{
  "artifacts": [
    {
      "file": "modernbert_ja_310m_decision_s512_wfp16.tflite",
      "sha256": "f87cda7c513d0b02ee2a98c76885cf048879e4c7e3d7e09e3ecde13e73cc61ec",
      "size_mb": 452.936
    }
  ],
  "benchmarks": [],
  "browser": {
    "backends": [
      {
        "backend": "wasm_xnnpack",
        "date": "2026-10-02",
        "env": {
          "browser": "chromium",
          "browser_version": "151.0.7922.34",
          "headless": true,
          "jspi": true,
          "litertjs_core_version": "2.5.3",
          "machine_label": "mac-studio-m4-max",
          "os": "macOS",
          "os_version": "27.0.0",
          "webgpu_adapter": {
            "architecture": "metal-3",
            "description": "",
            "device": "",
            "vendor": "apple"
          }
        },
        "full_delegation": null,
        "latency_p50_ms": null,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": false
      },
      {
        "backend": "webgpu_mldrift",
        "date": "2026-10-02",
        "env": {
          "browser": "chromium",
          "browser_version": "151.0.7922.34",
          "headless": true,
          "jspi": true,
          "litertjs_core_version": "2.5.3",
          "machine_label": "mac-studio-m4-max",
          "os": "macOS",
          "os_version": "27.0.0",
          "webgpu_adapter": {
            "architecture": "metal-3",
            "description": "",
            "device": "",
            "vendor": "apple"
          }
        },
        "full_delegation": null,
        "latency_p50_ms": null,
        "loads": false,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": false
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-10-02/modernbert-ja-310m-decision__modernbert_ja_310m_decision_s512_wfp16.json"
  },
  "conversion": {
    "command": "python scripts/export.py --seqs 256,512  (REPRODUCE.md: fp32 export of scripts/graph.py DecisionGraph at the fixed window, then the weight-only float16 cast; exports/modernbert_ja_310m_decision_s512_wfp16.tflite)",
    "quantization": "float16 weight storage (ai-edge-quantizer 0.8.0 FLOAT_CASTING, weight-only 16-bit on FULLY_CONNECTED): the 102 FULLY_CONNECTED weight tensors stored as float16 with a DEQUANTIZE to float32; the RoPE and band constants and all activations stay float32; 1,911 operators vs 1,809 in the fp32 export; no INT8 file is shipped",
    "tool": "litert-torch",
    "tool_version": "0.9.3 (torch 2.12.1; transformers 4.57.6; ai-edge-litert 2.1.6; ai-edge-quantizer 0.8.0 for the float16 weight storage)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [
      "DEQUANTIZE",
      "SQUARED_DIFFERENCE"
    ],
    "coverage_ops_pct": 91.9,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 1859,
      "unmatched": 52
    },
    "partitions": 155
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1911,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-mbja-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.decisionja.gate (android_gate/, Kotlin CompiledModel 2.2.0, Environment with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, fixture pair ids pushed as JSON, host float16 table lookup (android.util.Half for float16) outside the timed interval; timed interval = input writes + run + output readback per PAIR (one graph call per candidate).",
            "conditions: before thermal status 3, battery 44.7 C; after thermal status 3, battery 44.8 C; screen on (FLAG_KEEP_SCREEN_ON), USB powered.",
            "parity vs the reference (the card's transformers snippet, transformers 5.17.0 CPU fp32): same winning candidate on 60/60 questions (3/3 boundary = reference top-1 < 0.9); max |dprob| 0.02017; questions over 0.01: 2.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same pair ids with the float16 table: 160 violating of 180 compared pair logits, max |d| 0.11835193634033203.",
            "10-02 12:01:34.525 11230 11253 I tflite  : Replacing 1911 out of 1911 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 12:01:37.099 11230 11253 I DecisionJaGate: DECISION_GATE compiled in 2834.707916 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 230.960469,
          "loads": true,
          "max_abs_diff": 0.11835193634033203,
          "max_rel_diff": 0.6323969407113241,
          "metrics": {
            "argmax_equal_reference": 60,
            "battery_temp_c_end": 44.8,
            "battery_temp_c_start": 44.7,
            "boundary_argmax_equal_reference": 3,
            "boundary_questions": 3,
            "element_rule_elements": 180,
            "element_rule_violations": 160,
            "first_pair_ms": 102.916979,
            "fixtures": 60,
            "host_table_lookup_median_ms": 21.201197,
            "load_compile_ms": 2834.707916,
            "max_abs_dlogit_vs_reference": 0.11906,
            "max_abs_dprob_vs_reference": 0.02017,
            "median_request_ms": 691.1,
            "nonfinite_questions": 0,
            "pairs": 180,
            "questions": 60,
            "questions_dprob_gt_1e2": 2,
            "thermal_status_end": 3,
            "thermal_status_start": 3,
            "warm_max_pair_ms": 273.452968,
            "warm_median_pair_ms": 230.960469,
            "warm_min_pair_ms": 204.170364,
            "warmup_pairs_excluded": 5
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=DEFAULT (fp16 kernels); debug gate app com.decisionja.gate; first 60 JNLI requests",
          "total_ops": 1911,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/modernbert-ja-310m-decision__modernbert_ja_310m_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1911,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-mbja-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.decisionja.gate (android_gate/, Kotlin CompiledModel 2.2.0, Environment with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, fixture pair ids pushed as JSON, host float16 table lookup (android.util.Half for float16) outside the timed interval; timed interval = input writes + run + output readback per PAIR (one graph call per candidate).",
            "conditions: before thermal status 3, battery 44.8 C; after thermal status 3, battery 45.0 C; screen on (FLAG_KEEP_SCREEN_ON), USB powered.",
            "parity vs the reference (the card's transformers snippet, transformers 5.17.0 CPU fp32): same winning candidate on 60/60 questions (3/3 boundary = reference top-1 < 0.9); max |dprob| 0.02017; questions over 0.01: 2.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same pair ids with the float16 table: 160 violating of 180 compared pair logits, max |d| 0.11835193634033203.",
            "10-02 12:03:50.925 11983 12003 I tflite  : Replacing 1911 out of 1911 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 12:03:53.490 11983 12003 I DecisionJaGate: DECISION_GATE compiled in 2777.867707 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 252.711458,
          "loads": true,
          "max_abs_diff": 0.11835193634033203,
          "max_rel_diff": 0.6323969407113241,
          "metrics": {
            "argmax_equal_reference": 60,
            "battery_temp_c_end": 45.0,
            "battery_temp_c_start": 44.8,
            "boundary_argmax_equal_reference": 3,
            "boundary_questions": 3,
            "element_rule_elements": 180,
            "element_rule_violations": 160,
            "first_pair_ms": 100.581146,
            "fixtures": 60,
            "host_table_lookup_median_ms": 25.023073,
            "load_compile_ms": 2777.867707,
            "max_abs_dlogit_vs_reference": 0.11906,
            "max_abs_dprob_vs_reference": 0.02017,
            "median_request_ms": 747.9,
            "nonfinite_questions": 0,
            "pairs": 180,
            "questions": 60,
            "questions_dprob_gt_1e2": 2,
            "thermal_status_end": 3,
            "thermal_status_start": 3,
            "warm_max_pair_ms": 290.267135,
            "warm_median_pair_ms": 252.711458,
            "warm_min_pair_ms": 219.264531,
            "warmup_pairs_excluded": 5
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP16; debug gate app com.decisionja.gate; first 60 JNLI requests",
          "total_ops": 1911,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/modernbert-ja-310m-decision__modernbert_ja_310m_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1911,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-mbja-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.decisionja.gate (android_gate/, Kotlin CompiledModel 2.2.0, Environment with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, fixture pair ids pushed as JSON, host float16 table lookup (android.util.Half for float16) outside the timed interval; timed interval = input writes + run + output readback per PAIR (one graph call per candidate).",
            "conditions: before thermal status 3, battery 44.8 C; after thermal status 3, battery 44.8 C; screen on (FLAG_KEEP_SCREEN_ON), USB powered.",
            "parity vs the reference (the card's transformers snippet, transformers 5.17.0 CPU fp32): same winning candidate on 60/60 questions (3/3 boundary = reference top-1 < 0.9); max |dprob| 0.00446; questions over 0.01: 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same pair ids with the float16 table: 100 violating of 180 compared pair logits, max |d| 0.025213241577148438.",
            "10-02 12:02:27.016 11620 11638 I tflite  : Replacing 1911 out of 1911 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 12:02:29.508 11620 11638 I DecisionJaGate: DECISION_GATE compiled in 2705.541718 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 402.7025,
          "loads": true,
          "max_abs_diff": 0.025213241577148438,
          "max_rel_diff": 0.2966132881222582,
          "metrics": {
            "argmax_equal_reference": 60,
            "battery_temp_c_end": 44.8,
            "battery_temp_c_start": 44.8,
            "boundary_argmax_equal_reference": 3,
            "boundary_questions": 3,
            "element_rule_elements": 180,
            "element_rule_violations": 100,
            "first_pair_ms": 149.505989,
            "fixtures": 60,
            "host_table_lookup_median_ms": 23.053282,
            "load_compile_ms": 2705.541718,
            "max_abs_dlogit_vs_reference": 0.02715,
            "max_abs_dprob_vs_reference": 0.00446,
            "median_request_ms": 1207.0,
            "nonfinite_questions": 0,
            "pairs": 180,
            "questions": 60,
            "questions_dprob_gt_1e2": 0,
            "thermal_status_end": 3,
            "thermal_status_start": 3,
            "warm_max_pair_ms": 530.977864,
            "warm_median_pair_ms": 402.7025,
            "warm_min_pair_ms": 398.358594,
            "warmup_pairs_excluded": 5
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP16_WITH_FP32_ACCUM; debug gate app com.decisionja.gate; first 60 JNLI requests",
          "total_ops": 1911,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/modernbert-ja-310m-decision__modernbert_ja_310m_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1911,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-mbja-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.decisionja.gate (android_gate/, Kotlin CompiledModel 2.2.0, Environment with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, fixture pair ids pushed as JSON, host float16 table lookup (android.util.Half for float16) outside the timed interval; timed interval = input writes + run + output readback per PAIR (one graph call per candidate).",
            "conditions: before thermal status 0, battery 34.9 C; after thermal status 2, battery 44.8 C; screen on (FLAG_KEEP_SCREEN_ON), USB powered.",
            "parity vs the reference (the card's transformers snippet, transformers 5.17.0 CPU fp32): same winning candidate on 621/621 questions (123/123 boundary = reference top-1 < 0.9); max |dprob| 0.00079; questions over 0.01: 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same pair ids with the float16 table: 0 violating of 2357 compared pair logits, max |d| 3.814697265625e-05.",
            "conditions: the phone warmed from thermal status 0 (34.9 C) to 2 (44.8 C) over the 547 back-to-back requests (17 min); per-pair time rose from 183.6 ms (minimum) to 669.0 ms (maximum), so the median covers a thermally drifting run.",
            "10-02 10:59:32.688 28258 28279 I tflite  : Replacing 1911 out of 1911 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 10:59:34.693 28258 28279 I DecisionJaGate: DECISION_GATE compiled in 2197.237134 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 416.436562,
          "loads": true,
          "max_abs_diff": 3.814697265625e-05,
          "max_rel_diff": 0.01196046664418476,
          "metrics": {
            "argmax_equal_reference": 621,
            "battery_temp_c_end": 44.8,
            "battery_temp_c_start": 34.9,
            "boundary_argmax_equal_reference": 123,
            "boundary_questions": 123,
            "element_rule_elements": 2357,
            "element_rule_violations": 0,
            "first_pair_ms": 192.30224,
            "fixtures": 547,
            "host_table_lookup_median_ms": 26.437709,
            "load_compile_ms": 2197.237134,
            "max_abs_dlogit_vs_reference": 0.01274,
            "max_abs_dprob_vs_reference": 0.00079,
            "max_abs_dscore_or_dnoul_vs_reference": 0.00047,
            "median_request_ms": 2014.4,
            "nonfinite_questions": 0,
            "pairs": 2357,
            "questions": 621,
            "questions_dprob_gt_1e2": 0,
            "thermal_status_end": 2,
            "thermal_status_start": 0,
            "warm_max_pair_ms": 669.020313,
            "warm_median_pair_ms": 416.436562,
            "warm_min_pair_ms": 183.575938,
            "warmup_pairs_excluded": 5
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; debug gate app com.decisionja.gate",
          "total_ops": 1911,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/modernbert-ja-310m-decision__modernbert_ja_310m_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-mbja-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.decisionja.gate (android_gate/, Kotlin CompiledModel 2.2.0, Environment with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, fixture pair ids pushed as JSON, host float16 table lookup (android.util.Half for float16) outside the timed interval; timed interval = input writes + run + output readback per PAIR (one graph call per candidate).",
            "conditions: before thermal status 3, battery 46.5 C; after thermal status 3, battery 47.0 C; screen on (FLAG_KEEP_SCREEN_ON), USB powered.",
            "parity vs the reference (the card's transformers snippet, transformers 5.17.0 CPU fp32): same winning candidate on 621/621 questions (123/123 boundary = reference top-1 < 0.9); max |dprob| 0.01128; questions over 0.01: 1.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same pair ids with the float16 table: 1332 violating of 2357 compared pair logits, max |d| 0.06473922729492188.",
            "The whole graph ran as one DispatchDelegate node ('Replacing 1 out of 1 node(s) with delegate (DispatchDelegate)'); JIT compile 83.0 s; every winning candidate kept (621/621), max |dprob| 0.0113 on one question; the HTP computes in fp16, so the element rule against the CPU fp32 run is not met.",
            "10-02 11:32:25.630  3485  3505 I tflite  : Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 11:32:25.819  3485  3505 I DecisionJaGate: DECISION_GATE compiled in 83034.787624 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 142.595885,
          "loads": true,
          "max_abs_diff": 0.06473922729492188,
          "max_rel_diff": 15.810512959102347,
          "metrics": {
            "argmax_equal_reference": 621,
            "battery_temp_c_end": 47.0,
            "battery_temp_c_start": 46.5,
            "boundary_argmax_equal_reference": 123,
            "boundary_questions": 123,
            "element_rule_elements": 2357,
            "element_rule_violations": 1332,
            "first_pair_ms": 112.616042,
            "fixtures": 547,
            "host_table_lookup_median_ms": 25.51375,
            "jit_compile_s": 83.0,
            "load_compile_ms": 83034.787624,
            "max_abs_dlogit_vs_reference": 0.06562,
            "max_abs_dprob_vs_reference": 0.01128,
            "max_abs_dscore_or_dnoul_vs_reference": 0.01128,
            "median_request_ms": 705.8,
            "nonfinite_questions": 0,
            "pairs": 2357,
            "questions": 621,
            "questions_dprob_gt_1e2": 1,
            "thermal_status_end": 3,
            "thermal_status_start": 3,
            "warm_max_pair_ms": 151.699427,
            "warm_median_pair_ms": 142.595885,
            "warm_min_pair_ms": 110.642135,
            "warmup_pairs_excluded": 5
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "QualcommOptions htpPerformanceMode=BURST, JIT (first compile, no cache); debug gate app com.decisionja.gate",
          "total_ops": 1,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/modernbert-ja-310m-decision__modernbert_ja_310m_decision_s512_wfp16__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "modernbert-ja-310m",
    "id": "modernbert-ja-310m-decision__modernbert_ja_310m_decision_s512_wfp16",
    "license": "cc-by-sa-4.0",
    "source_url": "https://huggingface.co/mlboydaisuke/ModernBERT-Ja-310M-Decision-LiteRT",
    "task": "typed decisions in Japanese (question + state + candidates -> one logit per (context, candidate) pair; softmax over a question's candidates = choice / score / true-false; text-classification)"
  },
  "pitfalls": [
    "Run the GPU with explicit FP32 precision (GpuOptions precision = FP32 on Android, enforce_f32=True in Python). The shipped graphs carry two fp32-exact rewrites (mask constant -1e4; SafeLayerNorm 2^-k, k 5-6 in layers 14-24 and 2 on the final norm, because the residual reaches 2,694), so fp16 computation is finite, but it still moves answers: Mac Metal default precision keeps 619 of 621 winners (32 questions over 0.01, max 0.042); the Galaxy S26 GPU at default precision or FP16 keeps all 60 winners of a 60-request probe with 2 questions over 0.01 (max 0.0202); FP16_WITH_FP32_ACCUM stays within 0.0045 on those 60 but runs as slowly as explicit FP32.",
    "The Galaxy S26 Hexagon NPU (JIT, BURST) compiles both windows whole (29.5 s at S256, 83.0 s at S512) and keeps every winning candidate on 618 / 621 questions, but one question moves by 0.0113 (bar 0.01; the same with a float32 host table, so it is the HTP's fp16 arithmetic). The card lists the NPU with that number; the verified path is the GPU with explicit FP32. NPU warm medians 41.7 ms per 256-token pair and 142.6 ms per 512-token pair against 229.2 and 416.4 ms on the GPU with explicit FP32 on the same hot phone.",
    "Host contract (the author's serving code): context = '質問: {instructions}\\n状況: {state}' (a non-string state is json.dumps(ensure_ascii=False)); candidate = '{label}' or '{label} — {description}' (U+2014 with one space each side); choice labels = criteria keys, score labels = '0'..'K-1', noul labels = 'true' / 'false'; one pair per candidate '<s> context </s><s> candidate </s>' from tokenizer.json with truncation='only_first' at 512 (the context end is cut, never the candidate); pad id 3; softmax at temperature 1 over a question's candidates; confidence = 1 - H(p)/ln K.",
    "transformers 4.57.6 cannot load this repository's tokenizer class (TokenizersBackend, saved by transformers 5.17.0); the tokenizers library on tokenizer.json gives the reference ids on all 2,357 fixture pairs, and so does the base model's tokenizer.json.",
    "The checkpoint is F32, so the float16 token table is a rounding (max 6.1e-5, 104 of 78.6M values flush to zero): fp32 graph + float16 table differs from the reference by up to 0.00039 in probability, the shipped float16 weights + float16 table by 0.00079, every winner kept on all 621 questions.",
    "One graph call per candidate: a request costs one call per candidate across its questions (the 547 fixture requests took 2,357 calls); the S256 graph raises on a pair longer than 256 tokens (546 of the 547 fixture requests fit).",
    "The S26 GPU warms under load: the S512 explicit-FP32 run started at thermal status 0 (34.9 C) and ended at status 2 (44.8 C) after 17 minutes, per-pair time from 183.6 ms (minimum) to 669.0 ms (maximum, median 416.4 ms); every later row started hot.",
    "Japanese; per the author only Japanese and English were evaluated, business domains such as customer support were not in the training data, and wrong answers can come with high probability. The conversion reproduces the reference's answers, including the wrong ones (our JGLUE test subset: JNLI 238/250, JCommonsenseQA 233/250; the author's 16 items 14/16)."
  ],
  "schema_version": "1.2"
}
