{
  "artifacts": [
    {
      "file": "vela_L64_embeds_wfp16_shift_arith.tflite",
      "sha256": "d2c473a7de08955eb3cbc0eb566a797904b1eb99f35adbc7417e10a020ff91c2",
      "size_mb": 210.65
    }
  ],
  "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": 196.29,
        "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": 14.09,
        "loads": true,
        "max_rel_diff": 0.00173499573591404,
        "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_l64_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": 1644,
      "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": 1689,
          "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 FP16 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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.998274 (p01 0.999468, mean 0.999909), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999680 over 248 pairs, mask liveness moved, finite True; bar (cos >= 0.999 on every row, top-1 100 %, Spearman >= 0.999, live, finite): FAIL.",
            "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 11.502 ms (min 10.994, max 14.733); the 497-row parity run's median call was 11.341 ms; compile 1631 ms.",
            "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_gpu16_emb_L64 && 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 fp16 --burst 1 --libdir /data/local/tmp/vela_npu  --fixtures /data/local/tmp/vela_npu/fx_L64_embeds --count 498 --warmup 5 --rounds 20 --ab-fixture 000 --dump full --out /data/local/tmp/vela_npu/out_N1_gpu16_emb_L64 --model /data/local/tmp/vela_npu/vela_L64_embeds_wfp16_shift_arith.tflite 2>/data/local/tmp/vela_npu/N1_gpu16_emb_L64.stderr.txt",
            "VERBOSE: Replacing 1689 out of 1689 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 11.5015,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 0,
            "cos_mean_vs_reference": 0.9999090000967515,
            "cos_min_vs_reference": 0.9982741707222772,
            "cos_p01_vs_reference": 0.9994683861227154,
            "finite": 1,
            "fixtures": 497,
            "fully_accelerated": 1,
            "liveness_moved": 1,
            "load_compile_ms": 1630.6,
            "max_abs_vs_reference": 0.013198241591453552,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.006596320654352894,
            "pair_spearman_vs_reference": 0.9996797953207632,
            "pairs": 248,
            "parity_leg_call_ms_median": 11.3405,
            "partitions": 1,
            "runner_call_ms_max": 14.733,
            "runner_call_ms_median": 11.5015,
            "runner_call_ms_min": 10.994,
            "runner_rounds": 20,
            "runner_warmups": 5,
            "stsb_v2_spearman_delta": -0.0014181837803211028,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "C-API runner --accel gpu, precision fp16; host fp16 token table (inputs_embeds); the runner's GPU control run",
          "total_ops": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.999997 (p01 0.999999, mean 1.000000), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999998 over 248 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 17.015 ms (min 15.915, max 22.002); the 497-row parity run's median call was 16.970 ms; compile 2086 ms.",
            "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_gpu32_emb_L64 && 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_L64_embeds --count 498 --warmup 5 --rounds 20 --ab-fixture 000 --dump full --out /data/local/tmp/vela_npu/out_N1_gpu32_emb_L64 --model /data/local/tmp/vela_npu/vela_L64_embeds_wfp16_shift_arith.tflite 2>/data/local/tmp/vela_npu/N1_gpu32_emb_L64.stderr.txt",
            "VERBOSE: Replacing 1689 out of 1689 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 17.0155,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999997247628497,
            "cos_min_vs_reference": 0.9999973053189505,
            "cos_p01_vs_reference": 0.9999993359337173,
            "finite": 1,
            "fixtures": 497,
            "fully_accelerated": 1,
            "liveness_moved": 1,
            "load_compile_ms": 2085.9,
            "max_abs_vs_reference": 0.0004070401191711426,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00033945903770282726,
            "pair_spearman_vs_reference": 0.9999976397689491,
            "pairs": 248,
            "parity_leg_call_ms_median": 16.97,
            "partitions": 1,
            "runner_call_ms_max": 22.002,
            "runner_call_ms_median": 17.0155,
            "runner_call_ms_min": 15.915,
            "runner_rounds": 20,
            "runner_warmups": 5,
            "stsb_v2_spearman_delta": 0.0,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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_L64.json, 300 rows, 0 non-finite, compile 899.7 ms, warm median write + run + read 11.26 ms, ready: clean; rested 327s since 04:17:03; \"sha256 vela_L64_embeds_fp16.tflite phone=d2c473a7de08955e… MATCH\"; delegation: Replacing 1689 out of 1689 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": 899.7,
            "leg1_rows": 300.0,
            "leg1_warm_median_write_run_read_ms": 11.26,
            "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": 300.0,
            "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-31/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.998274 (p01 0.999468, mean 0.999909), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999680 over 248 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| 1.32e-02, min cos 0.99816056 over 498 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 493 rows (11.20 ms; min 10.51, max 13.26; 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.",
            "conditions: screen on (the gate activity over the keyguard), USB powered, one process per run; thermal status before the run: Thermal Status: 0; compile 870 ms.",
            "10-18 10:35:21.022 32270 32294 I tflite  : Replacing 1689 out of 1689 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 11.201354,
          "loads": true,
          "max_abs_diff": 0.013195380568504333,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 0,
            "cos_mean_vs_reference": 0.9999090000967515,
            "cos_min_vs_reference": 0.9982741707222772,
            "cos_p01_vs_reference": 0.9994683861227154,
            "finite": 1,
            "fixtures": 497,
            "gate_leg_rows": 498,
            "gate_leg_warm_max_ms": 13.258489,
            "gate_leg_warm_median_run_ms": 0.458073,
            "gate_leg_warm_median_write_run_read_ms": 11.201354,
            "gate_leg_warm_min_ms": 10.510208,
            "gate_leg_warm_rows": 493,
            "liveness_moved": 1,
            "load_compile_ms": 870.291979,
            "mac_same_graph_max_abs": 0.013195380568504333,
            "mac_same_graph_min_cos": 0.9981605621551178,
            "max_abs_vs_reference": 0.013198241591453552,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.006596320654352894,
            "pair_spearman_vs_reference": 0.9996797953207632,
            "pairs": 248,
            "partitions": 1,
            "stsb_v2_spearman_delta": -0.0014181837803211028,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.998756 (p01 0.999589, mean 0.999998), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999961 over 248 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| 1.10e-02, min cos 0.99896961 over 498 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 493 rows (28.80 ms; min 11.60, max 35.29; 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 1309 ms.",
            "10-18 11:03:09.906 32440 32462 I tflite  : Replacing 1689 out of 1689 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 28.803958,
          "loads": true,
          "max_abs_diff": 0.011042565107345581,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 0,
            "cos_mean_vs_reference": 0.9999979774414582,
            "cos_min_vs_reference": 0.9987558804702669,
            "cos_p01_vs_reference": 0.9995886872507413,
            "finite": 1,
            "fixtures": 497,
            "gate_leg_rows": 498,
            "gate_leg_warm_max_ms": 35.29349,
            "gate_leg_warm_median_run_ms": 0.494479,
            "gate_leg_warm_median_write_run_read_ms": 28.803958,
            "gate_leg_warm_min_ms": 11.601093,
            "gate_leg_warm_rows": 493,
            "liveness_moved": 1,
            "load_compile_ms": 1308.946979,
            "mac_same_graph_max_abs": 0.011042565107345581,
            "mac_same_graph_min_cos": 0.9989696113869101,
            "max_abs_vs_reference": 0.011021696031093597,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.001987209822388536,
            "pair_spearman_vs_reference": 0.9999606628158186,
            "pairs": 248,
            "partitions": 1,
            "stsb_v2_spearman_delta": 0.0005005354518780036,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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_L64.json, 300 rows, 0 non-finite, compile 882.0 ms, warm median write + run + read 12.25 ms, ready: clean; rested 284s since 05:16:14; \"sha256 vela_L64_embeds_fp16.tflite phone=d2c473a7de08955e… MATCH\"; delegation: Replacing 1689 out of 1689 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": 882.0,
            "leg1_rows": 300.0,
            "leg1_warm_median_write_run_read_ms": 12.246,
            "legs": 1.0,
            "nonfinite_rows": 0.0,
            "pipeline_row_cos_min": 0.998436,
            "pipeline_rows_under_0p999": 2.0,
            "pipeline_top1_agree": 297.0,
            "rows": 300.0,
            "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-31/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.999997 (p01 0.999999, mean 1.000000), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999998 over 248 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| 1.22e-06, min cos 1.00000000 over 498 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 61-token row 16.48 ms (min 16.39, max 19.44), the 32-token row 16.50 ms — the time is the window's, not the text's; compile 958 ms. The 493-row check run's warm median was 38.63 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 1516 ms.",
            "10-18 11:02:38.992 31568 31591 I tflite  : Replacing 1689 out of 1689 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 16.482760499999998,
          "loads": true,
          "max_abs_diff": 1.2218952178955078e-06,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999997247628497,
            "cos_min_vs_reference": 0.9999973053189505,
            "cos_p01_vs_reference": 0.9999993359337173,
            "finite": 1,
            "fixtures": 497,
            "gate_leg_rows": 498,
            "gate_leg_warm_max_ms": 46.483542,
            "gate_leg_warm_median_run_ms": 0.525521,
            "gate_leg_warm_median_write_run_read_ms": 38.629844,
            "gate_leg_warm_min_ms": 16.08625,
            "gate_leg_warm_rows": 493,
            "liveness_moved": 1,
            "load_compile_ms": 1516.222448,
            "mac_same_graph_max_abs": 1.2218952178955078e-06,
            "mac_same_graph_min_cos": 0.9999999999793941,
            "max_abs_vs_reference": 0.0004070401191711426,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00033945903770282726,
            "pair_spearman_vs_reference": 0.9999976397689491,
            "pairs": 248,
            "partitions": 1,
            "stsb_v2_spearman_delta": 0.0,
            "timing_leg_compile_ms": 958.428541,
            "timing_leg_long_max_ms": 19.444948,
            "timing_leg_long_median_ms": 16.482760499999998,
            "timing_leg_long_min_ms": 16.39401,
            "timing_leg_long_row_len": 61,
            "timing_leg_n": 20,
            "timing_leg_short_median_ms": 16.5034375,
            "timing_leg_short_row_len": 32,
            "timing_leg_warmups": 5,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.999530 (p01 0.999589, mean 0.999995), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999958 over 248 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 6.833 ms (min 6.781, max 7.333); the 497-row parity run's median call was 6.853 ms; compile 3880 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_L64 && 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_emb64 --fixtures /data/local/tmp/vela_npu/fx_L64_embeds --count 498 --warmup 5 --rounds 20 --ab-fixture 000 --dump full --out /data/local/tmp/vela_npu/out_N1_npu_emb_L64 --model /data/local/tmp/vela_npu/vela_L64_embeds_wfp16_shift_arith.tflite 2>/data/local/tmp/vela_npu/N1_npu_emb_L64.stderr.txt",
            "INFO: [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 1689 ops, from a total of 1689 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": 6.833,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999948519274587,
            "cos_min_vs_reference": 0.9995304416646922,
            "cos_p01_vs_reference": 0.9995894319792719,
            "finite": 1,
            "fixtures": 497,
            "fully_accelerated": 1,
            "liveness_moved": 1,
            "load_compile_ms": 3880.1,
            "max_abs_vs_reference": 0.004412218928337097,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.002162050098064583,
            "pair_spearman_vs_reference": 0.999957515841084,
            "pairs": 248,
            "parity_leg_call_ms_median": 6.853,
            "partitions": 1,
            "runner_call_ms_max": 7.333,
            "runner_call_ms_median": 6.833,
            "runner_call_ms_min": 6.781,
            "runner_rounds": 20,
            "runner_warmups": 5,
            "stsb_v2_spearman_delta": -0.00016684515062603822,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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_L64.json, 300 rows, 0 non-finite, compile 1521.7 ms, warm median write + run + read 55.27 ms, ready: clean; rested 295s since 02:52:25; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L64_embeds_fp16.tflite phone=d2c473a7de08955e… MATCH\"; delegation: Replacing 1689 out of 1689 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": 1521.7,
            "leg1_rows": 300.0,
            "leg1_warm_median_write_run_read_ms": 55.274,
            "legs": 1.0,
            "nonfinite_rows": 0.0,
            "pipeline_row_cos_min": 0.995571,
            "pipeline_rows_under_0p999": 9.0,
            "pipeline_top1_agree": 289.0,
            "rows": 300.0,
            "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-31/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.997840 (p01 0.999518, mean 0.999929), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999746 over 248 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| 1.29e-02, min cos 0.99820278 over 498 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 493 rows (54.61 ms; min 45.69, max 66.78; 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 (20.5 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 1620 ms.",
            "10-18 10:21:33.929 10909 10935 I tflite  : Replacing 1689 out of 1689 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 54.605306,
          "loads": true,
          "max_abs_diff": 0.012890204787254333,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 0,
            "cos_mean_vs_reference": 0.9999293006323305,
            "cos_min_vs_reference": 0.9978397329066844,
            "cos_p01_vs_reference": 0.9995181195448237,
            "finite": 1,
            "fixtures": 497,
            "gate_leg_rows": 498,
            "gate_leg_warm_max_ms": 66.782715,
            "gate_leg_warm_median_run_ms": 20.463582,
            "gate_leg_warm_median_write_run_read_ms": 54.605306,
            "gate_leg_warm_min_ms": 45.686849,
            "gate_leg_warm_rows": 493,
            "liveness_moved": 1,
            "load_compile_ms": 1619.979371,
            "mac_same_graph_max_abs": 0.012890204787254333,
            "mac_same_graph_min_cos": 0.9982027833848215,
            "max_abs_vs_reference": 0.012893065810203552,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.006352126454402196,
            "pair_spearman_vs_reference": 0.9997458817901881,
            "pairs": 248,
            "partitions": 1,
            "stsb_v2_spearman_delta": -0.0005005354518780036,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.999573 (p01 0.999626, mean 1.000021), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999955 over 248 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| 1.89e-03, min cos 0.99990099 over 498 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 493 rows (55.91 ms; min 43.40, max 66.51; 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 (23.3 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 3443 ms.",
            "10-18 10:50:57.791  1477  1532 I tflite  : Replacing 1689 out of 1689 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 55.912923,
          "loads": true,
          "max_abs_diff": 0.0018855780363082886,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 1.0000211879921332,
            "cos_min_vs_reference": 0.9995728583326953,
            "cos_p01_vs_reference": 0.9996259858284653,
            "finite": 1,
            "fixtures": 497,
            "gate_leg_rows": 498,
            "gate_leg_warm_max_ms": 66.508178,
            "gate_leg_warm_median_run_ms": 23.288046,
            "gate_leg_warm_median_write_run_read_ms": 55.912923,
            "gate_leg_warm_min_ms": 43.397746,
            "gate_leg_warm_rows": 493,
            "liveness_moved": 1,
            "load_compile_ms": 3442.536989,
            "mac_same_graph_max_abs": 0.0018855780363082886,
            "mac_same_graph_min_cos": 0.9999009888717575,
            "max_abs_vs_reference": 0.0016482751816511154,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.0018384297239890657,
            "pair_spearman_vs_reference": 0.9999551556100332,
            "pairs": 248,
            "partitions": 1,
            "stsb_v2_spearman_delta": 0.0011957235794863852,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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_L64.json, 300 rows, 0 non-finite, compile 4475.0 ms, warm median write + run + read 53.66 ms, ready: check run, not waited for clean (thermal 2, cap policy0:955000/1704000 policy4:712000/2367000 policy8:880000/2914000 , gpu_cooling 0); rested 14s since 04:09:40; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L64_embeds_fp16.tflite phone=d2c473a7de08955e… MATCH\"; delegation: Replacing 1689 out of 1689 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": 4475.0,
            "leg1_rows": 300.0,
            "leg1_warm_median_write_run_read_ms": 53.658,
            "legs": 1.0,
            "nonfinite_rows": 0.0,
            "pipeline_row_cos_min": 0.998766,
            "pipeline_rows_under_0p999": 1.0,
            "pipeline_top1_agree": 298.0,
            "rows": 300.0,
            "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-31/vela-1.0-encoder-307m-embedding__vela_l64_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": 1689,
          "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 497 device rows of this window (377 pair, 120 stsb): cosine min 0.999997 (p01 0.999999, mean 1.000000), top-1 retrieval 235/235 on non-near-tie queries (13 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999998 over 248 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| 1.74e-06, min cos 1.00000000 over 498 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 493 rows (75.92 ms; min 61.78, max 87.01; 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 (20.0 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 1516 ms.",
            "10-18 10:22:22.730 11642 11663 I tflite  : Replacing 1689 out of 1689 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 75.922119,
          "loads": true,
          "max_abs_diff": 1.7434358596801758e-06,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999997240857268,
            "cos_min_vs_reference": 0.9999971957472225,
            "cos_p01_vs_reference": 0.9999993268039495,
            "finite": 1,
            "fixtures": 497,
            "gate_leg_rows": 498,
            "gate_leg_warm_max_ms": 87.005494,
            "gate_leg_warm_median_run_ms": 19.97351,
            "gate_leg_warm_median_write_run_read_ms": 75.922119,
            "gate_leg_warm_min_ms": 61.777181,
            "gate_leg_warm_rows": 493,
            "liveness_moved": 1,
            "load_compile_ms": 1516.315268,
            "mac_same_graph_max_abs": 1.7434358596801758e-06,
            "mac_same_graph_min_cos": 0.9999999999590095,
            "max_abs_vs_reference": 0.00040756165981292725,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.0003392636022265805,
            "pair_spearman_vs_reference": 0.9999976397689491,
            "pairs": 248,
            "partitions": 1,
            "stsb_v2_spearman_delta": 0.0,
            "top1_agree": 235,
            "top1_near_tie": 13,
            "top1_queries": 235
          },
          "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": 1689,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-18/vela-1.0-encoder-307m-embedding__vela_l64_embeds_fp16__pixel-8a.json"
      }
    ]
  },
  "model": {
    "family": "vela-1.0-encoder-307m-embedding",
    "id": "vela-1.0-encoder-307m-embedding__vela_l64_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 [1,64,768] (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 [1,64] -> `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"
}
