{
  "artifacts": [
    {
      "file": "deberta_v3_large_decision_s256_wfp16.tflite",
      "sha256": "4514cd45d4ee6bf1025818bbecd140d4f744fa78d8309e22a17b89448c176e45",
      "size_mb": 679.76
    }
  ],
  "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_s256_wfp16.json"
  },
  "conversion": {
    "command": "python scripts/export.py --seq 256  (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_s256_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": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-03",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "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 5f9bc69), run by tools/decide_gate.sh (GATE_TEST=DebertaDecisionDeviceTest) on the Galaxy S26, one instrumentation process per (variant x backend); 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 (the log's descriptor sha256 prefix 1e2e59e2 = `git show 5f9bc69:catalog/dev/litert-community__Open-Decision-DeBERTa-v3-Large-LiteRT.hfmodels.json`), checked every file's sha256 at import and looks the token embeddings up on the host from the float16 table (CPU, 4 threads).",
            "ready litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT@7a276235 variant=s256_wfp16 profile=cpu inference=CPU prepare_ms=2788",
            "no import line: the load step reports import_ms=0, the files already in the SDK's store from the GPU run of the same variant just before, where they were imported with 'sha256 ok'",
            "RESULT step=load ok=true variant=s256_wfp16 backend=cpu import_ms=0 load_ms=2807 compile_ms=1862 profile=cpu window=256 head=256 max_options=128 notes=family=deberta_decision compile_ms=1862 on cpu (4 threads) | 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 256-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=s256_wfp16 backend=cpu rows=143 of 160 mismatches=0 window=256",
            "RESULT step=parity ok=true variant=s256_wfp16 backend=cpu rows=143 questions=371 argmax_match=371/371 answers_equal_author=371/371 max_prob_abs_err=6.865743171473593E-4 (worst fam-ood-47 q0) max_logit_abs_err_vs_author=0.0021677017211914062 dp_tol=0.01 forward_ms_median=631.75276 p90=652.560312 min=588.304167 max=661.984323",
            "RESULT step=demo ok=true variant=s256_wfp16 backend=cpu form=policy rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=7.364112837859871E-7 (worst b01) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=19/30 correct_mac=19/30 question_ms_median=654.6046329999999 p90=733.4785939999999 min=652.370051 max=737.960729 lookup_ms_median=0.558125",
            "RESULT step=demo ok=true variant=s256_wfp16 backend=cpu form=keys rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=1.2969376595450832E-6 (worst b19) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=22/30 correct_mac=22/30 question_ms_median=737.102551 p90=744.86901 min=730.3209380000001 max=764.898281 lookup_ms_median=0.566823",
            "RESULT step=timing ok=true variant=s256_wfp16 backend=cpu repeats=5 forwards_for_4q=1 decide_4q_total_ms_median=733.554531 four_decide_1q_total_ms_median=2930.190051 state_tokens=19 per_question_ms=734.269895,734.269895,734.269895,734.269895 answers={department=billing 0.7409549793857945, urgency=0.6637997150287804, refund_requested=0.9173052145575664, tone=neutral 0.44451788455769864}",
            "RESULT step=release ok=true variant=s256_wfp16 backend=cpu close_ms=149",
            "RESULT ok=true model=litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT variant=s256_wfp16 backend=cpu device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 failures=",
            "parity: the conversion run's 160 app-gate requests, the 143 that fit 256 tokens (371 questions), each packed into one forward as the author's decide() does, against the author's typed_decisions outputs (softmax at T = 1.05); demo: the 30 chat sentences of the sieve (r1 b_chat) through TypedDecisions.decide(text, question), options = the choices' descriptions (form policy) or keys (form keys), against the card's Python host on a Mac (CPU).",
            "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 and output readback on the model thread), after one warm forward. demo_*question_* = DecisionTiming.questionMs of one decide() (tokenize, forward, decode), after two warm calls.",
            "output_match null: the gate compares with the publisher's outputs under its own criterion (argmax and max |dp| <= dp_tol 0.01, in metrics) and keeps no raw per-element outputs, so the shared element rule cannot be recomputed. full_delegation null: litert 2.2.0 reports no per-op execution target to the SDK ('observed backend is UNKNOWN' in the load notes) and the gate's logcat filter keeps only the hfmodels tags, so no delegate line was captured.",
            "conditions (the gate's header and footer lines): thermal status 1 -> 1, battery 37.1 -> 38.4 C, battery level 78 %, display mWakefulness=Dozing at the start.",
            "source: hfmodels-android/litert/results/2026-10-03-1939-RFGL80R6A6H-litert2.2.0-decide-opendecision-s256_wfp16-cpu.log (committed), hfmodels-android/verification/2026-10-03-RFGL80R6A6H-BP4A.251205.006-deberta-decision.json."
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 631.75276,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answers_equal_author": 371.0,
            "argmax_equal_official": 371.0,
            "battery_temp_c_end": 38.4,
            "battery_temp_c_start": 37.1,
            "close_ms": 149.0,
            "cpu_threads": 4.0,
            "decide_4q_total_median_ms": 733.554531,
            "demo_keys_correct_device": 22.0,
            "demo_keys_correct_mac": 22.0,
            "demo_keys_encoded_length_equal_host": 30.0,
            "demo_keys_host_lookup_median_ms": 0.566823,
            "demo_keys_max_abs_dprob_vs_mac": 1.2969376595450832e-06,
            "demo_keys_question_max_ms": 764.898281,
            "demo_keys_question_median_ms": 737.102551,
            "demo_keys_question_min_ms": 730.3209380000001,
            "demo_keys_question_p90_ms": 744.86901,
            "demo_keys_same_answer_as_mac": 30.0,
            "demo_keys_sentences": 30.0,
            "demo_policy_correct_device": 19.0,
            "demo_policy_correct_mac": 19.0,
            "demo_policy_encoded_length_equal_host": 30.0,
            "demo_policy_host_lookup_median_ms": 0.558125,
            "demo_policy_max_abs_dprob_vs_mac": 7.364112837859871e-07,
            "demo_policy_question_max_ms": 737.960729,
            "demo_policy_question_median_ms": 654.6046329999999,
            "demo_policy_question_min_ms": 652.370051,
            "demo_policy_question_p90_ms": 733.4785939999999,
            "demo_policy_same_answer_as_mac": 30.0,
            "demo_policy_sentences": 30.0,
            "dp_tol": 0.01,
            "fixtures": 143.0,
            "forward_max_ms": 661.984323,
            "forward_median_ms": 631.75276,
            "forward_min_ms": 588.304167,
            "forward_p90_ms": 652.560312,
            "forwards_for_4q": 1.0,
            "four_decide_1q_total_median_ms": 2930.190051,
            "import_ms": 0.0,
            "load_compile_ms": 1862.0,
            "load_ms": 2807.0,
            "max_abs_dlogit_vs_author": 0.0021677017211914062,
            "max_abs_dprob_vs_official": 0.0006865743171473593,
            "max_options": 128.0,
            "prepare_ms": 2788.0,
            "questions": 371.0,
            "sequence_mismatches": 0.0,
            "sequence_requests": 143.0,
            "sequence_requests_total": 160.0,
            "thermal_status_end": 1,
            "thermal_status_start": 1,
            "timing_repeats": 5.0,
            "timing_state_tokens": 19.0,
            "window": 256.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "CpuOptions 4 threads; hfmodels-android SDK device test (EncoderDecisions deberta_decision)",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-03/open-decision-deberta-v3-large__deberta_v3_large_decision_s256_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-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": 298.967292,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "argmax_equal_official": 371,
            "fixtures": 143,
            "ids_equal_official_collator": 143,
            "latency_max_ms": 388.394218,
            "latency_min_ms": 176.177135,
            "max_abs_dprob_vs_official_T1_05": 0.0006861766536193636,
            "questions": 371,
            "spans_equal_official_collator": 143,
            "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_s256_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 2, battery 42.4 C; after thermal status 2, battery 43.0 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 1074/1074 questions (712/712 boundary = reference top-1 < 0.9); max |dprob| at T=1.05 0.00074; 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 3475 compared option logits, max |d| 1.8835067749023438e-05.",
            "10-02 00:17:30.733 12217 12240 I tflite  : Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 00:17:35.133 12217 12240 I DecisionGate: DECISION_GATE compiled in 4970.962602 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 407.64333350000004,
          "loads": true,
          "max_abs_diff": 1.8835067749023438e-05,
          "max_rel_diff": 0.0017240820296668558,
          "metrics": {
            "argmax_equal_official": 1074,
            "battery_temp_c_end": 43.0,
            "battery_temp_c_start": 42.4,
            "boundary_argmax_equal_official": 712,
            "boundary_questions": 712,
            "element_rule_elements": 3475,
            "element_rule_violations": 0,
            "first_call_ms": 428.14099,
            "fixtures": 419,
            "host_fp16_lookup_median_ms": 21.7,
            "load_compile_ms": 4970.962602,
            "max_abs_dprob_vs_official_T1_05": 0.00074,
            "max_abs_dscore_vs_official": 0.00079,
            "nonfinite_questions": 0,
            "questions": 1074,
            "questions_dprob_gt_1e2": 0,
            "thermal_status_end": 2,
            "thermal_status_start": 2,
            "warm_max_ms": 484.667813,
            "warm_median_ms": 407.593385,
            "warm_min_ms": 375.552187,
            "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_s256_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_s256_wfp16.tflite sha256 4514cd45d4ee6bf1025818bbecd140d4f744fa78d8309e22a17b89448c176e45 = the published deberta_v3_large_decision_s256_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=2026-10-04T07:10:59Z serial=RFGL80R6A6H device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 android=16",
            "# git=402a874 test=DebertaDecisionDeviceTest variant=s256_wfp16 backend=gpu thermal=Thermal Status: 0 battery_level= level: 79 battery_temp_x10C= temperature: 363 screen= mWakefulness=Dozing",
            "# gradle rc=0 end=2026-10-04T07:12:16Z thermal_end=Thermal Status: 2 battery_temp_end_x10C= temperature: 410",
            "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_s256_wfp16.tflite: 712780432 bytes, sha256 ok, from /data/local/tmp/hfmodels/open-decision/deberta_v3_large_decision_s256_wfp16.tflite",
            "ready litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT@7a276235 variant=s256_wfp16 profile=gpu inference=GPU prepare_ms=4048",
            "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=s256_wfp16 backend=gpu import_ms=1116 load_ms=4058 compile_ms=3497 profile=gpu window=256 head=256 max_options=128 notes=family=deberta_decision compile_ms=3497 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 256-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=s256_wfp16 backend=gpu rows=143 of 160 mismatches=0 window=256",
            "RESULT step=parity ok=true variant=s256_wfp16 backend=gpu rows=143 questions=371 argmax_match=371/371 answers_equal_author=371/371 max_prob_abs_err=6.861624558591184E-4 (worst fam-ood-47 q0) max_logit_abs_err_vs_author=0.00216519832611084 dp_tol=0.01 forward_ms_median=185.282708 p90=194.393646 min=177.676823 max=429.846667",
            "RESULT step=demo ok=true variant=s256_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=387.455989 p90=393.401822 min=330.927187 max=397.53062400000005 lookup_ms_median=1.928177",
            "RESULT step=demo ok=true variant=s256_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=331.49473900000004 p90=333.52302 min=309.39271 max=337.187187 lookup_ms_median=1.847552",
            "RESULT step=timing ok=true variant=s256_wfp16 backend=gpu repeats=5 forwards_for_4q=1 decide_4q_total_ms_median=312.234479 four_decide_1q_total_ms_median=1223.901719 state_tokens=19 per_question_ms=311.496041,311.496041,311.496041,311.496041 answers={department=billing 0.74095467372367, urgency=0.6637985189196726, refund_requested=0.9173051284357872, tone=neutral 0.4445183378688088}",
            "RESULT step=zoo_input ok=true variant=s256_wfp16 backend=gpu text_tokens=11 forward_tokens=64 window=256 first=1 warmups=2 timed=10 wall_ms_median=307.55333299999995 total_ms_median=307.47638 state_ms_median=1.490338 question_ms_median=305.863933 lookup_ms_median=1.704063 tokenize_ms_median=1.5704685 forward_build_ms_median=3.5157809999999996 graph_call_ms_median=301.0510415 answer=promise p=nothing:0.04098411157859003,promise:0.3786183967831869,request:0.2614846801469182,plan:0.3189128114913048 wall_ms=305.05,311.215052,305.773281,311.461615,309.580624,305.878854,309.227812,312.720104,305.753282,305.610833",
            "RESULT step=lookup ok=true variant=s256_wfp16 backend=gpu rows=256 hidden=1024 zoo_style=arith zoo_ms_median=16.793802 zoo_ms_min=16.762448 sdk_ms_median=0.553907 sdk_ms_min=0.552709 repeats=20 outputs_identical=true",
            "RESULT step=release ok=true variant=s256_wfp16 backend=gpu close_ms=77",
            "RESULT ok=true model=litert-community/Open-Decision-DeBERTa-v3-Large-LiteRT variant=s256_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 36.3 -> 41.0 C, battery level 79 %, display mWakefulness=Dozing at the start.",
            "source: hfmodels-android/litert/results/2026-10-04-1610-RFGL80R6A6H-litert2.2.0-decide-opendecision-zoo-s256_wfp16-gpu.log (main 3e1fcd8)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 185.282708,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "battery_level_start": 79,
            "battery_temp_c_end": 41.0,
            "battery_temp_c_start": 36.3,
            "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": 1.847552,
            "demo_keys_max": 337.187187,
            "demo_keys_max_prob_abs_err_vs_mac": 2.028996443748099e-06,
            "demo_keys_min": 309.39271,
            "demo_keys_p90": 333.52302,
            "demo_keys_question_ms_median": 331.49473900000004,
            "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": 1.928177,
            "demo_policy_max": 397.53062400000005,
            "demo_policy_max_prob_abs_err_vs_mac": 9.328052532620035e-07,
            "demo_policy_min": 330.927187,
            "demo_policy_p90": 393.401822,
            "demo_policy_question_ms_median": 387.455989,
            "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": 3497.0,
            "load_head": 256.0,
            "load_import_ms": 1116.0,
            "load_load_ms": 4058.0,
            "load_max_options": 128.0,
            "load_window": 256.0,
            "lookup_hidden": 1024.0,
            "lookup_repeats": 20.0,
            "lookup_rows": 256.0,
            "lookup_sdk_ms_median": 0.553907,
            "lookup_sdk_ms_min": 0.552709,
            "lookup_zoo_ms_median": 16.793802,
            "lookup_zoo_ms_min": 16.762448,
            "parity_answers_equal_author": 371.0,
            "parity_answers_equal_author_of": 371.0,
            "parity_argmax_match": 371.0,
            "parity_argmax_match_of": 371.0,
            "parity_dp_tol": 0.01,
            "parity_forward_ms_median": 185.282708,
            "parity_max": 429.846667,
            "parity_max_logit_abs_err_vs_author": 0.00216519832611084,
            "parity_max_prob_abs_err": 0.0006861624558591184,
            "parity_min": 177.676823,
            "parity_p90": 194.393646,
            "parity_rows": 143.0,
            "prepare_ms": 4048.0,
            "release_close_ms": 77.0,
            "sequence_mismatches": 0.0,
            "sequence_rows": 143.0,
            "sequence_window": 256.0,
            "thermal_status_end": 2,
            "thermal_status_start": 0,
            "timing_decide_4q_total_ms_median": 312.234479,
            "timing_forwards_for_4q": 1.0,
            "timing_four_decide_1q_total_ms_median": 1223.901719,
            "timing_repeats": 5.0,
            "timing_state_tokens": 19.0,
            "zoo_input_first": 1.0,
            "zoo_input_forward_build_ms_median": 3.5157809999999996,
            "zoo_input_forward_tokens": 64.0,
            "zoo_input_graph_call_ms_median": 301.0510415,
            "zoo_input_lookup_ms_median": 1.704063,
            "zoo_input_question_ms_median": 305.863933,
            "zoo_input_state_ms_median": 1.490338,
            "zoo_input_text_tokens": 11.0,
            "zoo_input_timed": 10.0,
            "zoo_input_tokenize_ms_median": 1.5704685,
            "zoo_input_total_ms_median": 307.47638,
            "zoo_input_wall_ms_median": 307.55333299999995,
            "zoo_input_warmups": 2.0,
            "zoo_input_window": 256.0
          },
          "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_s256_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": 1,
          "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 1, battery 40.8 C; after thermal status 2, battery 41.9 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 95/276 questions (54/192 boundary = reference top-1 < 0.9); max |dprob| at T=1.05 0.9986; questions over 0.01: 265.",
            "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: 893 violating of 893 compared option logits, max |d| 63199.58318400383.",
            "The whole graph ran as one DispatchDelegate node ('Partitioned subgraph<0>, selected 1785 ops, from a total of 1785 ops. resulted in 1 partitions', 'Replacing 1 out of 1 node(s) with delegate (DispatchDelegate)'); JIT compile 186.9 s. Outputs are finite but wrong: 95/276 questions keep the author's winning option, max |dprob| 0.9986 -> output_match false under the element rule.",
            "10-02 00:14:05.925 11137 11158 I litert  : [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 1785 ops, from a total of 1785 ops. resulted in 1 partitions.",
            "10-02 00:17:12.058 11137 11158 I tflite  : Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 00:17:12.487 11137 11158 I DecisionGate: DECISION_GATE compiled in 186850.614668 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 129.858073,
          "loads": true,
          "max_abs_diff": 63199.58318400383,
          "max_rel_diff": 16873786.359172937,
          "metrics": {
            "argmax_equal_official": 95,
            "battery_temp_c_end": 41.9,
            "battery_temp_c_start": 40.8,
            "boundary_argmax_equal_official": 54,
            "boundary_questions": 192,
            "element_rule_elements": 893,
            "element_rule_violations": 893,
            "first_call_ms": 135.808281,
            "fixtures": 100,
            "host_fp16_lookup_median_ms": 14.6,
            "jit_compile_s": 186.85,
            "load_compile_ms": 186850.614668,
            "max_abs_dprob_vs_official_T1_05": 0.9986,
            "max_abs_dscore_vs_official": 3.82997,
            "nonfinite_questions": 0,
            "questions": 276,
            "questions_dprob_gt_1e2": 265,
            "thermal_status_end": 2,
            "thermal_status_start": 1,
            "warm_max_ms": 140.277709,
            "warm_median_ms": 129.672135,
            "warm_min_ms": 124.212447,
            "warmup_rows_excluded": 5
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "QualcommOptions htpPerformanceMode=BURST, JIT (first compile, no cache); debug gate app com.decision.gate",
          "total_ops": 1,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/open-decision-deberta-v3-large__deberta_v3_large_decision_s256_wfp16__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "deberta-v3-large",
    "id": "open-decision-deberta-v3-large__deberta_v3_large_decision_s256_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 (256) 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"
}
