{
  "required": true,
  "paths": [
    {
      "id": "grpo_judge_advantage_via_custom_path",
      "fixed_input": {
        "sha256": "35ceda61dccc81731e96c435d31ae788035c8e00a875820651125d82bd376cc6",
        "description": "fixed-fixture.json: CPU float32 synthetic linear-softmax policy, identical weights, token identities, masks; full parameter gradients. No optimizer or remote execution."
      },
      "baseline_revision": "6daabe0838736557f0e802020b710073eb340287",
      "reference_revision": "6daabe0838736557f0e802020b710073eb340287",
      "implementation_revision": "d2c1edd7b28edd82aad3a5409ab5cde86c3d6b76d594957071d94bbfd142e513",
      "precision": "CPU torch.float32; no mixed precision; not a Spindle backend parity test",
      "normalization": "sum of masked token losses / 14 loss tokens; no reward mixing; no clipping in specified IS path",
      "errors": {
        "baseline_loss": {
          "max_abs_error": 0.0,
          "reference_max_abs": 0.00714289303869009,
          "relative_l2_error": 0.0,
          "cosine_similarity": 1.0
        },
        "baseline_gradient": {
          "max_abs_error": 0.0,
          "reference_max_abs": 0.0452161505818367,
          "relative_l2_error": 0.0,
          "cosine_similarity": 0.9999999999999998
        },
        "reference_loss": {
          "max_abs_error": 0.0,
          "reference_max_abs": 0.00714289303869009,
          "relative_l2_error": 0.0,
          "cosine_similarity": 1.0
        },
        "reference_gradient": {
          "max_abs_error": 0.0,
          "reference_max_abs": 0.0452161505818367,
          "relative_l2_error": 0.0,
          "cosine_similarity": 0.9999999999999998
        }
      },
      "lr_scaling": [],
      "scope": "Necessary-condition offline test only; cannot certify server precision, init, or optimizer."
    },
    {
      "id": "sdpo_self_teacher_advantage",
      "fixed_input": {
        "sha256": "35ceda61dccc81731e96c435d31ae788035c8e00a875820651125d82bd376cc6",
        "description": "fixed-fixture.json: CPU float32 synthetic linear-softmax policy, identical weights, token identities, masks; full parameter gradients. No optimizer or remote execution."
      },
      "baseline_revision": "6daabe0838736557f0e802020b710073eb340287",
      "reference_revision": "7c457fc1b1f636ae794eb0362ba37d4743b06fbc",
      "implementation_revision": "d2c1edd7b28edd82aad3a5409ab5cde86c3d6b76d594957071d94bbfd142e513",
      "precision": "CPU torch.float32; no mixed precision; not a Spindle backend parity test",
      "normalization": "sum of masked token losses / 14 loss tokens; no reward mixing; no clipping in specified IS path",
      "errors": {
        "baseline_loss": {
          "max_abs_error": 0.0,
          "reference_max_abs": 0.0,
          "relative_l2_error": null,
          "cosine_similarity": null
        },
        "baseline_gradient": {
          "max_abs_error": 0.0,
          "reference_max_abs": 0.0,
          "relative_l2_error": null,
          "cosine_similarity": null
        },
        "reference_loss": {
          "max_abs_error": 0.13540398702025414,
          "reference_max_abs": 0.09341718256473541,
          "relative_l2_error": 1.4494548358533836,
          "cosine_similarity": -1.0
        },
        "reference_gradient": {
          "max_abs_error": 0.0,
          "reference_max_abs": 0.04461999982595444,
          "relative_l2_error": 0.0,
          "cosine_similarity": 1.0
        }
      },
      "lr_scaling": [],
      "scope": "Necessary-condition offline test only; cannot certify server precision, init, or optimizer."
    }
  ]
}