{
  "artifacts": [
    {
      "file": "gliclass_edge_v3_s128_fp32.tflite",
      "sha256": "79036f538289fd406e6f33b6a1dda51758c77d5383c7704369248c491dd6dd59",
      "size_mb": 51.216
    }
  ],
  "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": 48.098,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": true
      },
      {
        "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": true,
        "latency_p50_ms": 5.635,
        "loads": true,
        "max_rel_diff": 1.8194276951746214e-05,
        "output_match": true,
        "provenance": "measured",
        "runs": true
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-10-02/gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32.json"
  },
  "conversion": {
    "command": "conversion/export.py --seq 128 --matmul rank4 --ln-shift audit --suffix sln (litert-torch convert of the classification-path graph; the fp16-weight file by ai-edge-quantizer FLOAT_CASTING in the same script); card 'Provenance, conversion and license', REPRODUCE.md of the Hub repo",
    "quantization": "float32 weights: 670 operators, no quantization (card 'Files': the default download).",
    "tool": "litert-torch",
    "tool_version": "0.9.3 (torch 2.12.1; transformers 5.17.0; gliclass 0.1.20; 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": [
      "SQUARED_DIFFERENCE"
    ],
    "coverage_ops_pct": 96.9,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 649,
      "unmatched": 21
    },
    "partitions": 22
  },
  "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 GliclassDeviceTest of the litert module (git 5f9bc69), run by tools/decide_gate.sh (GATE_TEST=GliclassDeviceTest) on the Galaxy S26, one instrumentation process per (variant x backend); the SDK (hfmodels-litert EncoderDecisions, family gliclass) resolved litert-community/GLiClass-Edge-v3.0-LiteRT through the dev descriptor catalog/dev/litert-community__GLiClass-Edge-v3.0-LiteRT.hfmodels.json (the log's descriptor sha256 prefix afe7d983 = `git show 5f9bc69:catalog/dev/litert-community__GLiClass-Edge-v3.0-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/GLiClass-Edge-v3.0-LiteRT@88c90950 variant=s128_fp32 profile=cpu inference=CPU prepare_ms=529",
            "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=s128_fp32 backend=cpu import_ms=0 load_ms=541 compile_ms=37 profile=cpu window=128 head=128 max_options=25 notes=family=gliclass compile_ms=37 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: tok_embeddings_fp16.bin (float16, 50370 rows x 384) feeds inputs_embeds | one gliclass request per question: labels = a choice's descriptions (the key when there is none), a score's levels, a noul's false / true descriptions or no / yes; prompt = the instructions plus one space, right before the text; a request longer than the 128-token window or with more than 25 labels is refused, never cut; probabilities are gliclass's single-label float32 softmax of the label logits (T=1)",
            "RESULT step=sequence ok=true variant=s128_fp32 backend=cpu rows=152 of 152 mismatches=0 window=128",
            "RESULT step=parity ok=true variant=s128_fp32 backend=cpu rows=152 argmax_match=152/152 single_label_equal_official=152/152 multi_label_equal_official=152/152 max_prob_abs_err=0.001450374722480774 (worst semif_135) max_logit_abs_err_vs_oracle=0.009016335010528564 dp_tol=0.01 forward_ms_median=11.875052 p90=12.972135 min=11.46625 max=19.368229",
            "RESULT step=demo ok=true variant=s128_fp32 backend=cpu form=described rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=0.0 (worst ) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=10/30 correct_mac=10/30 question_ms_median=14.823854999999998 p90=16.770105 min=12.695260000000001 max=17.720834 lookup_ms_median=0.15349",
            "RESULT step=demo ok=true variant=s128_fp32 backend=cpu form=keys rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=0.0 (worst ) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=10/30 correct_mac=10/30 question_ms_median=12.992500000000001 p90=14.497708 min=12.796615 max=14.971563 lookup_ms_median=0.134271",
            "RESULT step=timing ok=true variant=s128_fp32 backend=cpu repeats=5 decide_4q_total_ms_median=50.163177 four_decide_1q_total_ms_median=50.370885 state_tokens=21 per_question_ms=13.379114000000001,12.894478,12.462709,12.080677999999999 answers={department=billing 0.9799988865852356, urgency=0.11925152130424976, refund_requested=0.33966735005378723, tone=angry 0.7151762247085571}",
            "RESULT step=release ok=true variant=s128_fp32 backend=cpu close_ms=6",
            "RESULT ok=true model=litert-community/GLiClass-Edge-v3.0-LiteRT variant=s128_fp32 backend=cpu device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 failures=",
            "parity: the card's 152 shareable test requests (android/sample gate_fixtures.json) through the loaded graph, against the official gliclass 0.1.20 pipeline outputs (softmax top label and sigmoid >= 0.5 label set); demo: the 30 chat sentences of the sieve (r1 b_chat) through TypedDecisions.decide(text, question), labels = the choices' descriptions (form described) 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 0 -> 0, battery 29.2 -> 29.3 C, battery level 79 %, display mWakefulness=Dozing at the start.",
            "source: hfmodels-android/litert/results/2026-10-03-1938-RFGL80R6A6H-litert2.2.0-decide-gliclass-s128_fp32-cpu.log (committed), hfmodels-android/verification/2026-10-03-RFGL80R6A6H-BP4A.251205.006-gliclass.json."
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 11.875052,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "argmax_equal_official": 152.0,
            "battery_temp_c_end": 29.3,
            "battery_temp_c_start": 29.2,
            "close_ms": 6.0,
            "cpu_threads": 4.0,
            "decide_4q_total_median_ms": 50.163177,
            "demo_described_correct_device": 10.0,
            "demo_described_correct_mac": 10.0,
            "demo_described_encoded_length_equal_host": 30.0,
            "demo_described_host_lookup_median_ms": 0.15349,
            "demo_described_max_abs_dprob_vs_mac": 0.0,
            "demo_described_question_max_ms": 17.720834,
            "demo_described_question_median_ms": 14.823854999999998,
            "demo_described_question_min_ms": 12.695260000000001,
            "demo_described_question_p90_ms": 16.770105,
            "demo_described_same_answer_as_mac": 30.0,
            "demo_described_sentences": 30.0,
            "demo_keys_correct_device": 10.0,
            "demo_keys_correct_mac": 10.0,
            "demo_keys_encoded_length_equal_host": 30.0,
            "demo_keys_host_lookup_median_ms": 0.134271,
            "demo_keys_max_abs_dprob_vs_mac": 0.0,
            "demo_keys_question_max_ms": 14.971563,
            "demo_keys_question_median_ms": 12.992500000000001,
            "demo_keys_question_min_ms": 12.796615,
            "demo_keys_question_p90_ms": 14.497708,
            "demo_keys_same_answer_as_mac": 30.0,
            "demo_keys_sentences": 30.0,
            "dp_tol": 0.01,
            "fixtures": 152.0,
            "forward_max_ms": 19.368229,
            "forward_median_ms": 11.875052,
            "forward_min_ms": 11.46625,
            "forward_p90_ms": 12.972135,
            "four_decide_1q_total_median_ms": 50.370885,
            "import_ms": 0.0,
            "load_compile_ms": 37.0,
            "load_ms": 541.0,
            "max_abs_dlogit_vs_official": 0.009016335010528564,
            "max_abs_dprob_vs_official": 0.001450374722480774,
            "max_options": 25.0,
            "multi_label_equal_official": 152.0,
            "prepare_ms": 529.0,
            "sequence_mismatches": 0.0,
            "sequence_requests": 152.0,
            "sequence_requests_total": 152.0,
            "single_label_equal_official": 152.0,
            "thermal_status_end": 0,
            "thermal_status_start": 0,
            "timing_repeats": 5.0,
            "timing_state_tokens": 21.0,
            "window": 128.0
          },
          "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 gliclass)",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-03/gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 670,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliclass-app",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Android sample app (zoo main 7cf7a55 gliclass/, mirrored as android/sample/ in the Hub repo; Kotlin host GliclassTokenizer.kt + GliclassInputs.kt + GliclassClassifier.kt + GliclassDecoder.kt), debug build, gate launched by intent extra; one process for both windows (the smallest window that fits each request); graph ms = first input write through logits readback.",
            "on-device token ids, <<LABEL>> positions and padding equal the captured official inputs on 482/482 requests of this window; single-label 482/482 and multi-label 482/482 answers equal the official pipeline; max |dlogit| 0.009016.",
            "output_match null: the app gate is a text-to-answer check against the official pipeline, not a same-bytes element comparison (adapter rule); the debug-gate rows above carry the element rule.",
            "conditions: thermal status 1 before and after (state files beside the report), USB powered, top-app.",
            "10-02 12:55:10.912 23279 23375 I tflite  : Replacing 670 out of 670 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 12:55:10.995 23279 23375 I tflite  : Replacing 670 out of 670 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 7.8285420000000006,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "argmax_equal_official": 482.0,
            "fixtures": 482.0,
            "graph_max_ms": 14.302135999999999,
            "graph_median_ms": 7.8285420000000006,
            "graph_min_ms": 6.4238539999999995,
            "inputs_identical_to_official": 482.0,
            "label_set_equal_official": 482.0,
            "max_abs_dlogit_vs_official": 0.009016335010528564,
            "whole_gate_requests": 552.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "CpuOptions numThreads=4 (XNNPACK); Android sample app com.gliclass (zoo gliclass/, debug build, on-device tokenizer)",
          "total_ops": 670,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 670,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliclass-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.gliclass.gate (android_gate/ of the run dir, Kotlin CompiledModel 2.2.0, Environment.create(context, DispatchLibraryDir + CompilerPluginLibraryDir = nativeLibraryDir with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, the official token ids + <<LABEL>> positions pushed as JSON, host fp16 table lookup (android.util.Half) outside the timed interval; timed interval = the three input writes + run + logits readback per request, warm median after 5 warm-up requests. Source, logcat, state / .state_after.txt.",
            "conditions: before thermal status 0, battery 37.9 C; after thermal status 0, battery 38.0 C; screen on (FLAG_KEEP_SCREEN_ON, setShowWhenLocked), USB powered, gate activity top-resumed (cgroup top-app).",
            "parity vs the official gliclass 0.1.20 pipeline (CPU fp32, transformers 5.17.0): same top label (softmax argmax) on 482/482 requests and same multi-label set (sigmoid >= 0.5) on 482/482 (351/351 boundary = reference top-1 < 0.9); max |dlogit| 0.009016; max |dp| softmax 0.00145 / sigmoid 0.0018; non-finite requests 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over the valid label logits (4588 elements over 482 requests) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same fp16-table inputs (recomputed by scripts/litert_compat_records.py 2026-10-02): 0 violating elements, max |d| 0, max |d|/|ref| 0.",
            "Reference row: on-device CPU with the same inputs.",
            "10-02 13:50:36.902  7323  7348 I tflite  : Replacing 670 out of 670 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 13:50:36.926  7323  7348 I GliclassGate: GLICLASS_GATE compiled in 132.243073 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 6.767916,
          "loads": true,
          "max_abs_diff": 0.0,
          "max_rel_diff": 0.0,
          "metrics": {
            "argmax_equal_official": 482.0,
            "battery_temp_c_end": 38.0,
            "battery_temp_c_start": 37.9,
            "boundary_argmax_equal_official": 351.0,
            "boundary_fixtures": 351.0,
            "element_rule_elements": 4588.0,
            "element_rule_violations": 0.0,
            "finite_fixtures": 482.0,
            "first_call_ms": 15.201718,
            "fixtures": 482.0,
            "label_set_equal_official": 482.0,
            "load_compile_ms": 132.243073,
            "max_abs_dlogit_vs_official": 0.009016335010528564,
            "max_abs_dp_sigmoid_vs_official": 0.0017973155251483908,
            "max_abs_dp_softmax_vs_official": 0.0014503869105252643,
            "thermal_status_end": 0.0,
            "thermal_status_start": 0.0,
            "warm_max_ms": 10.852396,
            "warm_median_ms": 6.767916,
            "warm_min_ms": 6.376458,
            "warmup_rows_excluded": 5.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "CpuOptions numThreads=4 (XNNPACK); debug gate app com.gliclass.gate",
          "total_ops": 670,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 670,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliclass-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.gliclass.gate (android_gate/ of the run dir, Kotlin CompiledModel 2.2.0, Environment.create(context, DispatchLibraryDir + CompilerPluginLibraryDir = nativeLibraryDir with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, the official token ids + <<LABEL>> positions pushed as JSON, host fp16 table lookup (android.util.Half) outside the timed interval; timed interval = the three input writes + run + logits readback per request, warm median after 5 warm-up requests. Source, logcat, state / .state_after.txt.",
            "conditions: before thermal status 1, battery 40.2 C; after thermal status 1, battery 40.0 C; screen on (FLAG_KEEP_SCREEN_ON, setShowWhenLocked), USB powered, gate activity top-resumed (cgroup top-app).",
            "parity vs the official gliclass 0.1.20 pipeline (CPU fp32, transformers 5.17.0): same top label (softmax argmax) on 475/482 requests and same multi-label set (sigmoid >= 0.5) on 455/482 (344/351 boundary = reference top-1 < 0.9); max |dlogit| 0.3723; max |dp| softmax 0.0396 / sigmoid 0.0753; non-finite requests 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over the valid label logits (4588 elements over 482 requests) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same fp16-table inputs (recomputed by scripts/litert_compat_records.py 2026-10-02): 4337 violating elements, max |d| 0.3671, max |d|/|ref| 79.12.",
            "Default GPU precision on the shipped graph: finite outputs, 475/482 top labels and 455/482 label sets kept (reference softmax gaps <= 0.026, sigmoid margins <= 0.018 on the changed rows), max |dlogit| 0.37; without the LayerNorm pre-scale the same precision keeps 261/482 and 138/482 (run-2 run of the stock graph, not a catalog artifact).",
            "10-02 12:16:49.841 15441 15463 I tflite  : Replacing 670 out of 670 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 12:16:50.296 15441 15463 I GliclassGate: GLICLASS_GATE compiled in 568.353073 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 3.864323,
          "loads": true,
          "max_abs_diff": 0.3671194911003113,
          "max_rel_diff": 79.12491652089717,
          "metrics": {
            "argmax_equal_official": 475.0,
            "battery_temp_c_end": 40.0,
            "battery_temp_c_start": 40.2,
            "boundary_argmax_equal_official": 344.0,
            "boundary_fixtures": 351.0,
            "element_rule_elements": 4588.0,
            "element_rule_violations": 4337.0,
            "finite_fixtures": 482.0,
            "first_call_ms": 4.772343,
            "fixtures": 482.0,
            "label_set_equal_official": 455.0,
            "load_compile_ms": 568.353073,
            "max_abs_dlogit_vs_official": 0.37231284379959106,
            "max_abs_dp_sigmoid_vs_official": 0.07527854223295394,
            "max_abs_dp_softmax_vs_official": 0.03959111781434543,
            "thermal_status_end": 1.0,
            "thermal_status_start": 1.0,
            "warm_max_ms": 6.05599,
            "warm_median_ms": 3.864323,
            "warm_min_ms": 3.568854,
            "warmup_rows_excluded": 5.0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=DEFAULT (fp16 compute) — diagnostic probe, not a shipping setting; debug gate app com.gliclass.gate",
          "total_ops": 670,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 670,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliclass-app",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Android sample app (zoo main 7cf7a55 gliclass/, mirrored as android/sample/ in the Hub repo; Kotlin host GliclassTokenizer.kt + GliclassInputs.kt + GliclassClassifier.kt + GliclassDecoder.kt), debug build, gate launched by intent extra; one process for both windows (the smallest window that fits each request); graph ms = first input write through logits readback.",
            "on-device token ids, <<LABEL>> positions and padding equal the captured official inputs on 482/482 requests of this window; single-label 482/482 and multi-label 482/482 answers equal the official pipeline; max |dlogit| 0.009079.",
            "output_match null: the app gate is a text-to-answer check against the official pipeline, not a same-bytes element comparison (adapter rule); the debug-gate rows above carry the element rule.",
            "conditions: thermal status 1 before and after (state files beside the report), USB powered, top-app.",
            "10-02 12:54:21.100 22777 22866 I tflite  : Replacing 670 out of 670 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 12:54:21.604 22777 22866 I tflite  : Replacing 670 out of 670 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 5.860834,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "argmax_equal_official": 482.0,
            "fixtures": 482.0,
            "graph_max_ms": 6.605,
            "graph_median_ms": 5.860834,
            "graph_min_ms": 5.403281,
            "inputs_identical_to_official": 482.0,
            "label_set_equal_official": 482.0,
            "max_abs_dlogit_vs_official": 0.009079456329345703,
            "whole_gate_requests": 552.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; Android sample app com.gliclass (zoo gliclass/, debug build, on-device tokenizer)",
          "total_ops": 670,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 670,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliclass-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.gliclass.gate (android_gate/ of the run dir, Kotlin CompiledModel 2.2.0, Environment.create(context, DispatchLibraryDir + CompilerPluginLibraryDir = nativeLibraryDir with the Qualcomm dispatch/compiler libraries in jniLibs), one process per run, the official token ids + <<LABEL>> positions pushed as JSON, host fp16 table lookup (android.util.Half) outside the timed interval; timed interval = the three input writes + run + logits readback per request, warm median after 5 warm-up requests. Source, logcat, state / .state_after.txt.",
            "conditions: before thermal status 0, battery 37.3 C; after thermal status 0, battery 37.2 C; screen on (FLAG_KEEP_SCREEN_ON, setShowWhenLocked), USB powered, gate activity top-resumed (cgroup top-app).",
            "parity vs the official gliclass 0.1.20 pipeline (CPU fp32, transformers 5.17.0): same top label (softmax argmax) on 482/482 requests and same multi-label set (sigmoid >= 0.5) on 482/482 (351/351 boundary = reference top-1 < 0.9); max |dlogit| 0.009079; max |dp| softmax 0.00146 / sigmoid 0.00181; non-finite requests 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over the valid label logits (4588 elements over 482 requests) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same fp16-table inputs (recomputed by scripts/litert_compat_records.py 2026-10-02): 1 violating elements, max |d| 6.312e-05, max |d|/|ref| 0.01848.",
            "GPU explicit FP32 is the validated setting (the default precision keeps 475/482 top labels and 455/482 label sets, see the default-precision row).",
            "10-02 13:47:36.414  4816  4845 I tflite  : Replacing 670 out of 670 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "10-02 13:47:36.849  4816  4845 I GliclassGate: GLICLASS_GATE compiled in 547.790885 ms"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 5.805,
          "loads": true,
          "max_abs_diff": 6.312131881713867e-05,
          "max_rel_diff": 0.018484475848331197,
          "metrics": {
            "argmax_equal_official": 482.0,
            "battery_temp_c_end": 37.2,
            "battery_temp_c_start": 37.3,
            "boundary_argmax_equal_official": 351.0,
            "boundary_fixtures": 351.0,
            "element_rule_elements": 4588.0,
            "element_rule_violations": 1.0,
            "finite_fixtures": 482.0,
            "first_call_ms": 6.428125,
            "fixtures": 482.0,
            "label_set_equal_official": 482.0,
            "load_compile_ms": 547.790885,
            "max_abs_dlogit_vs_official": 0.009079456329345703,
            "max_abs_dp_sigmoid_vs_official": 0.0018099238581474886,
            "max_abs_dp_softmax_vs_official": 0.0014627833890153152,
            "thermal_status_end": 0.0,
            "thermal_status_start": 0.0,
            "warm_max_ms": 6.845261,
            "warm_median_ms": 5.805,
            "warm_min_ms": 5.41677,
            "warmup_rows_excluded": 5.0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; debug gate app com.gliclass.gate",
          "total_ops": 670,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-04",
          "decode_tokens_per_s": null,
          "delegated_ops": 670,
          "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 GliclassDeviceTest 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 gliclass) resolved litert-community/GLiClass-Edge-v3.0-LiteRT through the dev descriptor catalog/dev/litert-community__GLiClass-Edge-v3.0-LiteRT.hfmodels.json at that commit (the log's descriptor sha256 prefix afe7d983 = `git show 402a874:catalog/dev/litert-community__GLiClass-Edge-v3.0-LiteRT.hfmodels.json`); graph gliclass_edge_v3_s128_fp32.tflite sha256 79036f538289fd406e6f33b6a1dda51758c77d5383c7704369248c491dd6dd59 = the published gliclass_edge_v3_s128_fp32.tflite (litert-community/GLiClass-Edge-v3.0-LiteRT@88c90950; other records hold rows of the same name and digest).",
            "# hfmodels typed-decisions device check  date=2026-10-04T07:07:30Z serial=RFGL80R6A6H device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9 android=16",
            "# git=402a874 test=GliclassDeviceTest variant=s128_fp32 backend=gpu thermal=Thermal Status: 0 battery_level= level: 80 battery_temp_x10C= temperature: 367 screen= mWakefulness=Dozing",
            "# gradle rc=0 end=2026-10-04T07:07:35Z thermal_end=Thermal Status: 0 battery_temp_end_x10C= temperature: 366",
            "resolve litert-community/GLiClass-Edge-v3.0-LiteRT: commit=88c90950 via EXPLICIT_REVISION, descriptor=caller@explicit/hfmodels.json sha=afe7d983",
            "imported gliclass_edge_v3_s128_fp32.tflite: 53703480 bytes, sha256 ok, from /data/local/tmp/hfmodels/gliclass/gliclass_edge_v3_s128_fp32.tflite",
            "ready litert-community/GLiClass-Edge-v3.0-LiteRT@88c90950 variant=s128_fp32 profile=gpu inference=GPU prepare_ms=1185",
            "tflite  : Replacing 670 out of 670 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "RESULT step=load ok=true variant=s128_fp32 backend=gpu import_ms=173 load_ms=1195 compile_ms=687 profile=gpu window=128 head=128 max_options=25 notes=family=gliclass compile_ms=687 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: tok_embeddings_fp16.bin (float16, 50370 rows x 384) feeds inputs_embeds | one gliclass request per question: labels = a choice's descriptions (the key when there is none), a score's levels, a noul's false / true descriptions or no / yes; prompt = the instructions plus one space, right before the text; a request longer than the 128-token window or with more than 25 labels is refused, never cut; probabilities are gliclass's single-label float32 softmax of the label logits (T=1)",
            "RESULT step=sequence ok=true variant=s128_fp32 backend=gpu rows=152 of 152 mismatches=0 window=128",
            "RESULT step=parity ok=true variant=s128_fp32 backend=gpu rows=152 argmax_match=152/152 single_label_equal_official=152/152 multi_label_equal_official=152/152 max_prob_abs_err=0.0014627724885940552 (worst semif_135) max_logit_abs_err_vs_oracle=0.009079456329345703 dp_tol=0.01 forward_ms_median=6.165365 p90=6.259219 min=5.730521 max=7.515989",
            "RESULT step=demo ok=true variant=s128_fp32 backend=gpu form=described rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=2.950429916381836E-6 (worst b06) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=10/30 correct_mac=10/30 question_ms_median=8.487448 p90=11.481199 min=7.397551 max=11.786094 lookup_ms_median=0.225104",
            "RESULT step=demo ok=true variant=s128_fp32 backend=gpu form=keys rows=30 same_answer_as_mac=30/30 max_prob_abs_err_vs_mac=1.1473894119262695E-6 (worst b12) encoded_length_equal_host=30/30 dp_tol=0.01 correct_device=10/30 correct_mac=10/30 question_ms_median=9.000312 p90=9.171562 min=8.36375 max=9.299324 lookup_ms_median=0.456875",
            "RESULT step=timing ok=true variant=s128_fp32 backend=gpu repeats=5 decide_4q_total_ms_median=38.775938 four_decide_1q_total_ms_median=40.117448 state_tokens=21 per_question_ms=10.098959,9.781771,9.322709,9.359791 answers={department=billing 0.9799990653991699, urgency=0.11925135459750891, refund_requested=0.3396678566932678, tone=angry 0.7151787877082825}",
            "RESULT step=zoo_input ok=true variant=s128_fp32 backend=gpu text_tokens=52 forward_tokens=62 window=128 first=1 warmups=2 timed=10 wall_ms_median=12.125052 total_ms_median=12.0885675 state_ms_median=0.026458 question_ms_median=12.0189845 lookup_ms_median=0.454349 tokenize_ms_median=0.028958499999999998 forward_build_ms_median=4.8166925 graph_call_ms_median=7.1752865 answer=request p=nothing:0.17268067598342896,promise:0.034088000655174255,request:0.5787389278411865,plan:0.21449242532253265 wall_ms=12.029844,12.148438,12.085052,12.113073,12.281771,12.177083,12.113698,12.136406,12.017344,12.181771",
            "RESULT step=lookup ok=true variant=s128_fp32 backend=gpu rows=128 hidden=384 zoo_style=lut zoo_ms_median=2.0525 zoo_ms_min=1.68875 sdk_ms_median=0.121614 sdk_ms_min=0.121146 repeats=20 outputs_identical=true",
            "RESULT step=release ok=true variant=s128_fp32 backend=gpu close_ms=9",
            "RESULT ok=true model=litert-community/GLiClass-Edge-v3.0-LiteRT variant=s128_fp32 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 670 out of 670 node(s) with delegate (LITERT_CL)').",
            "conditions (the gate's header and footer): thermal status 0 -> 0, battery 36.7 -> 36.6 C, battery level 80 %, display mWakefulness=Dozing at the start.",
            "source: hfmodels-android/litert/results/2026-10-04-1607-RFGL80R6A6H-litert2.2.0-decide-gliclass-zoo-s128_fp32-gpu.log (main 3e1fcd8)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 6.165365,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "battery_level_start": 80,
            "battery_temp_c_end": 36.6,
            "battery_temp_c_start": 36.7,
            "demo_described_correct_device": 10.0,
            "demo_described_correct_device_of": 30.0,
            "demo_described_correct_mac": 10.0,
            "demo_described_correct_mac_of": 30.0,
            "demo_described_dp_tol": 0.01,
            "demo_described_encoded_length_equal_host": 30.0,
            "demo_described_encoded_length_equal_host_of": 30.0,
            "demo_described_lookup_ms_median": 0.225104,
            "demo_described_max": 11.786094,
            "demo_described_max_prob_abs_err_vs_mac": 2.950429916381836e-06,
            "demo_described_min": 7.397551,
            "demo_described_p90": 11.481199,
            "demo_described_question_ms_median": 8.487448,
            "demo_described_rows": 30.0,
            "demo_described_same_answer_as_mac": 30.0,
            "demo_described_same_answer_as_mac_of": 30.0,
            "demo_keys_correct_device": 10.0,
            "demo_keys_correct_device_of": 30.0,
            "demo_keys_correct_mac": 10.0,
            "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": 0.456875,
            "demo_keys_max": 9.299324,
            "demo_keys_max_prob_abs_err_vs_mac": 1.1473894119262695e-06,
            "demo_keys_min": 8.36375,
            "demo_keys_p90": 9.171562,
            "demo_keys_question_ms_median": 9.000312,
            "demo_keys_rows": 30.0,
            "demo_keys_same_answer_as_mac": 30.0,
            "demo_keys_same_answer_as_mac_of": 30.0,
            "load_compile_ms": 687.0,
            "load_head": 128.0,
            "load_import_ms": 173.0,
            "load_load_ms": 1195.0,
            "load_max_options": 25.0,
            "load_window": 128.0,
            "lookup_hidden": 384.0,
            "lookup_repeats": 20.0,
            "lookup_rows": 128.0,
            "lookup_sdk_ms_median": 0.121614,
            "lookup_sdk_ms_min": 0.121146,
            "lookup_zoo_ms_median": 2.0525,
            "lookup_zoo_ms_min": 1.68875,
            "parity_argmax_match": 152.0,
            "parity_argmax_match_of": 152.0,
            "parity_dp_tol": 0.01,
            "parity_forward_ms_median": 6.165365,
            "parity_max": 7.515989,
            "parity_max_logit_abs_err_vs_oracle": 0.009079456329345703,
            "parity_max_prob_abs_err": 0.0014627724885940552,
            "parity_min": 5.730521,
            "parity_multi_label_equal_official": 152.0,
            "parity_multi_label_equal_official_of": 152.0,
            "parity_p90": 6.259219,
            "parity_rows": 152.0,
            "parity_single_label_equal_official": 152.0,
            "parity_single_label_equal_official_of": 152.0,
            "prepare_ms": 1185.0,
            "release_close_ms": 9.0,
            "sequence_mismatches": 0.0,
            "sequence_rows": 152.0,
            "sequence_window": 128.0,
            "thermal_status_end": 0,
            "thermal_status_start": 0,
            "timing_decide_4q_total_ms_median": 38.775938,
            "timing_four_decide_1q_total_ms_median": 40.117448,
            "timing_repeats": 5.0,
            "timing_state_tokens": 21.0,
            "zoo_input_first": 1.0,
            "zoo_input_forward_build_ms_median": 4.8166925,
            "zoo_input_forward_tokens": 62.0,
            "zoo_input_graph_call_ms_median": 7.1752865,
            "zoo_input_lookup_ms_median": 0.454349,
            "zoo_input_question_ms_median": 12.0189845,
            "zoo_input_state_ms_median": 0.026458,
            "zoo_input_text_tokens": 52.0,
            "zoo_input_timed": 10.0,
            "zoo_input_tokenize_ms_median": 0.028958499999999998,
            "zoo_input_total_ms_median": 12.0885675,
            "zoo_input_wall_ms_median": 12.125052,
            "zoo_input_warmups": 2.0,
            "zoo_input_window": 128.0
          },
          "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 gliclass)",
          "total_ops": 670,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-04/gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "gliclass-edge-v3.0",
    "id": "gliclass-edge-v3.0__gliclass_edge_v3_s128_fp32",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/GLiClass-Edge-v3.0-LiteRT",
    "task": "zero-shot-classification"
  },
  "pitfalls": [
    "Weight storage does not set the GPU compute precision: request GpuOptions(precision = FP32). At the default precision the outputs stay finite but the answers move: Galaxy S26 475/482 top labels and 455/482 label sets, Mac Metal 479/469 (card 'Galaxy S26, gate app' and 'fp16, the GPU default precision and the NPU').",
    "The graph takes host-built inputs: inputs_embeds [1,128,384] from host_assets/tok_embeddings_fp16.bin ([50370,384] float16 rows upcast to float32, [PAD] 50283 rows at padded positions), attention_mask [1,128] (1 real / 0 pad), label_routing [1,25,128] (row k = 1 at the k-th <<LABEL>> token); output logits [1,1,1,25], slots past the request's labels are ignored; softmax (single-label) or sigmoid >= threshold (multi-label) runs on the host (card 'Host contract').",
    "The request string is <<LABEL>>label 1<<LABEL>>label 2…<<SEP>> + prompt + text with nothing between the prompt and the text; tokenize with the source tokenizer.json ([CLS] … [SEP], ByteLevel prefix space per piece between added tokens); at most 25 labels and 128 tokens, longer requests raise an error, nothing is truncated (card 'Host contract', 'Limits').",
    "The shipped graph computes 15 LayerNorms on x*2^-k with eps*2^-2k (k = 5 in layers 3-6, 6 in layers 7-9 and the final norm) and uses a -1e4 mask; this is exact in FP32 (bit-identical logits to the unmodified graph) and keeps fp16 computation finite but not exact: the Galaxy S26 Hexagon NPU (JIT, one partition, 2.51 ms) keeps 480/482 top labels and 462/482 label sets, every change a near tie, so the card reports the NPU and does not call it a validated path (card 'fp16, the GPU default precision and the NPU').",
    "The checkpoint is float32: the float16 host table rounds values by up to 1.2e-4 (max logit difference 0.0090 against 0.000036 with a float32 table, same answers on all 552 requests) (card 'fp16 …').",
    "Validated on one desktop (Apple M4 Max, ai-edge-litert 2.1.6 CPU and Metal explicit FP32) and one phone (Galaxy S26, LiteRT 2.2.0 GPU explicit FP32 and CPU); English only; conversion fidelity to the official gliclass pipeline (552/552 answers), not a task-accuracy claim — own-subset gold agreement 67/144 SemIf, 115/200 ag_news, 71/200 banking77 (25-way) is not the publisher's benchmark (card 'Agreement with the dataset labels', 'Limits')."
  ],
  "schema_version": "1.2"
}
