{
  "artifacts": [
    {
      "file": "phocinae_largha_150m_L256_fp16fc_embeds.tflite",
      "sha256": "2ba31415bd13dbbdfec6f99241cedc3fc7539a68ffbaab5230953e3db404e4fa",
      "size_mb": 88.387
    }
  ],
  "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": 454.525,
        "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": 17.063,
        "loads": true,
        "max_rel_diff": 0.0033851218060208478,
        "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_embeds.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": "host-lookup form: fp16 FULLY_CONNECTED weights read through DEQUANTIZE; the token table stays outside the graph (tables/tok_embeddings_fp16.bin, the publisher's float16 rows unchanged, [256000,384]) and the host passes inputs_embeds (card 'Conversion notes', 'Files')",
    "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"
    ],
    "coverage_ops_pct": 91.7,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 1751,
      "unmatched": 50
    },
    "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": 1801,
          "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 20/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0004, mean 4.8e-05 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: 389 violations (fp16 weights move logits by up to 0.0015; 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 = 38.77 ms (min 37.41, max 39.72); the 853-row parity run's warm median was 66.89 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 186 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:08:16.019 26561 26581 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:08:16.102 26561 26581 I tflite  : Replacing 1801 out of 1801 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 16:08:16.183 26561 26581 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:08:16.183 26561 26581 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": 38.767083,
          "loads": true,
          "max_abs_diff": 0.0014727115631103516,
          "max_rel_diff": 0.9260522099090017,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 389,
            "first_call_ms": 50.448803,
            "fixtures": 622,
            "load_compile_ms": 186.116094,
            "max_abs_dprob_vs_official": 0.0003830439767066318,
            "mean_abs_dprob_vs_official": 4.77275382496114e-05,
            "near_tie_argmax_equal": 20,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 90.719427,
            "parity_leg_warm_median_ms": 66.888333,
            "parity_leg_warm_min_ms": 36.96724,
            "partitions": 1,
            "rested_leg_max_ms": 39.718073,
            "rested_leg_median_ms": 38.767083,
            "rested_leg_min_ms": 37.406718,
            "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; host fp16 token table (inputs_embeds); debug gate app com.phocinae.gate",
          "total_ops": 1801,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_embeds__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1801,
          "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| 0.0158, mean 1.2e-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: 1982 violations (fp16 weights move logits by up to 0.0542; 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 = 15.33 ms (min 14.96, max 15.78); the 853-row parity run's warm median was 14.87 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 835 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:02:02.129 23672 23700 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:02:02.226 23672 23700 I tflite  : Replacing 1801 out of 1801 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 16:02:02.942 23672 23700 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:02:02.942 23672 23700 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": 15.331458,
          "loads": true,
          "max_abs_diff": 0.054184556007385254,
          "max_rel_diff": 37.612031386224935,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 1982,
            "first_call_ms": 22.327396,
            "fixtures": 622,
            "load_compile_ms": 834.969895,
            "max_abs_dprob_vs_official": 0.015765709631448188,
            "mean_abs_dprob_vs_official": 0.0011716419948953072,
            "near_tie_argmax_equal": 19,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 17.781771,
            "parity_leg_warm_median_ms": 14.870781,
            "parity_leg_warm_min_ms": 14.355989,
            "partitions": 1,
            "rested_leg_max_ms": 15.776823,
            "rested_leg_median_ms": 15.331458,
            "rested_leg_min_ms": 14.964427,
            "rested_leg_n": 20,
            "rested_leg_row_len": 181,
            "rested_leg_warmups": 5,
            "rows_dprob_gt_1e2": 3
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP16; host fp16 token table (inputs_embeds); debug gate app com.phocinae.gate",
          "total_ops": 1801,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_embeds__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1801,
          "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 20/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0027, mean 3.0e-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: 1446 violations (fp16 weights move logits by up to 0.0157; 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 = 20.36 ms (min 19.92, max 20.85); the 853-row parity run's warm median was 19.79 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 804 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 15:59:03.792 22134 22161 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 15:59:03.890 22134 22161 I tflite  : Replacing 1801 out of 1801 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 15:59:04.574 22134 22161 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 15:59:04.574 22134 22161 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": 20.357761,
          "loads": true,
          "max_abs_diff": 0.015710830688476562,
          "max_rel_diff": 5.57541412380122,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 1446,
            "first_call_ms": 30.5075,
            "fixtures": 622,
            "load_compile_ms": 804.250729,
            "max_abs_dprob_vs_official": 0.0026761315330231517,
            "mean_abs_dprob_vs_official": 0.0002951882459147474,
            "near_tie_argmax_equal": 20,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 22.105104,
            "parity_leg_warm_median_ms": 19.794688,
            "parity_leg_warm_min_ms": 19.373698,
            "partitions": 1,
            "rested_leg_max_ms": 20.850625,
            "rested_leg_median_ms": 20.357761,
            "rested_leg_min_ms": 19.921927,
            "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; host fp16 token table (inputs_embeds); debug gate app com.phocinae.gate",
          "total_ops": 1801,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_embeds__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1801,
          "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 20/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0004, mean 4.8e-05 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: 385 violations (fp16 weights move logits by up to 0.0015; the bar is on probabilities).",
            "latency_p50_ms = the parity run's warm median (28.87 ms over 853 rows, min 25.29, max 38.03); 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 918 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 15:55:29.787 19436 19465 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 15:55:29.895 19436 19465 I tflite  : Replacing 1801 out of 1801 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 15:55:30.608 19436 19465 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 15:55:30.609 19436 19465 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": 28.867031,
          "loads": true,
          "max_abs_diff": 0.0014727115631103516,
          "max_rel_diff": 0.9206625980819529,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 385,
            "first_call_ms": 32.332083,
            "fixtures": 622,
            "load_compile_ms": 917.83151,
            "max_abs_dprob_vs_official": 0.0003834208927294691,
            "mean_abs_dprob_vs_official": 4.772231888366135e-05,
            "near_tie_argmax_equal": 20,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 38.031197,
            "parity_leg_warm_median_ms": 28.867031,
            "parity_leg_warm_min_ms": 25.287344,
            "partitions": 1,
            "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; host fp16 token table (inputs_embeds); debug gate app com.phocinae.gate",
          "total_ops": 1801,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_embeds__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-10-10",
          "decode_tokens_per_s": null,
          "delegated_ops": 1801,
          "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 20/20 near ties (reference top-2 gap <= 0.02), max |dp| 0.0031, mean 3.9e-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: 1617 violations (fp16 weights move logits by up to 0.0100; 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 = 16.79 ms (min 16.58, max 17.24); the 853-row parity run's warm median was 16.49 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 3902 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:03:49.773 24498 24595 I litert  : [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 1801 ops, from a total of 1801 ops. resulted in 1 partitions.",
            "10-10 16:03:53.266 24498 24595 I litert  : [compiled_model.cc:1250] 1 compiler plugins were applied successfully: Qualcomm compiler plugin (ver 0.1.0)",
            "10-10 16:03:53.407 24498 24595 I tflite  : Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 16:03:53.437 24498 24595 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 16.793229,
          "loads": true,
          "max_abs_diff": 0.010011449456214905,
          "max_rel_diff": 6.8256320836966,
          "metrics": {
            "all_rows": 853,
            "argmax_equal_official": 602,
            "bar_pass": 1,
            "element_rule_elements": 2186,
            "element_rule_violations": 1617,
            "first_call_ms": 27.478333,
            "fixtures": 622,
            "load_compile_ms": 3902.481978,
            "max_abs_dprob_vs_official": 0.0031231408733266575,
            "mean_abs_dprob_vs_official": 0.0003946382496177136,
            "near_tie_argmax_equal": 20,
            "near_tie_rows": 20,
            "non_near_tie_rows": 602,
            "parity_leg_warm_max_ms": 19.538854,
            "parity_leg_warm_median_ms": 16.492448,
            "parity_leg_warm_min_ms": 16.118803,
            "partitions": 1,
            "rested_leg_max_ms": 17.241406,
            "rested_leg_median_ms": 16.793229,
            "rested_leg_min_ms": 16.581823,
            "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": "QualcommOptions JIT (HTP, BURST); host fp16 token table (inputs_embeds); debug gate app com.phocinae.gate",
          "total_ops": 1801,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_embeds__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "phocinae-largha-150m",
    "id": "phocinae-largha-150m__phocinae_largha_150m_l256_fp16fc_embeds",
    "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')."
  ],
  "schema_version": "1.2"
}
