{
  "artifacts": [
    {
      "file": "ema_decoder_multi_fp16.tflite",
      "sha256": "d94569d7b3eebf12cbaebdf4248d3bb8dd10c604ac75e59fdcde56dd6cd902b3",
      "size_mb": 6.059
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "TODO — the Hub card states the tool and the runtime but publishes no conversion command or script (the repo ships the graphs, config.json and the host loop ema_host.py)",
    "quantization": "fp16 weights (card 'Files' / 'Weights': decoder fp16 — lossless-grade on the CPU, native on the GPU; config.json forms.decoder = fp16)",
    "tool": "litert-torch",
    "tool_version": "0.9.4 (LiteRT 2.2.0 CompiledModel for the checks)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [
      "DEQUANTIZE",
      "LEAKY_RELU",
      "TRANSPOSE_CONV",
      "TANH"
    ],
    "coverage_ops_pct": 76.7,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 1720,
      "unmatched": 220
    },
    "partitions": 442
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 970,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-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": [
            "harness: the debug gate app com.mlboydaisuke.d1omni.gate (generic mode, Kotlin CompiledModel LiteRT 2.2.0; one process per run; a rows file from make_rows.py names the signature and its inputs as little-endian files; measured interval = write inputs + run + read outputs per call)",
            "file measured: the shipped ema_decoder_multi_fp16.tflite (6353648 B, sha256 d94569d7b3eebf12cbaebdf4248d3bb8dd10c604ac75e59fdcde56dd6cd902b3 = config.json = the Hub LFS sha256; the phone copy digested MATCH /), signatures ema_decoder_P48, ema_decoder_P120",
            "latency_p50_ms = the check run's warm median (no rested timing run exists for this cell): (mode generic, run p3, 8 decoder windows of the six texts) summary.warm_median_write_run_read_ms 99.552291 ms over 3 warm rows of 8 (min 96.337395, max 99.731459, first call 212.158958); compile_ms 36.708229; graph ema_decoder_multi_fp16.tflite (6353648 B)",
            "10-11 20:33:36.889 29273 29356 I tflite  : Replacing 970 out of 970 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 1 (ema_decoder_P120).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 58.15, min_corr 1.000000, max_abs 0.000719, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1_w1__output_0: corr 1.000000, SNR 66.9 dB, max |d| 0.000527; readme2_w1__output_0: corr 1.000000, SNR 58.5 dB, max |d| 0.000687; readme2_w2__output_0: corr 1.000000, SNR 63.9 dB, max |d| 0.000471; readme3_w1__output_0: corr 1.000000, SNR 58.2 dB, max |d| 0.000666; readme3_w2__output_0: corr 1.000000, SNR 64.6 dB, max |d| 0.000491; own1_w1__output_0: corr 1.000000, SNR 70.3 dB, max |d| 0.000556; own2_w1__output_0: corr 1.000000, SNR 58.3 dB, max |d| 0.000558; own3_w1__output_0: corr 1.000000, SNR 59.6 dB, max |d| 0.000719",
            "output_match null: the repo element rule cannot be settled from the scorer's per-row summary (max |d| = 0.000719 > 1e-5 and the SNR leaves room for the relative term); the bar (corr >= 0.999 and SNR >= 30 dB per row and output) is in metrics.bar_pass",
            "the all-CPU fallback's decoder figure in the ship notes (98.8 ms) is the fp32 single-bucket file's rested median; this fp16-weight file's CPU run is a check run (warm median 99.6 ms, 58.2 dB)"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 99.552291,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 36.708229,
            "gate_leg_first_call_ms": 212.158958,
            "gate_leg_rows": 8.0,
            "gate_leg_warm_max_ms": 99.731459,
            "gate_leg_warm_median_ms": 99.552291,
            "gate_leg_warm_min_ms": 96.337395,
            "gate_leg_warm_rows": 3.0,
            "max_abs_vs_mac_fp32": 0.0007190406322479248,
            "min_corr_vs_mac_fp32": 0.9999996656437216,
            "min_snr_db_vs_mac_fp32": 58.151,
            "parity_rows": 8.0,
            "partitions": 1.0,
            "threads": 4.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_decoder_P120 (the shipped file); CPU XNNPACK 4 threads; gate app",
          "total_ops": 970,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_decoder_multi_fp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 970,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-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": [
            "harness: the debug gate app com.mlboydaisuke.d1omni.gate (generic mode, Kotlin CompiledModel LiteRT 2.2.0; one process per run; a rows file from make_rows.py names the signature and its inputs as little-endian files; measured interval = write inputs + run + read outputs per call)",
            "file measured: the shipped ema_decoder_multi_fp16.tflite (6353648 B, sha256 d94569d7b3eebf12cbaebdf4248d3bb8dd10c604ac75e59fdcde56dd6cd902b3 = config.json = the Hub LFS sha256; the phone copy digested MATCH /), signatures ema_decoder_P48, ema_decoder_P120",
            "latency_p50_ms = (mode generic_timing, run p3, rested phone, fixture row readme1_w1): timing.t_decoder.per_call_ms_write_run_read median 30.506614 ms (min 29.329167, max 32.603593, n 20) after 3 warm-up calls on fixture row 0 of ema_decoder_P120.json; run-only median 0.438828 ms; compile_ms 771.618906; graph ema_decoder_multi_fp16.tflite (6353648 B); litert 2.2.0, device SM-S942Q, android API 36",
            "check run (mode generic, run p3): 8 fixture rows, warm median 30.598802 ms (a parity run, not the timing figure); compile_ms 767.410833",
            "10-11 20:36:10.045   989  1016 I tflite  : Replacing 970 out of 970 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (ema_decoder_P120).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 40.82, min_corr 0.999959, max_abs 0.0112, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1_w1__output_0: corr 0.999991, SNR 49.2 dB, max |d| 0.0112; readme2_w1__output_0: corr 0.999968, SNR 41.9 dB, max |d| 0.00924; readme2_w2__output_0: corr 0.999986, SNR 46.2 dB, max |d| 0.00675; readme3_w1__output_0: corr 0.999959, SNR 40.8 dB, max |d| 0.0107; readme3_w2__output_0: corr 0.999988, SNR 47.1 dB, max |d| 0.00694; own1_w1__output_0: corr 0.999995, SNR 53.6 dB, max |d| 0.00572; own2_w1__output_0: corr 0.999960, SNR 40.9 dB, max |d| 0.0071; own3_w1__output_0: corr 0.999968, SNR 42.0 dB, max |d| 0.00926",
            "output_match = the repo element rule |out - ref| <= max(1e-5, 1e-3 |ref|) against the Mac fp32 graph on the same inputs: provably violated — max |d| = 0.0112 with the row's SNR below 60 dB bounds sum ref^2 so tightly that no element can carry that difference under the rule (the bound: d_i^2 <= 1e-10 + 1e-6 ref_i^2 summed, with sum d^2 = sum ref^2 * 10^(-snr/10)); max_rel_diff not computed. The pass bar (corr >= 0.999 and SNR >= 30 dB per row and output) is in metrics.bar_pass",
            "the published assignment's decoder accelerator for the later windows; the fp32 single-bucket file at the same precision reads 31.0 ms / 40.8 dB"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 30.506614,
          "loads": true,
          "max_abs_diff": 0.011178404092788696,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 767.410833,
            "gate_leg_first_call_ms": 36.173855,
            "gate_leg_rows": 8.0,
            "gate_leg_warm_max_ms": 31.772708,
            "gate_leg_warm_median_ms": 30.598802,
            "gate_leg_warm_min_ms": 28.810781,
            "gate_leg_warm_rows": 3.0,
            "max_abs_vs_mac_fp32": 0.011178404092788696,
            "min_corr_vs_mac_fp32": 0.9999593387736163,
            "min_snr_db_vs_mac_fp32": 40.825,
            "parity_rows": 8.0,
            "partitions": 1.0,
            "rested_leg_max_ms": 32.603593,
            "rested_leg_median_ms": 30.506614,
            "rested_leg_min_ms": 29.329167,
            "rested_leg_n": 20.0,
            "rested_leg_warmups": 3.0,
            "run_only_median_ms": 0.438828
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_decoder_P120 (the shipped file); GPU default precision (fp16 activations); gate app",
          "total_ops": 970,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_decoder_multi_fp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": 970,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-ema-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": [
            "harness: the debug gate app com.mlboydaisuke.d1omni.gate (generic mode, Kotlin CompiledModel LiteRT 2.2.0; one process per run; a rows file from make_rows.py names the signature and its inputs as little-endian files; measured interval = write inputs + run + read outputs per call)",
            "file measured: the shipped ema_decoder_multi_fp16.tflite (6353648 B, sha256 d94569d7b3eebf12cbaebdf4248d3bb8dd10c604ac75e59fdcde56dd6cd902b3 = config.json = the Hub LFS sha256; the phone copy digested MATCH /), signatures ema_decoder_P48, ema_decoder_P120",
            "latency_p50_ms = (mode generic_timing, run p3, rested phone, fixture row readme1_w0): timing.t_decoder.per_call_ms_write_run_read median 14.624141000000002 ms (min 13.844218, max 17.374062, n 20) after 3 warm-up calls on fixture row 0 of ema_decoder_P48.json; run-only median 0.259219 ms; compile_ms 768.203073; graph ema_decoder_multi_fp16.tflite (6353648 B); litert 2.2.0, device SM-S942Q, android API 36",
            "check run (mode generic, run p3): 8 fixture rows, warm median 14.090156 ms (a parity run, not the timing figure); compile_ms 764.20927",
            "10-11 20:35:41.799 32702 32726 I tflite  : Replacing 970 out of 970 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (ema_decoder_P48).",
            "parity vs the Mac fp32 graph (expected outputs computed by make_rows.py with ai-edge-litert 2.2.0 XNNPACK on the work fixtures, scored by device_score.py / npu_score.py): — summary min_snr_db 46.45, min_corr 0.999985, max_abs 0.00903, bar {'corr': 0.999, 'snr_db': 30.0} -> PASS; rows: readme1_w0__output_0: corr 0.999985, SNR 46.4 dB, max |d| 0.00903; readme2_w0__output_0: corr 0.999989, SNR 47.8 dB, max |d| 0.00728; readme3_w0__output_0: corr 0.999987, SNR 46.8 dB, max |d| 0.00694; own1_w0__output_0: corr 0.999988, SNR 47.1 dB, max |d| 0.00759; own2_w0__output_0: corr 0.999986, SNR 46.7 dB, max |d| 0.00679; own2_w2__output_0: corr 0.999991, SNR 48.5 dB, max |d| 0.00563; own3_w0__output_0: corr 0.999989, SNR 47.6 dB, max |d| 0.00473; own3_w2__output_0: corr 0.999999, SNR 61.7 dB, max |d| 0.00177",
            "output_match = the repo element rule |out - ref| <= max(1e-5, 1e-3 |ref|) against the Mac fp32 graph on the same inputs: provably violated — max |d| = 0.00903 with the row's SNR below 60 dB bounds sum ref^2 so tightly that no element can carry that difference under the rule (the bound: d_i^2 <= 1e-10 + 1e-6 ref_i^2 summed, with sum d^2 = sum ref^2 * 10^(-snr/10)); max_rel_diff not computed. The pass bar (corr >= 0.999 and SNR >= 30 dB per row and output) is in metrics.bar_pass",
            "the published assignment's decoder accelerator for the first window of a piece (first audio); the fp32 single-bucket file at the same precision reads 14.6 ms / 46.4 dB"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 14.624141000000002,
          "loads": true,
          "max_abs_diff": 0.009034223854541779,
          "max_rel_diff": null,
          "metrics": {
            "bar_pass": 1.0,
            "compile_ms": 764.20927,
            "gate_leg_first_call_ms": 19.330834,
            "gate_leg_rows": 8.0,
            "gate_leg_warm_max_ms": 14.090833,
            "gate_leg_warm_median_ms": 14.090156,
            "gate_leg_warm_min_ms": 14.04375,
            "gate_leg_warm_rows": 3.0,
            "max_abs_vs_mac_fp32": 0.009034223854541779,
            "min_corr_vs_mac_fp32": 0.9999850644398058,
            "min_snr_db_vs_mac_fp32": 46.447,
            "parity_rows": 8.0,
            "partitions": 1.0,
            "rested_leg_max_ms": 17.374062,
            "rested_leg_median_ms": 14.624141000000002,
            "rested_leg_min_ms": 13.844218,
            "rested_leg_n": 20.0,
            "rested_leg_warmups": 3.0,
            "run_only_median_ms": 0.259219
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_decoder_P48 (the shipped file); GPU default precision (fp16 activations); gate app",
          "total_ops": 970,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_decoder_multi_fp16__galaxy-s26.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-11",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: the host loop hostloop_e2e.py --split --multi --threads 4 (ai-edge-litert 2.2.0 Interpreter, XNNPACK 4 threads; normalizer-tr frontend + the package's chunker; per-stage wall time of single calls: write inputs + run + read outputs) on the shipped multi-signature files (multi files, sha256 = config.json = the Hub LFS)",
            "Per text the piece's L / T, the bucket Lb / Tb it ran in, text_ms, sound_ms (= align + the four steps + the host Euler mixing), decode_ms per window, first_audio_ms, total_ms; forms {\"text\": \"fp32\", \"sound\": \"fp32\", \"decoder\": \"fp16\"}, split True, threads 4",
            "(verify_tflite.py --multi on the same files vs the package oracle fixtures/): per fixture frames_ok, |dh|, |ddur|, |dlat|, corr_lat, |daudio|, corr_audio, snr_db — readme1: frames_ok True, |dh| 1.91e-06, |ddur| 1.31e-06, corr_lat 1.000000000, SNR 56.40 dB, corr_audio 0.9999995; readme2: frames_ok True, |dh| 1.79e-06, |ddur| 2.62e-06, corr_lat 1.000000000, SNR 56.82 dB, corr_audio 0.9999995; readme3: frames_ok True, |dh| 1.43e-06, |ddur| 2.38e-06, corr_lat 1.000000000, SNR 56.84 dB, corr_audio 0.9999996; own1: frames_ok True, |dh| 1.19e-06, |ddur| 1.31e-06, corr_lat 1.000000000, SNR 57.11 dB, corr_audio 0.9999996; own2: frames_ok True, |dh| 1.43e-06, |ddur| 1.55e-06, corr_lat 1.000000000, SNR 56.57 dB, corr_audio 0.9999995; own3: frames_ok True, |dh| 1.31e-06, |ddur| 1.79e-06, corr_lat 1.000000000, SNR 56.57 dB, corr_audio 0.9999995",
            "output_match null: the comparison is against the upstream PyTorch package (the oracle), not a same-runtime CPU reference (shape); the pass bar passes on all six fixtures",
            "decode window w0 is the P48 signature (the piece's first 25 frames + 8 of context = 48, so audio starts early), the later windows are P120 (100 + 2 x 8); the host loop's decode_ms per window is the single call's write + run + read"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "audio_corr_vs_pytorch_own1": 0.9999996055832846,
            "audio_corr_vs_pytorch_own2": 0.9999995362701206,
            "audio_corr_vs_pytorch_own3": 0.9999994917052253,
            "audio_corr_vs_pytorch_readme1": 0.9999994641767198,
            "audio_corr_vs_pytorch_readme2": 0.9999995450154168,
            "audio_corr_vs_pytorch_readme3": 0.9999995646920609,
            "audio_snr_db_vs_pytorch_own1": 57.111,
            "audio_snr_db_vs_pytorch_own2": 56.569,
            "audio_snr_db_vs_pytorch_own3": 56.57,
            "audio_snr_db_vs_pytorch_readme1": 56.405,
            "audio_snr_db_vs_pytorch_readme2": 56.819,
            "audio_snr_db_vs_pytorch_readme3": 56.841,
            "bar_pass": 1.0,
            "decode_call_ms_own1_w0_P48": 28.4562,
            "decode_call_ms_own1_w1_P120": 65.5566,
            "decode_call_ms_own2_w0_P48": 28.7758,
            "decode_call_ms_own2_w1_P120": 65.2517,
            "decode_call_ms_own2_w2_P120": 28.4008,
            "decode_call_ms_own3_w0_P48": 28.4544,
            "decode_call_ms_own3_w1_P120": 63.854,
            "decode_call_ms_own3_w2_P120": 28.5333,
            "decode_call_ms_readme1_w0_P48": 28.7085,
            "decode_call_ms_readme1_w1_P120": 65.3782,
            "decode_call_ms_readme2_w0_P48": 28.5506,
            "decode_call_ms_readme2_w1_P120": 64.9004,
            "decode_call_ms_readme2_w2_P120": 64.9434,
            "decode_call_ms_readme3_w0_P48": 28.6464,
            "decode_call_ms_readme3_w1_P120": 65.0726,
            "decode_call_ms_readme3_w2_P120": 65.2272,
            "min_audio_snr_db_vs_pytorch": 56.405,
            "pipeline_first_audio_ms_own1": 39.892,
            "pipeline_first_audio_ms_own2": 45.382,
            "pipeline_first_audio_ms_own3": 44.886,
            "pipeline_first_audio_ms_readme1": 40.202,
            "pipeline_first_audio_ms_readme2": 57.595,
            "pipeline_first_audio_ms_readme3": 57.713,
            "pipeline_rtf_own1": 0.03935,
            "pipeline_rtf_own2": 0.02272,
            "pipeline_rtf_own3": 0.02681,
            "pipeline_rtf_readme1": 0.02999,
            "pipeline_rtf_readme2": 0.0219,
            "pipeline_rtf_readme3": 0.02271,
            "pipeline_total_ms_own1": 105.449,
            "pipeline_total_ms_own2": 139.037,
            "pipeline_total_ms_own3": 137.275,
            "pipeline_total_ms_readme1": 105.582,
            "pipeline_total_ms_readme2": 187.441,
            "pipeline_total_ms_readme3": 188.015,
            "threads": 4.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ema_decoder_P{48,120} (the shipped file); host loop, one call per fixture",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-11/ema-lightning__ema_decoder_multi_fp16__mac-studio-m4-max.json"
      }
    ]
  },
  "model": {
    "family": "ema-lightning",
    "id": "ema-lightning__ema_decoder_multi_fp16",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/EMA-Lightning-LiteRT",
    "task": "text-to-speech"
  },
  "pitfalls": [
    "Decoder contract (card 'Files'): z [1,64,P] + 7 masks [1,1,len_i] -> audio float32 [1, P·1920] (HiFi-GAN, 48 kHz); signatures ema_decoder_P48 and ema_decoder_P120, shapes fixed per signature; config.json decoder_stage_lengths gives the seven mask lengths per bucket (P48: 48, 384, 2304, 11521, 23042, 46084, 92168; P120: 120, 960, 5760, 28801, 57602, 115204, 230408).",
    "Windows: the host decodes 25 frames first (the first second, so audio starts early: the P48 signature) then 100 frames at a time (P120), with 8 frames of context on each side cut away after decoding; the pieces join sample-exactly (card 'How the host loop works' step 5).",
    "The decoder is where the GPU pays off on the Galaxy S26 (P120: 98.8 ms CPU 4 threads, 54.9 ms GPU fp32, 31.0 ms GPU fp16 at 40.8 dB; P48: 39.1 / 22.2 / 14.6 ms at 46.4 dB) and it is the wrong workload for the HTP: P48 took 518 s to compile and runs 127.7 ms (31 dB), P120 did not compile in 600 s (card 'Performance').",
    "Dynamic int8 weights on the decoder fail the parity bar (broadband int8 conv noise, 24 dB) and do not ship; the fp16 weights are lossless-grade (56.4 dB vs the PyTorch package over six sentences) (card 'Parity').",
    "The base card's RTF figures are the authors' own (an RTX 4090 and a cloud CPU) and the published Hub clips assets/sample-*.wav are not the reference (a different noise stream) (card 'Performance', 'Parity').",
    "Turkish only, one voice; fixed buckets — a piece longer than 256 letters / 512 frames is cut by the chunker first (card 'Limitations')."
  ],
  "schema_version": "1.2"
}
