{
  "artifacts": [
    {
      "file": "phocinae_largha_150m_L384_fp16fc_embeds.tflite",
      "sha256": "8a836463fe3cb7bf4231728013726b81aa50bbac9e7859d0a38b9f04639331e6",
      "size_mb": 89.074
    }
  ],
  "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": 701.555,
        "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": 20.103,
        "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_l384_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": "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 1552/1552 non-near-tie test rows and 42/45 near ties (reference top-2 gap <= 0.02), max |dp| 0.0158, mean 1.1e-03 over 2051 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 5486 option logits of the 1597 test rows against the PyTorch logits: 4944 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 294-token row, median of 20 calls after 5 warm-ups = 22.44 ms (min 21.78, max 23.25); the 2051-row parity run's warm median was 54.21 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 1321 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:24:40.829 31682 31706 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:24:40.937 31682 31706 I tflite  : Replacing 1801 out of 1801 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 16:24:42.121 31682 31706 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:24:42.121 31682 31706 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": 22.439323,
          "loads": true,
          "max_abs_diff": 0.054184556007385254,
          "max_rel_diff": 37.612031386224935,
          "metrics": {
            "all_rows": 2051,
            "argmax_equal_official": 1552,
            "bar_pass": 1,
            "element_rule_elements": 5486,
            "element_rule_violations": 4944,
            "first_call_ms": 28.233958,
            "fixtures": 1597,
            "load_compile_ms": 1321.233385,
            "max_abs_dprob_vs_official": 0.015765709631448188,
            "mean_abs_dprob_vs_official": 0.001062051447606741,
            "near_tie_argmax_equal": 42,
            "near_tie_rows": 45,
            "non_near_tie_rows": 1552,
            "parity_leg_warm_max_ms": 64.979531,
            "parity_leg_warm_median_ms": 54.214114,
            "parity_leg_warm_min_ms": 21.103073,
            "partitions": 1,
            "rested_leg_max_ms": 23.253854,
            "rested_leg_median_ms": 22.439323,
            "rested_leg_min_ms": 21.779583,
            "rested_leg_n": 20,
            "rested_leg_row_len": 294,
            "rested_leg_warmups": 5,
            "rows_dprob_gt_1e2": 5
          },
          "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_l384_fp16fc_embeds__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 1801,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-phocinae-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Zoo sample app com.phocinae debug gate (github.com/john-rocky/LiteRT-Models phocinae/, PR #75): the app's own tokenizer + builder + graphs on the 1,597 public test rows of window 384.",
            "the app's token ids and marker positions equal the reference host's on 1597/1597 and 1597/1597 rows; same answer on 1552/1552 non-near-tie rows and 42/45 near ties, max |dp| 0.0158, mean 1.1e-03; bar PASS.",
            "output_match = element rule over 5486 option logits against the PyTorch logits: 4944 violations (max |dlogit| 0.0542).",
            "latency_p50_ms = this 1,597-row run's warm median 22.93 ms (min 21.19, max 58.91; write + run + readback of the main graph); the GPU runs heat the phone inside the run (the FP32 run's minimum is the rested value; the rested runs are the 2026-10-10 results of this record).",
            "conditions: screen on (the activity shows over the keyguard and keeps the screen on), USB powered, airplane mode off, one process per run, a fresh install.",
            "10-11 13:58:26.228 24398 24450 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-11 13:58:26.363 24398 24450 I tflite  : Replacing 1801 out of 1801 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-11 13:58:27.097 24398 24450 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-11 13:58:27.097 24398 24450 I tflite  : Replacing 4 out of 4 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 22.93,
          "loads": true,
          "max_abs_diff": 0.054184556007385254,
          "max_rel_diff": 37.612031386224935,
          "metrics": {
            "argmax_equal_official": 1552,
            "bar_pass": 1,
            "element_rule_elements": 5486,
            "element_rule_violations": 4944,
            "fixtures": 1597,
            "ids_exact": 1597,
            "load_compile_ms": 921.150364,
            "markers_exact": 1597,
            "max_abs_dprob_vs_official": 0.015765709631448188,
            "mean_abs_dprob_vs_official": 0.001094955176385271,
            "near_tie_argmax_equal": 42,
            "near_tie_rows": 45,
            "non_near_tie_rows": 1552,
            "parity_leg_warm_max_ms": 58.91,
            "parity_leg_warm_median_ms": 22.93,
            "parity_leg_warm_min_ms": 21.19,
            "partitions": 1
          },
          "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); zoo sample app com.phocinae debug gate",
          "total_ops": 1801,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l384_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 1552/1552 non-near-tie test rows and 45/45 near ties (reference top-2 gap <= 0.02), max |dp| 0.0027, mean 2.7e-04 over 2051 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 5486 option logits of the 1597 test rows against the PyTorch logits: 3494 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 294-token row, median of 20 calls after 5 warm-ups = 30.00 ms (min 29.70, max 30.82); the 2051-row parity run's warm median was 48.89 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 820 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:19:45.830 29959 30003 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:19:45.916 29959 30003 I tflite  : Replacing 1801 out of 1801 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 16:19:46.627 29959 30003 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-10 16:19:46.627 29959 30003 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": 29.996458,
          "loads": true,
          "max_abs_diff": 0.015710830688476562,
          "max_rel_diff": 5.57541412380122,
          "metrics": {
            "all_rows": 2051,
            "argmax_equal_official": 1552,
            "bar_pass": 1,
            "element_rule_elements": 5486,
            "element_rule_violations": 3494,
            "first_call_ms": 35.871093,
            "fixtures": 1597,
            "load_compile_ms": 819.613802,
            "max_abs_dprob_vs_official": 0.0026761315330231517,
            "mean_abs_dprob_vs_official": 0.0002723596836087681,
            "near_tie_argmax_equal": 45,
            "near_tie_rows": 45,
            "non_near_tie_rows": 1552,
            "parity_leg_warm_max_ms": 75.557708,
            "parity_leg_warm_median_ms": 48.89151,
            "parity_leg_warm_min_ms": 28.624844,
            "partitions": 1,
            "rested_leg_max_ms": 30.816563,
            "rested_leg_median_ms": 29.996458,
            "rested_leg_min_ms": 29.697448,
            "rested_leg_n": 20,
            "rested_leg_row_len": 294,
            "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_l384_fp16fc_embeds__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 1801,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-phocinae-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Zoo sample app com.phocinae debug gate (github.com/john-rocky/LiteRT-Models phocinae/, PR #75): the app's own tokenizer + builder + graphs on the 1,597 public test rows of window 384.",
            "the app's token ids and marker positions equal the reference host's on 1597/1597 and 1597/1597 rows; same answer on 1552/1552 non-near-tie rows and 45/45 near ties, max |dp| 0.0008, mean 4.6e-05; bar PASS.",
            "output_match = element rule over 5486 option logits against the PyTorch logits: 867 violations (max |dlogit| 0.0024).",
            "latency_p50_ms = this 1,597-row run's warm median 68.76 ms (min 40.75, max 88.85; write + run + readback of the main graph); the GPU runs heat the phone inside the run (the FP32 run's minimum is the rested value; the rested runs are the 2026-10-10 results of this record).",
            "conditions: screen on (the activity shows over the keyguard and keeps the screen on), USB powered, airplane mode off, one process per run, a fresh install.",
            "10-11 14:00:20.428 25186 25237 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-11 14:00:20.542 25186 25237 I tflite  : Replacing 1801 out of 1801 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-11 14:00:21.300 25186 25237 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-11 14:00:21.300 25186 25237 I tflite  : Replacing 4 out of 4 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 68.76,
          "loads": true,
          "max_abs_diff": 0.0023908615112304688,
          "max_rel_diff": 1.0894130204207673,
          "metrics": {
            "argmax_equal_official": 1552,
            "bar_pass": 1,
            "element_rule_elements": 5486,
            "element_rule_violations": 867,
            "fixtures": 1597,
            "ids_exact": 1597,
            "load_compile_ms": 925.512187,
            "markers_exact": 1597,
            "max_abs_dprob_vs_official": 0.0008037586650582185,
            "mean_abs_dprob_vs_official": 4.572825618006555e-05,
            "near_tie_argmax_equal": 45,
            "near_tie_rows": 45,
            "non_near_tie_rows": 1552,
            "parity_leg_warm_max_ms": 88.85,
            "parity_leg_warm_median_ms": 68.76,
            "parity_leg_warm_min_ms": 40.75,
            "partitions": 1
          },
          "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); zoo sample app com.phocinae debug gate",
          "total_ops": 1801,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l384_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 1552/1552 non-near-tie test rows and 43/45 near ties (reference top-2 gap <= 0.02), max |dp| 0.0032, mean 3.9e-04 over 2051 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 5486 option logits of the 1597 test rows against the PyTorch logits: 3949 violations (fp16 weights move logits by up to 0.0129; the bar is on probabilities).",
            "latency_p50_ms = the rested run (thermal status 0, CPU / GPU caps off, 45 s rest): a 294-token row, median of 20 calls after 5 warm-ups = 19.69 ms (min 19.16, max 20.69); the 2051-row parity run's warm median was 19.82 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 11035 ms (NPU: JIT, cold cache on the first run of a graph).",
            "10-10 16:27:33.177 32544 32568 I litert  : [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 1801 ops, from a total of 1801 ops. resulted in 1 partitions.",
            "10-10 16:27:43.787 32544 32568 I litert  : [compiled_model.cc:1250] 1 compiler plugins were applied successfully: Qualcomm compiler plugin (ver 0.1.0)",
            "10-10 16:27:43.956 32544 32568 I tflite  : Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "10-10 16:27:44.002 32544 32568 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 19.69052,
          "loads": true,
          "max_abs_diff": 0.01288747787475586,
          "max_rel_diff": 6.8256320836966,
          "metrics": {
            "all_rows": 2051,
            "argmax_equal_official": 1552,
            "bar_pass": 1,
            "element_rule_elements": 5486,
            "element_rule_violations": 3949,
            "first_call_ms": 22.137968,
            "fixtures": 1597,
            "load_compile_ms": 11034.788642,
            "max_abs_dprob_vs_official": 0.0031869103701956947,
            "mean_abs_dprob_vs_official": 0.00038690939980504135,
            "near_tie_argmax_equal": 43,
            "near_tie_rows": 45,
            "non_near_tie_rows": 1552,
            "parity_leg_warm_max_ms": 21.851563,
            "parity_leg_warm_median_ms": 19.819947,
            "parity_leg_warm_min_ms": 18.758333,
            "partitions": 1,
            "rested_leg_max_ms": 20.685416,
            "rested_leg_median_ms": 19.69052,
            "rested_leg_min_ms": 19.158593,
            "rested_leg_n": 20,
            "rested_leg_row_len": 294,
            "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_l384_fp16fc_embeds__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 4,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-phocinae-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Zoo sample app com.phocinae debug gate (github.com/john-rocky/LiteRT-Models phocinae/, PR #75): the app's own tokenizer + builder + graphs on the 1,597 public test rows of window 384.",
            "the app's token ids and marker positions equal the reference host's on 1597/1597 and 1597/1597 rows; same answer on 1552/1552 non-near-tie rows and 43/45 near ties, max |dp| 0.0032, mean 4.0e-04; bar PASS.",
            "output_match = element rule over 5486 option logits against the PyTorch logits: 3949 violations (max |dlogit| 0.0129).",
            "latency_p50_ms = this 1,597-row run's warm median 19.77 ms (min 19.01, max 21.65; write + run + readback of the main graph); the GPU runs heat the phone inside the run (the FP32 run's minimum is the rested value; the rested runs are the 2026-10-10 results of this record).",
            "conditions: screen on (the activity shows over the keyguard and keeps the screen on), USB powered, airplane mode off, one process per run, a fresh install.",
            "10-11 14:22:51.296 31037 31091 I tflite  : Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "10-11 14:22:51.341 31037 31091 I litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully:",
            "10-11 14:22:51.342 31037 31091 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": 19.77,
          "loads": true,
          "max_abs_diff": 0.01288747787475586,
          "max_rel_diff": 6.8256320836966,
          "metrics": {
            "argmax_equal_official": 1552,
            "bar_pass": 1,
            "element_rule_elements": 5486,
            "element_rule_violations": 3949,
            "fixtures": 1597,
            "ids_exact": 1597,
            "load_compile_ms": 151.641563,
            "markers_exact": 1597,
            "max_abs_dprob_vs_official": 0.0031869103701956947,
            "mean_abs_dprob_vs_official": 0.00040227126542836993,
            "near_tie_argmax_equal": 43,
            "near_tie_rows": 45,
            "non_near_tie_rows": 1552,
            "parity_leg_warm_max_ms": 21.65,
            "parity_leg_warm_median_ms": 19.77,
            "parity_leg_warm_min_ms": 19.01,
            "partitions": 1
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "QualcommOptions JIT (HTP, BURST, cached compile); host fp16 token table (inputs_embeds); zoo sample app com.phocinae debug gate",
          "total_ops": 4,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-10/phocinae-largha-150m__phocinae_largha_150m_l384_fp16fc_embeds__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "phocinae-largha-150m",
    "id": "phocinae-largha-150m__phocinae_largha_150m_l384_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 403 at L384 (card 'Limits')."
  ],
  "schema_version": "1.2"
}
