565e-16,
        "envelope_gradient_error": 0.0,
        "eight_iteration_row_residual": 0.0021412147736108444,
        "eight_iteration_capacity_excess": 2.220446049250313e-16,
        "converged_row_residual": 0.0,
        "nothing_mass": 0.7
      },
      "checks": {
        "scalar_dual_reference": true,
        "inactive_cap_is_row_softmax": true,
        "zero_capacity_goes_to_nothing": true,
        "no_gradient_through_plan": true,
        "envelope_gradient": true,
        "caption_gradient_retained": true,
        "finite_eight_step_plan": true,
        "capacity_bound": true
      },
      "execution_device": "cuda",
      "execution_accelerator": "NVIDIA A100-SXM4-80GB"
    },
    {
      "seed": 43,
      "diagnostic_only": true,
      "evaluation_tier": "l1_mechanism",
      "metrics": {
        "scalar_dual_max_error": 1.6653345369377348e-16,
        "inactive_cap_softmax_error": 1.1102230246251565e-16,
        "envelope_gradient_error": 0.0,
        "eight_iteration_row_residual": 0.000527098141325899,
        "eight_iteration_capacity_excess": 0.0,
        "converged_row_residual": 2.220446049250313e-16,
        "nothing_mass": 0.7
      },
      "checks": {
        "scalar_dual_reference": true,
        "inactive_cap_is_row_softmax": true,
        "zero_capacity_goes_to_nothing": true,
        "no_gradient_through_plan": true,
        "envelope_gradient": true,
        "caption_gradient_retained": true,
        "finite_eight_step_plan": true,
        "capacity_bound": true
      },
      "execution_device": "cuda",
      "execution_accelerator": "NVIDIA A100-SXM4-80GB"
    },
    {
      "seed": 44,
      "diagnostic_only": true,
      "evaluation_tier": "l1_mechanism",
      "metrics": {
        "scalar_dual_max_error": 1.1102230246251565e-16,
        "inactive_cap_softmax_error": 1.1102230246251565e-16,
        "envelope_gradient_error": 0.0,
        "eight_iteration_row_residual": 0.0003125999529579593,
        "eight_iteration_capacity_excess": 0.0,
        "converged_row_residual": 0.0,
        "nothing_mass": 0.7
      },
      "checks": {
        "scalar_dual_reference": true,
        "inactive_cap_is_row_softmax": true,
        "zero_capacity_goes_to_nothing": true,
        "no_gradient_through_plan": true,
        "envelope_gradient": true,
        "caption_gradient_retained": true,
        "finite_eight_step_plan": true,
        "capacity_bound": true
      },
      "execution_device": "cuda",
      "execution_accelerator": "NVIDIA A100-SXM4-80GB"
    }
  ],
  "metrics": {
    "converged_row_residual": 7.401486830834377e-17,
    "eight_iteration_capacity_excess": 7.401486830834377e-17,
    "eight_iteration_row_residual": 0.0009936376226315675,
    "envelope_gradient_error": 0.0,
    "inactive_cap_softmax_error": 1.1102230246251565e-16,
    "nothing_mass": 0.6999999999999998,
    "scalar_dual_max_error": 1.295260195396016e-16
  },
  "diagnostic_only": true,
  "fixture_demo": true,
  "formal_comparison": false,
  "evaluation_tier": "l1_mechanism",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "claim_policy": "three-seed synthetic mechanism diagnostic; not a capability comparison"
  },
  "promotion_eligible": false,
  "evolve_integrated": false,
  "full_experiment_status": "not_reproduced_high_cost",
  "full_experiment_review_done": true,
  "automatic_retry": false,
  "automatic_implementation_queue": false,
  "provenance": {
    "commit": "3f6f422472dc6ed486d97160328560e22ce03cb9",
    "artifact_path": "docs/foundation-models/2611.00319-gavagai/metrics/component-cuda-seeds42-44.json",
    "dataset_fingerprint": "sha256:7920ad02125e8dbf755b2017c1ef238a044e58ed94ae7092be210a88e1df2fd8",
    "accelerator": "NVIDIA A100-SXM4-80GB",
    "executed_source_sha256": {
      "scripts/run_nov04_components.py": "61d35d574d9561e4611d2647174b8a6af07fabf8c30b6909e21e253542cf4336",
      "docs/research-audits/nov04-components/preflight.json": "443ad9d8319c1c32ee31f8ec1f99f046c6b3091d427512db071248fc8673dd47",
      "scripts/oct18_frozen_sources.py": "da08355049c1f6b5184ee659643c9be9decedf2041ef6a87a3dce2e4582e385a",
      "src/auto_research/agent_research/agentheal.py": "34c7a3137073075333d16d186185e439e3e5538a6172c33f686e2ce912e57699",
      "src/auto_research/agent_research/strategist.py": "f496322604ae6e54ffa74ff459bfdace27cc673514f86ac14dbfbd4678706e6a",
      "src/auto_research/foundation_models/gavagai.py": "ad8389f6d94d104bb9f76740381ec40dadb3005cde6d437c232e0345ab8e132f",
      "src/auto_research/foundation_models/proland.py": "d98512b9af75517dc5522f060244412bc2ef3535c732042d8eece14b6ca39f9d",
      "src/auto_research/post_training/rlid.py": "d349bdc43097604151e755f1ebed761f7b2f8136c0a76bf728ae95f266266805",
      "src/auto_research/post_training/specopd.py": "5602752775ce8cfeb7ca1045a5d5168be7e3f99820234f56fdf109f4a89d03de",
      "src/auto_research/reproductions/wesearch_components.py": "e516559381a2f3e0d96abd1a39e2075b78d3cee2b6b97468e4724a7642d9064c"
    },
    "raw_runner_artifact_sha256": "c841b9c5fe0be8bd68197d1bca22cc80c83ed62a404db7fc02523577eee413f4"
  }
}