{
  "schema": "cke.clef.native-e2e-evidence.v1",
  "date": "2026-11-14",
  "base_commit": "7325ab9fe",
  "host": "cke-amd-zen5-01",
  "hardware": "AMD Ryzen 9 9950X3D",
  "threads": 8,
  "request": {
    "id": "sentiment",
    "state": "The service was excellent.",
    "questions": {
      "sentiment": {
        "type": "choice",
        "instructions": "Classify sentiment.",
        "criteria": {
          "positive": "Positive",
          "negative": "Negative"
        }
      }
    }
  },
  "model": "Cloudflare/clef-flash",
  "revision": "fb4417d31c1d0f9e7e77f1141b40aaede7491e00",
  "tokens": 128,
  "production_framework_imports": "torch and transformers imports blocked during native inference and training",
  "backbone_dtype": "BF16 weights; provider-declared FP32 state and BF16 storage boundaries",
  "head_training_dtype": "FP32",
  "backbone_training": "FROZEN",
  "native_execution": {
    "schema": "cke.clef.native-diagnostic.v1",
    "execution": "PASS",
    "numerical_parity": "NOT_TESTED",
    "application_quality": "NOT_TESTED",
    "backbone_executor": "cke_generated",
    "head_executor": "cke_generated",
    "frontend": "pinned_schema_python_and_tokenizers",
    "model": "Cloudflare/clef-flash",
    "revision": "fb4417d31c1d0f9e7e77f1141b40aaede7491e00",
    "release_lock_sha256": "3e909060f2a07e7224482bcf8e36ac75e69bbc822f9fb35337bdfe5f1f53f20b",
    "request_sha256": "bc327750eeb532bae37ab4bcbf1344da73cf2076d884d62ca6e357958e5fd506",
    "tokens": 128,
    "input_ids": [
      248045,
      8678,
      198,
      4274,
      279,
      4434,
      1528,
      321,
      10485,
      13,
      95462,
      1396,
      2002,
      50407,
      13,
      8618,
      4087,
      1902,
      381,
      6681,
      799,
      314,
      421,
      2002,
      579,
      5252,
      2519,
      13,
      248046,
      198,
      248045,
      846,
      198,
      23852,
      25,
      198,
      760,
      2393,
      557,
      8800,
      13,
      271,
      3423,
      34651,
      426,
      32349,
      25,
      198,
      198,
      27806,
      220,
      16,
      198,
      887,
      25,
      25144,
      198,
      9222,
      25,
      5576,
      198,
      669,
      55530,
      25,
      220,
      1890,
      1386,
      25144,
      13,
      198,
      51535,
      1425,
      49555,
      25,
      198,
      32929,
      220,
      16,
      25,
      220,
      4754,
      4532,
      3147,
      37212,
      2129,
      1979,
      816,
      3147,
      40823,
      8934,
      198,
      32929,
      220,
      17,
      25,
      220,
      4754,
      4532,
      3147,
      34317,
      2129,
      1979,
      816,
      3147,
      29492,
      8934,
      198,
      4537,
      38945,
      198,
      198,
      248046,
      198,
      248045,
      74455,
      198,
      248068,
      271,
      248069,
      271,
      25310,
      3118,
      7295,
      34651,
      41402,
      23451,
      50,
      25
    ],
    "feature_sha256": "e68776a4c75c32335b8b57776688e700add5bbc3050c7f82571c0bf3f4eca7bd",
    "embedding_sha256": "361ac064cf98ea04750bdcdbce33b3341b6f2b4ff2de30cc43a20a9d5b2296dd",
    "backbone_bundle_sha256": "a19e9a407da43086b1597aac2f7c8d7afc14eb864c0a84b4d9ccaf6cf5f0f054",
    "head_package_sha256": "b6ab558b77db7fa78b78868989b58c52deab840d00da6428073bb5e521ea8cce",
    "native_math_libraries": {
      "CK_SLEEF_LIBRARY": {
        "path": "/data/cke/tools/sleef-runtime/root/usr/lib/x86_64-linux-gnu/libsleef.so.3.7.0",
        "sha256": "ff8f3513360e8e344ba60f5020b4350f94e043525f83072a1c4fa374f2d84d66"
      },
      "CK_MKL_LIBRARY": {
        "path": "/data/cke/tools/mkl-2024.2.0-probe/wheel/mkl-2024.2.0.data/data/lib/libmkl_rt.so.2",
        "sha256": "0f3c805b85c36f3aaf6792930b054d80a945af032f0a97c4926d7ad94a7d61d9"
      }
    },
    "results": [
      {
        "question_id": "sentiment",
        "candidate_ids": [
          "negative",
          "positive"
        ],
        "logits": [
          -4.680750846862793,
          0.1342177391052246
        ]
      }
    ],
    "elapsed_seconds": 223.3686696900986,
    "verify_and_backbone_seconds": 217.91963743115775,
    "limitations": [
      "request-shaped offline head build",
      "no native backbone training",
      "not a persistent serving endpoint"
    ]
  },
  "training": {
    "schema": "cke.clef.native-training-diagnostic.v1",
    "execution": "PASS",
    "native_report_sha256": "1c87d3ac80a9c9af8fc0c0a813554b9718870754e075e5b4aaef45fcbed6e354",
    "backbone_training": "FROZEN",
    "before_loss": 0.008074787445366383,
    "after_loss": 0.007803728338330984,
    "step_losses": [
      0.008074787445366383
    ],
    "steps": 1,
    "learning_rate": 1e-05,
    "trainable_elements": 121762820,
    "workspace_bytes": 2509897536,
    "workspace_budget_bytes": 4294967296,
    "inference_package_sha256": "226f0458df874d1974360cfb9277393fc57d362bb4fbfe2d7470eeb676a42619",
    "trained_head_sha256": "0bf05bf7135695ca56ef928870f3ec8e7298362a745f0e5f807f73fdd5b03489",
    "gradient_sha256": "b257084ba390b2bf65dfaf2c95b40f28e76c2894923f20e946278244e481abca",
    "reload_max_abs": 0.0,
    "numerical_gradient_parity": "NOT_TESTED",
    "connected_backbone_parity": "NOT_TESTED",
    "held_out_accuracy": "NOT_TESTED",
    "seconds": 30.664285435108468
  },
  "updated_parameter_tensor_count": 162,
  "training_oracle": {
    "schema": "cke.clef.native-training-oracle.v1",
    "native_report_sha256": "1c87d3ac80a9c9af8fc0c0a813554b9718870754e075e5b4aaef45fcbed6e354",
    "training_report_sha256": "fc470f98e981165fdbd09d7eb99f7b388cb768eac0d5dc46b366256a0c28907b",
    "torch_version": "2.13.0+cpu",
    "reference_source_sha256": "0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3",
    "oracle_script_sha256": "0a8d0d6b782a60db2f97e3b4d229c17639ecbe2b54408490970c782ffc01be59",
    "steps": 1,
    "reference_losses": [
      0.008074735291302204
    ],
    "gradient_max": {
      "parameter": "residual_scorer.3.weight",
      "max_abs": 1.2246891856193542e-07
    },
    "updated_weight_max": {
      "parameter": "type_embedding.weight",
      "max_abs": 2.384185791015625e-07
    },
    "gradient_atol": 3e-05,
    "update_atol": 3e-06,
    "gradient_status": "PASS",
    "update_status": "PASS",
    "parameter_tensors": 162,
    "parameter_elements": 121762820
  },
  "connected_prefill_oracle": {
    "schema": "cke.clef.native-oracle.v1",
    "native_report_sha256": "1c87d3ac80a9c9af8fc0c0a813554b9718870754e075e5b4aaef45fcbed6e354",
    "capture_sha256": "fea696d403eac42980b07bed5ff967073dc37a5951d3bd86e5fa92876d553f33",
    "reference_source_sha256": "0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3",
    "torch_version": "2.13.0+cpu",
    "oracle_script_sha256": "94a72d9658c2c562ecafb6f80568ddd04e8b248c5f0efa5a756e7e61d3fb2cdc",
    "backbone_reference_transformers": "5.10.2",
    "backbone_reference_schedule": "prefill",
    "backbone_reference_attention": "sdpa",
    "native_logits": [
      -4.680750846862793,
      0.1342177391052246
    ],
    "same_input_logits": [
      -4.680751800537109,
      0.13421738147735596
    ],
    "connected_reference_logits": [
      -4.677098274230957,
      0.12618112564086914
    ],
    "same_input_head": {
      "max_abs": 9.5367431640625e-07,
      "rmse": 7.202056630608185e-07,
      "reference_rms": 3.3111517425000345,
      "atol": 2e-06,
      "status": "PASS"
    },
    "connected_logits": {
      "max_abs": 0.008036613464355469,
      "rmse": 0.00624213276077562,
      "reference_rms": 3.308411245846123,
      "atol": 2e-06,
      "status": "FAIL"
    },
    "backbone_exact_storage": {
      "max_abs": 1.4375,
      "rmse": 0.024273943936513664,
      "reference_rms": 1.7142072946518132,
      "atol": 0,
      "status": "FAIL"
    },
    "note": "Connected 2e-6 check retains the strict logit baseline; no broader model tolerance is inferred. Zero-tolerance feature comparison is exact-storage evidence."
  },
  "connected_incremental_oracle": {
    "schema": "cke.clef.native-oracle.v1",
    "native_report_sha256": "1c87d3ac80a9c9af8fc0c0a813554b9718870754e075e5b4aaef45fcbed6e354",
    "capture_sha256": "6d85ca7281ef1add094e08f2aa2a70a15148be173fd7d2ba58a0bfa982e736de",
    "reference_source_sha256": "0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3",
    "torch_version": "2.13.0+cpu",
    "oracle_script_sha256": "94a72d9658c2c562ecafb6f80568ddd04e8b248c5f0efa5a756e7e61d3fb2cdc",
    "backbone_reference_transformers": "5.10.2",
    "backbone_reference_schedule": "incremental",
    "backbone_reference_attention": "sdpa",
    "native_logits": [
      -4.680750846862793,
      0.1342177391052246
    ],
    "same_input_logits": [
      -4.680751800537109,
      0.13421738147735596
    ],
    "connected_reference_logits": [
      -4.6752448081970215,
      0.12504327297210693
    ],
    "same_input_head": {
      "max_abs": 9.5367431640625e-07,
      "rmse": 7.202056630608185e-07,
      "reference_rms": 3.3111517425000345,
      "atol": 2e-06,
      "status": "PASS"
    },
    "connected_logits": {
      "max_abs": 0.009174466133117676,
      "rmse": 0.007565953033712733,
      "reference_rms": 3.307079514971539,
      "atol": 2e-06,
      "status": "FAIL"
    },
    "backbone_exact_storage": {
      "max_abs": 1.1875,
      "rmse": 0.025337947930965454,
      "reference_rms": 1.7145243837177928,
      "atol": 0,
      "status": "FAIL"
    },
    "note": "Connected 2e-6 check retains the strict logit baseline; no broader model tolerance is inferred. Zero-tolerance feature comparison is exact-storage evidence."
  },
  "local_contract_replays": {
    "attention_gate": {
      "first_divergence": "layer3.attn_gate_sigmoid_mul",
      "old_provider": "attn_gate_sigmoid_mul_forward",
      "required_provider": "attn_gate_sigmoid_mul_pytorch_bf16_storage",
      "old": {
        "max_abs": 0.004815340042114258,
        "differing": 4096
      },
      "existing_bf16_provider": {
        "max_abs": 0.0,
        "differing": 0
      }
    },
    "rotary": {
      "old_q": {
        "max_abs": 0.017657756805419922,
        "differing": 1007
      },
      "old_k": {
        "max_abs": 0.0050275325775146484,
        "differing": 252
      },
      "existing_bf16_q": {
        "max_abs": 0.0,
        "differing": 0
      },
      "existing_bf16_k": {
        "max_abs": 0.0,
        "differing": 0
      }
    },
    "recurrent_dt_gate": {
      "old": {
        "max_abs": 3.550667315721512e-09,
        "differing": 21
      },
      "existing_pytorch_provider": {
        "max_abs": 0.0,
        "differing": 0
      }
    }
  },
  "optimizer": {
    "name": "AdamW",
    "learning_rate": 1e-05,
    "beta1": 0.9,
    "beta2": 0.999,
    "epsilon": 0.0001,
    "weight_decay": 0.01
  },
  "native_onednn": {
    "version": "3.12",
    "sha256": "9476142d9b469adaa0d0fa07fb2a3c2bd0930c1063bc2779f942a86b40b94b8b"
  },
  "report_sha256": {
    "native-contract-fixed/result.json": "1c87d3ac80a9c9af8fc0c0a813554b9718870754e075e5b4aaef45fcbed6e354",
    "training-contract-final/result.json": "fc470f98e981165fdbd09d7eb99f7b388cb768eac0d5dc46b366256a0c28907b",
    "training-contract-final/oracle-comparison.json": "9c9d72701628c07e3e0e8294c122739f2b6d2f882af72fe18b1dbee9a5e6db20",
    "comparison-contract-fixed-sdpa.json": "37f2b8137b4667151c49339aa135e5a71c651a51b020d3466a25814a8b790268",
    "comparison-contract-fixed-incremental-sdpa.json": "e05bb4eb710349937cbe1a801a573c09f13e6369e5f72909b66445769e5bc97f"
  },
  "tests": {
    "numerical_contracts": {
      "run": 92,
      "skipped": 1,
      "failed": 0
    },
    "ryzen_frontend_evaluation_dataset": {
      "run": 25,
      "skipped": 1,
      "failed": 0
    },
    "dsl_policy": "PASS",
    "registered_nightly_frontend": "PASS",
    "training_ledger_freshness": "PASS",
    "staged_v7_fast": "PASS overall; Nanbeige coherence diagnostic FAIL is nonblocking under the existing matrix policy",
    "staged_v8_fast": "PASS overall; Nanbeige coherence diagnostic FAIL is nonblocking under the existing matrix policy",
    "precommit_resources": {
      "threads": 4,
      "minimum_available_ram_gib": 16,
      "minimum_free_swap_gib": 0,
      "reason": "31 GiB available RAM; swap was full; all regression gates retained"
    }
  },
  "limitations": [
    "Connected numerical parity FAIL remains explicit under unchanged strict logit bound.",
    "One 128-token decision request; no held-out quality claim.",
    "Offline request-shaped head builds, no persistent application endpoint.",
    "One native AdamW update at released geometry; longer/resume tests here are synthetic existing regressions.",
    "Headset/TTS/LoRA are outside this change.",
    "Times include verification, builds and shared-host contention; not latency benchmarks."
  ],
  "private_replay_directory": "/data/cke/private/clef-native-e2e-20261114",
  "state_storage_diagnosis": {
    "native_state_dtype": "fp32",
    "reference_cache_dtype": "bf16",
    "raw_max_abs": 0.007483482360839844,
    "after_diagnostic_bf16_roundtrip_max_abs": 0.0,
    "after_diagnostic_bf16_roundtrip_differing": 0,
    "elements": 524288
  },
  "fp32_cache_reference_diagnostic": {
    "schema": "cke.clef.native-oracle.v1",
    "native_report_sha256": "1c87d3ac80a9c9af8fc0c0a813554b9718870754e075e5b4aaef45fcbed6e354",
    "capture_sha256": "d54a89203944bcf3cdb64f45d7e6122eb8eeecb414650a8b213d68b7cc084cb1",
    "reference_source_sha256": "0e304cf7c6500e8bb59bef7e2afd2c6373f82596dfb3b57d1aa93c175e2dc3a3",
    "torch_version": "2.13.0+cpu",
    "oracle_script_sha256": "94a72d9658c2c562ecafb6f80568ddd04e8b248c5f0efa5a756e7e61d3fb2cdc",
    "backbone_reference_transformers": "5.10.2",
    "backbone_reference_schedule": "incremental",
    "backbone_reference_attention": "sdpa",
    "native_logits": [
      -4.680750846862793,
      0.1342177391052246
    ],
    "same_input_logits": [
      -4.680751800537109,
      0.13421738147735596
    ],
    "connected_reference_logits": [
      -4.678879737854004,
      0.13497495651245117
    ],
    "same_input_head": {
      "max_abs": 9.5367431640625e-07,
      "rmse": 7.202056630608185e-07,
      "reference_rms": 3.3111517425000345,
      "atol": 2e-06,
      "status": "PASS"
    },
    "connected_logits": {
      "max_abs": 0.0018711090087890625,
      "rmse": 0.0014273099040815393,
      "reference_rms": 3.3098439419545365,
      "atol": 2e-06,
      "status": "FAIL"
    },
    "backbone_exact_storage": {
      "max_abs": 1.5,
      "rmse": 0.023975634923971544,
      "reference_rms": 1.7144595174097836,
      "atol": 0,
      "status": "FAIL"
    },
    "note": "Connected 2e-6 check retains the strict logit baseline; no broader model tolerance is inferred. Zero-tolerance feature comparison is exact-storage evidence.",
    "diagnostic_overrides": [
      "LinearAttentionLayer recurrent cache allocated FP32 instead of released BF16. Diagnostic only; not unmodified released execution."
    ]
  },
  "state_diagnosis_scope": "The released incremental cache stores BF16 whereas CKE declares FP32 state. Rounding the captured native first-step state to BF16 reproduces all 524288 reference values. A separately labelled FP32-cache reference diagnostic still fails connected parity; this isolates one cause, not all remaining divergence. No production state rounding or oracle substitution was added.",
  "composition_scope_correction": {
    "scope": "native_components_with_host_handoff",
    "generated_composition": "NOT_IMPLEMENTED",
    "end_to_end_certification": "NOT_TESTED",
    "note": "Historical component measurements remain unchanged. This evidence does not certify one composed generated backbone/head inference or training entry."
  }
}
