{
  "artifacts": [
    {
      "file": "vela_L128_embeds_wfp16_shift_arith.tflite",
      "sha256": "2713eab61b798e5d86362a25f7510dacfee109c55143bd25b7c53b4c04733ebd",
      "size_mb": 210.791
    }
  ],
  "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": 399.72,
        "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": 19.13,
        "loads": true,
        "max_rel_diff": 0.0028173834156779076,
        "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_l128_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": "2026-10-31",
          "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_L128.json, 172 rows, 0 non-finite, compile 912.3 ms, warm median write + run + read 13.65 ms, ready: check run, not waited for clean (thermal 0, cap policy6:4608000/4742400 , kgsl_pwrlevel 0); rested 11s since 04:22:44; \"sha256 vela_L128_embeds_fp16.tflite phone=2713eab61b798e5d… 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": 912.3,
            "leg1_rows": 172.0,
            "leg1_warm_median_write_run_read_ms": 13.648,
            "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": 172.0,
            "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_l128_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 273 device rows of this window (73 long128, 200 pair): cosine min 0.999559 (p01 0.999580, mean 0.999964), top-1 retrieval 135/135 on non-near-tie queries (1 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999886 over 136 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| 6.50e-03, min cos 0.99978575 over 274 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 269 rows (49.40 ms; min 19.29, max 55.99; 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 1386 ms.",
            "10-18 11:01:15.605 29767 29789 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": 49.399167,
          "loads": true,
          "max_abs_diff": 0.006504744291305542,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999638400837169,
            "cos_min_vs_reference": 0.9995591191725124,
            "cos_p01_vs_reference": 0.9995799824835226,
            "finite": 1,
            "fixtures": 273,
            "gate_leg_rows": 274,
            "gate_leg_warm_max_ms": 55.988542,
            "gate_leg_warm_median_run_ms": 0.540573,
            "gate_leg_warm_median_write_run_read_ms": 49.399167,
            "gate_leg_warm_min_ms": 19.29224,
            "gate_leg_warm_rows": 269,
            "liveness_moved": 1,
            "load_compile_ms": 1385.955156,
            "mac_same_graph_max_abs": 0.006504744291305542,
            "mac_same_graph_min_cos": 0.999785747636375,
            "max_abs_vs_reference": 0.005914874374866486,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.0015912691737867135,
            "pair_spearman_vs_reference": 0.9998855016459138,
            "pairs": 136,
            "partitions": 1,
            "top1_agree": 135,
            "top1_near_tie": 1,
            "top1_queries": 135
          },
          "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_l128_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_L128.json, 172 rows, 0 non-finite, compile 880.8 ms, warm median write + run + read 19.53 ms, ready: check run, not waited for clean (thermal 0, cap policy6:4396800/4742400 , kgsl_pwrlevel 0); rested 10s since 05:21:11; \"sha256 vela_L128_embeds_fp16.tflite phone=2713eab61b798e5d… 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": 880.8,
            "leg1_rows": 172.0,
            "leg1_warm_median_write_run_read_ms": 19.527,
            "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": 172.0,
            "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_l128_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 273 device rows of this window (73 long128, 200 pair): cosine min 0.999991 (p01 0.999998, mean 1.000000), top-1 retrieval 135/135 on non-near-tie queries (1 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999981 over 136 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| 1.33e-05, min cos 1.00000000 over 274 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 128-token row 24.10 ms (min 23.29, max 27.80), the 32-token row 23.88 ms — the time is the window's, not the text's; compile 998 ms. The 269-row check run's warm median was 47.06 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 1581 ms.",
            "10-18 11:00:46.724 28832 28854 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": 24.098672,
          "loads": true,
          "max_abs_diff": 1.3276934623718262e-05,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999996459327429,
            "cos_min_vs_reference": 0.9999905447624844,
            "cos_p01_vs_reference": 0.9999982197243977,
            "finite": 1,
            "fixtures": 273,
            "gate_leg_rows": 274,
            "gate_leg_warm_max_ms": 66.063959,
            "gate_leg_warm_median_run_ms": 0.496823,
            "gate_leg_warm_median_write_run_read_ms": 47.057917,
            "gate_leg_warm_min_ms": 23.214791,
            "gate_leg_warm_rows": 269,
            "liveness_moved": 1,
            "load_compile_ms": 1581.045312,
            "mac_same_graph_max_abs": 1.3276934623718262e-05,
            "mac_same_graph_min_cos": 0.9999999994748638,
            "max_abs_vs_reference": 0.0016837641596794128,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00022348530188520677,
            "pair_spearman_vs_reference": 0.9999809169409857,
            "pairs": 136,
            "partitions": 1,
            "timing_leg_compile_ms": 998.371823,
            "timing_leg_long_max_ms": 27.799219,
            "timing_leg_long_median_ms": 24.098672,
            "timing_leg_long_min_ms": 23.289479,
            "timing_leg_long_row_len": 128,
            "timing_leg_n": 20,
            "timing_leg_short_median_ms": 23.8798435,
            "timing_leg_short_row_len": 32,
            "timing_leg_warmups": 5,
            "top1_agree": 135,
            "top1_near_tie": 1,
            "top1_queries": 135
          },
          "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_l128_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 273 device rows of this window (73 long128, 200 pair): cosine min 0.999548 (p01 0.999574, mean 0.999989), top-1 retrieval 135/135 on non-near-tie queries (1 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999847 over 136 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 13.062 ms (min 12.823, max 14.917); the 273-row parity run's median call was 13.020 ms; compile 8991 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_L128 && 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_emb128 --fixtures /data/local/tmp/vela_npu/fx_L128_embeds --count 274 --warmup 5 --rounds 20 --ab-fixture 000 --dump full --out /data/local/tmp/vela_npu/out_N1_npu_emb_L128 --model /data/local/tmp/vela_npu/vela_L128_embeds_wfp16_shift_arith.tflite 2>/data/local/tmp/vela_npu/N1_npu_emb_L128.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": 13.062,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999887815754674,
            "cos_min_vs_reference": 0.9995476415958039,
            "cos_p01_vs_reference": 0.9995740886464325,
            "finite": 1,
            "fixtures": 273,
            "fully_accelerated": 1,
            "liveness_moved": 1,
            "load_compile_ms": 8990.7,
            "max_abs_vs_reference": 0.0033208802342414856,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.0014350486626426573,
            "pair_spearman_vs_reference": 0.9998473355278851,
            "pairs": 136,
            "parity_leg_call_ms_median": 13.02,
            "partitions": 1,
            "runner_call_ms_max": 14.917,
            "runner_call_ms_median": 13.062,
            "runner_call_ms_min": 12.823,
            "runner_rounds": 20,
            "runner_warmups": 5,
            "top1_agree": 135,
            "top1_near_tie": 1,
            "top1_queries": 135
          },
          "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_l128_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_L128.json, 172 rows, 0 non-finite, compile 1414.6 ms, warm median write + run + read 72.61 ms, ready: clean; rested 12s since 02:57:47; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L128_embeds_fp16.tflite phone=2713eab61b798e5d… 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": 1414.6,
            "leg1_rows": 172.0,
            "leg1_warm_median_write_run_read_ms": 72.612,
            "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": 172.0,
            "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_l128_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 273 device rows of this window (73 long128, 200 pair): cosine min 0.997503 (p01 0.999446, mean 0.999900), top-1 retrieval 135/135 on non-near-tie queries (1 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999642 over 136 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| 1.62e-02, min cos 0.99742651 over 274 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 269 rows (67.64 ms; min 57.51, max 94.91; 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 (23.0 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 3501 ms.",
            "10-18 10:45:52.859 28920 28942 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": 67.639405,
          "loads": true,
          "max_abs_diff": 0.016150400042533875,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 0,
            "cos_mean_vs_reference": 0.9998998774315495,
            "cos_min_vs_reference": 0.9975030341292324,
            "cos_p01_vs_reference": 0.9994455857622528,
            "finite": 1,
            "fixtures": 273,
            "gate_leg_rows": 274,
            "gate_leg_warm_max_ms": 94.911214,
            "gate_leg_warm_median_run_ms": 22.960652,
            "gate_leg_warm_median_write_run_read_ms": 67.639405,
            "gate_leg_warm_min_ms": 57.507569,
            "gate_leg_warm_rows": 269,
            "liveness_moved": 1,
            "load_compile_ms": 3501.313357,
            "mac_same_graph_max_abs": 0.016150400042533875,
            "mac_same_graph_min_cos": 0.9974265120221282,
            "max_abs_vs_reference": 0.01610507071018219,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.004268028603291896,
            "pair_spearman_vs_reference": 0.9996421926434808,
            "pairs": 136,
            "partitions": 1,
            "top1_agree": 135,
            "top1_near_tie": 1,
            "top1_queries": 135
          },
          "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_l128_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 273 device rows of this window (73 long128, 200 pair): cosine min 0.999475 (p01 0.999596, mean 1.000007), top-1 retrieval 135/135 on non-near-tie queries (1 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999900 over 136 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| 4.53e-03, min cos 0.99977482 over 274 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 269 rows (71.89 ms; min 61.46, max 113.33; 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 3600 ms.",
            "10-18 10:46:35.872 29542 29564 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": 71.886922,
          "loads": true,
          "max_abs_diff": 0.004531152546405792,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 1.000007410840859,
            "cos_min_vs_reference": 0.9994752714396143,
            "cos_p01_vs_reference": 0.9995961858613396,
            "finite": 1,
            "fixtures": 273,
            "gate_leg_rows": 274,
            "gate_leg_warm_max_ms": 113.32841,
            "gate_leg_warm_median_run_ms": 23.332113,
            "gate_leg_warm_median_write_run_read_ms": 71.886922,
            "gate_leg_warm_min_ms": 61.459961,
            "gate_leg_warm_rows": 269,
            "liveness_moved": 1,
            "load_compile_ms": 3600.096762,
            "mac_same_graph_max_abs": 0.004531152546405792,
            "mac_same_graph_min_cos": 0.9997748222356836,
            "max_abs_vs_reference": 0.0034587658010423183,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00292787736725586,
            "pair_spearman_vs_reference": 0.9998998139401746,
            "pairs": 136,
            "partitions": 1,
            "top1_agree": 135,
            "top1_near_tie": 1,
            "top1_queries": 135
          },
          "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_l128_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_L128.json, 172 rows, 0 non-finite, compile 3478.7 ms, warm median write + run + read 72.83 ms, ready: check run, not waited for clean (thermal 1, cap policy0:1328000/1704000 policy4:1065000/2367000 policy8:1164000/2914000 , gpu_cooling 0); rested 12s since 04:10:27; front com.google.android.apps.weather (a system app: left alone); \"sha256 vela_L128_embeds_fp16.tflite phone=2713eab61b798e5d… 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": 3478.7,
            "leg1_rows": 172.0,
            "leg1_warm_median_write_run_read_ms": 72.833,
            "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": 172.0,
            "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_l128_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 273 device rows of this window (73 long128, 200 pair): cosine min 0.999990 (p01 0.999998, mean 1.000000), top-1 retrieval 135/135 on non-near-tie queries (1 near ties, reference top-2 gap <= 0.01), pair Spearman 0.999981 over 136 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| 5.10e-06, min cos 1.00000000 over 274 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 269 rows (91.46 ms; min 89.78, max 97.25; 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 (22.0 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 4038 ms.",
            "10-18 10:47:18.564 30187 30209 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": 91.464844,
          "loads": true,
          "max_abs_diff": 5.103647708892822e-06,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1,
            "cos_mean_vs_reference": 0.9999996479409399,
            "cos_min_vs_reference": 0.9999904388297984,
            "cos_p01_vs_reference": 0.9999982267620934,
            "finite": 1,
            "fixtures": 273,
            "gate_leg_rows": 274,
            "gate_leg_warm_max_ms": 97.251384,
            "gate_leg_warm_median_run_ms": 21.982178,
            "gate_leg_warm_median_write_run_read_ms": 91.464844,
            "gate_leg_warm_min_ms": 89.782511,
            "gate_leg_warm_rows": 269,
            "liveness_moved": 1,
            "load_compile_ms": 4038.229413,
            "mac_same_graph_max_abs": 5.103647708892822e-06,
            "mac_same_graph_min_cos": 0.9999999999144049,
            "max_abs_vs_reference": 0.0016919374465942383,
            "nonfinite_rows": 0,
            "pair_max_abs_dcos": 0.00022304454501642645,
            "pair_spearman_vs_reference": 0.9999809169409857,
            "pairs": 136,
            "partitions": 1,
            "top1_agree": 135,
            "top1_near_tie": 1,
            "top1_queries": 135
          },
          "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_l128_embeds_fp16__pixel-8a.json"
      }
    ]
  },
  "model": {
    "family": "vela-1.0-encoder-307m-embedding",
    "id": "vela-1.0-encoder-307m-embedding__vela_l128_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 [1,128,768] (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 [1,128] -> `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"
}
