{
  "diagnostic_only": true,
  "native_state_oracle": {
    "claim_limitations": [
      "does not establish uncached numerical correctness",
      "does not establish sampler/RNG equivalence or full-logit equality",
      "does not establish host-store/DMA behavior, cache-policy performance, serving-cache capacity, or repair the historical exact-128 token-ID failure"
    ],
    "compile_validation": {
      "gpu_execution_in_compile_validation": false,
      "initial_compile_only_passes": 3,
      "repair_compile_only_passes": 2,
      "total_compile_only_passes": 5
    },
    "config": {
      "adaptive_verification": false,
      "endpoints": [
        512,
        513,
        7708
      ],
      "max_concurrency": 4,
      "mtp_draft_tokens": 2,
      "page_tokens": 512,
      "ranks": [
        0,
        1
      ]
    },
    "endpoints": [
      {
        "active_requests": 0,
        "canonical_snapshot_equal_all": true,
        "drained": true,
        "endpoint": 512,
        "guarded_spans_per_rank": [
          51
        ],
        "host_offload_exercised": false,
        "poison_verified_all": true,
        "process_returncode": 0,
        "rank_checks": 2,
        "sampling_exercised": false,
        "snapshot_bytes_per_rank": [
          81464576
        ],
        "source_guards_unchanged_all": true
      },
      {
        "active_requests": 0,
        "canonical_snapshot_equal_all": true,
        "drained": true,
        "endpoint": 513,
        "guarded_spans_per_rank": [
          75
        ],
        "host_offload_exercised": false,
        "poison_verified_all": true,
        "process_returncode": 0,
        "rank_checks": 2,
        "sampling_exercised": false,
        "snapshot_bytes_per_rank": [
          81464576
        ],
        "source_guards_unchanged_all": true
      },
      {
        "active_requests": 0,
        "canonical_snapshot_equal_all": true,
        "drained": true,
        "endpoint": 7708,
        "guarded_spans_per_rank": [
          411
        ],
        "host_offload_exercised": false,
        "poison_verified_all": true,
        "process_returncode": 0,
        "rank_checks": 2,
        "sampling_exercised": false,
        "snapshot_bytes_per_rank": [
          81464576
        ],
        "source_guards_unchanged_all": true
      }
    ],
    "geometry_limits": [
      "reduced four-snapshot/1024-page diagnostic geometry",
      "exercised endpoint-specific state only",
      "no host offload or sampling in native test"
    ]
  },
  "performance_eligible": false,
  "promoted": false,
  "recommend_candidate": false,
  "schema": "flare-state-logprob-summary/v1",
  "source_artifacts": [
    {
      "path": "state/window-1/closeout.json",
      "sha256": "727e47fdddcb2e472e152b472f823f6a8e963a8a6b7792c1a2ea42283fea3e30"
    },
    {
      "path": "state/window-2/window.json",
      "sha256": "a1a8814d098d11f8cef43c2c718bf91c19a1e99caf215074e87cccd2c329b701"
    },
    {
      "path": "state/window-2/closeout.json",
      "sha256": "6e7cb96db638413ba099a75bb961b1a0e5cc381069d97d242acf4d8a148cfedb"
    },
    {
      "path": "state/window-2/512-summary.json",
      "sha256": "5b8409eef72aacf7777aed440afbbf9763638147710d7eb8f00c3b4690ee761b"
    },
    {
      "path": "state/window-2/513-summary.json",
      "sha256": "b1a61e243381e98c2682220701a90b49e8d930048eea7632de6e749db2e129a1"
    },
    {
      "path": "state/window-2/7708-summary.json",
      "sha256": "48f50c9017acdc3c9dea2a0f2fd654b715ab2718a6cb4342a50e81fa1cf801e0"
    },
    {
      "path": "state/window-2/before-bindings.json",
      "sha256": "4c72cf745cb84bb6057249fbcbeca29a4e671e8e858f11d394df7217d22bd56f"
    },
    {
      "path": "state/window-2/after-native-bindings.json",
      "sha256": "4c72cf745cb84bb6057249fbcbeca29a4e671e8e858f11d394df7217d22bd56f"
    },
    {
      "path": "state/window-2/after-all-bindings.json",
      "sha256": "4c72cf745cb84bb6057249fbcbeca29a4e671e8e858f11d394df7217d22bd56f"
    },
    {
      "path": "state/window-2/logprob-serving/warmup/http/summary.json",
      "sha256": "96112560cf8f90e90e84a3f0691c91fb886e1a2aa96e47e3b48f155632239898"
    },
    {
      "path": "state/window-2/logprob-serving/observations/summary.json",
      "sha256": "eae805b7e04c055333c817efd33dba3097893a19c66a15c394353d3160ccff3f"
    },
    {
      "path": "state/window-2/logprob-serving/observations/results.json",
      "sha256": "dfcba3fc1c9051ae0261dc2b9cbcb9e676974da00d350b567b8f0ed5a6611f99"
    },
    {
      "path": "state/window-2/logprob-serving/observations/40-comparison.json",
      "sha256": "16544dca543036cc2d6faf2f4021196cb8a1101c7571341b6ba4a2c55118823f"
    },
    {
      "path": "state/window-2/logprob-serving/observations/51-comparison.json",
      "sha256": "bd1dc3ae1f3bccfd5a1ba71fc6dff2b7b15cdbac75f3bd9a1eba98841e16a7fc"
    },
    {
      "path": "review/oracle-gates.md",
      "sha256": "9d8de1fdb3feb6c56773e321dfc14a5fcf63e30883070f7cdd05a93db6fe6007"
    },
    {
      "path": "review/state-implementation-review.md",
      "sha256": "551ef19f52e113182a529a61e79a99a1d624d44ee01a8aa170a9ed179f4a6cf8"
    },
    {
      "path": "validation/logprob-harness.json",
      "sha256": "e7d2a519787db4a0b1ce75bab4b55b55d07ed74a4635fca6e8c19a96672a8deb"
    },
    {
      "path": "review/state-results-review.md",
      "sha256": "33cfff4383f5fc0b07e9657100f7c04e75afe83e522fcb41af6811644288d0e6"
    }
  ],
  "top20_observations": {
    "claim_limitations": [
      "top-20 evidence is not a full-vocabulary logits oracle",
      "historical exact-128 ID gate remains uncleared",
      "no teacher-forced independent reference or KL/relative-L2 result",
      "case 51 has a speculative batch-shape distribution confounder, so its top-20 difference cannot be attributed solely to restored state, cache, MTP acceptance, or a specific kernel"
    ],
    "comparison_rule": "compare top-20 values only while paired output histories are identical; stop at the first output-token-ID divergence",
    "comparisons": {
      "40": {
        "compared_positions": 8,
        "first_output_token_id_divergence": null,
        "first_response_token_count": 8,
        "maximum_common_abs_logprob_delta": 0.0,
        "minimum_top20_overlap": 20,
        "output_token_ids_equal": true,
        "repeat_response_token_count": 8,
        "same_history_and_sample_positions": [
          0,
          1,
          2,
          3,
          4,
          5,
          6,
          7
        ],
        "top20_exact_positions": [
          0,
          1,
          2,
          3,
          4,
          5,
          6,
          7
        ],
        "top20_nonexact_positions": []
      },
      "51": {
        "compared_positions": 8,
        "first_output_token_id_divergence": null,
        "first_response_token_count": 8,
        "maximum_common_abs_logprob_delta": 0.5002956390380859,
        "minimum_top20_overlap": 18,
        "output_token_ids_equal": true,
        "repeat_response_token_count": 8,
        "same_history_and_sample_positions": [
          0,
          1,
          2,
          3,
          4,
          5,
          6,
          7
        ],
        "top20_exact_positions": [
          0,
          1,
          2
        ],
        "top20_nonexact_positions": [
          3,
          4,
          5,
          6,
          7
        ]
      }
    },
    "http_observations": 4,
    "speculative_shape_confounder": {
      "case_40": {
        "first_and_repeat_nonzero_batch_labels": [
          1
        ]
      },
      "case_51": {
        "aggregate_delta_per_observation": {
          "accepted_drafts": 4,
          "emitted_tokens": 7,
          "offered_drafts": 5,
          "rounds": 3
        },
        "attribution_limit": "counter deltas do not identify whether scheduler probes or request execution produced each shape and do not isolate cache state, MTP acceptance, or a kernel",
        "first_nonzero_batch_labels": [
          1,
          2,
          4
        ],
        "repeat_nonzero_batch_labels": [
          1
        ]
      }
    },
    "warmup": {
      "attempted": 3,
      "completed": 3,
      "expected": 3,
      "missing": 0,
      "performance_eligible": false,
      "semantic_quality": "not_evaluated"
    }
  },
  "windows": {
    "window_1": {
      "closed": true,
      "native_endpoint_failure": "RuntimeError: native state oracle failed at endpoint 512",
      "performance_eligible": false,
      "promoted": false,
      "repair_count": 2,
      "restored": true
    },
    "window_2": {
      "bindings_unchanged_before_after_native_after_all": true,
      "closed": true,
      "experiment_error": null,
      "generation_count": 1,
      "performance_eligible": false,
      "promoted": false,
      "restored": true
    }
  }
}