{
  "artifacts": [
    {
      "file": "vela_L512_embeds_wfp16_shift_arith.tflite",
      "sha256": "fbb052ef8008953a4c16d7bca68fd1f88d8af69e7d8b9253adf2355bed84415e",
      "size_mb": 212.853
    }
  ],
  "benchmarks": [],
  "browser": {
    "backends": [
      {
        "backend": "wasm_xnnpack",
        "date": "2026-10-18",
        "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": 1789.8,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": true
      },
      {
        "backend": "webgpu_mldrift",
        "date": "2026-10-18",
        "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": 38.43,
        "loads": true,
        "max_rel_diff": 0.008672780672525941,
        "output_match": true,
        "provenance": "measured",
        "runs": true
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16.json"
  },
  "conversion": {
    "command": "the conversion scripts and the full record are in github.com/john-rocky/hf-to-litertlm vela_work/ (card 'Licence and source'); the card names the tools and the graph, not a command line",
    "quantization": "host-table form: fp16 FULLY_CONNECTED weights read through DEQUANTIZE; the token table stays outside the graph (tok_embeddings_fp16.bin, the checkpoint's float16 rows) and the host passes inputs_embeds (card 'Files', 'Conversion')",
    "tool": "litert-torch",
    "tool_version": "0.10.0 (litert-converter 0.5.0.dev20260912; ai-edge-litert 2.2.0 for the desktop checks)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [
      "DEQUANTIZE",
      "SQUARED_DIFFERENCE"
    ],
    "coverage_ops_pct": 92.1,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 1645,
      "unmatched": 45
    },
    "partitions": 134
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-18",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-vela-npu-runner",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL SM-S942Q)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850 (Adreno GPU, Hexagon HTP)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "the LiteRT C-API runner vela_npu_runner (the r17 runner with the Qualcomm QNN HTP libraries pushed to /data/local/tmp/vela_npu), one process per run: the GPU at FP32 precision through the same runner (the control for the HTP runs).",
            "parity vs the publisher's fp32 stack (sentence-transformers 5.1.2 + transformers 4.57.6, CPU, SDPA math) on the 140 device rows of this window (20 edge512, 40 long256, 80 long512): cosine min 0.999974 (p01 0.999997, mean 0.999999), top-1 retrieval 70/70 on non-near-tie queries (0 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999895 over 70 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): PASS.",
            "output_match null: this harness compares the dumped vectors to the fp32 reference only (no same-graph CPU dump); the per-row cosine metrics carry the comparison.",
            "latency_p50_ms = the runner's latency phase: 5 warm-ups then 20 rounds of one call (fixture 000, host fp16 token table (inputs_embeds)), median 133.264 ms (min 130.162, max 138.795); the 140-row parity run's median call was 122.566 ms; compile 1875 ms.",
            "conditions: adb shell, screen state as left by the chain, USB powered, one process per run; thermal status before the run: Thermal Status: 1. Command: cd /data/local/tmp/vela_npu && rm -rf /data/local/tmp/vela_npu/out_N1_gpu32_emb_L512 && LD_LIBRARY_PATH=/data/local/tmp/vela_npu ADSP_LIBRARY_PATH='/data/local/tmp/vela_npu;/vendor/dsp/cdsp;/vendor/lib/rfsa/adsp;/system/lib/rfsa/adsp;/dsp' ./vela_npu_runner --accel gpu --gpu-precision fp32 --burst 1 --libdir /data/local/tmp/vela_npu  --fixtures /data/local/tmp/vela_npu/fx_L512_embeds --count 141 --warmup 5 --rounds 20 --ab-fixture 000 --dump full --out /data/local/tmp/vela_npu/out_N1_gpu32_emb_L512 --model /data/local/tmp/vela_npu/vela_L512_embeds_wfp16_shift_arith.tflite 2>/data/local/tmp/vela_npu/N1_gpu32_emb_L512.stderr.txt",
            "VERBOSE: Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 133.26350000000002,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999992672380982,
            "cos_min_vs_reference": 0.9999738314339787,
            "cos_p01_vs_reference": 0.9999965912785455,
            "finite": 1,
            "fixtures": 140,
            "fully_accelerated": 1,
            "liveness_moved": 1,
            "load_compile_ms": 1874.9,
            "max_abs_vs_reference": 0.0012090019881725311,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00036840677819949263,
            "pair_spearman_vs_reference": 0.9998950223077595,
            "pairs": 70,
            "parity_leg_call_ms_median": 122.566,
            "partitions": 1,
            "runner_call_ms_max": 138.795,
            "runner_call_ms_median": 133.26350000000002,
            "runner_call_ms_min": 130.162,
            "runner_rounds": 20,
            "runner_warmups": 5,
            "top1_agree": 70,
            "top1_near_tie": 0,
            "top1_queries": 70
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "C-API runner --accel gpu, precision fp32; host fp16 token table (inputs_embeds); the runner's GPU control run",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-31",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL SM-S942Q)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850 (Adreno GPU, Hexagon HTP)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Harness: the gate app (LiteRT 2.2.0 Kotlin CompiledModel) driven by a benchmark script in generic mode, one cold process per run, the rows pushed as raw tensors (the host fp16 token-table lookup done on the Mac before the push)",
            "run (2026-10-31): rows file vela_scifact_L512.json, 3142 rows, 0 non-finite, compile 967.9 ms, warm median write + run + read 127.88 ms, ready: check run, not waited for clean (thermal 1, cap policy0:2227200/3628800 policy6:2227200/4742400 , kgsl_pwrlevel 0); rested 11s since 04:24:55; \"sha256 vela_L512_embeds_fp16.tflite phone=fbb052ef8008953a… MATCH\"; delegation: Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "the work line (registered 2026-10-30 22:24 JST before any number): SciFact (mteb/scifact revision d56462d0, CC BY-NC 2.0: 300 test queries over the corpus) embedded on the phone in the four window files together, nDCG@10 and recall@10 each >= the publisher's stack in fp32 (sentence-transformers, the same texts) − 0.005; scored on the Mac: line {\"ndcg_ge_ref_minus_0p005\": true, \"recall_ge_ref_minus_0p005\": true, \"verdict\": \"PASS\"}; the verdict belongs to the four files together, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "leg1_compile_ms": 967.9,
            "leg1_rows": 3142.0,
            "leg1_warm_median_write_run_read_ms": 127.877,
            "legs": 1.0,
            "nonfinite_rows": 0.0,
            "pipeline_row_cos_min": 0.995885,
            "pipeline_rows_under_0p999": 11.0,
            "pipeline_top1_agree": 288.0,
            "rows": 3142.0,
            "scifact_line_pass": 1.0,
            "scifact_ndcg_at_10": 0.488134,
            "scifact_ndcg_at_10_reference": 0.485806,
            "scifact_ndcg_delta": 0.002328,
            "scifact_queries": 300.0,
            "scifact_recall_at_10": 0.607222,
            "scifact_recall_at_10_reference": 0.608333,
            "scifact_recall_delta": -0.001111
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=DEFAULT (fp16); host fp16 token table (inputs_embeds); gate app com.mlboydaisuke.d1omni.gate (generic mode); the SciFact texts of this window, scored as one retrieval run over the four window files",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-31/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-18",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL SM-S942Q)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850 (Adreno GPU, Hexagon HTP)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "gate app com.mlboydaisuke.d1omni.gate in generic mode (Kotlin CompiledModel, LiteRT 2.2.0), one cold process per run, the fixture rows pushed as raw tensors, the host token-table lookup (embeds form) done on the Mac before the push; measured interval = input writes + run + output read-back per row.",
            "parity vs the publisher's fp32 stack (sentence-transformers 5.1.2 + transformers 4.57.6, CPU, SDPA math) on the 140 device rows of this window (20 edge512, 40 long256, 80 long512): cosine min 0.997381 (p01 0.999390, mean 0.999818), top-1 retrieval 70/70 on non-near-tie queries (0 near ties, reference top-2 gap <= 0.01), pair Spearman 0.995451 over 70 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): FAIL.",
            "output_match = the Mac CPU run of the same file on the same rows (ai-edge-litert 2.2.0 XNNPACK, the chain's mac-baseline step): max |out - mac| 1.01e-02, min cos 0.99743004 over 141 rows; rule max_abs <= 1e-4 and cos >= 0.9999 on the unit-norm 768-d vectors -> False. max_rel_diff is not computed by the chain (null).",
            "latency_p50_ms = this check run's warm median over 136 rows (100.55 ms; min 58.51, max 124.79; write + run + read-back per text, the gate app's own clock) on a working phone (thermal status before the run: Thermal Status: 2) — INFORMATIONAL: no rested timing run was run on this device for this form, and marks these numbers as check-run medians.",
            "conditions: screen on (the gate activity over the keyguard), USB powered, one process per run; thermal status before the run: Thermal Status: 2; compile 1545 ms.",
            "10-18 10:56:34.733 24091 24115 I tflite  : Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 100.54872399999999,
          "loads": true,
          "max_abs_diff": 0.010116569697856903,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 0,
            "cos_mean_vs_reference": 0.9998176348494313,
            "cos_min_vs_reference": 0.9973808537154156,
            "cos_p01_vs_reference": 0.9993902367127422,
            "finite": 1,
            "fixtures": 140,
            "gate_leg_rows": 141,
            "gate_leg_warm_max_ms": 124.789896,
            "gate_leg_warm_median_run_ms": 0.9720575,
            "gate_leg_warm_median_write_run_read_ms": 100.54872399999999,
            "gate_leg_warm_min_ms": 58.514427,
            "gate_leg_warm_rows": 136,
            "liveness_moved": 1,
            "load_compile_ms": 1544.640989,
            "mac_same_graph_max_abs": 0.010116569697856903,
            "mac_same_graph_min_cos": 0.997430035772439,
            "max_abs_vs_reference": 0.01015893742442131,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.006688202357850059,
            "pair_spearman_vs_reference": 0.9954509666695825,
            "pairs": 70,
            "partitions": 1,
            "top1_agree": 70,
            "top1_near_tie": 0,
            "top1_queries": 70
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP16 (the delegate default); host fp16 token table (inputs_embeds); gate app com.mlboydaisuke.d1omni.gate (generic mode)",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-18",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL SM-S942Q)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850 (Adreno GPU, Hexagon HTP)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "gate app com.mlboydaisuke.d1omni.gate in generic mode (Kotlin CompiledModel, LiteRT 2.2.0), one cold process per run, the fixture rows pushed as raw tensors, the host token-table lookup (embeds form) done on the Mac before the push; measured interval = input writes + run + output read-back per row.",
            "parity vs the publisher's fp32 stack (sentence-transformers 5.1.2 + transformers 4.57.6, CPU, SDPA math) on the 140 device rows of this window (20 edge512, 40 long256, 80 long512): cosine min 0.999421 (p01 0.999476, mean 0.999971), top-1 retrieval 70/70 on non-near-tie queries (0 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999475 over 70 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): PASS.",
            "output_match = the Mac CPU run of the same file on the same rows (ai-edge-litert 2.2.0 XNNPACK, the chain's mac-baseline step): max |out - mac| 3.55e-03, min cos 0.99980876 over 141 rows; rule max_abs <= 1e-4 and cos >= 0.9999 on the unit-norm 768-d vectors -> False. max_rel_diff is not computed by the chain (null).",
            "latency_p50_ms = the timing run (ready=timing: thermal status before the run Thermal Status: 0; 5 warm-ups then 20 calls on one row, one text per call, write + run + read-back): the 510-token row 83.86 ms (min 82.48, max 86.66), the 197-token row 83.73 ms — the time is the window's, not the text's; compile 934 ms. The 136-row check run's warm median was 152.30 ms.",
            "conditions: screen on (the gate activity over the keyguard), USB powered, one process per run; thermal status before the run: Thermal Status: 2; compile 1479 ms.",
            "10-18 10:55:59.730 23299 23326 I tflite  : Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 83.86,
          "loads": true,
          "max_abs_diff": 0.0035481639206409454,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999705853180109,
            "cos_min_vs_reference": 0.9994206258803751,
            "cos_p01_vs_reference": 0.9994759049758716,
            "finite": 1,
            "fixtures": 140,
            "gate_leg_rows": 141,
            "gate_leg_warm_max_ms": 170.182083,
            "gate_leg_warm_median_run_ms": 0.953646,
            "gate_leg_warm_median_write_run_read_ms": 152.3037765,
            "gate_leg_warm_min_ms": 82.488177,
            "gate_leg_warm_rows": 136,
            "liveness_moved": 1,
            "load_compile_ms": 1478.701614,
            "mac_same_graph_max_abs": 0.0035481639206409454,
            "mac_same_graph_min_cos": 0.9998087641486526,
            "max_abs_vs_reference": 0.0034506507217884064,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.0020715468917842372,
            "pair_spearman_vs_reference": 0.9994751115387979,
            "pairs": 70,
            "partitions": 1,
            "timing_leg_compile_ms": 933.895156,
            "timing_leg_long_max_ms": 86.657083,
            "timing_leg_long_median_ms": 83.86,
            "timing_leg_long_min_ms": 82.480417,
            "timing_leg_long_row_len": 510,
            "timing_leg_n": 20,
            "timing_leg_short_median_ms": 83.731979,
            "timing_leg_short_row_len": 197,
            "timing_leg_warmups": 5,
            "top1_agree": 70,
            "top1_near_tie": 0,
            "top1_queries": 70
          },
          "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); gate app com.mlboydaisuke.d1omni.gate (generic mode)",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-31",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL SM-S942Q)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850 (Adreno GPU, Hexagon HTP)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Harness: the gate app (LiteRT 2.2.0 Kotlin CompiledModel) driven by a benchmark script in generic mode, one cold process per run, the rows pushed as raw tensors (the host fp16 token-table lookup done on the Mac before the push)",
            "run (2026-10-31): rows file vela_scifact_L512.json, 2897 rows, 0 non-finite, compile 945.0 ms, warm median write + run + read 209.02 ms, ready: check run, not waited for clean (thermal 1, cap policy0:2496000/3628800 policy6:2438400/4742400 , kgsl_pwrlevel 0); rested 11s since 05:24:25; \"sha256 vela_L512_embeds_fp16.tflite phone=fbb052ef8008953a… MATCH\"; delegation: Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "run (2026-10-31): rows file vela_scifact_L512_t2897.json, 245 rows, 0 non-finite, compile 1053.1 ms, warm median write + run + read 122.86 ms, ready: check run, not waited for clean (thermal 1, cap none, kgsl_pwrlevel 0); rested 38s since 06:26:50; \"sha256 vela_L512_embeds_fp16.tflite phone=fbb052ef8008953a… MATCH\"; delegation: Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "the work line (registered 2026-10-30 22:24 JST before any number): SciFact (mteb/scifact revision d56462d0, CC BY-NC 2.0: 300 test queries over the corpus) embedded on the phone in the four window files together, nDCG@10 and recall@10 each >= the publisher's stack in fp32 (sentence-transformers, the same texts) − 0.005; scored on the Mac: line {\"ndcg_ge_ref_minus_0p005\": true, \"recall_ge_ref_minus_0p005\": true, \"verdict\": \"PASS\"}; the verdict belongs to the four files together, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "leg1_compile_ms": 945.0,
            "leg1_rows": 2897.0,
            "leg1_warm_median_write_run_read_ms": 209.021,
            "leg2_compile_ms": 1053.1,
            "leg2_rows": 245.0,
            "leg2_warm_median_write_run_read_ms": 122.863,
            "legs": 2.0,
            "nonfinite_rows": 0.0,
            "pipeline_row_cos_min": 0.998436,
            "pipeline_rows_under_0p999": 2.0,
            "pipeline_top1_agree": 297.0,
            "rows": 3142.0,
            "scifact_line_pass": 1.0,
            "scifact_ndcg_at_10": 0.485758,
            "scifact_ndcg_at_10_reference": 0.485806,
            "scifact_ndcg_delta": -4.8e-05,
            "scifact_queries": 300.0,
            "scifact_recall_at_10": 0.608333,
            "scifact_recall_at_10_reference": 0.608333,
            "scifact_recall_delta": 0.0
          },
          "output_match": null,
          "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); gate app com.mlboydaisuke.d1omni.gate (generic mode); the SciFact texts of this window, scored as one retrieval run over the four window files",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-31/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-18",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL SM-S942Q)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850 (Adreno GPU, Hexagon HTP)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "gate app com.mlboydaisuke.d1omni.gate in generic mode (Kotlin CompiledModel, LiteRT 2.2.0), one cold process per run, the fixture rows pushed as raw tensors, the host token-table lookup (embeds form) done on the Mac before the push; measured interval = input writes + run + output read-back per row.",
            "parity vs the publisher's fp32 stack (sentence-transformers 5.1.2 + transformers 4.57.6, CPU, SDPA math) on the 140 device rows of this window (20 edge512, 40 long256, 80 long512): cosine min 0.999974 (p01 0.999997, mean 0.999999), top-1 retrieval 70/70 on non-near-tie queries (0 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999895 over 70 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): PASS.",
            "output_match = the Mac CPU run of the same file on the same rows (ai-edge-litert 2.2.0 XNNPACK, the chain's mac-baseline step): max |out - mac| 2.07e-06, min cos 1.00000000 over 141 rows; rule max_abs <= 1e-4 and cos >= 0.9999 on the unit-norm 768-d vectors -> True. max_rel_diff is not computed by the chain (null).",
            "latency_p50_ms = the timing run (ready=timing: thermal status before the run Thermal Status: 0; 5 warm-ups then 20 calls on one row, one text per call, write + run + read-back): the 510-token row 119.55 ms (min 118.74, max 122.63), the 197-token row 119.41 ms — the time is the window's, not the text's; compile 1037 ms. The 136-row check run's warm median was 137.32 ms.",
            "conditions: screen on (the gate activity over the keyguard), USB powered, one process per run; thermal status before the run: Thermal Status: 2; compile 1173 ms.",
            "10-18 10:55:24.879 22424 22465 I tflite  : Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 119.5477865,
          "loads": true,
          "max_abs_diff": 2.0675361156463623e-06,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999992672380982,
            "cos_min_vs_reference": 0.9999738314339787,
            "cos_p01_vs_reference": 0.9999965912785455,
            "finite": 1,
            "fixtures": 140,
            "gate_leg_rows": 141,
            "gate_leg_warm_max_ms": 163.444427,
            "gate_leg_warm_median_run_ms": 0.9538545,
            "gate_leg_warm_median_write_run_read_ms": 137.3237755,
            "gate_leg_warm_min_ms": 118.754427,
            "gate_leg_warm_rows": 136,
            "liveness_moved": 1,
            "load_compile_ms": 1173.124531,
            "mac_same_graph_max_abs": 2.0675361156463623e-06,
            "mac_same_graph_min_cos": 0.9999999999187538,
            "max_abs_vs_reference": 0.0012090019881725311,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00036840677819949263,
            "pair_spearman_vs_reference": 0.9998950223077595,
            "pairs": 70,
            "partitions": 1,
            "timing_leg_compile_ms": 1037.080989,
            "timing_leg_long_max_ms": 122.626458,
            "timing_leg_long_median_ms": 119.5477865,
            "timing_leg_long_min_ms": 118.744115,
            "timing_leg_long_row_len": 510,
            "timing_leg_n": 20,
            "timing_leg_short_median_ms": 119.4138025,
            "timing_leg_short_row_len": 197,
            "timing_leg_warmups": 5,
            "top1_agree": 70,
            "top1_near_tie": 0,
            "top1_queries": 70
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; host fp16 token table (inputs_embeds); gate app com.mlboydaisuke.d1omni.gate (generic mode)",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-10-18",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-vela-npu-runner",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL SM-S942Q)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850 (Adreno GPU, Hexagon HTP)",
            "vendor_sdk": "Qualcomm QNN HTP libraries (v81) in /data/local/tmp/vela_npu"
          },
          "error": null,
          "evidence": [
            "the LiteRT C-API runner vela_npu_runner (the r17 runner with the Qualcomm QNN HTP libraries pushed to /data/local/tmp/vela_npu), one process per run: the HTP through the Qualcomm compiler plugin (JIT, a fresh cache dir per run).",
            "parity vs the publisher's fp32 stack (sentence-transformers 5.1.2 + transformers 4.57.6, CPU, SDPA math) on the 140 device rows of this window (20 edge512, 40 long256, 80 long512): cosine min 0.999517 (p01 0.999573, mean 0.999994), top-1 retrieval 70/70 on non-near-tie queries (0 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999125 over 70 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): PASS.",
            "output_match null: this harness compares the dumped vectors to the fp32 reference only (no same-graph CPU dump); the per-row cosine metrics carry the comparison.",
            "latency_p50_ms = the runner's latency phase: 5 warm-ups then 20 rounds of one call (fixture 000, host fp16 token table (inputs_embeds)), median 68.400 ms (min 67.327, max 69.823); the 140-row parity run's median call was 68.280 ms; compile 33746 ms (JIT, cold cache).",
            "conditions: adb shell, screen state as left by the chain, USB powered, one process per run; thermal status before the run: Thermal Status: 0. Command: cd /data/local/tmp/vela_npu && rm -rf /data/local/tmp/vela_npu/out_N1_npu_emb_L512 && LD_LIBRARY_PATH=/data/local/tmp/vela_npu ADSP_LIBRARY_PATH='/data/local/tmp/vela_npu;/vendor/dsp/cdsp;/vendor/lib/rfsa/adsp;/system/lib/rfsa/adsp;/dsp' ./vela_npu_runner --accel npu+cpu  --burst 1 --libdir /data/local/tmp/vela_npu --cache-dir /data/local/tmp/vela_npu/cache_emb512 --fixtures /data/local/tmp/vela_npu/fx_L512_embeds --count 141 --warmup 5 --rounds 20 --ab-fixture 000 --dump full --out /data/local/tmp/vela_npu/out_N1_npu_emb_L512 --model /data/local/tmp/vela_npu/vela_L512_embeds_wfp16_shift_arith.tflite 2>/data/local/tmp/vela_npu/N1_npu_emb_L512.stderr.txt",
            "INFO: [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 1690 ops, from a total of 1690 ops. resulted in 1 partitions.",
            "VERBOSE: Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 68.4005,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999939714526391,
            "cos_min_vs_reference": 0.9995169025872218,
            "cos_p01_vs_reference": 0.9995729007424946,
            "finite": 1,
            "fixtures": 140,
            "fully_accelerated": 1,
            "liveness_moved": 1,
            "load_compile_ms": 33745.8,
            "max_abs_vs_reference": 0.0032675452530384064,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.0015145541556194075,
            "pair_spearman_vs_reference": 0.9991251858979966,
            "pairs": 70,
            "parity_leg_call_ms_median": 68.28,
            "partitions": 1,
            "runner_call_ms_max": 69.823,
            "runner_call_ms_median": 68.4005,
            "runner_call_ms_min": 67.327,
            "runner_rounds": 20,
            "runner_warmups": 5,
            "top1_agree": 70,
            "top1_near_tie": 0,
            "top1_queries": 70
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "QualcommOptions JIT (HTP, burst), cold compile cache per run; host fp16 token table (inputs_embeds); C-API runner --accel npu+cpu",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__galaxy-s26.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-31",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Pixel 8a",
            "machine_label": "pixel-8a-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL Pixel 8a)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Google Tensor G3 (Mali-G715 GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Harness: the gate app (LiteRT 2.2.0 Kotlin CompiledModel) driven by a benchmark script in generic mode, one cold process per run, the rows pushed as raw tensors (the host fp16 token-table lookup done on the Mac before the push)",
            "run (2026-10-31): rows file vela_scifact_L512.json, 2602 rows, 0 non-finite, compile 4739.5 ms, warm median write + run + read 346.81 ms, ready: check run, not waited for clean (thermal 2, cap policy0:955000/1704000 policy4:712000/2367000 policy8:880000/2914000 , gpu_cooling 0); rested 15s since 03:02:34; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L512_embeds_fp16.tflite phone=fbb052ef8008953a… MATCH\"; delegation: no LITERT_CL line in this run's captured logcat (the buffer turned over during the long run); the other run of this window and precision printed it",
            "run (2026-10-31): rows file vela_scifact_L512_r2602.json, 540 rows, 0 non-finite, compile 1565.2 ms, warm median write + run + read 214.21 ms, ready: clean; rested 298s since 04:02:27; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L512_embeds_fp16.tflite phone=fbb052ef8008953a… MATCH\"; delegation: Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "the work line (registered 2026-10-30 22:24 JST before any number): SciFact (mteb/scifact revision d56462d0, CC BY-NC 2.0: 300 test queries over the corpus) embedded on the phone in the four window files together, nDCG@10 and recall@10 each >= the publisher's stack in fp32 (sentence-transformers, the same texts) − 0.005; scored on the Mac: line {\"ndcg_ge_ref_minus_0p005\": true, \"recall_ge_ref_minus_0p005\": true, \"verdict\": \"PASS\"}; the verdict belongs to the four files together, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "leg1_compile_ms": 4739.5,
            "leg1_rows": 2602.0,
            "leg1_warm_median_write_run_read_ms": 346.808,
            "leg2_compile_ms": 1565.2,
            "leg2_rows": 540.0,
            "leg2_warm_median_write_run_read_ms": 214.213,
            "legs": 2.0,
            "nonfinite_rows": 0.0,
            "pipeline_row_cos_min": 0.995571,
            "pipeline_rows_under_0p999": 9.0,
            "pipeline_top1_agree": 289.0,
            "rows": 3142.0,
            "scifact_line_pass": 1.0,
            "scifact_ndcg_at_10": 0.486991,
            "scifact_ndcg_at_10_reference": 0.485806,
            "scifact_ndcg_delta": 0.001185,
            "scifact_queries": 300.0,
            "scifact_recall_at_10": 0.607222,
            "scifact_recall_at_10_reference": 0.608333,
            "scifact_recall_delta": -0.001111
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=DEFAULT (fp16); host fp16 token table (inputs_embeds); gate app com.mlboydaisuke.d1omni.gate (generic mode); the SciFact texts of this window, scored as one retrieval run over the four window files",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-31/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__pixel-8a.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-18",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Pixel 8a",
            "machine_label": "pixel-8a-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL Pixel 8a)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Google Tensor G3 (Mali-G715 GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "gate app com.mlboydaisuke.d1omni.gate in generic mode (Kotlin CompiledModel, LiteRT 2.2.0), one cold process per run, the fixture rows pushed as raw tensors, the host token-table lookup (embeds form) done on the Mac before the push; measured interval = input writes + run + output read-back per row.",
            "parity vs the publisher's fp32 stack (sentence-transformers 5.1.2 + transformers 4.57.6, CPU, SDPA math) on the 140 device rows of this window (20 edge512, 40 long256, 80 long512): cosine min 0.996044 (p01 0.999329, mean 0.999865), top-1 retrieval 70/70 on non-near-tie queries (0 near ties, reference top-2 gap <= 0.01), pair Spearman 0.992442 over 70 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): FAIL.",
            "output_match = the Mac CPU run of the same file on the same rows (ai-edge-litert 2.2.0 XNNPACK, the chain's mac-baseline step): max |out - mac| 1.17e-02, min cos 0.99602276 over 141 rows; rule max_abs <= 1e-4 and cos >= 0.9999 on the unit-norm 768-d vectors -> False. max_rel_diff is not computed by the chain (null).",
            "latency_p50_ms = this check run's warm median over 136 rows (209.03 ms; min 206.44, max 212.79; write + run + read-back per text, the gate app's own clock) on a working phone (thermal status before the run: Thermal Status: 0) — INFORMATIONAL: no rested timing run was run on this device for this form, and marks these numbers as check-run medians; on Mali the GPU work completes inside the read-back, so run_ms alone (16.9 ms) is not the time.",
            "conditions: screen on (the gate activity over the keyguard), USB powered, one process per run; thermal status before the run: Thermal Status: 0; compile 1403 ms.",
            "10-18 10:30:51.031 16872 16897 I tflite  : Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 209.0309655,
          "loads": true,
          "max_abs_diff": 0.011711552739143372,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 0,
            "cos_mean_vs_reference": 0.9998650859435518,
            "cos_min_vs_reference": 0.9960444885769689,
            "cos_p01_vs_reference": 0.9993293357928408,
            "finite": 1,
            "fixtures": 140,
            "gate_leg_rows": 141,
            "gate_leg_warm_max_ms": 212.788371,
            "gate_leg_warm_median_run_ms": 16.9336345,
            "gate_leg_warm_median_write_run_read_ms": 209.0309655,
            "gate_leg_warm_min_ms": 206.440592,
            "gate_leg_warm_rows": 136,
            "liveness_moved": 1,
            "load_compile_ms": 1403.39974,
            "mac_same_graph_max_abs": 0.011711552739143372,
            "mac_same_graph_min_cos": 0.9960227553545654,
            "max_abs_vs_reference": 0.011828728020191193,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00796806261604388,
            "pair_spearman_vs_reference": 0.9924416061586911,
            "pairs": 70,
            "partitions": 1,
            "top1_agree": 70,
            "top1_near_tie": 0,
            "top1_queries": 70
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP16 (the delegate default); host fp16 token table (inputs_embeds); gate app com.mlboydaisuke.d1omni.gate (generic mode)",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__pixel-8a.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-18",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Pixel 8a",
            "machine_label": "pixel-8a-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL Pixel 8a)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Google Tensor G3 (Mali-G715 GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "gate app com.mlboydaisuke.d1omni.gate in generic mode (Kotlin CompiledModel, LiteRT 2.2.0), one cold process per run, the fixture rows pushed as raw tensors, the host token-table lookup (embeds form) done on the Mac before the push; measured interval = input writes + run + output read-back per row.",
            "parity vs the publisher's fp32 stack (sentence-transformers 5.1.2 + transformers 4.57.6, CPU, SDPA math) on the 140 device rows of this window (20 edge512, 40 long256, 80 long512): cosine min 0.999568 (p01 0.999586, mean 1.000051), top-1 retrieval 70/70 on non-near-tie queries (0 near ties, reference top-2 gap <= 0.01), pair Spearman 0.998880 over 70 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): FAIL.",
            "output_match = the Mac CPU run of the same file on the same rows (ai-edge-litert 2.2.0 XNNPACK, the chain's mac-baseline step): max |out - mac| 2.34e-03, min cos 0.99996382 over 141 rows; rule max_abs <= 1e-4 and cos >= 0.9999 on the unit-norm 768-d vectors -> False. max_rel_diff is not computed by the chain (null).",
            "latency_p50_ms = this check run's warm median over 136 rows (230.85 ms; min 228.13, max 235.99; write + run + read-back per text, the gate app's own clock) on a working phone (thermal status before the run: Thermal Status: 1) — INFORMATIONAL: no rested timing run was run on this device for this form, and marks these numbers as check-run medians; on Mali the GPU work completes inside the read-back, so run_ms alone (17.3 ms) is not the time.",
            "conditions: screen on (the gate activity over the keyguard), USB powered, one process per run; thermal status before the run: Thermal Status: 1; compile 1479 ms.",
            "10-18 10:31:40.108 17719 17742 I tflite  : Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 230.8465575,
          "loads": true,
          "max_abs_diff": 0.0023419559001922607,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 0,
            "cos_mean_vs_reference": 1.0000512965257367,
            "cos_min_vs_reference": 0.9995684370483066,
            "cos_p01_vs_reference": 0.9995863304254685,
            "finite": 1,
            "fixtures": 140,
            "gate_leg_rows": 141,
            "gate_leg_warm_max_ms": 235.991862,
            "gate_leg_warm_median_run_ms": 17.313924,
            "gate_leg_warm_median_write_run_read_ms": 230.8465575,
            "gate_leg_warm_min_ms": 228.126465,
            "gate_leg_warm_rows": 136,
            "liveness_moved": 1,
            "load_compile_ms": 1478.539918,
            "mac_same_graph_max_abs": 0.0023419559001922607,
            "mac_same_graph_min_cos": 0.9999638227341267,
            "max_abs_vs_reference": 0.002042725682258606,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.001499423237191877,
            "pair_spearman_vs_reference": 0.9988802379494356,
            "pairs": 70,
            "partitions": 1,
            "top1_agree": 70,
            "top1_near_tie": 0,
            "top1_queries": 70
          },
          "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); gate app com.mlboydaisuke.d1omni.gate (generic mode)",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__pixel-8a.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-31",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Pixel 8a",
            "machine_label": "pixel-8a-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL Pixel 8a)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Google Tensor G3 (Mali-G715 GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Harness: the gate app (LiteRT 2.2.0 Kotlin CompiledModel) driven by a benchmark script in generic mode, one cold process per run, the rows pushed as raw tensors (the host fp16 token-table lookup done on the Mac before the push)",
            "run (2026-10-31): rows file vela_scifact_L512_a.json, 1048 rows, 0 non-finite, compile 4932.4 ms, warm median write + run + read 354.81 ms, ready: check run, not waited for clean (thermal 2, cap policy0:955000/1704000 policy4:712000/2367000 policy8:880000/2914000 , gpu_cooling 1); rested 15s since 04:16:06; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L512_embeds_fp16.tflite phone=fbb052ef8008953a… MATCH\"; delegation: Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "run (2026-10-31): rows file vela_scifact_L512_b.json, 1048 rows, 0 non-finite, compile 3622.4 ms, warm median write + run + read 353.79 ms, ready: check run, not waited for clean (thermal 1, cap policy0:1425000/1704000 policy4:1065000/2367000 policy8:1164000/2914000 , gpu_cooling 0); rested 241s since 04:22:57; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L512_embeds_fp16.tflite phone=fbb052ef8008953a… MATCH\"; delegation: Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "run (2026-10-31): rows file vela_scifact_L512_c.json, 1046 rows, 0 non-finite, compile 1503.4 ms, warm median write + run + read 322.58 ms, ready: check run, not waited for clean (thermal 1, cap policy8:2850000/2914000 , gpu_cooling 0); rested 230s since 04:36:14; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L512_embeds_fp16.tflite phone=fbb052ef8008953a… MATCH\"; delegation: Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "the work line (registered 2026-10-30 22:24 JST before any number): SciFact (mteb/scifact revision d56462d0, CC BY-NC 2.0: 300 test queries over the corpus) embedded on the phone in the four window files together, nDCG@10 and recall@10 each >= the publisher's stack in fp32 (sentence-transformers, the same texts) − 0.005; scored on the Mac: line {\"ndcg_ge_ref_minus_0p005\": true, \"recall_ge_ref_minus_0p005\": true, \"verdict\": \"PASS\"}; the verdict belongs to the four files together, so output_match stays null"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "leg1_compile_ms": 4932.4,
            "leg1_rows": 1048.0,
            "leg1_warm_median_write_run_read_ms": 354.812,
            "leg2_compile_ms": 3622.4,
            "leg2_rows": 1048.0,
            "leg2_warm_median_write_run_read_ms": 353.787,
            "leg3_compile_ms": 1503.4,
            "leg3_rows": 1046.0,
            "leg3_warm_median_write_run_read_ms": 322.582,
            "legs": 3.0,
            "nonfinite_rows": 0.0,
            "pipeline_row_cos_min": 0.998766,
            "pipeline_rows_under_0p999": 1.0,
            "pipeline_top1_agree": 298.0,
            "rows": 3142.0,
            "scifact_line_pass": 1.0,
            "scifact_ndcg_at_10": 0.485749,
            "scifact_ndcg_at_10_reference": 0.485806,
            "scifact_ndcg_delta": -5.7e-05,
            "scifact_queries": 300.0,
            "scifact_recall_at_10": 0.608333,
            "scifact_recall_at_10_reference": 0.608333,
            "scifact_recall_delta": 0.0
          },
          "output_match": null,
          "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); gate app com.mlboydaisuke.d1omni.gate (generic mode); the SciFact texts of this window, scored as one retrieval run over the four window files",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-31/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__pixel-8a.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-18",
          "decode_tokens_per_s": null,
          "delegated_ops": 1690,
          "env": {
            "device": "Pixel 8a",
            "machine_label": "pixel-8a-vela-gate",
            "os_build": "Android SDK 36 (Build.VERSION.SDK_INT read by the gate app; Build.MODEL Pixel 8a)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Google Tensor G3 (Mali-G715 GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "gate app com.mlboydaisuke.d1omni.gate in generic mode (Kotlin CompiledModel, LiteRT 2.2.0), one cold process per run, the fixture rows pushed as raw tensors, the host token-table lookup (embeds form) done on the Mac before the push; measured interval = input writes + run + output read-back per row.",
            "parity vs the publisher's fp32 stack (sentence-transformers 5.1.2 + transformers 4.57.6, CPU, SDPA math) on the 140 device rows of this window (20 edge512, 40 long256, 80 long512): cosine min 0.999974 (p01 0.999997, mean 0.999999), top-1 retrieval 70/70 on non-near-tie queries (0 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999895 over 70 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): PASS.",
            "output_match = the Mac CPU run of the same file on the same rows (ai-edge-litert 2.2.0 XNNPACK, the chain's mac-baseline step): max |out - mac| 1.50e-06, min cos 1.00000000 over 141 rows; rule max_abs <= 1e-4 and cos >= 0.9999 on the unit-norm 768-d vectors -> True. max_rel_diff is not computed by the chain (null).",
            "latency_p50_ms = this check run's warm median over 136 rows (408.31 ms; min 387.80, max 420.78; write + run + read-back per text, the gate app's own clock) on a working phone (thermal status before the run: Thermal Status: 1) — INFORMATIONAL: no rested timing run was run on this device for this form, and marks these numbers as check-run medians; on Mali the GPU work completes inside the read-back, so run_ms alone (21.9 ms) is not the time.",
            "conditions: screen on (the gate activity over the keyguard), USB powered, one process per run; thermal status before the run: Thermal Status: 1; compile 3636 ms.",
            "10-18 10:32:31.924 18567 18595 I tflite  : Replacing 1690 out of 1690 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 408.3141075,
          "loads": true,
          "max_abs_diff": 1.4975666999816895e-06,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999992588054425,
            "cos_min_vs_reference": 0.9999739275010664,
            "cos_p01_vs_reference": 0.9999965715239946,
            "finite": 1,
            "fixtures": 140,
            "gate_leg_rows": 141,
            "gate_leg_warm_max_ms": 420.780314,
            "gate_leg_warm_median_run_ms": 21.880249,
            "gate_leg_warm_median_write_run_read_ms": 408.3141075,
            "gate_leg_warm_min_ms": 387.802531,
            "gate_leg_warm_rows": 136,
            "liveness_moved": 1,
            "load_compile_ms": 3635.727541,
            "mac_same_graph_max_abs": 1.4975666999816895e-06,
            "mac_same_graph_min_cos": 0.9999999999565403,
            "max_abs_vs_reference": 0.0012089908123016357,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00036859505531505743,
            "pair_spearman_vs_reference": 0.9998950223077595,
            "pairs": 70,
            "partitions": 1,
            "top1_agree": 70,
            "top1_near_tie": 0,
            "top1_queries": 70
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; host fp16 token table (inputs_embeds); gate app com.mlboydaisuke.d1omni.gate (generic mode)",
          "total_ops": 1690,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16__pixel-8a.json"
      }
    ]
  },
  "model": {
    "family": "vela-1.0-encoder-307m-embedding",
    "id": "vela-1.0-encoder-307m-embedding__vela_l512_embeds_fp16",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/Vela-1.0-Encoder-307M-Embedding-LiteRT",
    "task": "sentence-similarity"
  },
  "pitfalls": [
    "The LayerNorms of layers 12-21 and the final norm are computed on x*2^-6 (eps*2^-12), layers 9-11 on x*2^-2: the sum of squares of their inputs reaches 4e7 in float32, beyond float16's 65,504, and the rewrite keeps float16-storage GPU paths finite; exact in float32 (card 'Conversion').",
    "`inputs_embeds` float32 [1,512,768] (the host looks each token's row up in tok_embeddings_fp16.bin, the checkpoint's float16 table, 256,000 x 768) + `attention_mask` float32 [1,512] -> `embedding` float32 [1,768]; tokenize with the repo's tokenizer.json (`<bos>` = 2, `<eos>` = 1), pad the rows beyond the text with attention_mask 0 (card 'Files').",
    "The `_embeds_` files exist because of the GPU delegate: it does not take a float16 token table inside the graph, so this form takes the looked-up rows from the host and runs fully on the GPU (Metal: every op accelerated) and on the NPU; the pooling is written as multiply / sum / rsqrt because the delegate also refuses a masked select (`SELECT_V2` with a broadcast condition) (card 'Conversion').",
    "GPU fp16 arithmetic (the GPU defaults on Metal, Adreno and Mali) keeps every top-1 and the pair ranking but drops the per-row cosine floor to 0.991-0.998 on about 1 % of the texts, so the card's GPU numbers are at FP32; on the S26, FP16_WITH_FP32_ACCUM passes the full bar on L128-L512 at 84 ms (L512) if a 0.9994 floor is enough (card 'Performance').",
    "An int8 token table was tried and rejected: its per-row rounding error (0.75 % per row on average, 2.1 % on `<bos>`) moves short sentences by up to 0.7 % in cosine (floor 0.9929 over 3,918 texts) and no set of exactly-kept rows restores the 0.999 floor; int8 FC weights fail harder (floor 0.964); everything shipped is float16 (card 'Conversion').",
    "Every file is a static window: a text of n tokens (including `<bos>` and `<eos>`) runs in the smallest window with L >= n, and the time is the window's, not the text's; the NPU JIT compile on the S26 takes 4-34 s per window on the first run (cache it) (card 'Files', 'Performance').",
    "Strip the text before tokenizing: sentence-transformers does, and a trailing space adds a token (card 'Usage').",
    "The checks measure agreement with the publisher's own stack (cosine per text, top-1 retrieval, pair Spearman) on PAWS-X and STS-B v2 test rows, not accuracy against gold labels; the phone subsets are 140-497 texts per window, the Mac runs the full 3,918-4,184 (card 'Parity')."
  ],
  "schema_version": "1.2"
}
