{
  "artifacts": [
    {
      "file": "phocinae_largha_150m_L256_fp16fc_int8emb.tflite",
      "sha256": "b080faf19eb9c7d7447fac1c7ddb080330b141257548977653579d183413780b",
      "size_mb": 185.059
    }
  ],
  "benchmarks": [],
  "browser": {
    "backends": [
      {
        "backend": "wasm_xnnpack",
        "date": "2026-10-11",
        "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": 450.27,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": true
      },
      {
        "backend": "webgpu_mldrift",
        "date": "2026-10-11",
        "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": true,
        "latency_p50_ms": 16.978,
        "loads": true,
        "max_rel_diff": 0.00027210438317542537,
        "output_match": true,
        "provenance": "measured",
        "runs": true
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-10-11/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_int8emb.json"
  },
  "conversion": {
    "command": "conversion/ scripts of the Hub repo (export.py per window, quantize.py for the fp16 FC weights and the table forms, tables.py for the host fp16 table); REPRODUCE.md of the Hub repo",
    "quantization": "int8 EMBEDDING_LOOKUP table inside the graph (input_ids form) + fp16 FULLY_CONNECTED weights read through DEQUANTIZE; the int8 table costs max |Δp| 0.0066 to 0.0107 against the reference (card 'Conversion notes'); the L256 graph has 1,804 operators, among them 1 EMBEDDING_LOOKUP and 99 FULLY_CONNECTED",
    "tool": "litert-torch",
    "tool_version": "0.10.0 (ai-edge-litert 2.2.0 for the desktop checks)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [
      "DEQUANTIZE",
      "SQUARED_DIFFERENCE",
      "RESHAPE",
      "EMBEDDING_LOOKUP"
    ],
    "coverage_ops_pct": 91.6,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 1752,
      "unmatched": 52
    },
    "partitions": 150
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1802,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-phocinae-gate",
            "os_build": "Android 36 build BP4A.251205.006.S942QOPS1AZH9",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.phocinae.gate (android/gate/, Kotlin CompiledModel 2.2.0), one cold process per run, fixture ids + markers pushed as JSON, the host token-table lookup (embeds form) outside the measured interval; measured interval = input writes + run + output readback per row.",
            "parity vs the publisher's fp32 PyTorch (phocinae-server 0.1.5 Engine; scripts/device_compare.py): same answer on 602/602 non-near-tie test rows and 19/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0066, mean 7.1e-04 over 853 rows (test + reversed-order + control); bar (argmax 100 % on non-near-tie rows, max |dp| <= 0.02, mean <= 0.002): PASS.",
            "output_match = element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over 2186 option logits of the 622 test rows against the PyTorch logits: 1815 violations (fp16 weights move logits by up to 0.0297; the bar is on probabilities).",
            "latency_p50_ms = the rested run (thermal status 0, CPU / GPU caps off, 45 s rest): a 181-token row, median of 20 calls after 5 warm-ups = 38.96 ms (min 36.90, max 40.43); the 853-row parity run's warm median was 58.03 ms (a long run caps the GPU to 461 MHz and the CPU to ~1.5 GHz by its end).",
            "conditions: screen on (gate activity over the keyguard), USB powered, airplane mode off, one process per run; compile 193 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:06:08.247 25851 25872 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:06:08.330 25851 25872 I tflite  : Replacing 1802 out of 1804 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (main).",
            "10-10 16:06:08.408 25851 25872 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:06:08.408 25851 25872 I tflite  : Replacing 4 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": 38.96224,
          "loads": true,
          "max_abs_diff": 0.02974066138267517,
          "max_rel_diff": 6.092772586094159,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 1815,
            "first_call_ms": 59.90552,
            "fixtures": 622,
            "load_compile_ms": 192.982083,
            "max_abs_dprob_vs_official": 0.006649477079292754,
            "mean_abs_dprob_vs_official": 0.0007081294014530018,
            "near_tie_argmax_equal": 19,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 88.280677,
            "parity_leg_warm_median_ms": 58.030677,
            "parity_leg_warm_min_ms": 36.466146,
            "partitions": 2,
            "rested_leg_max_ms": 40.429115,
            "rested_leg_median_ms": 38.96224,
            "rested_leg_min_ms": 36.89599,
            "rested_leg_n": 20,
            "rested_leg_row_len": 181,
            "rested_leg_warmups": 5,
            "rows_dprob_gt_1e2": 0,
            "threads": 4
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "CPU XNNPACK 4 threads; int8 EMBEDDING_LOOKUP table in the graph (input_ids); debug gate app com.phocinae.gate",
          "total_ops": 1804,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_int8emb__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1804,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-phocinae-gate",
            "os_build": "Android 36 build BP4A.251205.006.S942QOPS1AZH9",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.phocinae.gate (android/gate/, Kotlin CompiledModel 2.2.0), one cold process per run, fixture ids + markers pushed as JSON, the host token-table lookup (embeds form) outside the measured interval; measured interval = input writes + run + output readback per row.",
            "parity vs the publisher's fp32 PyTorch (phocinae-server 0.1.5 Engine; scripts/device_compare.py): same answer on 602/602 non-near-tie test rows and 18/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0141, mean 1.4e-03 over 853 rows (test + reversed-order + control); bar (argmax 100 % on non-near-tie rows, max |dp| <= 0.02, mean <= 0.002): PASS.",
            "output_match = element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over 2186 option logits of the 622 test rows against the PyTorch logits: 1994 violations (fp16 weights move logits by up to 0.0427; the bar is on probabilities).",
            "latency_p50_ms = the rested run (thermal status 0, CPU / GPU caps off, 45 s rest): a 181-token row, median of 20 calls after 5 warm-ups = 15.82 ms (min 15.58, max 17.04); the 853-row parity run's warm median was 15.51 ms (a long run caps the GPU to 461 MHz and the CPU to ~1.5 GHz by its end).",
            "conditions: screen on (gate activity over the keyguard), USB powered, airplane mode off, one process per run; compile 1821 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:00:23.947 22851 22873 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:00:24.045 22851 22873 I tflite  : Replacing 1804 out of 1804 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 16:00:25.736 22851 22873 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:00:25.736 22851 22873 I tflite  : Replacing 4 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 15.82375,
          "loads": true,
          "max_abs_diff": 0.04265248775482178,
          "max_rel_diff": 40.178727114210986,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 1994,
            "first_call_ms": 30.865052,
            "fixtures": 622,
            "load_compile_ms": 1820.779843,
            "max_abs_dprob_vs_official": 0.014109275330179805,
            "mean_abs_dprob_vs_official": 0.0013708386609003742,
            "near_tie_argmax_equal": 18,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 17.403854,
            "parity_leg_warm_median_ms": 15.514479,
            "parity_leg_warm_min_ms": 14.382917,
            "partitions": 1,
            "rested_leg_max_ms": 17.038073,
            "rested_leg_median_ms": 15.82375,
            "rested_leg_min_ms": 15.584844,
            "rested_leg_n": 20,
            "rested_leg_row_len": 181,
            "rested_leg_warmups": 5,
            "rows_dprob_gt_1e2": 7
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP16; int8 EMBEDDING_LOOKUP table in the graph (input_ids); debug gate app com.phocinae.gate",
          "total_ops": 1804,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_int8emb__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1804,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-phocinae-gate",
            "os_build": "Android 36 build BP4A.251205.006.S942QOPS1AZH9",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.phocinae.gate (android/gate/, Kotlin CompiledModel 2.2.0), one cold process per run, fixture ids + markers pushed as JSON, the host token-table lookup (embeds form) outside the measured interval; measured interval = input writes + run + output readback per row.",
            "parity vs the publisher's fp32 PyTorch (phocinae-server 0.1.5 Engine; scripts/device_compare.py): same answer on 602/602 non-near-tie test rows and 19/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0080, mean 7.6e-04 over 853 rows (test + reversed-order + control); bar (argmax 100 % on non-near-tie rows, max |dp| <= 0.02, mean <= 0.002): PASS.",
            "output_match = element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over 2186 option logits of the 622 test rows against the PyTorch logits: 1829 violations (fp16 weights move logits by up to 0.0331; the bar is on probabilities).",
            "latency_p50_ms = the rested run (thermal status 0, CPU / GPU caps off, 45 s rest): a 181-token row, median of 20 calls after 5 warm-ups = 20.82 ms (min 20.40, max 21.04); the 853-row parity run's warm median was 20.46 ms (a long run caps the GPU to 461 MHz and the CPU to ~1.5 GHz by its end).",
            "conditions: screen on (gate activity over the keyguard), USB powered, airplane mode off, one process per run; compile 1789 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 15:57:09.198 21345 21374 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 15:57:09.298 21345 21374 I tflite  : Replacing 1804 out of 1804 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 15:57:10.956 21345 21374 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 15:57:10.956 21345 21374 I tflite  : Replacing 4 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 20.815052,
          "loads": true,
          "max_abs_diff": 0.033109426498413086,
          "max_rel_diff": 9.537053182214473,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 1829,
            "first_call_ms": 32.72,
            "fixtures": 622,
            "load_compile_ms": 1789.143906,
            "max_abs_dprob_vs_official": 0.007950708273498752,
            "mean_abs_dprob_vs_official": 0.0007611683086933379,
            "near_tie_argmax_equal": 19,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 35.026354,
            "parity_leg_warm_median_ms": 20.457448,
            "parity_leg_warm_min_ms": 19.224948,
            "partitions": 1,
            "rested_leg_max_ms": 21.043229,
            "rested_leg_median_ms": 20.815052,
            "rested_leg_min_ms": 20.399532,
            "rested_leg_n": 20,
            "rested_leg_row_len": 181,
            "rested_leg_warmups": 5,
            "rows_dprob_gt_1e2": 0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP16_WITH_FP32_ACCUM; int8 EMBEDDING_LOOKUP table in the graph (input_ids); debug gate app com.phocinae.gate",
          "total_ops": 1804,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_int8emb__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1804,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-phocinae-gate",
            "os_build": "Android 36 build BP4A.251205.006.S942QOPS1AZH9",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.phocinae.gate (android/gate/, Kotlin CompiledModel 2.2.0), one cold process per run, fixture ids + markers pushed as JSON, the host token-table lookup (embeds form) outside the measured interval; measured interval = input writes + run + output readback per row.",
            "parity vs the publisher's fp32 PyTorch (phocinae-server 0.1.5 Engine; scripts/device_compare.py): same answer on 602/602 non-near-tie test rows and 19/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0066, mean 7.1e-04 over 853 rows (test + reversed-order + control); bar (argmax 100 % on non-near-tie rows, max |dp| <= 0.02, mean <= 0.002): PASS.",
            "output_match = element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over 2186 option logits of the 622 test rows against the PyTorch logits: 1815 violations (fp16 weights move logits by up to 0.0297; the bar is on probabilities).",
            "latency_p50_ms = the rested run (thermal status 0, CPU / GPU caps off, 45 s rest): a 181-token row, median of 20 calls after 5 warm-ups = 26.85 ms (min 26.27, max 27.52); the 853-row parity run's warm median was 26.45 ms (a long run caps the GPU to 461 MHz and the CPU to ~1.5 GHz by its end).",
            "conditions: screen on (gate activity over the keyguard), USB powered, airplane mode off, one process per run; compile 1935 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 15:54:12.668 18286 18344 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 15:54:12.776 18286 18344 I tflite  : Replacing 1804 out of 1804 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 15:54:14.492 18286 18344 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 15:54:14.493 18286 18344 I tflite  : Replacing 4 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 26.851146,
          "loads": true,
          "max_abs_diff": 0.029742270708084106,
          "max_rel_diff": 6.092530786835223,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 1815,
            "first_call_ms": 48.177187,
            "fixtures": 622,
            "load_compile_ms": 1934.73104,
            "max_abs_dprob_vs_official": 0.006649785373360806,
            "mean_abs_dprob_vs_official": 0.0007081356582820448,
            "near_tie_argmax_equal": 19,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 29.955937,
            "parity_leg_warm_median_ms": 26.450573,
            "parity_leg_warm_min_ms": 25.465469,
            "partitions": 1,
            "rested_leg_max_ms": 27.521614,
            "rested_leg_median_ms": 26.851146,
            "rested_leg_min_ms": 26.266927,
            "rested_leg_n": 20,
            "rested_leg_row_len": 181,
            "rested_leg_warmups": 5,
            "rows_dprob_gt_1e2": 0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; int8 EMBEDDING_LOOKUP table in the graph (input_ids); debug gate app com.phocinae.gate",
          "total_ops": 1804,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_int8emb__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1803,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-phocinae-gate",
            "os_build": "Android 36 build BP4A.251205.006.S942QOPS1AZH9",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.phocinae.gate (android/gate/, Kotlin CompiledModel 2.2.0), one cold process per run, fixture ids + markers pushed as JSON, the host token-table lookup (embeds form) outside the measured interval; measured interval = input writes + run + output readback per row.",
            "parity vs the publisher's fp32 PyTorch (phocinae-server 0.1.5 Engine; scripts/device_compare.py): same answer on 602/602 non-near-tie test rows and 19/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0090, mean 7.7e-04 over 853 rows (test + reversed-order + control); bar (argmax 100 % on non-near-tie rows, max |dp| <= 0.02, mean <= 0.002): PASS.",
            "output_match = element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over 2186 option logits of the 622 test rows against the PyTorch logits: 1889 violations (fp16 weights move logits by up to 0.0283; the bar is on probabilities).",
            "latency_p50_ms = the parity run's warm median (17.51 ms over 853 rows, min 16.72, max 19.78); no rested run for this form (the parity run heats the phone).",
            "conditions: screen on (gate activity over the keyguard), USB powered, airplane mode off, one process per run; compile 3928 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:05:17.293 25246 25271 I litert  : [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 1803 ops, from a total of 1804 ops. resulted in 2 partitions.",
            "10-10 16:05:20.809 25246 25271 I litert  : [compiled_model.cc:1250] 1 compiler plugins were applied successfully: Qualcomm compiler plugin (ver 0.1.0)",
            "10-10 16:05:21.012 25246 25271 I tflite  : Replacing 2 out of 3 node(s) with delegate (DispatchDelegate) node, yielding 3 partitions for subgraph 0 (main).",
            "10-10 16:05:21.040 25246 25271 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": 17.514896,
          "loads": true,
          "max_abs_diff": 0.0283358097076416,
          "max_rel_diff": 3.8296047395879387,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 1889,
            "first_call_ms": 28.381979,
            "fixtures": 622,
            "load_compile_ms": 3927.937029,
            "max_abs_dprob_vs_official": 0.009007529199987763,
            "mean_abs_dprob_vs_official": 0.000770895468040231,
            "near_tie_argmax_equal": 19,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 19.779011,
            "parity_leg_warm_median_ms": 17.514896,
            "parity_leg_warm_min_ms": 16.717136,
            "partitions": 2,
            "rows_dprob_gt_1e2": 0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "QualcommOptions JIT (HTP, BURST); int8 EMBEDDING_LOOKUP table in the graph (input_ids); debug gate app com.phocinae.gate",
          "total_ops": 1804,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_int8emb__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "phocinae-largha-150m",
    "id": "phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_int8emb",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/Phocinae-Largha-150M-LiteRT",
    "task": "text-classification"
  },
  "pitfalls": [
    "Every graph computes 21 LayerNorms on x·2⁻⁶ with the epsilon scaled to match: bit-identical in fp32; without it the sums of squares inside these norms leave the float16 range, so fp16 arithmetic and the NPU need it (card 'Conversion notes').",
    "Inputs: `input_ids` int32 [1,L] (int8emb) or `inputs_embeds` float32 [1,L,384] (embeds), `attention_mask` float32 [1,L], `qtype_onehot` float32 [1,3] in the order choice, score, noul; outputs `token_logits` float32 [1,L] and `pooled_cls` float32 [1,384]. Gather token_logits at each option's <mask> marker, divide by the question type's temperature from rl_agent_config.json (choice 0.8660, score 0.8081, noul 0.6625) and softmax; the act head takes pooled_cls + four features of the pre-temperature softmax and returns act_logits (card 'Raw CompiledModel contract').",
    "GPU FP16 (float16 arithmetic) passes the bar with less margin: it flipped 1 to 2 of the 20 near ties at L256 and its max |Δp| reaches 0.0188 at L512 (Mac Metal default, embeds); GPU FP32, FP16_WITH_FP32_ACCUM and the NPU keep max |Δp| at or below 0.0137 (card 'Limits').",
    "Dynamic int8 FULLY_CONNECTED fails the bar (max |Δp| 0.0835) and does not ship; the GPU delegate refuses a float16 table inside the graph, so the float16 table ships for host lookup instead (card 'Limits', 'Conversion notes').",
    "A graph computes every position of its window, so a smaller window runs faster; the host cuts the state to fit the window while the publisher's runtime cuts at 512 (card 'Conversion notes', 'Limits').",
    "The checks measure agreement with the publisher's fp32 runtime (phocinae-server) on the public typed-decisions test split, not accuracy against gold labels; measured on one Apple M4 Max and one Galaxy S26 only, not through the hfmodels-android SDK (card 'Checks', 'Limits').",
    "Of the 2,000 test decisions the state is cut on 1,378 at L256 (card 'Limits').",
    "The int8emb form was not run on the NPU at L384 or L512; at L256 it ran with the int8 lookup on the CPU, in a second partition (card 'Limits')."
  ],
  "schema_version": "1.2"
}
