{
  "artifacts": [
    {
      "file": "deberta_v3_large_decision_s512_wfp16.tflite",
      "sha256": "238f0cc4dca7b6f18677950e9261d07401a3de9857d76f2a7450fada7f97a5d9",
      "size_mb": 775.76
    }
  ],
  "benchmarks": [],
  "browser": {
    "backends": [
      {
        "backend": "wasm_xnnpack",
        "date": "2026-10-02",
        "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": null,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": false
      },
      {
        "backend": "webgpu_mldrift",
        "date": "2026-10-02",
        "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": null,
        "loads": false,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": false
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-10-02/open-decision-deberta-v3-large__deberta_v3_large_decision_s512_wfp16.json"
  },
  "conversion": {
    "command": "python scripts/export.py --seq 512  (REPRODUCE.md section 3: fp32 export of the shaped-DeBERTa graph with the rank-4 matmul lowering, then the weight-only float16 cast of that export; exports/deberta_v3_large_decision_s512_wfp16.tflite)",
    "quantization": "float16 weight storage (ai-edge-quantizer 0.8.0 FLOAT_CASTING, weight-only 16-bit): the 146 FULLY_CONNECTED weight tensors stored as float16 with a DEQUANTIZE to float32; the 48 relative-position constants and all activations stay float32; 1,785 operators vs 1,639 in the fp32 export (card 'Files'); no INT8 file is shipped",
    "tool": "litert-torch",
    "tool_version": "0.9.3 (torch 2.12.1; transformers 4.57.6; ai-edge-litert 2.1.6; ai-edge-quantizer 0.8.0 for the float16 weight storage)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [
      "DEQUANTIZE",
      "SQUARED_DIFFERENCE"
    ],
    "coverage_ops_pct": 89.1,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 1736,
      "unmatched": 49
    },
    "partitions": 196
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1785,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-opendecision-app",
            "os_build": "Android 16 (SDK 36, build BP4A.251205.006)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Android sample app (zoo main 59aa580 opendecision/, Kotlin host: DecisionTokenizer.kt + DecisionInputs.kt + DecisionModel.kt), debug build, launched with --ez gate true --es accel GPU; one process for both windows (the smallest window that fits each request); graph_ms = first input write through output readback.",
            "output_match null: the app gate records per-question max |dprob| and argmax equality against the oracle, not raw logits (adapter rule).",
            "conditions: thermal status 2 before and after on the adjacent debug-gate runs (the app gate itself recorded no thermal state); the S26 had run the GPU/NPU checks for ~30 min before."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 619.77177,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "argmax_equal_official": 49,
            "fixtures": 17,
            "ids_equal_official_collator": 17,
            "latency_max_ms": 1196.355051,
            "latency_min_ms": 583.440677,
            "max_abs_dprob_vs_official_T1_05": 0.00043716773646840235,
            "questions": 49,
            "spans_equal_official_collator": 17,
            "thermal_status_whole_gate": 2,
            "whole_gate_argmax_equal": 420,
            "whole_gate_max_abs_dprob": 0.0006861766536193636,
            "whole_gate_questions": 420,
            "whole_gate_requests": 160
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; Android sample app com.opendecision (zoo opendecision/, debug build, on-device tokenizer)",
          "total_ops": 1785,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/open-decision-deberta-v3-large__deberta_v3_large_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1785,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-opendecision-gate",
            "os_build": "Android 16 (SDK 36, build BP4A.251205.006)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.decision.gate (android_gate/, Kotlin CompiledModel 2.2.0, Environment with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, fixture ids + spans pushed as JSON, host fp16 table lookup (android.util.Half) outside the timed interval; timed interval = input writes + run + output readback per request.",
            "conditions: before thermal status 0, battery 31.0 C; after thermal status 3, battery 44.9 C; screen on (FLAG_KEEP_SCREEN_ON), USB powered, no cool-start wait (the S26 came warm from the earlier runs where stated in notes).",
            "parity vs the author's implementation (CPU fp32 typed_decisions package): same winning option on 1302/1302 questions (754/754 boundary = reference top-1 < 0.9); max |dprob| at T=1.05 0.00083; questions over 0.01: 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same ids: 0 violating of 9741 compared option logits, max |d| 5.0067901611328125e-05.",
            "conditions: the phone warmed from thermal status 0 (31.0 C) to 3 (44.9 C) over the 500 back-to-back requests (6 min); per-request times rose from ~584 ms (rows 1-2) to ~1,024 ms (rows 497-499), so the median covers a thermally drifting run (min 575.5 ms).",
            "10-01 23:56:49.151  5929  6013 I tflite  : Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-01 23:56:52.266  5929  6013 I DecisionGate: DECISION_GATE compiled in 3689.534634 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 697.48927,
          "loads": true,
          "max_abs_diff": 5.0067901611328125e-05,
          "max_rel_diff": 0.0017240820296668558,
          "metrics": {
            "argmax_equal_official": 1302,
            "battery_temp_c_end": 44.9,
            "battery_temp_c_start": 31.0,
            "boundary_argmax_equal_official": 754,
            "boundary_questions": 754,
            "element_rule_elements": 9741,
            "element_rule_violations": 0,
            "first_call_ms": 620.938021,
            "fixtures": 500,
            "host_fp16_lookup_median_ms": 37.5,
            "load_compile_ms": 3689.534634,
            "max_abs_dprob_vs_official_T1_05": 0.00083,
            "max_abs_dscore_vs_official": 0.00079,
            "nonfinite_questions": 0,
            "questions": 1302,
            "questions_dprob_gt_1e2": 0,
            "thermal_status_end": 3,
            "thermal_status_start": 0,
            "warm_max_ms": 1201.669062,
            "warm_median_ms": 697.48927,
            "warm_min_ms": 575.546718,
            "warmup_rows_excluded": 5
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; debug gate app com.decision.gate",
          "total_ops": 1785,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/open-decision-deberta-v3-large__deberta_v3_large_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-04",
          "decode_tokens_per_s": null,
          "delegated_ops": 1785,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-hfmodels-decide-gate",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "hfmodels-android SDK device check: androidTest DebertaDecisionDeviceTest of the litert module (git 29bc04c), run by tools/decide_gate.sh on the Galaxy S26, one instrumentation process per run; the SDK (hfmodels-litert EncoderDecisions, family deberta_decision) resolved litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT through the dev descriptor catalog/dev/litert-community__Open-Decision-DeBERTa-v3-Large-LiteRT.hfmodels.json at that commit (the log's descriptor sha256 prefix 1e2e59e2 = `git show 29bc04c:catalog/dev/litert-community__Open-Decision-DeBERTa-v3-Large-LiteRT.hfmodels.json`); graph deberta_v3_large_decision_s512_wfp16.tflite sha256 238f0cc4dca7b6f18677950e9261d07401a3de9857d76f2a7450fada7f97a5d9 = the published deberta_v3_large_decision_s512_wfp16.tflite (litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT@7a276235; other records hold rows of the same name and digest).",
            "# hfmodels typed-decisions device check  date=2026-10-04T01:37:08Z serial=RFGL80R6A6H device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 android=16",
            "# git=29bc04c test=DebertaDecisionDeviceTest variant=s512_wfp16 backend=gpu thermal=Thermal Status: 0 battery_level= level: 80 battery_temp_x10C= temperature: 367 screen= mWakefulness=Dozing",
            "# gradle rc=0 end=2026-10-04T01:41:20Z thermal_end=Thermal Status: 2 battery_temp_end_x10C= temperature: 418",
            "resolve litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT: commit=7a276235 via EXPLICIT_REVISION, descriptor=caller@explicit/hfmodels.json sha=1e2e59e2",
            "imported deberta_v3_large_decision_s512_wfp16.tflite: 813443728 bytes, sha256 ok, from /data/local/tmp/hfmodels/open-decision/deberta_v3_large_decision_s512_wfp16.tflite",
            "ready litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT@7a276235 variant=s512_wfp16 profile=gpu inference=GPU prepare_ms=5078",
            "litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully: ",
            "tflite  : Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "RESULT step=load ok=true variant=s512_wfp16 backend=gpu import_ms=1101 load_ms=5088 compile_ms=4520 profile=gpu window=512 head=512 max_options=128 notes=family=deberta_decision compile_ms=4520 on gpu (precision fp32) | observed backend is UNKNOWN: litert 2.2.0 reports no per-op execution target; 'initialized' is the accelerator the CompiledModel was created with | host token lookup: word_embeddings_fp16.bin (float16, 128100 rows x 1024) feeds inputs_embeds | the questions of one decide() in one forward as the author's decide() does, split in question order when they do not fit the window or the option slots: options = a choice's descriptions (the key when there is none), a score's levels, a noul's no / yes; the state is cut to its first 256 tokens; a request longer than the 512-token window or with more than 128 options is refused; probabilities are the softmax of the logits / 1.05 (the author's calibration), extras.max_probability is the author's confidence",
            "RESULT step=sequence ok=true variant=s512_wfp16 backend=gpu rows=160 of 160 mismatches=0 window=512",
            "RESULT step=parity ok=true variant=s512_wfp16 backend=gpu rows=160 questions=420 argmax_match=420/420 answers_equal_author=420/420 max_prob_abs_err=6.861624558591184E-4 (worst fam-ood-47 q0) max_logit_abs_err_vs_author=0.004940509796142578 dp_tol=0.01 forward_ms_median=1027.440625 p90=1057.908594 min=571.566198 max=1289.380155",
            "RESULT step=demo ok=true variant=s512_wfp16 backend=gpu form=policy rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=9.328052532620035E-7 (worst b10) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=19/30 correct_mac=19/30 question_ms_median=1203.294478 p90=1215.3562490000002 min=1031.077499 max=1220.047343 lookup_ms_median=3.180313",
            "RESULT step=demo ok=true variant=s512_wfp16 backend=gpu form=keys rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=2.028996443748099E-6 (worst b22) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=22/30 correct_mac=22/30 question_ms_median=1200.626406 p90=1225.3238020000001 min=1188.83427 max=1298.910989 lookup_ms_median=3.034583",
            "RESULT step=timing ok=true variant=s512_wfp16 backend=gpu repeats=2 forwards_for_4q=1 decide_4q_total_ms_median=1285.523437 four_decide_1q_total_ms_median=4918.670884 state_tokens=19 per_question_ms=1209.198333,1209.198333,1209.198333,1209.198333 answers={department=billing 0.74095467372367, urgency=0.6637985189196726, refund_requested=0.9173051284357872, tone=neutral 0.4445183378688088}",
            "RESULT step=lookup ok=true variant=s512_wfp16 backend=gpu rows=256 hidden=1024 zoo_style=arith zoo_ms_median=21.120729 zoo_ms_min=20.554271 sdk_ms_median=0.687761 sdk_ms_min=0.687239 repeats=20 outputs_identical=true",
            "RESULT step=release ok=true variant=s512_wfp16 backend=gpu close_ms=119",
            "RESULT ok=true model=litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT variant=s512_wfp16 backend=gpu device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 failures=",
            "latency_p50_ms = the parity step's forward_ms_median: one LiteRtDecisionModel.scores call per captured request (host float16 table lookup, input writes, CompiledModel run, output readback on the model thread), after one warm forward.",
            "output_match null: the gate compares with the publisher's outputs under its own criterion (argmax and max |dp| <= 0.01, in metrics) and keeps no raw per-element outputs. full_delegation from the delegate line the gate's logcat filter now keeps ('Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL)').",
            "conditions (the gate's header and footer): thermal status 0 -> 2, battery 36.7 -> 41.8 C, battery level 80 %, display mWakefulness=Dozing at the start.",
            "source: hfmodels-android/litert/results/2026-10-04-1037-RFGL80R6A6H-litert2.2.0-decide-opendecision-s512_wfp16-gpu.log (main 3e1fcd8)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 1027.440625,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "battery_level_start": 80,
            "battery_temp_c_end": 41.8,
            "battery_temp_c_start": 36.7,
            "demo_keys_correct_device": 22.0,
            "demo_keys_correct_device_of": 30.0,
            "demo_keys_correct_mac": 22.0,
            "demo_keys_correct_mac_of": 30.0,
            "demo_keys_dp_tol": 0.01,
            "demo_keys_encoded_length_equal_host": 30.0,
            "demo_keys_encoded_length_equal_host_of": 30.0,
            "demo_keys_lookup_ms_median": 3.034583,
            "demo_keys_max": 1298.910989,
            "demo_keys_max_prob_abs_err_vs_mac": 2.028996443748099e-06,
            "demo_keys_min": 1188.83427,
            "demo_keys_p90": 1225.3238020000001,
            "demo_keys_question_ms_median": 1200.626406,
            "demo_keys_rows": 30.0,
            "demo_keys_same_answer_as_mac": 30.0,
            "demo_keys_same_answer_as_mac_of": 30.0,
            "demo_policy_correct_device": 19.0,
            "demo_policy_correct_device_of": 30.0,
            "demo_policy_correct_mac": 19.0,
            "demo_policy_correct_mac_of": 30.0,
            "demo_policy_dp_tol": 0.01,
            "demo_policy_encoded_length_equal_host": 30.0,
            "demo_policy_encoded_length_equal_host_of": 30.0,
            "demo_policy_lookup_ms_median": 3.180313,
            "demo_policy_max": 1220.047343,
            "demo_policy_max_prob_abs_err_vs_mac": 9.328052532620035e-07,
            "demo_policy_min": 1031.077499,
            "demo_policy_p90": 1215.3562490000002,
            "demo_policy_question_ms_median": 1203.294478,
            "demo_policy_rows": 30.0,
            "demo_policy_same_answer_as_mac": 30.0,
            "demo_policy_same_answer_as_mac_of": 30.0,
            "load_compile_ms": 4520.0,
            "load_head": 512.0,
            "load_import_ms": 1101.0,
            "load_load_ms": 5088.0,
            "load_max_options": 128.0,
            "load_window": 512.0,
            "lookup_hidden": 1024.0,
            "lookup_repeats": 20.0,
            "lookup_rows": 256.0,
            "lookup_sdk_ms_median": 0.687761,
            "lookup_sdk_ms_min": 0.687239,
            "lookup_zoo_ms_median": 21.120729,
            "lookup_zoo_ms_min": 20.554271,
            "parity_answers_equal_author": 420.0,
            "parity_answers_equal_author_of": 420.0,
            "parity_argmax_match": 420.0,
            "parity_argmax_match_of": 420.0,
            "parity_dp_tol": 0.01,
            "parity_forward_ms_median": 1027.440625,
            "parity_max": 1289.380155,
            "parity_max_logit_abs_err_vs_author": 0.004940509796142578,
            "parity_max_prob_abs_err": 0.0006861624558591184,
            "parity_min": 571.566198,
            "parity_p90": 1057.908594,
            "parity_rows": 160.0,
            "prepare_ms": 5078.0,
            "release_close_ms": 119.0,
            "sequence_mismatches": 0.0,
            "sequence_rows": 160.0,
            "sequence_window": 512.0,
            "thermal_status_end": 2,
            "thermal_status_start": 0,
            "timing_decide_4q_total_ms_median": 1285.523437,
            "timing_forwards_for_4q": 1.0,
            "timing_four_decide_1q_total_ms_median": 4918.670884,
            "timing_repeats": 2.0,
            "timing_state_tokens": 19.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; hfmodels-android SDK device test (EncoderDecisions deberta_decision)",
          "total_ops": 1785,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-04/open-decision-deberta-v3-large__deberta_v3_large_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-05",
          "decode_tokens_per_s": null,
          "delegated_ops": 1785,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-hfmodels-decide-gate",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "hfmodels-android SDK device check: androidTest DebertaDecisionDeviceTest of the litert module (git 4706851), run by tools/decide_gate.sh on the Galaxy S26, one instrumentation process per run; the SDK (hfmodels-litert EncoderDecisions, family deberta_decision) resolved litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT through the dev descriptor catalog/dev/litert-community__Open-Decision-DeBERTa-v3-Large-LiteRT.hfmodels.json at that commit (the log's descriptor sha256 prefix b9226bd1 = `git show 4706851:catalog/dev/litert-community__Open-Decision-DeBERTa-v3-Large-LiteRT.hfmodels.json`); graph deberta_v3_large_decision_s512_wfp16.tflite sha256 238f0cc4dca7b6f18677950e9261d07401a3de9857d76f2a7450fada7f97a5d9 = the published deberta_v3_large_decision_s512_wfp16.tflite (litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT@7a276235; other records hold rows of the same name and digest).",
            "# hfmodels typed-decisions device check  date=2026-10-05T02:13:57Z serial=RFGL80R6A6H device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 android=16",
            "# git=4706851 test=DebertaDecisionDeviceTest variant=s512_wfp16 backend=gpu thermal=Thermal Status: 0 battery_level= level: 79 battery_temp_x10C= temperature: 375 screen= mWakefulness=Dozing",
            "# gradle rc=0 end=2026-10-05T02:15:50Z thermal_end=Thermal Status: 2 battery_temp_end_x10C= temperature: 414",
            "resolve litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT: commit=7a276235 via EXPLICIT_REVISION, descriptor=caller@explicit/hfmodels.json sha=b9226bd1",
            "imported deberta_v3_large_decision_s512_wfp16.tflite: 813443728 bytes, sha256 ok, from /data/local/tmp/hfmodels/open-decision-r21/deberta_v3_large_decision_s512_wfp16.tflite",
            "ready litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT@7a276235 variant=s512_wfp16 profile=gpu inference=GPU prepare_ms=5714",
            "litert  : [compiled_model.cc:1250] 0 compiler plugins were applied successfully: ",
            "tflite  : Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "RESULT step=load ok=true variant=s512_wfp16 backend=gpu import_ms=989 load_ms=5725 compile_ms=5076 profile=gpu window=512 head=512 max_options=128 descriptor_patch=none npu_libs_ready=true notes=family=deberta_decision compile_ms=5076 on gpu (precision fp32) | observed backend is UNKNOWN: litert 2.2.0 reports no per-op execution target; 'initialized' is the accelerator the CompiledModel was created with | host token lookup: word_embeddings_fp16.bin (float16, 128100 rows x 1024) feeds inputs_embeds | the questions of one decide() in one forward as the author's decide() does, split in question order when they do not fit the window or the option slots: options = a choice's descriptions (the key when there is none), a score's levels, a noul's no / yes; the state is cut to its first 256 tokens; a request longer than the 512-token window or with more than 128 options is refused; probabilities are the softmax of the logits / 1.05 (the author's calibration), extras.max_probability is the author's confidence",
            "RESULT step=sequence ok=true variant=s512_wfp16 backend=gpu rows=160 of 160 mismatches=0 window=512",
            "RESULT step=parity ok=true variant=s512_wfp16 backend=gpu skipped=true rows=160 (run argument parity=0: the demo is the graph's first work after the warm-up)",
            "RESULT step=demo ok=true variant=s512_wfp16 backend=gpu form=policy rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=9.328052532620035E-7 (worst b10) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=19/30 correct_mac=19/30 question_ms_median=594.641043 p90=600.0148439999999 min=588.0835930000001 max=605.296823 lookup_ms_median=3.938229",
            "RESULT step=demo ok=true variant=s512_wfp16 backend=gpu form=keys rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=2.028996443748099E-6 (worst b22) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=22/30 correct_mac=22/30 question_ms_median=1157.3343750000001 p90=1285.0666669999998 min=588.9578630000001 max=1289.613332 lookup_ms_median=3.845938",
            "RESULT step=timing ok=true variant=s512_wfp16 backend=gpu repeats=5 forwards_for_4q=1 decide_4q_total_ms_median=1043.04177 four_decide_1q_total_ms_median=4129.436769 state_tokens=19 per_question_ms=1038.000988,1038.000988,1038.000988,1038.000988 answers={department=billing 0.74095467372367, urgency=0.6637985189196726, refund_requested=0.9173051284357872, tone=neutral 0.4445183378688088}",
            "RESULT step=zoo_input ok=true variant=s512_wfp16 backend=gpu text_tokens=11 forward_tokens=64 window=512 first=1 warmups=2 timed=10 wall_ms_median=957.898229 total_ms_median=957.822682 state_ms_median=1.39237 question_ms_median=956.2674999999999 lookup_ms_median=3.6247395 tokenize_ms_median=1.418281 forward_build_ms_median=3.440443 graph_call_ms_median=1026.9220305 answer=promise p=nothing:0.04098411157859003,promise:0.3786183967831869,request:0.2614846801469182,plan:0.3189128114913048 wall_ms=956.893593,958.732969,953.482656,957.063489,957.026406,956.794375,1029.262604,1037.175,1037.136457,1030.949583",
            "RESULT step=lookup ok=true variant=s512_wfp16 backend=gpu rows=256 hidden=1024 zoo_style=arith zoo_ms_median=18.591041 zoo_ms_min=18.395417 sdk_ms_median=0.700677 sdk_ms_min=0.638489 repeats=20 outputs_identical=true",
            "RESULT step=release ok=true variant=s512_wfp16 backend=gpu close_ms=81",
            "RESULT ok=true model=litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT variant=s512_wfp16 backend=gpu device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 failures=",
            "latency_p50_ms null: the run skipped the parity step (run argument parity=0) so that the demo was the graph's first work after the warm-up; the demo medians are metrics.",
            "output_match null: the gate compares with the publisher's outputs under its own criterion (argmax and max |dp| <= 0.01, in metrics) and keeps no raw per-element outputs. full_delegation from the delegate line the gate's logcat filter now keeps ('Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL)').",
            "conditions (the gate's header and footer): thermal status 0 -> 2, battery 37.5 -> 41.4 C, battery level 79 %, display mWakefulness=Dozing at the start.",
            "source: hfmodels-android/litert/results/2026-10-05-1113-RFGL80R6A6H-litert2.2.0-decide-opendecision-cold-parity0-s512_wfp16-gpu.log (main 3e1fcd8)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "battery_level_start": 79,
            "battery_temp_c_end": 41.4,
            "battery_temp_c_start": 37.5,
            "demo_keys_correct_device": 22.0,
            "demo_keys_correct_device_of": 30.0,
            "demo_keys_correct_mac": 22.0,
            "demo_keys_correct_mac_of": 30.0,
            "demo_keys_dp_tol": 0.01,
            "demo_keys_encoded_length_equal_host": 30.0,
            "demo_keys_encoded_length_equal_host_of": 30.0,
            "demo_keys_lookup_ms_median": 3.845938,
            "demo_keys_max": 1289.613332,
            "demo_keys_max_prob_abs_err_vs_mac": 2.028996443748099e-06,
            "demo_keys_min": 588.9578630000001,
            "demo_keys_p90": 1285.0666669999998,
            "demo_keys_question_ms_median": 1157.3343750000001,
            "demo_keys_rows": 30.0,
            "demo_keys_same_answer_as_mac": 30.0,
            "demo_keys_same_answer_as_mac_of": 30.0,
            "demo_policy_correct_device": 19.0,
            "demo_policy_correct_device_of": 30.0,
            "demo_policy_correct_mac": 19.0,
            "demo_policy_correct_mac_of": 30.0,
            "demo_policy_dp_tol": 0.01,
            "demo_policy_encoded_length_equal_host": 30.0,
            "demo_policy_encoded_length_equal_host_of": 30.0,
            "demo_policy_lookup_ms_median": 3.938229,
            "demo_policy_max": 605.296823,
            "demo_policy_max_prob_abs_err_vs_mac": 9.328052532620035e-07,
            "demo_policy_min": 588.0835930000001,
            "demo_policy_p90": 600.0148439999999,
            "demo_policy_question_ms_median": 594.641043,
            "demo_policy_rows": 30.0,
            "demo_policy_same_answer_as_mac": 30.0,
            "demo_policy_same_answer_as_mac_of": 30.0,
            "load_compile_ms": 5076.0,
            "load_head": 512.0,
            "load_import_ms": 989.0,
            "load_load_ms": 5725.0,
            "load_max_options": 128.0,
            "load_window": 512.0,
            "lookup_hidden": 1024.0,
            "lookup_repeats": 20.0,
            "lookup_rows": 256.0,
            "lookup_sdk_ms_median": 0.700677,
            "lookup_sdk_ms_min": 0.638489,
            "lookup_zoo_ms_median": 18.591041,
            "lookup_zoo_ms_min": 18.395417,
            "parity_rows": 160.0,
            "prepare_ms": 5714.0,
            "release_close_ms": 81.0,
            "sequence_mismatches": 0.0,
            "sequence_rows": 160.0,
            "sequence_window": 512.0,
            "thermal_status_end": 2,
            "thermal_status_start": 0,
            "timing_decide_4q_total_ms_median": 1043.04177,
            "timing_forwards_for_4q": 1.0,
            "timing_four_decide_1q_total_ms_median": 4129.436769,
            "timing_repeats": 5.0,
            "timing_state_tokens": 19.0,
            "zoo_input_first": 1.0,
            "zoo_input_forward_build_ms_median": 3.440443,
            "zoo_input_forward_tokens": 64.0,
            "zoo_input_graph_call_ms_median": 1026.9220305,
            "zoo_input_lookup_ms_median": 3.6247395,
            "zoo_input_question_ms_median": 956.2674999999999,
            "zoo_input_state_ms_median": 1.39237,
            "zoo_input_text_tokens": 11.0,
            "zoo_input_timed": 10.0,
            "zoo_input_tokenize_ms_median": 1.418281,
            "zoo_input_total_ms_median": 957.822682,
            "zoo_input_wall_ms_median": 957.898229,
            "zoo_input_warmups": 2.0,
            "zoo_input_window": 512.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; hfmodels-android SDK device test (EncoderDecisions deberta_decision); parity step skipped (parity=0), the demo first after the warm-up",
          "total_ops": 1785,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-05/open-decision-deberta-v3-large__deberta_v3_large_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-04",
          "decode_tokens_per_s": null,
          "delegated_ops": 1785,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-hfmodels-decide-gate",
            "os_build": "Android 16 (build BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "hfmodels-android SDK device check: androidTest DebertaDecisionDeviceTest of the litert module (git 402a874), run by tools/decide_gate.sh on the Galaxy S26, one instrumentation process per run; the SDK (hfmodels-litert EncoderDecisions, family deberta_decision) resolved litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT through the dev descriptor catalog/dev/litert-community__Open-Decision-DeBERTa-v3-Large-LiteRT.hfmodels.json at that commit (the log's descriptor sha256 prefix a092c779 = `git show 402a874:catalog/dev/litert-community__Open-Decision-DeBERTa-v3-Large-LiteRT.hfmodels.json`); graph deberta_v3_large_decision_s512_wfp16.tflite sha256 238f0cc4dca7b6f18677950e9261d07401a3de9857d76f2a7450fada7f97a5d9 = the published deberta_v3_large_decision_s512_wfp16.tflite (litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT@7a276235; other records hold rows of the same name and digest).",
            "# hfmodels typed-decisions device check  date=2026-10-04T06:57:09Z serial=RFGL80R6A6H device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 android=16",
            "# git=402a874 test=DebertaDecisionDeviceTest variant=s512_wfp16 backend=gpu thermal=Thermal Status: 0 battery_level= level: 79 battery_temp_x10C= temperature: 369 screen= mWakefulness=Dozing",
            "# gradle rc=0 end=2026-10-04T07:02:14Z thermal_end=Thermal Status: 2 battery_temp_end_x10C= temperature: 419",
            "resolve litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT: commit=7a276235 via EXPLICIT_REVISION, descriptor=caller@explicit/hfmodels.json sha=a092c779",
            "imported deberta_v3_large_decision_s512_wfp16.tflite: 813443728 bytes, sha256 ok, from /data/local/tmp/hfmodels/open-decision/deberta_v3_large_decision_s512_wfp16.tflite",
            "ready litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT@7a276235 variant=s512_wfp16 profile=gpu inference=GPU prepare_ms=4914",
            "tflite  : Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "RESULT step=load ok=true variant=s512_wfp16 backend=gpu import_ms=1113 load_ms=4925 compile_ms=4356 profile=gpu window=512 head=512 max_options=128 notes=family=deberta_decision compile_ms=4356 on gpu (precision fp32) | observed backend is UNKNOWN: litert 2.2.0 reports no per-op execution target; 'initialized' is the accelerator the CompiledModel was created with | host token lookup: word_embeddings_fp16.bin (float16, 128100 rows x 1024) feeds inputs_embeds | the questions of one decide() in one forward as the author's decide() does, split in question order when they do not fit the window or the option slots: options = a choice's descriptions (the key when there is none), a score's levels, a noul's no / yes; the state is cut to its first 256 tokens; a request longer than the 512-token window or with more than 128 options is refused; probabilities are the softmax of the logits / 1.05 (the author's calibration), extras.max_probability is the author's confidence",
            "RESULT step=sequence ok=true variant=s512_wfp16 backend=gpu rows=160 of 160 mismatches=0 window=512",
            "RESULT step=parity ok=true variant=s512_wfp16 backend=gpu rows=160 questions=420 argmax_match=420/420 answers_equal_author=420/420 max_prob_abs_err=6.861624558591184E-4 (worst fam-ood-47 q0) max_logit_abs_err_vs_author=0.004940509796142578 dp_tol=0.01 forward_ms_median=1031.236406 p90=1200.968542 min=576.186979 max=1289.640781",
            "RESULT step=demo ok=true variant=s512_wfp16 backend=gpu form=policy rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=9.328052532620035E-7 (worst b10) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=19/30 correct_mac=19/30 question_ms_median=1206.4008849999998 p90=1221.117082 min=1044.615364 max=1225.7235930000002 lookup_ms_median=3.174479",
            "RESULT step=demo ok=true variant=s512_wfp16 backend=gpu form=keys rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=2.028996443748099E-6 (worst b22) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=22/30 correct_mac=22/30 question_ms_median=1278.974948 p90=1292.4499480000002 min=1197.498853 max=1322.97651 lookup_ms_median=3.041875",
            "RESULT step=timing ok=true variant=s512_wfp16 backend=gpu repeats=5 forwards_for_4q=1 decide_4q_total_ms_median=1292.511719 four_decide_1q_total_ms_median=5151.086925 state_tokens=19 per_question_ms=1280.524948,1280.524948,1280.524948,1280.524948 answers={department=billing 0.74095467372367, urgency=0.6637985189196726, refund_requested=0.9173051284357872, tone=neutral 0.4445183378688088}",
            "RESULT step=zoo_input ok=true variant=s512_wfp16 backend=gpu text_tokens=11 forward_tokens=64 window=512 first=1 warmups=2 timed=10 wall_ms_median=1288.8432025 total_ms_median=1288.7717185000001 state_ms_median=1.3258595 question_ms_median=1287.2986455 lookup_ms_median=3.083594 tokenize_ms_median=1.466406 forward_build_ms_median=3.4091665 graph_call_ms_median=1277.1552864999999 answer=promise p=nothing:0.04098411157859003,promise:0.3786183967831869,request:0.2614846801469182,plan:0.3189128114913048 wall_ms=1279.518542,1290.890103,1288.879843,1286.619427,1286.570677,1293.281718,1288.806562,1300.661041,1325.844375,1287.759635",
            "RESULT step=lookup ok=true variant=s512_wfp16 backend=gpu rows=256 hidden=1024 zoo_style=arith zoo_ms_median=21.272448 zoo_ms_min=21.234896 sdk_ms_median=0.69349 sdk_ms_min=0.692187 repeats=20 outputs_identical=true",
            "RESULT step=release ok=true variant=s512_wfp16 backend=gpu close_ms=105",
            "RESULT ok=true model=litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT variant=s512_wfp16 backend=gpu device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 failures=",
            "latency_p50_ms = the parity step's forward_ms_median: one LiteRtDecisionModel.scores call per captured request (host float16 table lookup, input writes, CompiledModel run, output readback on the model thread), after one warm forward.",
            "output_match null: the gate compares with the publisher's outputs under its own criterion (argmax and max |dp| <= 0.01, in metrics) and keeps no raw per-element outputs. full_delegation from the delegate line the gate's logcat filter now keeps ('Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL)').",
            "conditions (the gate's header and footer): thermal status 0 -> 2, battery 36.9 -> 41.9 C, battery level 79 %, display mWakefulness=Dozing at the start.",
            "source: hfmodels-android/litert/results/2026-10-04-1557-RFGL80R6A6H-litert2.2.0-decide-opendecision-cold-s512_wfp16-gpu.log (main 3e1fcd8)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 1031.236406,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "battery_level_start": 79,
            "battery_temp_c_end": 41.9,
            "battery_temp_c_start": 36.9,
            "demo_keys_correct_device": 22.0,
            "demo_keys_correct_device_of": 30.0,
            "demo_keys_correct_mac": 22.0,
            "demo_keys_correct_mac_of": 30.0,
            "demo_keys_dp_tol": 0.01,
            "demo_keys_encoded_length_equal_host": 30.0,
            "demo_keys_encoded_length_equal_host_of": 30.0,
            "demo_keys_lookup_ms_median": 3.041875,
            "demo_keys_max": 1322.97651,
            "demo_keys_max_prob_abs_err_vs_mac": 2.028996443748099e-06,
            "demo_keys_min": 1197.498853,
            "demo_keys_p90": 1292.4499480000002,
            "demo_keys_question_ms_median": 1278.974948,
            "demo_keys_rows": 30.0,
            "demo_keys_same_answer_as_mac": 30.0,
            "demo_keys_same_answer_as_mac_of": 30.0,
            "demo_policy_correct_device": 19.0,
            "demo_policy_correct_device_of": 30.0,
            "demo_policy_correct_mac": 19.0,
            "demo_policy_correct_mac_of": 30.0,
            "demo_policy_dp_tol": 0.01,
            "demo_policy_encoded_length_equal_host": 30.0,
            "demo_policy_encoded_length_equal_host_of": 30.0,
            "demo_policy_lookup_ms_median": 3.174479,
            "demo_policy_max": 1225.7235930000002,
            "demo_policy_max_prob_abs_err_vs_mac": 9.328052532620035e-07,
            "demo_policy_min": 1044.615364,
            "demo_policy_p90": 1221.117082,
            "demo_policy_question_ms_median": 1206.4008849999998,
            "demo_policy_rows": 30.0,
            "demo_policy_same_answer_as_mac": 30.0,
            "demo_policy_same_answer_as_mac_of": 30.0,
            "load_compile_ms": 4356.0,
            "load_head": 512.0,
            "load_import_ms": 1113.0,
            "load_load_ms": 4925.0,
            "load_max_options": 128.0,
            "load_window": 512.0,
            "lookup_hidden": 1024.0,
            "lookup_repeats": 20.0,
            "lookup_rows": 256.0,
            "lookup_sdk_ms_median": 0.69349,
            "lookup_sdk_ms_min": 0.692187,
            "lookup_zoo_ms_median": 21.272448,
            "lookup_zoo_ms_min": 21.234896,
            "parity_answers_equal_author": 420.0,
            "parity_answers_equal_author_of": 420.0,
            "parity_argmax_match": 420.0,
            "parity_argmax_match_of": 420.0,
            "parity_dp_tol": 0.01,
            "parity_forward_ms_median": 1031.236406,
            "parity_max": 1289.640781,
            "parity_max_logit_abs_err_vs_author": 0.004940509796142578,
            "parity_max_prob_abs_err": 0.0006861624558591184,
            "parity_min": 576.186979,
            "parity_p90": 1200.968542,
            "parity_rows": 160.0,
            "prepare_ms": 4914.0,
            "release_close_ms": 105.0,
            "sequence_mismatches": 0.0,
            "sequence_rows": 160.0,
            "sequence_window": 512.0,
            "thermal_status_end": 2,
            "thermal_status_start": 0,
            "timing_decide_4q_total_ms_median": 1292.511719,
            "timing_forwards_for_4q": 1.0,
            "timing_four_decide_1q_total_ms_median": 5151.086925,
            "timing_repeats": 5.0,
            "timing_state_tokens": 19.0,
            "zoo_input_first": 1.0,
            "zoo_input_forward_build_ms_median": 3.4091665,
            "zoo_input_forward_tokens": 64.0,
            "zoo_input_graph_call_ms_median": 1277.1552864999999,
            "zoo_input_lookup_ms_median": 3.083594,
            "zoo_input_question_ms_median": 1287.2986455,
            "zoo_input_state_ms_median": 1.3258595,
            "zoo_input_text_tokens": 11.0,
            "zoo_input_timed": 10.0,
            "zoo_input_tokenize_ms_median": 1.466406,
            "zoo_input_total_ms_median": 1288.7717185000001,
            "zoo_input_wall_ms_median": 1288.8432025,
            "zoo_input_warmups": 2.0,
            "zoo_input_window": 512.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; hfmodels-android SDK device test (EncoderDecisions deberta_decision); started after a 5-minute cool-down",
          "total_ops": 1785,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-04/open-decision-deberta-v3-large__deberta_v3_large_decision_s512_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-opendecision-gate",
            "os_build": "Android 16 (SDK 36, build BP4A.251205.006)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": "Scudo ERROR: internal map failure (error desc=Out of memory)",
          "evidence": [
            "JIT compile of the S512 graph on the Hexagon HTP aborted 93.6 s after 'compiling ... on npu' with SIGABRT inside libQnnHtpPrepare.so GraphPrepare::sequencing_stage -> operator new -> malloc (host-side memory; the GPU compiles the same 1785 ops as one partition).",
            "10-02 00:04:36.065  8191  8232 I DecisionGate: compiling deberta_v3_large_decision_s512_wfp16.tflite on npu fp32",
            "10-02 00:04:36.457  8191  8232 I litert  : [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 1785 ops, from a total of 1785 ops. resulted in 1 partitions.",
            "10-02 00:06:09.641  8191  8232 F libc    : Fatal signal 6 (SIGABRT), code -1 (SI_QUEUE) in tid 8232 (DecisionGate), pid 8191 (m.decision.gate)",
            "10-02 00:06:10.931  8449  8449 F DEBUG   : Abort message: 'Scudo ERROR: internal map failure (error desc=Out of memory)'",
            "10-02 00:06:10.931  8449  8449 F DEBUG   :       #17 pc 00000000030ec050  /data/app/~~5uBIgH4pC_a1qEFiEYN5MQ==/com.decision.gate-coUwnWQN9ARmSXz4TteWHw==/lib/arm64/libQnnHtpPrepare.so (GraphPrepare::sequencing_stage(VtcmCacheInstance&, int&, bool, std::__ndk1::vector<hnnx::tagx<1u>, std::__ndk1::all"
          ],
          "failure_class": "engine_create_failed",
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": false,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "battery_temp_c_start": 31.0,
            "compile_abort_after_s": 93.6,
            "thermal_status_start": 0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": false,
          "signature": "QualcommOptions htpPerformanceMode=BURST, JIT (no AOT); debug gate app com.decision.gate",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/open-decision-deberta-v3-large__deberta_v3_large_decision_s512_wfp16__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "deberta-v3-large",
    "id": "open-decision-deberta-v3-large__deberta_v3_large_decision_s512_wfp16",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT",
    "task": "typed decisions (state + choice / score / noul questions -> one calibrated distribution per question; text-classification)"
  },
  "pitfalls": [
    "Run the GPU with explicit FP32 precision (GpuOptions precision = FP32 on Android, enforce_f32=True in Python). At the default precision every output is non-finite on Mac Metal (the attention mask constant -3.4e38 is -inf in fp16) and, on the Galaxy S26 with a finite-mask build, finite but wrong (216/274 answers).",
    "In fp16 computation two parts fail: the mask constant (NaN) and the LayerNorm sums of squares (1024 x 28.5^2 overflows fp16). The card documents the fp32-exact fix (-1e4 mask + SafeLayerNorm 2^-3); it is not applied to the shipped graphs, and the resulting NPU graph is not shipped (274/276 answers on the S26 Hexagon NPU at 110 ms).",
    "The S512 graph cannot be JIT-compiled for the Galaxy S26 NPU (Hexagon HTP prepare runs out of host memory 94 s in); the S256 graph compiles in 170-200 s as one DispatchDelegate node.",
    "Host contract: [CLS] [STATE] state[:256] ([Q] instructions ([OPT] option)+)+ [SEP] from the source tokenizer.json (keep the Precompiled charsmap; spm.model differs on exotic characters); the state is cut at 256 tokens, the sequence must fit the window (512) and the options the 128 slots; q_routing / o_routing rows are 1/len over the question / option text tokens; softmax at temperature 1.05 within each question.",
    "English only (banking77 / SST-5 / BoolQ training domains, per the author); the conversion reproduces the author's answers, including the wrong ones (author's in-domain accuracy 0.854, OOD 0.690).",
    "The S26 GPU warms under load: over 500 back-to-back 512-token requests the thermal status rose from 0 to 3 and per-request time from ~584 ms to ~1,024 ms (median 697.5 ms); the S256 graph measured 407.6 ms on an already warm phone."
  ],
  "schema_version": "1.2"
}
