Qhub-AI
/

QAdapt

diff --git a/code/benchmarks/__init__.py b/code/benchmarks/__init__.py
new file mode 100644
index 0000000..81f0abb
--- /dev/null
+++ b/code/benchmarks/__init__.py
@@ -0,0 +1,4 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Benchmark dataset integrations."""
diff --git a/code/benchmarks/google_qec.py b/code/benchmarks/google_qec.py
new file mode 100644
index 0000000..0fda551
--- /dev/null
+++ b/code/benchmarks/google_qec.py
@@ -0,0 +1,298 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Google Quantum AI QEC benchmark dataset integration.
+
+The source dataset is the Zenodo record for "Quantum error correction below the
+surface code threshold". This module deliberately treats the data as an
+external benchmark archive: the files are multi-GB zip archives with their own
+README files and are not committed to this repository.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import shutil
+import urllib.request
+import zipfile
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Iterable, Sequence
+
+
+GOOGLE_QEC_RECORD_ID = 13273331
+GOOGLE_QEC_RECORD_URL = f"https://zenodo.org/api/records/{GOOGLE_QEC_RECORD_ID}"
+GOOGLE_QEC_RECORD_HTML = f"https://zenodo.org/records/{GOOGLE_QEC_RECORD_ID}"
+
+DEFAULT_BENCHMARK_KEY = "google_105Q_surface_code_d3_d5_d7.zip"
+
+
+@dataclass(frozen=True)
+class GoogleQECFile:
+    key: str
+    size_bytes: int
+    md5: str
+    url: str
+    code_family: str
+    distances: tuple[int, ...]
+
+
+@dataclass(frozen=True)
+class GoogleQECManifest:
+    record_id: int
+    title: str
+    license_id: str
+    record_url: str
+    files: tuple[GoogleQECFile, ...]
+
+    def by_key(self) -> dict[str, GoogleQECFile]:
+        return {entry.key: entry for entry in self.files}
+
+
+@dataclass(frozen=True)
+class DownloadItem:
+    entry: GoogleQECFile
+    path: Path
+    exists: bool
+
+
+@dataclass(frozen=True)
+class DownloadPlan:
+    root: Path
+    items: tuple[DownloadItem, ...]
+    required_bytes: int
+
+
+@dataclass(frozen=True)
+class GoogleQECIndex:
+    root: Path
+    manifest_path: Path | None
+    archives: dict[str, Path]
+    extracted_dirs: dict[str, Path]
+
+
+def _infer_code_family(key: str) -> str:
+    if "surface_code" in key:
+        return "surface"
+    if "repetition_code" in key:
+        return "repetition"
+    return "unknown"
+
+
+def _infer_distances(key: str) -> tuple[int, ...]:
+    stem = key.removesuffix(".zip")
+    values = []
+    for part in stem.split("_"):
+        if len(part) > 1 and part[0] == "d" and part[1:].isdigit():
+            values.append(int(part[1:]))
+    return tuple(values)
+
+
+def parse_zenodo_record(record: dict) -> GoogleQECManifest:
+    """Parse the Zenodo API response into a stable local manifest."""
+
+    files = []
+    for file_info in record.get("files", []):
+        checksum = str(file_info.get("checksum", ""))
+        if not checksum.startswith("md5:"):
+            raise ValueError(f"Unsupported checksum for {file_info.get('key')!r}: {checksum!r}")
+        key = str(file_info["key"])
+        files.append(
+            GoogleQECFile(
+                key=key,
+                size_bytes=int(file_info["size"]),
+                md5=checksum.split(":", 1)[1],
+                url=str(file_info["links"]["self"]),
+                code_family=_infer_code_family(key),
+                distances=_infer_distances(key),
+            )
+        )
+
+    metadata = record.get("metadata", {})
+    license_info = metadata.get("license") or {}
+    return GoogleQECManifest(
+        record_id=int(record["id"]),
+        title=str(metadata.get("title", record.get("title", ""))),
+        license_id=str(license_info.get("id", "")),
+        record_url=str(record.get("links", {}).get("self_html", GOOGLE_QEC_RECORD_HTML)),
+        files=tuple(sorted(files, key=lambda entry: entry.size_bytes)),
+    )
+
+
+def fetch_zenodo_manifest(url: str = GOOGLE_QEC_RECORD_URL, timeout: float = 60.0) -> GoogleQECManifest:
+    """Fetch and parse the official Zenodo record."""
+
+    with urllib.request.urlopen(url, timeout=timeout) as response:
+        payload = json.loads(response.read().decode("utf-8"))
+    return parse_zenodo_record(payload)
+
+
+def build_download_plan(
+    manifest: GoogleQECManifest,
+    root: Path,
+    keys: Sequence[str] | None = None,
+) -> DownloadPlan:
+    """Build a concrete download plan without performing network or disk writes."""
+
+    selected_keys = tuple(keys) if keys else (DEFAULT_BENCHMARK_KEY,)
+    by_key = manifest.by_key()
+    missing = [key for key in selected_keys if key not in by_key]
+    if missing:
+        raise KeyError(f"Unknown Google QEC benchmark file(s): {missing}")
+
+    root = Path(root)
+    items = []
+    required = 0
+    for key in selected_keys:
+        entry = by_key[key]
+        path = root / entry.key
+        exists = path.exists()
+        items.append(DownloadItem(entry=entry, path=path, exists=exists))
+        if not exists:
+            required += entry.size_bytes
+    return DownloadPlan(root=root, items=tuple(items), required_bytes=required)
+
+
+def ensure_sufficient_space(path: Path, required_bytes: int, margin: float = 1.10) -> None:
+    """Raise before starting a large download if the filesystem is too full."""
+
+    if required_bytes <= 0:
+        return
+    usage = shutil.disk_usage(path)
+    needed = int(required_bytes * float(margin))
+    if usage.free < needed:
+        raise RuntimeError(
+            f"Not enough free space under {path}: need at least {needed:,} bytes "
+            f"including margin, found {usage.free:,} bytes"
+        )
+
+
+def _md5_file(path: Path, chunk_size: int = 16 * 1024 * 1024) -> str:
+    digest = hashlib.md5()
+    with path.open("rb") as f:
+        while True:
+            chunk = f.read(chunk_size)
+            if not chunk:
+                break
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def verify_archive(path: Path, entry: GoogleQECFile) -> None:
+    if path.stat().st_size != entry.size_bytes:
+        raise RuntimeError(
+            f"Size mismatch for {path}: expected {entry.size_bytes}, got {path.stat().st_size}"
+        )
+    got = _md5_file(path)
+    if got != entry.md5:
+        raise RuntimeError(f"MD5 mismatch for {path}: expected {entry.md5}, got {got}")
+
+
+def build_download_request(entry: GoogleQECFile, resume_from: int = 0) -> urllib.request.Request:
+    """Build a request for a benchmark archive, optionally using HTTP Range."""
+
+    headers = {}
+    if int(resume_from) > 0:
+        headers["Range"] = f"bytes={int(resume_from)}-"
+    return urllib.request.Request(entry.url, headers=headers)
+
+
+def download_entry(entry: GoogleQECFile, path: Path, force: bool = False) -> Path:
+    """Download one benchmark archive and verify size + md5."""
+
+    path.parent.mkdir(parents=True, exist_ok=True)
+    if path.exists() and not force:
+        verify_archive(path, entry)
+        return path
+
+    tmp_path = path.with_suffix(path.suffix + ".part")
+    if force and tmp_path.exists():
+        tmp_path.unlink()
+
+    resume_from = tmp_path.stat().st_size if tmp_path.exists() else 0
+    if resume_from >= entry.size_bytes:
+        tmp_path.replace(path)
+        verify_archive(path, entry)
+        return path
+
+    request = build_download_request(entry, resume_from=resume_from)
+    with urllib.request.urlopen(request, timeout=60.0) as response:
+        status = getattr(response, "status", None) or response.getcode()
+        mode = "ab" if resume_from > 0 and status == 206 else "wb"
+        if mode == "wb":
+            resume_from = 0
+        with tmp_path.open(mode) as out:
+            while True:
+                chunk = response.read(16 * 1024 * 1024)
+                if not chunk:
+                    break
+                out.write(chunk)
+    tmp_path.replace(path)
+    verify_archive(path, entry)
+    return path
+
+
+def extract_archive(path: Path, output_dir: Path | None = None) -> Path:
+    """Extract a downloaded benchmark zip next to the archive by default."""
+
+    target = output_dir or path.with_suffix("")
+    target.mkdir(parents=True, exist_ok=True)
+    with zipfile.ZipFile(path) as zf:
+        zf.extractall(target)
+    return target
+
+
+class GoogleQECBenchmarkStore:
+    """Local project store for Google QEC benchmark archives."""
+
+    def __init__(self, root: Path | str = "benchmarks/google_qec"):
+        self.root = Path(root)
+
+    @property
+    def manifest_path(self) -> Path:
+        return self.root / "manifest.json"
+
+    def write_manifest(self, manifest: GoogleQECManifest) -> Path:
+        self.root.mkdir(parents=True, exist_ok=True)
+        payload = asdict(manifest)
+        self.manifest_path.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
+        return self.manifest_path
+
+    def index(self) -> GoogleQECIndex:
+        archives = {path.name: path for path in sorted(self.root.glob("*.zip"))}
+        extracted_dirs = {
+            path.name: path
+            for path in sorted(self.root.iterdir()) if path.is_dir() and path.name != "__pycache__"
+        } if self.root.exists() else {}
+        manifest_path = self.manifest_path if self.manifest_path.exists() else None
+        return GoogleQECIndex(
+            root=self.root,
+            manifest_path=manifest_path,
+            archives=archives,
+            extracted_dirs=extracted_dirs,
+        )
+
+    def download(
+        self,
+        manifest: GoogleQECManifest,
+        keys: Sequence[str] | None = None,
+        *,
+        force: bool = False,
+        extract: bool = False,
+        check_space: bool = True,
+    ) -> DownloadPlan:
+        self.root.mkdir(parents=True, exist_ok=True)
+        plan = build_download_plan(manifest, self.root, keys)
+        if check_space:
+            ensure_sufficient_space(self.root, plan.required_bytes)
+        self.write_manifest(manifest)
+        for item in plan.items:
+            archive_path = download_entry(item.entry, item.path, force=force)
+            if extract:
+                extract_archive(archive_path)
+        return plan
+
+
+def benchmark_keys(files: Iterable[GoogleQECFile]) -> list[str]:
+    return [entry.key for entry in sorted(files, key=lambda entry: (entry.code_family, entry.size_bytes))]
diff --git a/code/data/datapipe_stim.py b/code/data/datapipe_stim.py
index f8a13a4..d7c0a1b 100644
--- a/code/data/datapipe_stim.py
+++ b/code/data/datapipe_stim.py
@@ -46,11 +46,13 @@ class QCDataPipePreDecoder_Memory_inference(Dataset):
         measure_basis='X',
         code_rotation='XV',  # <--- NEW: surface code orientation
         noise_model=None,  # Optional explicit NoiseModel (overrides p_error when provided)
+        seed=None,
     ):
         self.distance = int(distance)
         self.n_rounds = max(int(n_rounds), 1)
         self.num_samples = int(num_samples)
         self.measure_basis = str(measure_basis).upper()
+        self.seed = None if seed is None else int(seed)
         self.code_rotation = code_rotation.upper() if code_rotation else 'XV'
 
         if error_mode != "circuit_level_surface_custom":
@@ -98,7 +100,7 @@ class QCDataPipePreDecoder_Memory_inference(Dataset):
                 add_boundary_detectors=True,  # Required for proper PyMatching decoding
             )
             self.circ_X.set_error_rates()
-            meas_X = self.circ_X.stim_circuit.compile_sampler().sample(shots=self.nX)
+            meas_X = self.circ_X.stim_circuit.compile_sampler(seed=self.seed).sample(shots=self.nX)
             # drop final D*D data-qubit measurements and reshape to (shots, n_rounds, D^2-1)
             self.meas_X = (
                 torch.from_numpy(meas_X[..., :-(D * D)]
@@ -129,7 +131,7 @@ class QCDataPipePreDecoder_Memory_inference(Dataset):
                 add_boundary_detectors=True,  # Required for proper PyMatching decoding
             )
             self.circ_Z.set_error_rates()
-            meas_Z = self.circ_Z.stim_circuit.compile_sampler().sample(shots=self.nZ)
+            meas_Z = self.circ_Z.stim_circuit.compile_sampler(seed=None if self.seed is None else self.seed + 1).sample(shots=self.nZ)
             self.meas_Z = (
                 torch.from_numpy(meas_Z[..., :-(D * D)]
                                 ).to(torch.uint8).view(self.nZ, self.n_rounds,
@@ -154,7 +156,7 @@ class QCDataPipePreDecoder_Memory_inference(Dataset):
                 add_boundary_detectors=True,  # Required for proper PyMatching decoding
             )
             self.circ.set_error_rates()
-            meas = self.circ.stim_circuit.compile_sampler().sample(shots=self.num_samples)
+            meas = self.circ.stim_circuit.compile_sampler(seed=self.seed).sample(shots=self.num_samples)
             self.meas = (
                 torch.from_numpy(meas[..., :-(D * D)]
                                 ).to(torch.uint8).view(self.num_samples, self.n_rounds,
diff --git a/code/data/factory.py b/code/data/factory.py
index 343f995..0c49dad 100644
--- a/code/data/factory.py
+++ b/code/data/factory.py
@@ -25,7 +25,7 @@ import torch
 class DatapipeFactory:
     """
     Factory for creating datapipes.
-    
+
     Training: Uses Torch on-the-fly generation in training.train
     Inference: Uses Stim-based datapipes
     """
@@ -54,7 +54,7 @@ class DatapipeFactory:
     def _create_surface_datapipe(cfg):
         """
         Datapipe for training - on-the-fly generation only.
-        
+
         Returns (None, None) to signal on-the-fly mode - train.py will create
         the generators directly.
         """
@@ -128,6 +128,9 @@ class DatapipeFactory:
                     dataset_kwargs["strict_noise"] = False
             else:
                 dataset_cls = QCDataPipePreDecoder_Memory_inference
+                seed = getattr(cfg, "seed", None)
+                if seed is not None:
+                    dataset_kwargs["seed"] = int(seed)
 
             test_dataset = dataset_cls(
                 distance=cfg.distance,
@@ -214,7 +217,7 @@ class DatapipeFactory:
 def inspect_sample(sample, label: str):
     """
     Inspect a sample from the datapipe for debugging.
-    
+
     Args:
         sample: dict from the datapipe with keys 'x_syn_diff', 'z_syn_diff', 'trainX'
         label: "X" or "Z" (what we expect for this sample)
diff --git a/code/examples/infer_ood.py b/code/examples/infer_ood.py
new file mode 100644
index 0000000..f3b5d96
--- /dev/null
+++ b/code/examples/infer_ood.py
@@ -0,0 +1,128 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Run released pre-decoders on the fixed training-axis OOD grid."""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+CODE_ROOT = Path(__file__).resolve().parents[1]
+if str(CODE_ROOT) not in sys.path:
+    sys.path.insert(0, str(CODE_ROOT))
+
+from scripts.experiments.unknown_noise.generate_unknown_axismix_grid_u1p2_5p0_configs import (  # noqa: E402
+    write_axismix_grid_configs,
+)
+from scripts.qadapt_example_utils import (  # noqa: E402
+    InferenceJob,
+    add_common_inference_args,
+    build_paired_command,
+    parse_gpus,
+    run_jobs,
+)
+
+
+PAPER_DISTANCES = (3, 5, 7, 9)
+PAPER_MULTIPLIERS = (1.2, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0)
+
+
+def parse_distances(value: str) -> list[int]:
+    result = [int(item.strip()) for item in value.split(",") if item.strip()]
+    if not result or result != sorted(set(result)):
+        raise argparse.ArgumentTypeError(
+            "distances must be a non-empty, increasing comma-separated list"
+        )
+    return result
+
+
+def parse_multipliers(value: str) -> list[float]:
+    result = [float(item.strip()) for item in value.split(",") if item.strip()]
+    if not result or result != sorted(set(result)) or any(item <= 0 for item in result):
+        raise argparse.ArgumentTypeError(
+            "multipliers must be a non-empty, increasing comma-separated list "
+            "of positive numbers"
+        )
+    return result
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--distances",
+        type=parse_distances,
+        default=list(PAPER_DISTANCES),
+        help="Comma-separated distances; defaults to the v2 paper's d=3,5,7,9 grid.",
+    )
+    parser.add_argument("--n-rounds", type=int, default=9)
+    parser.add_argument(
+        "--base-config",
+        default="conf/examples/qadapt_v2/T0.yaml",
+        help="Base config whose model architecture and T0 noise define the OOD grid.",
+    )
+    parser.add_argument(
+        "--multipliers",
+        type=parse_multipliers,
+        default=list(PAPER_MULTIPLIERS),
+        help="Comma-separated OOD multipliers; defaults to the v2 paper's 1.2--5.0 grid.",
+    )
+    parser.add_argument(
+        "--generated-config-dir",
+        type=Path,
+        default=Path("outputs/generated_configs/ood"),
+    )
+    parser.add_argument(
+        "--manifest",
+        type=Path,
+        default=Path("outputs/generated_configs/ood/manifest.json"),
+    )
+    add_common_inference_args(
+        parser,
+        default_output_dir=Path("outputs/examples/released_models/ood"),
+    )
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    _, manifest = write_axismix_grid_configs(
+        base_config=args.base_config,
+        output_dir=args.generated_config_dir,
+        manifest=args.manifest,
+        grid_multipliers=args.multipliers,
+    )
+    jobs = []
+    for distance in args.distances:
+        for environment in manifest["environments"]:
+            config_file = args.generated_config_dir / environment["config_filename"]
+            label = (
+                f"d{distance}_{environment['env_key']}_"
+                f"{environment['multiplier_key']}"
+            )
+            output_path = args.output_dir / f"d{distance}" / f"{label}.json"
+            jobs.append(
+                InferenceJob(
+                    label=label,
+                    command=build_paired_command(
+                        args,
+                        config_file=config_file,
+                        output_path=output_path,
+                        distance=distance,
+                        n_rounds=args.n_rounds,
+                    ),
+                    output_path=output_path,
+                )
+            )
+    run_jobs(
+        jobs,
+        gpus=parse_gpus(args.gpus),
+        parallelism=args.parallelism,
+        resume=args.resume,
+        dry_run=args.dry_run,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/code/examples/infer_t0_t4.py b/code/examples/infer_t0_t4.py
new file mode 100644
index 0000000..d3fd27b
--- /dev/null
+++ b/code/examples/infer_t0_t4.py
@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Run released pre-decoders on the five T0-T4 simulated noise tasks."""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+CODE_ROOT = Path(__file__).resolve().parents[1]
+if str(CODE_ROOT) not in sys.path:
+    sys.path.insert(0, str(CODE_ROOT))
+
+from scripts.qadapt_example_utils import (  # noqa: E402
+    InferenceJob,
+    TASK_CONFIGS,
+    add_common_inference_args,
+    build_paired_command,
+    parse_gpus,
+    run_jobs,
+)
+
+
+TASK_BY_ID = {
+    f"T{index}": (task_key, config_name)
+    for index, (task_key, config_name) in enumerate(TASK_CONFIGS)
+}
+
+
+def parse_distances(value: str) -> list[int]:
+    result = [int(item.strip()) for item in value.split(",") if item.strip()]
+    if not result or result != sorted(set(result)):
+        raise argparse.ArgumentTypeError(
+            "distances must be a non-empty, increasing comma-separated list"
+        )
+    return result
+
+
+def parse_tasks(value: str) -> list[str]:
+    result = [item.strip().upper() for item in value.split(",") if item.strip()]
+    if not result or len(result) != len(set(result)):
+        raise argparse.ArgumentTypeError(
+            "tasks must be a non-empty comma-separated subset of T0,T1,T2,T3,T4"
+        )
+    unknown = [item for item in result if item not in TASK_BY_ID]
+    if unknown:
+        raise argparse.ArgumentTypeError(f"unknown task(s): {','.join(unknown)}")
+    return result
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--distances",
+        type=parse_distances,
+        default=[9],
+        help=(
+            "Comma-separated distances. Use 7,9 with --tasks T0 for the "
+            "paper's mapped-noise geometry; the default is release coverage at d=9."
+        ),
+    )
+    parser.add_argument(
+        "--tasks",
+        type=parse_tasks,
+        default=list(TASK_BY_ID),
+        help="Comma-separated task subset; defaults to T0,T1,T2,T3,T4.",
+    )
+    parser.add_argument("--n-rounds", type=int, default=9)
+    add_common_inference_args(
+        parser,
+        default_output_dir=Path("outputs/examples/released_models/t0_t4"),
+    )
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    jobs = []
+    for distance in args.distances:
+        for task_id in args.tasks:
+            task_key, config_name = TASK_BY_ID[task_id]
+            label = f"d{distance}_{task_key}"
+            output_path = args.output_dir / f"d{distance}" / f"{task_key}.json"
+            jobs.append(
+                InferenceJob(
+                    label=label,
+                    command=build_paired_command(
+                        args,
+                        config_name=config_name,
+                        output_path=output_path,
+                        distance=distance,
+                        n_rounds=args.n_rounds,
+                    ),
+                    output_path=output_path,
+                )
+            )
+    run_jobs(
+        jobs,
+        gpus=parse_gpus(args.gpus),
+        parallelism=args.parallelism,
+        resume=args.resume,
+        dry_run=args.dry_run,
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/code/examples/infer_willow.py b/code/examples/infer_willow.py
new file mode 100644
index 0000000..808c1f2
--- /dev/null
+++ b/code/examples/infer_willow.py
@@ -0,0 +1,118 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Reproduce the paper's d=3/d=5/d=7 native-length Google Willow evaluation."""
+
+from __future__ import annotations
+
+import argparse
+import os
+import shlex
+import sys
+from pathlib import Path
+
+CODE_ROOT = Path(__file__).resolve().parents[1]
+if str(CODE_ROOT) not in sys.path:
+    sys.path.insert(0, str(CODE_ROOT))
+
+from scripts.qadapt_example_utils import (  # noqa: E402
+    add_common_inference_args,
+    checkpoint_specs,
+    parse_gpus,
+)
+
+
+def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--benchmark-root",
+        type=Path,
+        default=Path("benchmarks/google_qec/google_105Q_surface_code_d3_d5_d7"),
+    )
+    parser.add_argument(
+        "--distances",
+        nargs="+",
+        type=int,
+        default=[3, 5, 7],
+        help="Paper default: d=3, d=5 and d=7.",
+    )
+    parser.add_argument(
+        "--rounds",
+        nargs="+",
+        type=int,
+        default=[10, 30, 50],
+        help="Paper default: 10, 30 and 50 native syndrome-extraction rounds.",
+    )
+    parser.add_argument("--config-name", default="examples/qadapt_v2/T0")
+    add_common_inference_args(
+        parser,
+        default_output_dir=Path("outputs/examples/released_models/willow"),
+        default_num_samples=50000,
+    )
+    return parser.parse_args(argv)
+
+
+def main(argv: list[str] | None = None) -> int:
+    args = parse_args(argv)
+    output_path = args.output_dir / "results.json"
+    if args.resume and output_path.is_file():
+        print(f"[resume] output exists: {output_path}")
+        return 0
+
+    selected_gpus = parse_gpus(args.gpus)
+    bases = ["X", "Z"] if args.basis == "both" else [args.basis]
+    specs = checkpoint_specs(args)
+    command_preview = [
+        str(args.python),
+        "-m",
+        "scripts.providers.google_qec_decoder_benchmark",
+        "--benchmark-root",
+        str(args.benchmark_root),
+        "--distances",
+        *(str(value) for value in args.distances),
+        "--rounds",
+        *(str(value) for value in args.rounds),
+        "--config-name",
+        args.config_name,
+        "--bases",
+        *bases,
+        "--models",
+        *(spec.name for spec in specs),
+        "--max-shots",
+        str(args.num_samples),
+        "--batch-size",
+        str(args.batch_size),
+        "--latency-shots",
+        str(args.latency_num_samples),
+        "--output",
+        str(output_path),
+    ]
+    if args.dry_run:
+        print(
+            f"[dry-run] gpu={selected_gpus[0]} seed={args.seed} "
+            + shlex.join(command_preview)
+        )
+        for spec in specs:
+            print(
+                f"[dry-run] model {spec.name}: "
+                f"model_id={spec.model_id} checkpoint={spec.checkpoint}"
+            )
+        return 0
+
+    os.environ["CUDA_VISIBLE_DEVICES"] = selected_gpus[0]
+    from scripts.providers import google_qec_decoder_benchmark as benchmark
+
+    benchmark.DEFAULT_MODELS = {
+        spec.name: benchmark.BenchmarkModel(
+            spec.name,
+            spec.model_id,
+            spec.checkpoint,
+        )
+        for spec in specs
+    }
+    benchmark.DEFAULT_BENCHMARK_ROOT = args.benchmark_root
+    return benchmark.main(command_preview[3:])
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/code/export/safetensors_utils.py b/code/export/safetensors_utils.py
index 76868bd..8306036 100644
--- a/code/export/safetensors_utils.py
+++ b/code/export/safetensors_utils.py
@@ -85,8 +85,12 @@ def save_safetensors(
         "num_filters": json.dumps(list(spec.num_filters)),
         "kernel_size": json.dumps(list(spec.kernel_size)),
     }
+    if hasattr(model, "htnet_ablation"):
+        metadata["htnet_ablation"] = json.dumps(model.htnet_ablation, sort_keys=True)
     if extra_metadata:
         for k, v in extra_metadata.items():
+            if str(k) in metadata and str(v) != metadata[str(k)]:
+                raise ValueError(f"extra_metadata cannot override architecture metadata: {k}")
             metadata[str(k)] = str(v)
 
     save_file(model.state_dict(), path, metadata=metadata)
@@ -123,6 +127,15 @@ def load_safetensors(
         model_id = metadata["model_id"]
 
     cfg = _build_minimal_cfg(model_id)
+    if "htnet_ablation" in metadata:
+        flags = json.loads(metadata["htnet_ablation"])
+        expected = {"branch_mixer", "axis_channel_gate", "input_skip_head"}
+        if (not isinstance(flags, dict) or set(flags) != expected
+                or any(type(value) is not bool for value in flags.values())):
+            raise ValueError("htnet_ablation metadata must contain three boolean architecture flags")
+        if cfg.model.version not in ("htnet", "predecoder_st_fusion_v2"):
+            raise ValueError("htnet_ablation metadata requires an HTNet model")
+        cfg.htnet_ablation = OmegaConf.create(flags)
     model = ModelFactory.create_model(cfg).to(device)
 
     if metadata.get("quant_format") == "fp16":
diff --git a/code/model/checkpoint_loader.py b/code/model/checkpoint_loader.py
new file mode 100644
index 0000000..2387be2
--- /dev/null
+++ b/code/model/checkpoint_loader.py
@@ -0,0 +1,48 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Load one explicitly identified pre-decoder checkpoint."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import torch
+
+
+def load_model_checkpoint(
+    cfg: Any,
+    *,
+    checkpoint: Path,
+    model_id: int,
+    distributed: Any,
+) -> torch.nn.Module:
+    """Load a ``.pt`` or ``.safetensors`` checkpoint for one public model ID."""
+
+    path = Path(checkpoint).expanduser().resolve()
+    if not path.is_file():
+        raise FileNotFoundError(f"Checkpoint not found: {path}")
+
+    if path.suffix.lower() != ".safetensors":
+        from workflows.run import _load_model
+
+        cfg.model_checkpoint_file = str(path)
+        return _load_model(cfg, distributed)
+
+    from export.safetensors_utils import load_safetensors
+
+    model, metadata = load_safetensors(
+        str(path),
+        model_id=None,
+        device=str(distributed.device),
+    )
+    embedded_model_id = metadata.get("model_id")
+    if embedded_model_id is not None and str(embedded_model_id) != str(model_id):
+        raise ValueError(
+            f"SafeTensors model_id mismatch for {path}: "
+            f"CLI requested {model_id}, file metadata contains {embedded_model_id}"
+        )
+    cfg.enable_fp16 = metadata.get("quant_format") == "fp16"
+    cfg.model_checkpoint_file = str(path)
+    return model
diff --git a/code/model/factory.py b/code/model/factory.py
index cd436a7..dabdd81 100644
--- a/code/model/factory.py
+++ b/code/model/factory.py
@@ -1,5 +1,6 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
+# Modified in 2026 for the QAdapt Hugging Face release: added HTNet dispatch.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -38,6 +39,9 @@ class ModelFactory:
             from model.predecoder import PreDecoderModelMemory_v1
             model = PreDecoderModelMemory_v1(cfg)
             return model
+        elif cfg.model.version == "htnet":
+            from model.qadapt import HTnet
+            return HTnet(cfg)
         elif cfg.model.version == "predecoder_memory_v2":
             from model.predecoder import PreDecoderModelMemory_v2
             model = PreDecoderModelMemory_v2(cfg)
diff --git a/code/model/qadapt.py b/code/model/qadapt.py
new file mode 100644
index 0000000..9448250
--- /dev/null
+++ b/code/model/qadapt.py
@@ -0,0 +1,270 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""HTNet inference architectures. V2 A101 and legacy V1 are selected by checkpoint metadata."""
+import torch
+from torch import nn
+
+class AdaptiveBranchFusion3D(nn.Module):
+
+    def __init__(self, channels, reduction, activation):
+        super(AdaptiveBranchFusion3D, self).__init__()
+        self.num_branches = 3
+        hidden_channels = max(1, channels // (reduction + 2))
+        self.pool = nn.AdaptiveAvgPool3d(1)
+        self.weight_net = nn.Sequential(
+            nn.Conv3d(channels * self.num_branches, hidden_channels, kernel_size=1),
+            activation,
+            nn.Conv3d(hidden_channels, channels * self.num_branches, kernel_size=1),
+        )
+        nn.init.zeros_(self.weight_net[-1].weight)
+        nn.init.zeros_(self.weight_net[-1].bias)
+
+    def forward(self, spatial, temporal, joint):
+        batch_size, channels = spatial.shape[:2]
+        pooled = torch.cat(
+            [self.pool(spatial), self.pool(temporal), self.pool(joint)],
+            dim=1,
+        )
+        weights = self.weight_net(pooled).view(
+            batch_size, self.num_branches, channels, 1, 1, 1
+        )
+        weights = torch.softmax(weights, dim=1)
+        fused = (
+            weights[:, 0] * spatial
+            + weights[:, 1] * temporal
+            + weights[:, 2] * joint
+        )
+        return fused * self.num_branches
+
+
+class ChannelGate3D(nn.Module):
+
+    def __init__(self, channels, reduction, activation):
+        super(ChannelGate3D, self).__init__()
+        hidden_channels = max(1, channels // reduction)
+        self.net = nn.Sequential(
+            nn.AdaptiveAvgPool3d(1),
+            nn.Conv3d(channels, hidden_channels, kernel_size=1),
+            activation,
+            nn.Conv3d(hidden_channels, channels, kernel_size=1),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x):
+        return x * self.net(x)
+
+
+class AxisChannelGate3D(nn.Module):
+
+    def __init__(self, channels, reduction, activation):
+        super(AxisChannelGate3D, self).__init__()
+        hidden_channels = max(1, channels // reduction)
+        self.channel_net = nn.Sequential(
+            nn.AdaptiveAvgPool3d(1),
+            nn.Conv3d(channels, hidden_channels, kernel_size=1),
+            activation,
+            nn.Conv3d(hidden_channels, channels, kernel_size=1),
+        )
+        self.temporal_conv = nn.Conv3d(
+            1,
+            1,
+            kernel_size=(3, 1, 1),
+            padding=(1, 0, 0),
+        )
+        self.spatial_conv = nn.Conv3d(
+            1,
+            1,
+            kernel_size=(1, 3, 3),
+            padding=(0, 1, 1),
+        )
+
+    def forward(self, x):
+        channel_logits = self.channel_net(x)
+        temporal_logits = self.temporal_conv(x.mean(dim=(1, 3, 4), keepdim=True))
+        spatial_logits = self.spatial_conv(x.mean(dim=(1, 2), keepdim=True))
+        return x * torch.sigmoid(channel_logits + temporal_logits + spatial_logits)
+
+
+class STFusionBlockV2(nn.Module):
+
+    def __init__(
+        self,
+        channels,
+        expand_channels,
+        joint_groups,
+        norm_groups,
+        se_reduction,
+        dropout_p,
+        activation_name,
+        use_branch_mixer=True,
+        use_axis_channel_gate=True,
+    ):
+        super(STFusionBlockV2, self).__init__()
+        if expand_channels % joint_groups != 0:
+            raise ValueError(
+                f"expand_channels ({expand_channels}) must be divisible by joint_groups ({joint_groups})"
+            )
+        if channels % norm_groups != 0 or expand_channels % norm_groups != 0:
+            raise ValueError(
+                "channels and expand_channels must be divisible by norm_groups "
+                f"(got channels={channels}, expand_channels={expand_channels}, norm_groups={norm_groups})"
+            )
+
+        activation = self._get_activation(activation_name)
+        self.pre = nn.Sequential(
+            nn.GroupNorm(num_groups=norm_groups, num_channels=channels),
+            nn.Conv3d(channels, expand_channels, kernel_size=1),
+            activation,
+        )
+        self.spatial = nn.Conv3d(
+            expand_channels,
+            expand_channels,
+            kernel_size=(1, 3, 3),
+            padding=(0, 1, 1),
+            groups=expand_channels,
+        )
+        self.temporal = nn.Conv3d(
+            expand_channels,
+            expand_channels,
+            kernel_size=(3, 1, 1),
+            padding=(1, 0, 0),
+            groups=expand_channels,
+        )
+        self.joint = nn.Sequential(
+            nn.GroupNorm(num_groups=norm_groups, num_channels=expand_channels),
+            nn.Conv3d(
+                expand_channels,
+                expand_channels,
+                kernel_size=3,
+                padding=1,
+                groups=joint_groups,
+            ),
+        )
+        self.branch_fusion = AdaptiveBranchFusion3D(
+            channels=expand_channels,
+            reduction=se_reduction,
+            activation=self._get_activation(activation_name),
+        )
+        self.branch_mixer = (
+            nn.Sequential(
+                nn.Conv3d(
+                    expand_channels,
+                    expand_channels,
+                    kernel_size=1,
+                    groups=joint_groups,
+                ),
+                self._get_activation(activation_name),
+            )
+            if use_branch_mixer
+            else nn.Identity()
+        )
+        self.project = nn.Sequential(
+            nn.Conv3d(expand_channels, channels, kernel_size=1),
+            self._get_activation(activation_name),
+        )
+        gate_type = AxisChannelGate3D if use_axis_channel_gate else ChannelGate3D
+        self.gate = gate_type(
+            channels=channels,
+            reduction=se_reduction,
+            activation=self._get_activation(activation_name),
+        )
+        self.dropout = nn.Dropout3d(p=dropout_p)
+
+    def _get_activation(self, name):
+        if name == "relu":
+            return nn.ReLU()
+        elif name == "gelu":
+            return nn.GELU(approximate='tanh')
+        elif name == "leakyrelu":
+            return nn.LeakyReLU()
+        else:
+            raise ValueError(f"Unsupported activation: {name}")
+
+    def forward(self, x):
+        residual = x
+        y = self.pre(x)
+        y = self.branch_fusion(self.spatial(y), self.temporal(y), self.joint(y))
+        y = self.branch_mixer(y)
+        y = self.project(y)
+        y = self.gate(y)
+        return residual + self.dropout(y)
+
+
+class HTnet(nn.Module):
+
+    def __init__(self, cfg):
+        super().__init__()
+
+        self.distance = cfg.distance
+        self.n_rounds = cfg.n_rounds
+        self.dropout_p = cfg.model.dropout_p
+
+        input_channels = cfg.model.input_channels
+        out_channels = cfg.model.out_channels
+        channels = int(cfg.model.channels)
+        expand_channels = int(cfg.model.expand_channels)
+        num_blocks = int(cfg.model.num_blocks)
+        joint_groups = int(cfg.model.joint_groups)
+        norm_groups = int(cfg.model.norm_groups)
+        se_reduction = int(cfg.model.se_reduction)
+        activation_name = cfg.model.activation
+        ablation_cfg = getattr(cfg, "htnet_ablation", None)
+        use_branch_mixer = bool(getattr(ablation_cfg, "branch_mixer", True))
+        use_axis_channel_gate = bool(
+            getattr(ablation_cfg, "axis_channel_gate", True)
+        )
+        self.use_input_skip_head = bool(
+            getattr(ablation_cfg, "input_skip_head", True)
+        )
+
+        self.htnet_ablation = {
+            "branch_mixer": use_branch_mixer,
+            "axis_channel_gate": use_axis_channel_gate,
+            "input_skip_head": self.use_input_skip_head,
+        }
+
+        self.stem = nn.Sequential(
+            nn.Conv3d(input_channels, channels, kernel_size=3, padding=1),
+            nn.GroupNorm(num_groups=norm_groups, num_channels=channels),
+            self._get_activation(activation_name),
+        )
+        self.blocks = nn.Sequential(*[
+            STFusionBlockV2(
+                channels=channels,
+                expand_channels=expand_channels,
+                joint_groups=joint_groups,
+                norm_groups=norm_groups,
+                se_reduction=se_reduction,
+                dropout_p=self.dropout_p,
+                activation_name=activation_name,
+                use_branch_mixer=use_branch_mixer,
+                use_axis_channel_gate=use_axis_channel_gate,
+            )
+            for _ in range(num_blocks)
+        ])
+        self.head_norm = nn.GroupNorm(num_groups=norm_groups, num_channels=channels)
+        self.head_hidden = nn.Conv3d(
+            channels + input_channels if self.use_input_skip_head else channels,
+            channels,
+            kernel_size=1,
+        )
+        self.head_activation = self._get_activation(activation_name)
+        self.head_out = nn.Conv3d(channels, out_channels, kernel_size=1)
+
+    def _get_activation(self, name):
+        if name == "relu":
+            return nn.ReLU()
+        elif name == "gelu":
+            return nn.GELU(approximate='tanh')
+        elif name == "leakyrelu":
+            return nn.LeakyReLU()
+        else:
+            raise ValueError(f"Unsupported activation: {name}")
+
+    def forward(self, x):
+        y = self.blocks(self.stem(x))
+        y = self.head_norm(y)
+        if self.use_input_skip_head:
+            y = torch.cat([y, x], dim=1)
+        y = self.head_activation(self.head_hidden(y))
+        return self.head_out(y)
diff --git a/code/model/registry.py b/code/model/registry.py
index a17bf9f..96e31af 100644
--- a/code/model/registry.py
+++ b/code/model/registry.py
@@ -1,5 +1,6 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
+# Modified in 2026 for the QAdapt Hugging Face release: added model ID 111.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -52,6 +53,12 @@ class PublicModelSpec:
     kernel_size: List[int]
     receptive_field: int
     model_version: str = "predecoder_memory_v1"
+    channels: Optional[int] = None
+    expand_channels: Optional[int] = None
+    num_blocks: Optional[int] = None
+    joint_groups: Optional[int] = None
+    norm_groups: Optional[int] = None
+    se_reduction: Optional[int] = None
     # Non-convolutional models (e.g. the cascade/bottleneck model "B") are not
     # described by num_filters/kernel_size. For those, `model_overrides` carries
     # the full `model.*` block that should be written into the merged config.
@@ -86,6 +93,21 @@ _MODEL_SPECS: Dict[Union[int, str], PublicModelSpec] = {
             kernel_size=[3, 3, 3, 3],
             receptive_field=compute_receptive_field([3, 3, 3, 3]),
         ),
+    # QAdapt: three HTNet blocks with an effective receptive field of nine.
+    111:
+        PublicModelSpec(
+            model_id=111,
+            num_filters=[112, 112, 112, 112, 4],
+            kernel_size=[3, 3, 3, 3],
+            receptive_field=compute_receptive_field([3, 3, 3, 3]),
+            model_version="htnet",
+            channels=112,
+            expand_channels=168,
+            num_blocks=3,
+            joint_groups=6,
+            norm_groups=8,
+            se_reduction=4,
+        ),
     # Model 2: 4 conv layers, k=3, wider
     2:
         PublicModelSpec(
@@ -152,13 +174,17 @@ def _normalize_model_id(model_id: Union[int, str]) -> Union[int, str]:
 
 
 def get_model_spec(model_id: Union[int, str]) -> PublicModelSpec:
-    """Return the public model spec for a given model_id (1..5 or "B")."""
+    """Return a public model spec, including QAdapt model_id 111."""
     try:
         key = _normalize_model_id(model_id)
     except Exception as e:
-        raise ValueError(f"model_id must be one of [1..5] or 'B', got: {model_id!r}") from e
+        raise ValueError(
+            f"model_id must be one of [1..5], 111, or 'B', got: {model_id!r}"
+        ) from e
     if key == 0:
         raise ValueError("model_id=0 is not supported in the public release")
     if key not in _MODEL_SPECS:
-        raise ValueError(f"model_id must be one of [1..5] or 'B', got: {model_id!r}")
+        raise ValueError(
+            f"model_id must be one of [1..5], 111, or 'B', got: {model_id!r}"
+        )
     return _MODEL_SPECS[key]
diff --git a/code/scripts/__init__.py b/code/scripts/__init__.py
new file mode 100644
index 0000000..ec0797d
--- /dev/null
+++ b/code/scripts/__init__.py
@@ -0,0 +1,4 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Developer and experiment command modules."""
diff --git a/code/scripts/config_paths.py b/code/scripts/config_paths.py
new file mode 100644
index 0000000..18aef86
--- /dev/null
+++ b/code/scripts/config_paths.py
@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Shared helpers for Hydra config names stored below ``conf/``."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any, Mapping
+
+
+CODE_ROOT = Path(__file__).resolve().parents[1]
+REPO_ROOT = CODE_ROOT.parent
+CONF_ROOT = REPO_ROOT / "conf"
+
+
+def rel(path: str | Path) -> Path:
+    path = Path(path)
+    return path if path.is_absolute() else REPO_ROOT / path
+
+
+def config_path(config_name: str | Path) -> Path:
+    """Return the YAML path for a Hydra config name below ``conf/``.
+
+    Configs are grouped in nested preset and experiment directories. For callers
+    that still pass a historical basename, return its unique recursive match.
+    """
+    raw = str(config_name)
+    if raw.endswith(".yaml"):
+        raw = raw[:-5]
+    direct = CONF_ROOT / f"{raw}.yaml"
+    if direct.exists() or "/" in raw or "\\" in raw:
+        return direct
+    matches = sorted(CONF_ROOT.rglob(f"{raw}.yaml"))
+    if len(matches) == 1:
+        return matches[0]
+    return direct
+
+
+def config_name_from_path(path: str | Path) -> str:
+    """Return the Hydra config name for a YAML path when it is below a ``conf/`` dir."""
+    path = rel(path)
+    try:
+        relative = path.relative_to(CONF_ROOT)
+    except ValueError:
+        parts = path.parts
+        if "conf" not in parts:
+            return path.stem
+        conf_index = len(parts) - 1 - list(reversed(parts)).index("conf")
+        relative = Path(*parts[conf_index + 1 :])
+    return relative.with_suffix("").as_posix()
+
+
+def config_basename(config_name: str | Path) -> str:
+    """Return the final component of a Hydra config name."""
+    raw = str(config_name)
+    if raw.endswith(".yaml"):
+        raw = raw[:-5]
+    return Path(raw).name
+
+
+def config_lookup_with_basename(
+    environments: list[Mapping[str, Any]] | tuple[Mapping[str, Any], ...],
+) -> dict[str, dict[str, Any]]:
+    """Map both full config names and historical basenames to manifest rows."""
+    lookup: dict[str, dict[str, Any]] = {}
+    for env in environments:
+        item = dict(env)
+        full = str(item["config_name"])
+        lookup[full] = item
+        lookup.setdefault(config_basename(full), item)
+    return lookup
diff --git a/code/scripts/download_google_qec_benchmark.py b/code/scripts/download_google_qec_benchmark.py
new file mode 100644
index 0000000..a02de79
--- /dev/null
+++ b/code/scripts/download_google_qec_benchmark.py
@@ -0,0 +1,89 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Download Google Quantum AI QEC benchmark archives from Zenodo."""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+from benchmarks.google_qec import (
+    DEFAULT_BENCHMARK_KEY,
+    GoogleQECBenchmarkStore,
+    benchmark_keys,
+    build_download_plan,
+    fetch_zenodo_manifest,
+)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path("benchmarks/google_qec"),
+        help="Directory for manifest and downloaded zip archives.",
+    )
+    parser.add_argument(
+        "--file",
+        action="append",
+        dest="files",
+        help=(
+            "Zenodo file key to download. May be repeated. "
+            f"Default: {DEFAULT_BENCHMARK_KEY}"
+        ),
+    )
+    parser.add_argument("--all", action="store_true", help="Download all Google QEC archives.")
+    parser.add_argument("--list", action="store_true", help="List available archives and exit.")
+    parser.add_argument("--manifest-only", action="store_true", help="Only write manifest.json.")
+    parser.add_argument("--extract", action="store_true", help="Extract downloaded zip archives.")
+    parser.add_argument("--force", action="store_true", help="Re-download archives that already exist.")
+    parser.add_argument("--skip-space-check", action="store_true", help="Skip free-space guard.")
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = _parse_args()
+    manifest = fetch_zenodo_manifest()
+    store = GoogleQECBenchmarkStore(args.output_dir)
+
+    if args.list:
+        for entry in manifest.files:
+            gib = entry.size_bytes / (1024**3)
+            distances = ",".join(str(d) for d in entry.distances) or "unknown"
+            print(f"{entry.key}\t{gib:.2f} GiB\t{entry.code_family}\td={distances}")
+        return 0
+
+    if args.all:
+        keys = benchmark_keys(manifest.files)
+    else:
+        keys = tuple(args.files) if args.files else (DEFAULT_BENCHMARK_KEY,)
+
+    store.write_manifest(manifest)
+    plan = build_download_plan(manifest, args.output_dir, keys)
+    print(f"Google QEC Zenodo record: {manifest.record_url}")
+    print(f"Output directory: {args.output_dir}")
+    for item in plan.items:
+        status = "exists" if item.exists else "download"
+        gib = item.entry.size_bytes / (1024**3)
+        print(f"  [{status}] {item.entry.key} ({gib:.2f} GiB, md5={item.entry.md5})")
+
+    if args.manifest_only:
+        print(f"Wrote manifest: {store.manifest_path}")
+        return 0
+
+    store.download(
+        manifest,
+        keys,
+        force=args.force,
+        extract=args.extract,
+        check_space=not args.skip_space_check,
+    )
+    print("Download complete.")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/code/scripts/experiments/__init__.py b/code/scripts/experiments/__init__.py
new file mode 100644
index 0000000..45fad83
--- /dev/null
+++ b/code/scripts/experiments/__init__.py
@@ -0,0 +1,4 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Experiment orchestration modules."""
diff --git a/code/scripts/experiments/unknown_noise/__init__.py b/code/scripts/experiments/unknown_noise/__init__.py
new file mode 100644
index 0000000..d5bbf76
--- /dev/null
+++ b/code/scripts/experiments/unknown_noise/__init__.py
@@ -0,0 +1,4 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Unknown-noise experiment configuration and comparison modules."""
diff --git a/code/scripts/experiments/unknown_noise/generate_unknown_axismix_grid_u1p2_5p0_configs.py b/code/scripts/experiments/unknown_noise/generate_unknown_axismix_grid_u1p2_5p0_configs.py
new file mode 100644
index 0000000..fb3d72d
--- /dev/null
+++ b/code/scripts/experiments/unknown_noise/generate_unknown_axismix_grid_u1p2_5p0_configs.py
@@ -0,0 +1,348 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Generate fixed multiplier-grid training-axis mixed OOD noise configs."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from itertools import combinations
+from pathlib import Path
+from typing import Any, Mapping, Sequence
+
+from omegaconf import OmegaConf
+
+CODE_ROOT = Path(__file__).resolve().parents[3]
+REPO_ROOT = CODE_ROOT.parent
+if str(CODE_ROOT) not in sys.path:
+    sys.path.insert(0, str(CODE_ROOT))
+
+from qec.noise_model import NoiseModel  # noqa: E402
+from scripts.config_paths import config_name_from_path  # noqa: E402
+
+
+DEFAULT_BASE_CONFIG = "conf/examples/qadapt/config_qadapt_t0_base.yaml"
+DESIGN_LABEL = "training-axis fixed multiplier grid OOD stress test"
+DEFAULT_PREFIX = "config_unknown_axismix_grid_u1p2_5p0"
+DEFAULT_OUTPUT_DIR = "outputs/generated_configs/ood"
+DEFAULT_MANIFEST = "outputs/generated_configs/ood/manifest.json"
+AXIS_ORDER = ("meas_all", "cnot_all", "idle_all", "z_bias")
+GRID_MULTIPLIERS = (1.2, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0)
+
+CNOT_KEYS = (
+    "p_cnot_IX",
+    "p_cnot_IY",
+    "p_cnot_IZ",
+    "p_cnot_XI",
+    "p_cnot_XX",
+    "p_cnot_XY",
+    "p_cnot_XZ",
+    "p_cnot_YI",
+    "p_cnot_YX",
+    "p_cnot_YY",
+    "p_cnot_YZ",
+    "p_cnot_ZI",
+    "p_cnot_ZX",
+    "p_cnot_ZY",
+    "p_cnot_ZZ",
+)
+
+AXES: dict[str, tuple[str, ...]] = {
+    "meas_all": ("p_meas_X", "p_meas_Z"),
+    "cnot_all": CNOT_KEYS,
+    "idle_all": (
+        "p_idle_cnot_X",
+        "p_idle_cnot_Y",
+        "p_idle_cnot_Z",
+        "p_idle_spam_X",
+        "p_idle_spam_Y",
+        "p_idle_spam_Z",
+    ),
+    "z_bias": (
+        "p_prep_X",
+        "p_meas_X",
+        "p_idle_cnot_Z",
+        "p_idle_spam_Z",
+        "p_cnot_IZ",
+        "p_cnot_XZ",
+        "p_cnot_YZ",
+        "p_cnot_ZI",
+        "p_cnot_ZX",
+        "p_cnot_ZY",
+        "p_cnot_ZZ",
+    ),
+}
+
+
+def rel(path: str | Path) -> Path:
+    path = Path(path)
+    return path if path.is_absolute() else REPO_ROOT / path
+
+
+def _plain_mapping(value: Any) -> dict[str, float]:
+    raw = OmegaConf.to_container(value, resolve=True) if hasattr(value, "items") else value
+    if raw is None:
+        raise ValueError("base config does not contain data.noise_model")
+    return {str(key): float(item) for key, item in dict(raw).items()}
+
+
+def load_base_noise_model(base_config: str | Path) -> dict[str, float]:
+    cfg = OmegaConf.load(rel(base_config))
+    noise_cfg = getattr(getattr(cfg, "data", None), "noise_model", None)
+    noise = _plain_mapping(noise_cfg)
+    return NoiseModel.from_config_dict(noise).to_config_dict()
+
+
+def multiplier_key(multiplier: float) -> str:
+    return f"m{float(multiplier):.1f}".replace(".", "p")
+
+
+def _axis_signature(active_axes: Sequence[str]) -> str:
+    return "+".join(active_axes)
+
+
+def _default_env_specs() -> list[dict[str, Any]]:
+    specs = []
+    for size in (2, 3, 4):
+        for active_axes in combinations(AXIS_ORDER, size):
+            env_index = len(specs)
+            specs.append(
+                {
+                    "env_index": env_index,
+                    "env_key": f"e{env_index:02d}",
+                    "active_axes": tuple(active_axes),
+                    "axis_signature": _axis_signature(active_axes),
+                    "combination_size": size,
+                    "contains_z_bias": "z_bias" in active_axes,
+                    "contains_cnot_z_bias": "cnot_all" in active_axes and "z_bias" in active_axes,
+                    "purpose": f"{size}-axis fixed multiplier grid composite",
+                }
+            )
+    return specs
+
+
+DEFAULT_ENV_SPECS: list[dict[str, Any]] = _default_env_specs()
+
+
+def _normalize_spec(raw_spec: Mapping[str, Any]) -> dict[str, Any]:
+    env_index = int(raw_spec["env_index"])
+    active_axes = tuple(str(axis) for axis in raw_spec["active_axes"])
+    if not 2 <= len(active_axes) <= 4:
+        raise ValueError(f"grid env must activate 2, 3, or 4 axes, got {active_axes}")
+    unknown = [axis for axis in active_axes if axis not in AXIS_ORDER]
+    if unknown:
+        raise ValueError(f"unknown grid axes: {unknown}")
+    if len(set(active_axes)) != len(active_axes):
+        raise ValueError(f"duplicate active axes: {active_axes}")
+    return {
+        "env_index": env_index,
+        "env_key": str(raw_spec.get("env_key", f"e{env_index:02d}")),
+        "active_axes": active_axes,
+        "axis_signature": str(raw_spec.get("axis_signature", _axis_signature(active_axes))),
+        "combination_size": len(active_axes),
+        "contains_z_bias": "z_bias" in active_axes,
+        "contains_cnot_z_bias": "cnot_all" in active_axes and "z_bias" in active_axes,
+        "purpose": str(raw_spec.get("purpose", f"{len(active_axes)}-axis fixed multiplier grid composite")),
+    }
+
+
+def _parameter_multipliers(
+    base_noise: Mapping[str, float],
+    active_axes: Sequence[str],
+    multiplier: float,
+) -> dict[str, float]:
+    multipliers = {key: 1.0 for key in base_noise}
+    for axis_name in active_axes:
+        if axis_name not in AXES:
+            raise ValueError(f"unknown training noise axis: {axis_name}")
+        for key in AXES[axis_name]:
+            if key not in base_noise:
+                raise ValueError(f"axis {axis_name} references missing noise parameter {key}")
+            multipliers[key] = max(multipliers[key], float(multiplier))
+    return multipliers
+
+
+def _probability_totals(noise: Mapping[str, float]) -> dict[str, float]:
+    return {
+        "cnot_total": sum(value for key, value in noise.items() if key.startswith("p_cnot_")),
+        "idle_cnot_total": sum(value for key, value in noise.items() if key.startswith("p_idle_cnot_")),
+        "idle_spam_total": sum(value for key, value in noise.items() if key.startswith("p_idle_spam_")),
+    }
+
+
+def generate_axismix_grid_noise_models(
+    base_noise: Mapping[str, float],
+    env_specs: Sequence[Mapping[str, Any]] = DEFAULT_ENV_SPECS,
+    *,
+    grid_multipliers: Sequence[float] = GRID_MULTIPLIERS,
+) -> list[dict[str, Any]]:
+    if not grid_multipliers:
+        raise ValueError("grid_multipliers must not be empty")
+    base = NoiseModel.from_config_dict(dict(base_noise)).to_config_dict()
+    generated = []
+    for raw_spec in env_specs:
+        spec = _normalize_spec(raw_spec)
+        for multiplier_index, multiplier in enumerate(grid_multipliers):
+            multiplier = float(multiplier)
+            if multiplier < 0:
+                raise ValueError(f"multiplier must be non-negative, got {multiplier}")
+            param_multipliers = _parameter_multipliers(base, spec["active_axes"], multiplier)
+            axis_multipliers = {
+                axis: (multiplier if axis in spec["active_axes"] else 1.0)
+                for axis in AXIS_ORDER
+            }
+            noise = {
+                key: float(base_value) * float(param_multipliers[key])
+                for key, base_value in base.items()
+            }
+            validated = NoiseModel.from_config_dict(noise)
+            noise = validated.to_config_dict()
+            generated.append(
+                {
+                    **spec,
+                    "multiplier_index": multiplier_index,
+                    "multiplier": multiplier,
+                    "multiplier_key": multiplier_key(multiplier),
+                    "axis_multipliers": axis_multipliers,
+                    "parameter_multipliers": param_multipliers,
+                    "noise_model": {key: float(value) for key, value in noise.items()},
+                    "probability_totals": _probability_totals(noise),
+                    "noise_model_sha256": validated.sha256(),
+                }
+            )
+    return generated
+
+
+def _render_config(base_cfg: Any, noise_model: Mapping[str, float], *, header: str) -> str:
+    cfg = OmegaConf.create(OmegaConf.to_container(base_cfg, resolve=True))
+    cfg.data.noise_model = dict(noise_model)
+    return header + OmegaConf.to_yaml(cfg, resolve=True)
+
+
+def _config_name(prefix: str, env_index: int, multiplier: float) -> str:
+    return f"{prefix}_e{int(env_index):02d}_{multiplier_key(multiplier)}"
+
+
+def write_axismix_grid_configs(
+    *,
+    base_config: str | Path = DEFAULT_BASE_CONFIG,
+    output_dir: str | Path = DEFAULT_OUTPUT_DIR,
+    prefix: str = DEFAULT_PREFIX,
+    manifest: str | Path = DEFAULT_MANIFEST,
+    env_specs: Sequence[Mapping[str, Any]] = DEFAULT_ENV_SPECS,
+    grid_multipliers: Sequence[float] = GRID_MULTIPLIERS,
+) -> tuple[list[Path], dict[str, Any]]:
+    base_path = rel(base_config)
+    if not base_path.exists():
+        raise FileNotFoundError(base_path)
+    out_dir = rel(output_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    base_cfg = OmegaConf.load(base_path)
+    base_noise = load_base_noise_model(base_path)
+    generated = generate_axismix_grid_noise_models(
+        base_noise,
+        env_specs,
+        grid_multipliers=grid_multipliers,
+    )
+
+    paths = []
+    environments = []
+    for item in generated:
+        config_name = _config_name(prefix, int(item["env_index"]), float(item["multiplier"]))
+        filename = f"{config_name}.yaml"
+        path = out_dir / filename
+        axis_json = json.dumps(item["axis_multipliers"], sort_keys=True)
+        header = (
+            "# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.\n"
+            "# SPDX-License-Identifier: Apache-2.0\n"
+            "\n"
+            "# Auto-generated training-axis fixed multiplier grid OOD noise environment.\n"
+            f"# design: {DESIGN_LABEL}\n"
+            f"# base_config: {base_path.name}\n"
+            f"# env_key: {item['env_key']}\n"
+            f"# active_axes: {item['axis_signature']}\n"
+            f"# multiplier: {float(item['multiplier']):.6g}\n"
+            f"# axis_multipliers: {axis_json}\n"
+            f"# noise_model_sha256: {item['noise_model_sha256']}\n\n"
+        )
+        path.write_text(
+            _render_config(base_cfg, item["noise_model"], header=header),
+            encoding="utf-8",
+        )
+        paths.append(path)
+        environments.append(
+            {
+                "env_index": int(item["env_index"]),
+                "env_key": item["env_key"],
+                "multiplier_index": int(item["multiplier_index"]),
+                "multiplier_key": item["multiplier_key"],
+                "multiplier": float(item["multiplier"]),
+                "config_name": config_name_from_path(path),
+                "config_filename": filename,
+                "active_axes": list(item["active_axes"]),
+                "axis_signature": item["axis_signature"],
+                "axis_multipliers": item["axis_multipliers"],
+                "combination_size": int(item["combination_size"]),
+                "contains_z_bias": bool(item["contains_z_bias"]),
+                "contains_cnot_z_bias": bool(item["contains_cnot_z_bias"]),
+                "parameter_multipliers": item["parameter_multipliers"],
+                "probability_totals": item["probability_totals"],
+                "noise_model_sha256": item["noise_model_sha256"],
+            }
+        )
+
+    env_count = len({int(item["env_index"]) for item in generated})
+    manifest_payload = {
+        "design": DESIGN_LABEL,
+        "base_config": str(base_config),
+        "prefix": prefix,
+        "axis_order": list(AXIS_ORDER),
+        "grid_multipliers": [float(value) for value in grid_multipliers],
+        "num_envs": env_count,
+        "num_configs": len(generated),
+        "axes": {name: list(keys) for name, keys in AXES.items()},
+        "environments": environments,
+    }
+    manifest_path = rel(manifest)
+    manifest_path.parent.mkdir(parents=True, exist_ok=True)
+    manifest_path.write_text(
+        json.dumps(manifest_payload, indent=2, sort_keys=True),
+        encoding="utf-8",
+    )
+    manifest_payload["manifest_path"] = str(manifest_path)
+    return paths, manifest_payload
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--base-config", default=DEFAULT_BASE_CONFIG)
+    parser.add_argument("--output-dir", default=DEFAULT_OUTPUT_DIR)
+    parser.add_argument("--prefix", default=DEFAULT_PREFIX)
+    parser.add_argument("--manifest", default=DEFAULT_MANIFEST)
+    parser.add_argument(
+        "--grid-multipliers",
+        default=",".join(str(value) for value in GRID_MULTIPLIERS),
+        help="Comma-separated multiplier grid.",
+    )
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    grid = [float(item.strip()) for item in args.grid_multipliers.split(",") if item.strip()]
+    paths, manifest = write_axismix_grid_configs(
+        base_config=args.base_config,
+        output_dir=args.output_dir,
+        prefix=args.prefix,
+        manifest=args.manifest,
+        grid_multipliers=grid,
+    )
+    print(f"[write] {manifest['manifest_path']}")
+    print(f"[write] {len(paths)} configs")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/code/scripts/paired_inference_compare.py b/code/scripts/paired_inference_compare.py
new file mode 100644
index 0000000..cf4ca4b
--- /dev/null
+++ b/code/scripts/paired_inference_compare.py
@@ -0,0 +1,943 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Paired inference comparison on one shared inference dataset.
+
+This script compares pure PyMatching with one or more predecoder models on the
+same samples for each measurement basis. Samples are generated by Stim unless
+``--stim-samples-dir`` points to external ``.dets`` artifacts. It is
+intentionally separate from the Hydra workflow so the standard train/inference
+entry points stay unchanged.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import math
+import os
+import random
+import sys
+import time
+from dataclasses import dataclass
+from itertools import combinations
+from pathlib import Path
+from types import SimpleNamespace
+from typing import Any
+
+import numpy as np
+import pymatching
+import torch
+from omegaconf import OmegaConf
+from torch.utils.data import DataLoader
+
+CODE_ROOT = Path(__file__).resolve().parents[1]
+REPO_ROOT = CODE_ROOT.parent
+if str(CODE_ROOT) not in sys.path:
+    sys.path.insert(0, str(CODE_ROOT))
+
+from scripts.config_paths import config_path  # noqa: E402
+from data.factory import DatapipeFactory  # noqa: E402
+from evaluation.logical_error_rate import (  # noqa: E402
+    PreDecoderMemoryEvalModule,
+    _build_stab_maps,
+)
+from training.utils import dict_to_device  # noqa: E402
+from workflows.config_validator import (  # noqa: E402
+    apply_public_defaults_and_model,
+    validate_public_config,
+)
+from model.checkpoint_loader import load_model_checkpoint  # noqa: E402
+
+
+@dataclass(frozen=True)
+class ModelSpec:
+    name: str
+    model_id: int
+    checkpoint: Path
+
+
+@dataclass(frozen=True)
+class ComparisonSpec:
+    candidate: str
+    baseline: str
+
+
+@dataclass(frozen=True)
+class FactorialContrastSpec:
+    name: str
+    cell_11: str
+    cell_10: str
+    cell_01: str
+    cell_00: str
+
+
+@dataclass
+class SyndromeDensityAccumulator:
+    """Stream shot-level syndrome-density moments without storing every sample."""
+
+    shots: int = 0
+    syndrome_ones: int = 0
+    syndrome_elements: int = 0
+    shot_density_sum: float = 0.0
+    shot_density_sum_squares: float = 0.0
+
+    def update(self, syndromes: np.ndarray) -> None:
+        values = np.asarray(syndromes, dtype=np.uint8)
+        if values.ndim == 1:
+            values = values.reshape(1, -1)
+        if values.ndim != 2 or values.shape[1] == 0:
+            raise ValueError(
+                "syndromes must be a non-empty-width 2D array, "
+                f"got shape={values.shape}"
+            )
+        ones_per_shot = np.count_nonzero(values, axis=1).astype(np.float64)
+        densities = ones_per_shot / float(values.shape[1])
+        self.shots += int(values.shape[0])
+        self.syndrome_ones += int(ones_per_shot.sum())
+        self.syndrome_elements += int(values.size)
+        self.shot_density_sum += float(densities.sum())
+        self.shot_density_sum_squares += float(np.square(densities).sum())
+
+    def statistics(self, prefix: str) -> dict[str, float | int]:
+        if not prefix:
+            raise ValueError("density prefix must not be empty")
+        center = (
+            float(self.syndrome_ones / self.syndrome_elements)
+            if self.syndrome_elements
+            else float("nan")
+        )
+        if self.shots > 1:
+            numerator = self.shot_density_sum_squares - (
+                self.shot_density_sum * self.shot_density_sum / self.shots
+            )
+            variance = max(0.0, numerator / (self.shots - 1))
+            standard_error = float(np.sqrt(variance / self.shots))
+        else:
+            standard_error = 0.0 if self.shots == 1 else float("nan")
+        margin = 1.96 * standard_error
+        return {
+            f"{prefix}_density_shots": self.shots,
+            f"{prefix}_syndrome_ones": self.syndrome_ones,
+            f"{prefix}_syndrome_elements": self.syndrome_elements,
+            f"{prefix}_density_shot_sum": self.shot_density_sum,
+            f"{prefix}_density_shot_sum_squares": self.shot_density_sum_squares,
+            f"{prefix}_syndrome_density": center,
+            f"{prefix}_density_standard_error": standard_error,
+            f"{prefix}_density_ci95_low": max(0.0, center - margin),
+            f"{prefix}_density_ci95_high": min(1.0, center + margin),
+        }
+
+
+def combine_density_statistics(
+    rows: list[dict[str, Any]],
+    prefix: str,
+) -> dict[str, float | int]:
+    """Combine density sufficient statistics using detector-element weighting."""
+
+    accumulator = SyndromeDensityAccumulator()
+    for row in rows:
+        accumulator.shots += int(row.get(f"{prefix}_density_shots", 0))
+        accumulator.syndrome_ones += int(row.get(f"{prefix}_syndrome_ones", 0))
+        accumulator.syndrome_elements += int(
+            row.get(f"{prefix}_syndrome_elements", 0)
+        )
+        accumulator.shot_density_sum += float(
+            row.get(f"{prefix}_density_shot_sum", 0.0)
+        )
+        accumulator.shot_density_sum_squares += float(
+            row.get(f"{prefix}_density_shot_sum_squares", 0.0)
+        )
+    return accumulator.statistics(prefix)
+
+
+def density_reduction_statistics(
+    input_density: float,
+    residual_density: float,
+) -> dict[str, float]:
+    input_value = float(input_density)
+    residual_value = float(residual_density)
+    delta = residual_value - input_value
+    if input_value > 0 and math.isfinite(input_value):
+        reduction_fraction = (input_value - residual_value) / input_value
+    else:
+        reduction_fraction = float("nan")
+    if residual_value > 0 and math.isfinite(residual_value):
+        reduction_factor = input_value / residual_value
+    elif input_value > 0 and residual_value == 0:
+        reduction_factor = float("inf")
+    else:
+        reduction_factor = float("nan")
+    return {
+        "density_delta": delta,
+        "density_reduction_fraction": reduction_fraction,
+        "density_reduction_factor": reduction_factor,
+    }
+
+
+def model_density_statistics(
+    input_accumulator: SyndromeDensityAccumulator,
+    residual_accumulator: SyndromeDensityAccumulator,
+) -> dict[str, float | int]:
+    input_stats = input_accumulator.statistics("input")
+    residual_stats = residual_accumulator.statistics("residual")
+    return {
+        **input_stats,
+        **residual_stats,
+        **density_reduction_statistics(
+            float(input_stats["input_syndrome_density"]),
+            float(residual_stats["residual_syndrome_density"]),
+        ),
+    }
+
+
+def parse_model_spec(value: str) -> ModelSpec:
+    parts = value.split(":", 2)
+    if len(parts) != 3:
+        raise argparse.ArgumentTypeError(
+            "--model must be formatted as name:model_id:/path/to/checkpoint"
+        )
+    name, model_id_raw, checkpoint_raw = parts
+    if not name:
+        raise argparse.ArgumentTypeError("model name must not be empty")
+    try:
+        model_id = int(model_id_raw)
+    except ValueError as exc:
+        raise argparse.ArgumentTypeError(f"invalid model_id: {model_id_raw}") from exc
+    checkpoint = Path(checkpoint_raw).expanduser()
+    if not checkpoint.is_absolute():
+        checkpoint = REPO_ROOT / checkpoint
+    return ModelSpec(name=name, model_id=model_id, checkpoint=checkpoint)
+
+
+def parse_comparison_spec(value: str) -> ComparisonSpec:
+    parts = value.split(":", 1)
+    if len(parts) != 2 or not all(part.strip() for part in parts):
+        raise argparse.ArgumentTypeError(
+            "--paired-comparison must be formatted as candidate:baseline"
+        )
+    candidate, baseline = (part.strip() for part in parts)
+    if candidate == baseline:
+        raise argparse.ArgumentTypeError("candidate and baseline must be different methods")
+    return ComparisonSpec(candidate=candidate, baseline=baseline)
+
+
+def parse_factorial_contrast_spec(value: str) -> FactorialContrastSpec:
+    parts = [part.strip() for part in value.split(":")]
+    if len(parts) != 5 or not all(parts):
+        raise argparse.ArgumentTypeError(
+            "--factorial-contrast must be formatted as "
+            "name:cell_11:cell_10:cell_01:cell_00"
+        )
+    name, cell_11, cell_10, cell_01, cell_00 = parts
+    if len({cell_11, cell_10, cell_01, cell_00}) != 4:
+        raise argparse.ArgumentTypeError("factorial contrast cells must be four distinct methods")
+    return FactorialContrastSpec(name, cell_11, cell_10, cell_01, cell_00)
+
+
+def factorial_contrast_statistics(
+    cell_11_errors: np.ndarray,
+    cell_10_errors: np.ndarray,
+    cell_01_errors: np.ndarray,
+    cell_00_errors: np.ndarray,
+) -> dict[str, float | int]:
+    masks = [
+        np.asarray(errors, dtype=np.bool_).reshape(-1)
+        for errors in (cell_11_errors, cell_10_errors, cell_01_errors, cell_00_errors)
+    ]
+    shapes = {mask.shape for mask in masks}
+    if len(shapes) != 1:
+        raise ValueError(f"factorial contrast masks must have one shape: {sorted(shapes)}")
+    samples = int(masks[0].size)
+    if samples == 0:
+        raise ValueError("factorial contrast masks must not be empty")
+    contrast = (
+        masks[0].astype(np.int8)
+        - masks[1].astype(np.int8)
+        - masks[2].astype(np.int8)
+        + masks[3].astype(np.int8)
+    )
+    interaction = float(contrast.mean())
+    standard_error = (
+        float(contrast.std(ddof=1) / np.sqrt(samples)) if samples > 1 else 0.0
+    )
+    margin = 1.96 * standard_error
+    result: dict[str, float | int] = {
+        "samples": samples,
+        "interaction_ler": interaction,
+        "standard_error": standard_error,
+        "ci95_low": max(-2.0, interaction - margin),
+        "ci95_high": min(2.0, interaction + margin),
+    }
+    result.update(
+        {
+            f"contrast_count_{value:+d}": int(np.count_nonzero(contrast == value))
+            for value in range(-2, 3)
+        }
+    )
+    return result
+
+
+def paired_error_statistics(
+    candidate_errors: np.ndarray,
+    baseline_errors: np.ndarray,
+) -> dict[str, float | int]:
+    candidate = np.asarray(candidate_errors, dtype=np.bool_).reshape(-1)
+    baseline = np.asarray(baseline_errors, dtype=np.bool_).reshape(-1)
+    if candidate.shape != baseline.shape:
+        raise ValueError(
+            f"paired error masks must have the same shape: {candidate.shape} != {baseline.shape}"
+        )
+    samples = int(candidate.size)
+    if samples == 0:
+        raise ValueError("paired error masks must not be empty")
+
+    candidate_only = int(np.count_nonzero(candidate & ~baseline))
+    baseline_only = int(np.count_nonzero(~candidate & baseline))
+    both = int(np.count_nonzero(candidate & baseline))
+    neither = samples - candidate_only - baseline_only - both
+    differences = candidate.astype(np.int8) - baseline.astype(np.int8)
+    delta = float(differences.mean())
+    standard_error = (
+        float(differences.std(ddof=1) / np.sqrt(samples)) if samples > 1 else 0.0
+    )
+    margin = 1.96 * standard_error
+    return {
+        "samples": samples,
+        "candidate_only_errors": candidate_only,
+        "baseline_only_errors": baseline_only,
+        "both_errors": both,
+        "neither_errors": neither,
+        "delta_ler": delta,
+        "standard_error": standard_error,
+        "ci95_low": max(-1.0, delta - margin),
+        "ci95_high": min(1.0, delta + margin),
+    }
+
+
+def paired_error_comparison(
+    method_a: str,
+    errors_a: np.ndarray,
+    method_b: str,
+    errors_b: np.ndarray,
+    *,
+    basis: str,
+) -> dict[str, Any]:
+    """Summarize two shot-aligned logical-error masks."""
+    stats = paired_error_statistics(errors_a, errors_b)
+    return {
+        "basis": basis,
+        "method_a": method_a,
+        "method_b": method_b,
+        "samples": stats["samples"],
+        "both_error": stats["both_errors"],
+        "a_only_error": stats["candidate_only_errors"],
+        "b_only_error": stats["baseline_only_errors"],
+        "neither_error": stats["neither_errors"],
+        "ler_delta_a_minus_b": stats["delta_ler"],
+        "paired_standard_error": stats["standard_error"],
+        "ler_delta_ci95_normal": [stats["ci95_low"], stats["ci95_high"]],
+    }
+
+
+def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Compare PyMatching and replaceable predecoder models on identical samples."
+    )
+    config_group = parser.add_mutually_exclusive_group()
+    config_group.add_argument(
+        "--config-name", default="examples/qadapt/config_qadapt_t0_base"
+    )
+    config_group.add_argument(
+        "--config-file", type=Path, help="Explicit YAML path, including generated OOD configs."
+    )
+    parser.add_argument("--distance", type=int, default=9)
+    parser.add_argument("--n-rounds", type=int, default=9)
+    parser.add_argument("--num-samples", type=int, default=262144)
+    parser.add_argument("--latency-num-samples", type=int, default=10000)
+    parser.add_argument("--batch-size", type=int, default=2048)
+    parser.add_argument("--num-workers", type=int, default=0)
+    parser.add_argument("--seed", type=int, default=12345)
+    parser.add_argument("--device", default=None)
+    parser.add_argument(
+        "--basis",
+        choices=("both", "X", "Z"),
+        default="both",
+        help="Measurement basis to evaluate.",
+    )
+    parser.add_argument(
+        "--stim-samples-dir",
+        default=None,
+        help=(
+            "Optional directory containing samples_X.dets/metadata_X.json and/or "
+            "samples_Z.dets/metadata_Z.json. When omitted, Stim generates samples."
+        ),
+    )
+    parser.add_argument(
+        "--model",
+        action="append",
+        type=parse_model_spec,
+        required=True,
+        help=(
+            "Repeatable model spec: name:model_id:/path/to/checkpoint "
+            "(.pt or .safetensors)."
+        ),
+    )
+    parser.add_argument(
+        "--paired-comparison",
+        action="append",
+        type=parse_comparison_spec,
+        default=[],
+        help="Repeatable paired comparison: candidate:baseline.",
+    )
+    parser.add_argument(
+        "--factorial-contrast",
+        action="append",
+        type=parse_factorial_contrast_spec,
+        default=[],
+        help="Repeatable contrast: name:cell_11:cell_10:cell_01:cell_00.",
+    )
+    parser.add_argument(
+        "--output",
+        default="outputs/examples/released_models/paired_inference.json",
+        help="JSON output path. A CSV summary is written next to it.",
+    )
+    parser.add_argument(
+        "--residual-output-dir",
+        default=None,
+        help=(
+            "Optional directory for full residual detector tensors. One uint8 "
+            "PyTorch tensor is written per basis and model."
+        ),
+    )
+    return parser.parse_args(argv)
+
+
+def set_all_seeds(seed: int) -> None:
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+
+def resolve_stim_samples_dir(args: argparse.Namespace) -> Path | None:
+    value = getattr(args, "stim_samples_dir", None) or os.environ.get(
+        "PREDECODER_STIM_SAMPLES_DIR"
+    )
+    if not value:
+        return None
+    path = Path(value).expanduser()
+    return path if path.is_absolute() else REPO_ROOT / path
+
+
+def build_cfg(args: argparse.Namespace, model: ModelSpec, basis: str) -> Any:
+    explicit_path = getattr(args, "config_file", None)
+    cfg_path = (
+        Path(explicit_path).expanduser()
+        if explicit_path is not None
+        else config_path(args.config_name)
+    )
+    cfg = OmegaConf.load(cfg_path)
+    cfg.model_id = model.model_id
+    cfg.distance = args.distance
+    cfg.n_rounds = args.n_rounds
+    cfg.workflow.task = "inference"
+
+    spec = validate_public_config(cfg)
+    cfg = apply_public_defaults_and_model(cfg, spec)
+    cfg.model_checkpoint_file = str(model.checkpoint)
+    cfg.test.meas_basis_test = basis
+    cfg.test.num_samples = int(args.num_samples)
+    cfg.test.latency_num_samples = int(args.latency_num_samples)
+    cfg.test.batch_size = int(args.batch_size)
+    cfg.test.dataloader_num_workers = int(args.num_workers)
+    stim_samples_dir = resolve_stim_samples_dir(args)
+    if stim_samples_dir:
+        cfg.test.stim_samples_dir = str(stim_samples_dir)
+    return cfg
+
+
+def make_dataset(cfg: Any, seed: int):
+    py_state = random.getstate()
+    np_state = np.random.get_state()
+    torch_state = torch.get_rng_state()
+    cuda_state = torch.cuda.get_rng_state_all() if torch.cuda.is_available() else None
+    try:
+        set_all_seeds(seed)
+        cfg.seed = int(seed)
+        return DatapipeFactory.create_datapipe_inference(cfg)
+    finally:
+        random.setstate(py_state)
+        np.random.set_state(np_state)
+        torch.set_rng_state(torch_state)
+        if cuda_state is not None:
+            torch.cuda.set_rng_state_all(cuda_state)
+
+
+def time_single_shot(matcher: pymatching.Matching, syndromes: np.ndarray, n_rounds: int) -> float:
+    n_rounds = max(int(n_rounds), 1)
+    if syndromes.size == 0:
+        return float("nan")
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+    warmup_n = min(50, len(syndromes))
+    for i in range(warmup_n):
+        matcher.decode(np.asarray(syndromes[i], dtype=np.uint8))
+
+    times = []
+    for row in syndromes:
+        start = time.perf_counter()
+        matcher.decode(np.asarray(row, dtype=np.uint8))
+        times.append(time.perf_counter() - start)
+    return float(np.mean(times) / n_rounds * 1e6)
+
+
+def build_matcher(dataset) -> tuple[pymatching.Matching, int]:
+    circuit = dataset.circ.stim_circuit
+    det_model = circuit.detector_error_model(decompose_errors=True, approximate_disjoint_errors=True)
+    return pymatching.Matching.from_detector_error_model(det_model), int(circuit.num_observables)
+
+
+def evaluate_pymatching(
+    matcher: pymatching.Matching,
+    dets_and_obs: np.ndarray,
+    num_obs: int,
+    latency_samples: int,
+    n_rounds: int,
+) -> tuple[dict[str, float | int], np.ndarray]:
+    dets = np.ascontiguousarray(dets_and_obs[:, :-num_obs], dtype=np.uint8)
+    obs = np.ascontiguousarray(dets_and_obs[:, -num_obs:], dtype=np.uint8)
+    pred = matcher.decode_batch(dets).reshape(obs.shape)
+    error_mask = np.asarray(pred != obs, dtype=np.bool_).reshape(obs.shape[0], -1).any(axis=1)
+    errors = int(error_mask.sum())
+    total = int(obs.shape[0])
+    latency_rows = dets[: min(latency_samples, len(dets))]
+    input_density = SyndromeDensityAccumulator()
+    input_density.update(dets)
+    return {
+        "logical_errors": errors,
+        "samples": total,
+        "ler": float(errors / total) if total else float("nan"),
+        "latency_us_per_round": time_single_shot(matcher, latency_rows, n_rounds),
+        **input_density.statistics("input"),
+    }, error_mask
+
+
+def evaluate_model(
+    model: torch.nn.Module,
+    cfg: Any,
+    dataset,
+    matcher: pymatching.Matching,
+    num_obs: int,
+    device: torch.device,
+    latency_samples: int,
+    n_rounds: int,
+    residual_tensor_path: Path | None = None,
+) -> tuple[dict[str, Any], np.ndarray]:
+    maps = _build_stab_maps(int(cfg.distance), getattr(cfg, "rotation", "XV"))
+    module = PreDecoderMemoryEvalModule(model, cfg, maps, device).to(device)
+    module.eval()
+    loader = DataLoader(
+        dataset,
+        batch_size=int(cfg.test.batch_size),
+        shuffle=False,
+        num_workers=int(cfg.test.dataloader_num_workers),
+        pin_memory=(device.type == "cuda"),
+    )
+
+    logical_errors = 0
+    total = 0
+    residual_chunks: list[np.ndarray] = []
+    saved_residual_chunks: list[np.ndarray] = []
+    error_chunks: list[np.ndarray] = []
+    residual_count = 0
+    input_density = SyndromeDensityAccumulator()
+    residual_density = SyndromeDensityAccumulator()
+
+    with torch.no_grad():
+        for batch in loader:
+            batch = dict_to_device(batch, device)
+            dets_and_obs = batch["dets_and_obs"]
+            dets_only = dets_and_obs[:, :-num_obs]
+            gt_obs = dets_and_obs[:, -num_obs:].to(torch.int64).cpu()
+
+            output = module(dets_only)
+            pre_l = output[:, 0].to(torch.int64).cpu()
+            residual = output[:, 1:].to(torch.uint8).cpu().numpy()
+            input_density.update(dets_only.to(torch.uint8).cpu().numpy())
+            residual_density.update(residual)
+            if residual_tensor_path is not None:
+                saved_residual_chunks.append(
+                    np.ascontiguousarray(residual, dtype=np.uint8)
+                )
+            pred_obs = torch.from_numpy(matcher.decode_batch(residual)).reshape(gt_obs.shape)
+            final_l = (pre_l.reshape(gt_obs.shape) + pred_obs).remainder(2)
+
+            error_mask = (final_l != gt_obs).reshape(gt_obs.shape[0], -1).any(dim=1)
+            logical_errors += int(error_mask.sum().item())
+            total += int(gt_obs.shape[0])
+            error_chunks.append(error_mask.numpy())
+
+            if residual_count < latency_samples:
+                take = min(latency_samples - residual_count, residual.shape[0])
+                residual_chunks.append(np.ascontiguousarray(residual[:take], dtype=np.uint8))
+                residual_count += take
+
+    residual_rows = (
+        np.concatenate(residual_chunks, axis=0) if residual_chunks else np.empty((0, 0), dtype=np.uint8)
+    )
+    all_errors = np.concatenate(error_chunks) if error_chunks else np.empty(0, dtype=np.bool_)
+    result: dict[str, Any] = {
+        "logical_errors": logical_errors,
+        "samples": total,
+        "ler": float(logical_errors / total) if total else float("nan"),
+        "latency_us_per_round": time_single_shot(matcher, residual_rows, n_rounds),
+        **model_density_statistics(input_density, residual_density),
+    }
+    if residual_tensor_path is not None:
+        residual_tensor_path.parent.mkdir(parents=True, exist_ok=True)
+        saved_residual = (
+            np.concatenate(saved_residual_chunks, axis=0)
+            if saved_residual_chunks
+            else np.empty((0, 0), dtype=np.uint8)
+        )
+        torch.save(torch.from_numpy(saved_residual), residual_tensor_path)
+        result.update(
+            residual_tensor_path=str(residual_tensor_path),
+            residual_tensor_rows=int(saved_residual.shape[0]),
+            residual_tensor_detectors=int(saved_residual.shape[1]),
+            residual_tensor_dtype="torch.uint8",
+        )
+    return result, all_errors
+
+
+def build_paired_comparison_rows(
+    error_masks_by_basis: dict[str, dict[str, np.ndarray]],
+    comparisons: list[ComparisonSpec],
+) -> list[dict[str, Any]]:
+    rows: list[dict[str, Any]] = []
+    basis_order = [basis for basis in ("X", "Z") if basis in error_masks_by_basis]
+    for comparison in comparisons:
+        candidate_chunks = []
+        baseline_chunks = []
+        for basis in basis_order:
+            masks = error_masks_by_basis[basis]
+            missing = {
+                method
+                for method in (comparison.candidate, comparison.baseline)
+                if method not in masks
+            }
+            if missing:
+                raise KeyError(f"paired comparison methods missing for {basis}: {sorted(missing)}")
+            candidate = masks[comparison.candidate]
+            baseline = masks[comparison.baseline]
+            rows.append(
+                {
+                    "basis": basis,
+                    "candidate": comparison.candidate,
+                    "baseline": comparison.baseline,
+                    **paired_error_statistics(candidate, baseline),
+                }
+            )
+            candidate_chunks.append(candidate)
+            baseline_chunks.append(baseline)
+        if len(basis_order) > 1:
+            rows.append(
+                {
+                    "basis": "both",
+                    "candidate": comparison.candidate,
+                    "baseline": comparison.baseline,
+                    **paired_error_statistics(
+                        np.concatenate(candidate_chunks),
+                        np.concatenate(baseline_chunks),
+                    ),
+                }
+            )
+    return rows
+
+
+def build_factorial_contrast_rows(
+    error_masks_by_basis: dict[str, dict[str, np.ndarray]],
+    contrasts: list[FactorialContrastSpec],
+) -> list[dict[str, Any]]:
+    rows: list[dict[str, Any]] = []
+    basis_order = [basis for basis in ("X", "Z") if basis in error_masks_by_basis]
+    for contrast in contrasts:
+        chunks = {field: [] for field in ("cell_11", "cell_10", "cell_01", "cell_00")}
+        for basis in basis_order:
+            masks = error_masks_by_basis[basis]
+            methods = {
+                field: getattr(contrast, field)
+                for field in ("cell_11", "cell_10", "cell_01", "cell_00")
+            }
+            missing = set(methods.values()) - set(masks)
+            if missing:
+                raise KeyError(f"factorial contrast methods missing for {basis}: {sorted(missing)}")
+            stats = factorial_contrast_statistics(*(masks[methods[field]] for field in chunks))
+            rows.append(
+                {
+                    "basis": basis,
+                    "name": contrast.name,
+                    **methods,
+                    **stats,
+                }
+            )
+            for field, method in methods.items():
+                chunks[field].append(masks[method])
+        if len(basis_order) > 1:
+            rows.append(
+                {
+                    "basis": "both",
+                    "name": contrast.name,
+                    "cell_11": contrast.cell_11,
+                    "cell_10": contrast.cell_10,
+                    "cell_01": contrast.cell_01,
+                    "cell_00": contrast.cell_00,
+                    **factorial_contrast_statistics(
+                        *(np.concatenate(chunks[field]) for field in chunks)
+                    ),
+                }
+            )
+    return rows
+
+
+def mean_metric(rows: list[dict[str, Any]], name: str) -> float:
+    values = [float(row[name]) for row in rows if row.get(name) is not None]
+    return float(np.mean(values)) if values else float("nan")
+
+
+def main() -> None:
+    args = parse_args()
+    stim_samples_dir = resolve_stim_samples_dir(args)
+    if stim_samples_dir is not None:
+        # DatapipeFactory historically gives the environment variable priority.
+        # Synchronize it so an explicit CLI path cannot be silently shadowed.
+        os.environ["PREDECODER_STIM_SAMPLES_DIR"] = str(stim_samples_dir)
+    output_path = Path(args.output)
+    if not output_path.is_absolute():
+        output_path = REPO_ROOT / output_path
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    residual_output_dir = (
+        Path(args.residual_output_dir) if args.residual_output_dir else None
+    )
+    if residual_output_dir is not None and not residual_output_dir.is_absolute():
+        residual_output_dir = REPO_ROOT / residual_output_dir
+
+    for spec in args.model:
+        if not spec.checkpoint.exists():
+            raise FileNotFoundError(f"Checkpoint not found for {spec.name}: {spec.checkpoint}")
+    available_methods = {"pymatching", *(spec.name for spec in args.model)}
+    for comparison in args.paired_comparison:
+        missing = {comparison.candidate, comparison.baseline} - available_methods
+        if missing:
+            raise ValueError(f"Unknown paired comparison methods: {sorted(missing)}")
+    for contrast in args.factorial_contrast:
+        missing = {
+            contrast.cell_11,
+            contrast.cell_10,
+            contrast.cell_01,
+            contrast.cell_00,
+        } - available_methods
+        if missing:
+            raise ValueError(f"Unknown factorial contrast methods: {sorted(missing)}")
+
+    device = torch.device(args.device or ("cuda:0" if torch.cuda.is_available() else "cpu"))
+    dist = SimpleNamespace(rank=0, world_size=1, device=device)
+    bases = ["X", "Z"] if args.basis == "both" else [args.basis]
+
+    model_cfgs = {spec.name: build_cfg(args, spec, basis=bases[0]) for spec in args.model}
+    models = {}
+    for spec in args.model:
+        print(f"[load] {spec.name}: model_id={spec.model_id}, checkpoint={spec.checkpoint}")
+        model = load_model_checkpoint(
+            model_cfgs[spec.name],
+            checkpoint=spec.checkpoint,
+            model_id=spec.model_id,
+            distributed=dist,
+        )
+        model.eval()
+        models[spec.name] = model
+
+    rows: list[dict[str, Any]] = []
+    error_masks_by_basis: dict[str, dict[str, np.ndarray]] = {}
+    paired_comparisons: list[dict[str, Any]] = []
+    sample_metadata: dict[str, Any] = {}
+    for basis_index, basis in enumerate(bases):
+        dataset_cfg = build_cfg(args, args.model[0], basis=basis)
+        dataset_seed = int(args.seed) + basis_index
+        print(f"[data] basis={basis}, seed={dataset_seed}, samples={args.num_samples}")
+        dataset = make_dataset(dataset_cfg, dataset_seed)
+        if hasattr(dataset, "metadata"):
+            sample_metadata[basis] = dict(dataset.metadata)
+        matcher, num_obs = build_matcher(dataset)
+        dets_and_obs = np.asarray(dataset.dets_and_obs, dtype=np.uint8)
+
+        baseline, baseline_errors = evaluate_pymatching(
+            matcher,
+            dets_and_obs,
+            num_obs,
+            int(args.latency_num_samples),
+            int(args.n_rounds),
+        )
+        error_masks_by_basis[basis] = {"pymatching": baseline_errors}
+        baseline_row = {
+            "basis": basis,
+            "method": "pymatching",
+            "model_id": "",
+            "checkpoint": "",
+            **baseline,
+            "speedup_vs_pymatching": 1.0,
+        }
+        rows.append(baseline_row)
+        basis_errors = {"pymatching": baseline_errors}
+        print(
+            f"[result] {basis} pymatching ler={baseline['ler']:.6f}, "
+            f"latency={baseline['latency_us_per_round']:.3f} us/round"
+        )
+
+        for spec in args.model:
+            cfg = build_cfg(args, spec, basis=basis)
+            residual_tensor_path = (
+                residual_output_dir / f"{basis}_{spec.name}_residual_detectors.pt"
+                if residual_output_dir is not None
+                else None
+            )
+            result, model_errors = evaluate_model(
+                models[spec.name],
+                cfg,
+                dataset,
+                matcher,
+                num_obs,
+                device,
+                int(args.latency_num_samples),
+                int(args.n_rounds),
+                residual_tensor_path,
+            )
+            error_masks_by_basis[basis][spec.name] = model_errors
+            speedup = float(baseline["latency_us_per_round"]) / float(result["latency_us_per_round"])
+            row = {
+                "basis": basis,
+                "method": spec.name,
+                "model_id": spec.model_id,
+                "checkpoint": str(spec.checkpoint),
+                **result,
+                "speedup_vs_pymatching": speedup,
+            }
+            rows.append(row)
+            basis_errors[spec.name] = model_errors
+            print(
+                f"[result] {basis} {spec.name} ler={result['ler']:.6f}, "
+                f"latency={result['latency_us_per_round']:.3f} us/round, speedup={speedup:.3f}x"
+            )
+        for method_a, method_b in combinations(basis_errors, 2):
+            paired_comparisons.append(
+                paired_error_comparison(
+                    method_a,
+                    basis_errors[method_a],
+                    method_b,
+                    basis_errors[method_b],
+                    basis=basis,
+                )
+            )
+
+    if args.paired_comparison:
+        paired_comparisons = build_paired_comparison_rows(
+            error_masks_by_basis,
+            args.paired_comparison,
+        )
+    factorial_contrasts = build_factorial_contrast_rows(
+        error_masks_by_basis,
+        args.factorial_contrast,
+    )
+    methods = sorted({row["method"] for row in rows})
+    summary = []
+    for method in methods:
+        method_rows = [row for row in rows if row["method"] == method]
+        summary_row: dict[str, Any] = {
+            "method": method,
+            "ler_avg": mean_metric(method_rows, "ler"),
+            "latency_us_per_round_avg": mean_metric(method_rows, "latency_us_per_round"),
+            "speedup_vs_pymatching_avg": mean_metric(method_rows, "speedup_vs_pymatching"),
+            **combine_density_statistics(method_rows, "input"),
+        }
+        if any(row.get("residual_syndrome_elements") for row in method_rows):
+            residual_stats = combine_density_statistics(method_rows, "residual")
+            summary_row.update(residual_stats)
+            summary_row.update(
+                density_reduction_statistics(
+                    float(summary_row["input_syndrome_density"]),
+                    float(residual_stats["residual_syndrome_density"]),
+                )
+            )
+        summary.append(summary_row)
+
+    payload = {
+        "config_name": args.config_name,
+        "distance": args.distance,
+        "n_rounds": args.n_rounds,
+        "num_samples": args.num_samples,
+        "latency_num_samples": args.latency_num_samples,
+        "seed": args.seed,
+        "device": str(device),
+        "sample_source": "stim_files" if stim_samples_dir else "generated",
+        "stim_samples_dir": str(stim_samples_dir) if stim_samples_dir else None,
+        "sample_metadata": sample_metadata,
+        "rows": rows,
+        "summary": summary,
+        "paired_comparisons": paired_comparisons,
+        "factorial_contrasts": factorial_contrasts,
+    }
+    output_path.write_text(json.dumps(payload, indent=2), encoding="utf-8")
+
+    csv_path = output_path.with_suffix(".csv")
+    fieldnames = [
+        "basis",
+        "method",
+        "model_id",
+        "logical_errors",
+        "samples",
+        "ler",
+        "latency_us_per_round",
+        "speedup_vs_pymatching",
+        "input_density_shots",
+        "input_syndrome_ones",
+        "input_syndrome_elements",
+        "input_density_shot_sum",
+        "input_density_shot_sum_squares",
+        "input_syndrome_density",
+        "input_density_standard_error",
+        "input_density_ci95_low",
+        "input_density_ci95_high",
+        "residual_density_shots",
+        "residual_syndrome_ones",
+        "residual_syndrome_elements",
+        "residual_density_shot_sum",
+        "residual_density_shot_sum_squares",
+        "residual_syndrome_density",
+        "residual_density_standard_error",
+        "residual_density_ci95_low",
+        "residual_density_ci95_high",
+        "density_delta",
+        "density_reduction_fraction",
+        "density_reduction_factor",
+        "residual_tensor_path",
+        "residual_tensor_rows",
+        "residual_tensor_detectors",
+        "residual_tensor_dtype",
+        "checkpoint",
+    ]
+    with csv_path.open("w", newline="", encoding="utf-8") as f:
+        writer = csv.DictWriter(f, fieldnames=fieldnames)
+        writer.writeheader()
+        for row in rows:
+            writer.writerow({field: row.get(field, "") for field in fieldnames})
+
+    print(f"[write] {output_path}")
+    print(f"[write] {csv_path}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/code/scripts/providers/__init__.py b/code/scripts/providers/__init__.py
new file mode 100644
index 0000000..f89f0e8
--- /dev/null
+++ b/code/scripts/providers/__init__.py
@@ -0,0 +1,4 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""External benchmark and circuit-data command modules."""
diff --git a/code/scripts/providers/google_qec_decoder_benchmark.py b/code/scripts/providers/google_qec_decoder_benchmark.py
new file mode 100644
index 0000000..2a48c01
--- /dev/null
+++ b/code/scripts/providers/google_qec_decoder_benchmark.py
@@ -0,0 +1,1390 @@
+#!/usr/bin/env python3
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Benchmark PyMatching and released pre-decoders on Google Willow QEC data."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import math
+import sys
+import time
+from dataclasses import asdict, dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from types import SimpleNamespace
+from typing import Any, Iterable, Mapping, Sequence
+
+import numpy as np
+import pymatching
+import stim
+import torch
+from omegaconf import OmegaConf
+
+CODE_ROOT = Path(__file__).resolve().parents[2]
+REPO_ROOT = CODE_ROOT.parent
+if str(CODE_ROOT) not in sys.path:
+    sys.path.insert(0, str(CODE_ROOT))
+
+from evaluation.logical_error_rate import (  # noqa: E402
+    PreDecoderMemoryEvalModule,
+    _build_stab_maps,
+)
+from qec.surface_code.memory_circuit import SurfaceCode  # noqa: E402
+from scripts.config_paths import config_path  # noqa: E402
+from scripts.paired_inference_compare import (  # noqa: E402
+    SyndromeDensityAccumulator,
+    model_density_statistics,
+)
+from workflows.config_validator import (  # noqa: E402
+    apply_public_defaults_and_model,
+    validate_public_config,
+)
+from model.checkpoint_loader import load_model_checkpoint  # noqa: E402
+
+
+DEFAULT_BENCHMARK_ROOT = (
+    REPO_ROOT / "benchmarks/google_qec/google_105Q_surface_code_d3_d5_d7"
+)
+
+
+@dataclass(frozen=True)
+class BenchmarkModel:
+    name: str
+    model_id: int
+    checkpoint: Path
+
+
+# The public wrapper injects explicitly named checkpoint paths before parsing.
+# Keep the backend free of internal training-output defaults.
+DEFAULT_MODELS: dict[str, BenchmarkModel] = {}
+
+
+def maybe_compile_model(
+    model: torch.nn.Module,
+    *,
+    enabled: bool,
+    mode: str = "default",
+) -> torch.nn.Module:
+    """Optionally compile one cached model with dynamic detector dimensions."""
+
+    return torch.compile(model, mode=mode, dynamic=True) if enabled else model
+
+
+@dataclass(frozen=True)
+class GoogleQECCase:
+    path: Path
+    patch: str
+    distance: int
+    basis: str
+    rounds: int
+    shots: int
+
+
+REQUIRED_CASE_FILES = (
+    "circuit_ideal.stim",
+    "circuit_noisy_si1000.stim",
+    "detection_events.b8",
+    "obs_flips_actual.b8",
+)
+
+
+def discover_cases(
+    root: Path,
+    *,
+    distances: set[int] | None = None,
+    rounds: set[int] | None = None,
+    bases: set[str] | None = None,
+    patches: set[str] | None = None,
+) -> list[GoogleQECCase]:
+    """Discover complete Google benchmark cases selected by metadata."""
+
+    root = Path(root)
+    cases = []
+    for metadata_path in root.glob("d*_at_q*/[XZ]/r*/metadata.json"):
+        metadata = json.loads(metadata_path.read_text())
+        case_dir = metadata_path.parent
+        patch = case_dir.parents[1].name
+        distance = int(metadata["distance"])
+        basis = str(metadata["basis"]).upper()
+        n_rounds = int(metadata["rounds"])
+        if distances is not None and distance not in distances:
+            continue
+        if rounds is not None and n_rounds not in rounds:
+            continue
+        if bases is not None and basis not in bases:
+            continue
+        if patches is not None and patch not in patches:
+            continue
+        missing = [name for name in REQUIRED_CASE_FILES if not (case_dir / name).is_file()]
+        if missing:
+            raise FileNotFoundError(f"Incomplete Google QEC case {case_dir}: missing {missing}")
+        cases.append(
+            GoogleQECCase(
+                path=case_dir,
+                patch=patch,
+                distance=distance,
+                basis=basis,
+                rounds=n_rounds,
+                shots=int(metadata["shots"]),
+            )
+        )
+    return sorted(cases, key=lambda case: (case.distance, case.patch, case.basis, case.rounds))
+
+
+def _google_to_xv_coordinate(
+    coordinate: Sequence[float],
+    *,
+    min_difference: int,
+    min_sum: int,
+) -> tuple[int, int]:
+    if len(coordinate) < 2:
+        raise ValueError(f"Google coordinate must contain x and y, got {coordinate!r}")
+    x = float(coordinate[0])
+    y = float(coordinate[1])
+    if not x.is_integer() or not y.is_integer():
+        raise ValueError(f"Google coordinate must be integral, got {coordinate!r}")
+    x_int = int(x)
+    y_int = int(y)
+    return (
+        x_int - y_int - int(min_difference) + 1,
+        x_int + y_int - int(min_sum) + 1,
+    )
+
+
+def build_detector_permutation(
+    circuit: stim.Circuit,
+    metadata: Mapping[str, Any],
+) -> np.ndarray:
+    """Return indices that map Google detector columns to the model's XV order.
+
+    Google emits each bulk round in physical measurement-qubit order. The
+    predecoder consumes initial-boundary, X-block, Z-block, ..., final-boundary
+    order, with stabilizers indexed by the repository's XV patch convention.
+    """
+
+    distance = int(metadata["distance"])
+    rounds = int(metadata["rounds"])
+    basis = str(metadata["basis"]).upper()
+    if basis not in {"X", "Z"}:
+        raise ValueError(f"basis must be X or Z, got {basis!r}")
+    if distance < 3 or distance % 2 == 0:
+        raise ValueError(f"distance must be an odd integer >= 3, got {distance}")
+    if rounds < 1:
+        raise ValueError(f"rounds must be positive, got {rounds}")
+
+    half = (distance * distance - 1) // 2
+    expected_detectors = 2 * rounds * half
+    if int(circuit.num_detectors) != expected_detectors:
+        raise ValueError(
+            "detector count mismatch: "
+            f"circuit has {circuit.num_detectors}, expected {expected_detectors} "
+            f"for d={distance}, rounds={rounds}"
+        )
+
+    data_coordinates = [tuple(item) for item in metadata["data_qubit_coords"]]
+    if len(data_coordinates) != distance * distance:
+        raise ValueError(
+            f"data coordinate count mismatch: {len(data_coordinates)} != {distance * distance}"
+        )
+    min_difference = min(int(x) - int(y) for x, y in data_coordinates)
+    min_sum = min(int(x) + int(y) for x, y in data_coordinates)
+    transformed_data = {
+        _google_to_xv_coordinate(
+            coordinate,
+            min_difference=min_difference,
+            min_sum=min_sum,
+        )
+        for coordinate in data_coordinates
+    }
+    odd_coordinates = range(1, 2 * distance, 2)
+    expected_data = {(x, y) for x in odd_coordinates for y in odd_coordinates}
+    if transformed_data != expected_data:
+        raise ValueError("Google data-qubit coordinates do not form the expected rotated patch")
+
+    code = SurfaceCode(distance, first_bulk_syndrome_type="X", rotated_type="V")
+    x_indices = {
+        tuple(map(int, code.xcheck_qubits_dict[int(qubit)]["coord"])): index
+        for index, qubit in enumerate(code.xcheck_qubits)
+    }
+    z_indices = {
+        tuple(map(int, code.zcheck_qubits_dict[int(qubit)]["coord"])): index
+        for index, qubit in enumerate(code.zcheck_qubits)
+    }
+    detector_coordinates = circuit.get_detector_coordinates()
+    if len(detector_coordinates) != expected_detectors:
+        raise ValueError(
+            "detector coordinate count mismatch: "
+            f"{len(detector_coordinates)} != {expected_detectors}"
+        )
+
+    canonical_to_source = np.full(expected_detectors, -1, dtype=np.int64)
+    boundary_start = expected_detectors - half
+    for source_index in range(expected_detectors):
+        raw_coordinate = detector_coordinates[source_index]
+        if len(raw_coordinate) < 3:
+            raise ValueError(f"detector {source_index} has no spatial/time coordinate")
+        # Initial and bulk detectors end in their stabilizer coordinate. Google
+        # final-boundary detectors list data coordinates first and the previous
+        # ancilla/stabilizer coordinate last, so the last coordinate triple is
+        # the uniform choice for every phase.
+        model_coordinate = _google_to_xv_coordinate(
+            raw_coordinate[-3:-1],
+            min_difference=min_difference,
+            min_sum=min_sum,
+        )
+        if model_coordinate in x_indices:
+            stabilizer_type = "X"
+            stabilizer_index = x_indices[model_coordinate]
+        elif model_coordinate in z_indices:
+            stabilizer_type = "Z"
+            stabilizer_index = z_indices[model_coordinate]
+        else:
+            raise ValueError(
+                f"detector {source_index} coordinate {raw_coordinate!r} maps to "
+                f"unknown XV stabilizer {model_coordinate}"
+            )
+
+        if source_index < half:
+            if stabilizer_type != basis:
+                raise ValueError(
+                    f"initial detector {source_index} is {stabilizer_type}, expected {basis}"
+                )
+            canonical_index = stabilizer_index
+        elif source_index >= boundary_start:
+            if stabilizer_type != basis:
+                raise ValueError(
+                    f"boundary detector {source_index} is {stabilizer_type}, expected {basis}"
+                )
+            canonical_index = boundary_start + stabilizer_index
+        else:
+            bulk_offset = source_index - half
+            bulk_round = bulk_offset // (2 * half)
+            type_offset = 0 if stabilizer_type == "X" else half
+            canonical_index = half + bulk_round * 2 * half + type_offset + stabilizer_index
+
+        if canonical_to_source[canonical_index] != -1:
+            raise ValueError(
+                f"duplicate detector mapping for canonical index {canonical_index}"
+            )
+        canonical_to_source[canonical_index] = source_index
+
+    if np.any(canonical_to_source < 0):
+        missing = np.flatnonzero(canonical_to_source < 0).tolist()
+        raise ValueError(f"incomplete detector mapping; missing canonical indices {missing}")
+    return canonical_to_source
+
+
+def google_to_canonical(data: np.ndarray, canonical_to_source: np.ndarray) -> np.ndarray:
+    rows = np.asarray(data)
+    permutation = np.asarray(canonical_to_source, dtype=np.int64)
+    if rows.ndim != 2 or rows.shape[1] != permutation.size:
+        raise ValueError(
+            f"Google detector shape {rows.shape} is incompatible with permutation "
+            f"width {permutation.size}"
+        )
+    return np.ascontiguousarray(rows[:, permutation])
+
+
+def canonical_to_google(data: np.ndarray, canonical_to_source: np.ndarray) -> np.ndarray:
+    rows = np.asarray(data)
+    permutation = np.asarray(canonical_to_source, dtype=np.int64)
+    if rows.ndim != 2 or rows.shape[1] != permutation.size:
+        raise ValueError(
+            f"canonical detector shape {rows.shape} is incompatible with permutation "
+            f"width {permutation.size}"
+        )
+    restored = np.empty_like(rows)
+    restored[:, permutation] = rows
+    return np.ascontiguousarray(restored)
+
+
+def verify_bulk_data_fault_equivalence(
+    circuit: stim.Circuit,
+    metadata: Mapping[str, Any],
+) -> dict[str, Any]:
+    """Compare all inter-cycle physical X/Y/Z faults with CSS signatures."""
+
+    distance = int(metadata["distance"])
+    basis = str(metadata["basis"]).upper()
+    if basis not in {"X", "Z"}:
+        raise ValueError(f"basis must be X or Z, got {basis!r}")
+
+    data_coordinates = [tuple(map(int, item)) for item in metadata["data_qubit_coords"]]
+    if len(data_coordinates) != distance * distance:
+        raise ValueError(
+            f"data coordinate count mismatch: {len(data_coordinates)} != {distance * distance}"
+        )
+    qubit_coordinates = {
+        int(qubit): tuple(map(int, coordinate))
+        for qubit, coordinate in circuit.get_final_qubit_coordinates().items()
+    }
+    coordinate_to_qubit = {coordinate: qubit for qubit, coordinate in qubit_coordinates.items()}
+    missing_qubits = [coordinate for coordinate in data_coordinates if coordinate not in coordinate_to_qubit]
+    if missing_qubits:
+        raise ValueError(f"data coordinates missing from circuit: {missing_qubits}")
+    data_qubits = {coordinate_to_qubit[coordinate] for coordinate in data_coordinates}
+
+    cycle_boundaries = []
+    for instruction_index in range(len(circuit)):
+        instruction = circuit[instruction_index]
+        if instruction.name != "Y":
+            continue
+        targets = {
+            int(target.value)
+            for target in instruction.targets_copy()
+            if target.is_qubit_target
+        }
+        if targets == data_qubits:
+            cycle_boundaries.append(instruction_index)
+    expected_boundaries = int(metadata["rounds"]) - 1
+    if len(cycle_boundaries) != expected_boundaries:
+        raise ValueError(
+            "inter-cycle boundary count mismatch: "
+            f"{len(cycle_boundaries)} != {expected_boundaries}"
+        )
+
+    permutation = build_detector_permutation(circuit, metadata)
+    maps = _build_stab_maps(distance, "XV")
+    hx = maps["Hx_i32"].to(torch.uint8).cpu().numpy()
+    hz = maps["Hz_i32"].to(torch.uint8).cpu().numpy()
+    half = (distance * distance - 1) // 2
+    min_difference = min(x - y for x, y in data_coordinates)
+    min_sum = min(x + y for x, y in data_coordinates)
+    mismatches = []
+    error_names = {"X": "X_ERROR", "Y": "Y_ERROR", "Z": "Z_ERROR"}
+
+    for pair_index, boundary_index in enumerate(cycle_boundaries):
+        insertion_index = boundary_index + 1
+        pair_start = half + pair_index * 2 * half
+        for coordinate in data_coordinates:
+            qubit = coordinate_to_qubit[coordinate]
+            model_x, model_y = _google_to_xv_coordinate(
+                coordinate,
+                min_difference=min_difference,
+                min_sum=min_sum,
+            )
+            row = (model_x - 1) // 2
+            column = (model_y - 1) // 2
+            data_index = row * distance + column
+            has_local_hadamard = (row + column) % 2 == 1
+
+            for physical_pauli, error_name in error_names.items():
+                if physical_pauli == "Y":
+                    css_components = {"x", "z"}
+                elif physical_pauli == "X":
+                    css_components = {"z" if has_local_hadamard else "x"}
+                else:
+                    css_components = {"x" if has_local_hadamard else "z"}
+
+                faulty = circuit[:insertion_index]
+                faulty.append(error_name, [qubit], 1.0)
+                faulty += circuit[insertion_index:]
+                google_detectors, observables = faulty.compile_detector_sampler().sample(
+                    shots=1,
+                    separate_observables=True,
+                )
+                actual_detectors = google_to_canonical(
+                    np.asarray(google_detectors, dtype=np.uint8),
+                    permutation,
+                )[0]
+                actual_observable = int(np.asarray(observables, dtype=np.uint8)[0, 0])
+
+                expected_detectors = np.zeros(int(circuit.num_detectors), dtype=np.uint8)
+                if "z" in css_components:
+                    expected_detectors[pair_start : pair_start + half] ^= hx[:, data_index]
+                if "x" in css_components:
+                    expected_detectors[pair_start + half : pair_start + 2 * half] ^= hz[:, data_index]
+                expected_observable = int(
+                    (basis == "X" and "z" in css_components and row == 0)
+                    or (basis == "Z" and "x" in css_components and column == 0)
+                )
+                if not np.array_equal(actual_detectors, expected_detectors) or (
+                    actual_observable != expected_observable
+                ):
+                    mismatches.append(
+                        {
+                            "bulk_pair_index": pair_index,
+                            "coordinate": list(coordinate),
+                            "qubit": qubit,
+                            "physical_pauli": physical_pauli,
+                            "local_hadamard": has_local_hadamard,
+                            "css_components": sorted(css_components),
+                            "actual_detector_indices": np.flatnonzero(actual_detectors).tolist(),
+                            "expected_detector_indices": np.flatnonzero(expected_detectors).tolist(),
+                            "actual_observable": actual_observable,
+                            "expected_observable": expected_observable,
+                        }
+                    )
+
+    return {
+        "distance": distance,
+        "basis": basis,
+        "bulk_pair_indices": list(range(len(cycle_boundaries))),
+        "faults_checked": 3 * len(data_coordinates) * len(cycle_boundaries),
+        "mismatches": mismatches,
+    }
+
+
+
+def verify_final_data_fault_equivalence(
+    circuit: stim.Circuit,
+    metadata: Mapping[str, Any],
+) -> dict[str, Any]:
+    """Compare Google final-measurement fault signatures with CSS-frame signatures.
+
+    An X immediately before the final data-qubit measurement flips exactly one
+    physical measurement result.  For every data qubit this checks that the
+    resulting Google detector/observable signature, after canonicalization,
+    equals the CSS parity-check column and logical-string parity used by the
+    predecoder.
+    """
+
+    distance = int(metadata["distance"])
+    basis = str(metadata["basis"]).upper()
+    if basis not in {"X", "Z"}:
+        raise ValueError(f"basis must be X or Z, got {basis!r}")
+
+    data_coordinates = [tuple(map(int, item)) for item in metadata["data_qubit_coords"]]
+    if len(data_coordinates) != distance * distance:
+        raise ValueError(
+            f"data coordinate count mismatch: {len(data_coordinates)} != {distance * distance}"
+        )
+    qubit_coordinates = {
+        int(qubit): tuple(map(int, coordinate))
+        for qubit, coordinate in circuit.get_final_qubit_coordinates().items()
+    }
+    coordinate_to_qubit = {coordinate: qubit for qubit, coordinate in qubit_coordinates.items()}
+    missing_qubits = [coordinate for coordinate in data_coordinates if coordinate not in coordinate_to_qubit]
+    if missing_qubits:
+        raise ValueError(f"data coordinates missing from circuit: {missing_qubits}")
+    data_qubits = {coordinate_to_qubit[coordinate] for coordinate in data_coordinates}
+
+    final_measurement_index = None
+    for instruction_index in range(len(circuit) - 1, -1, -1):
+        instruction = circuit[instruction_index]
+        if instruction.name not in {"M", "MX", "MY"}:
+            continue
+        measured_qubits = {
+            int(target.value)
+            for target in instruction.targets_copy()
+            if target.is_qubit_target
+        }
+        if measured_qubits == data_qubits:
+            final_measurement_index = instruction_index
+            break
+    if final_measurement_index is None:
+        raise ValueError("could not find the final all-data-qubit measurement")
+
+    permutation = build_detector_permutation(circuit, metadata)
+    maps = _build_stab_maps(distance, "XV")
+    parity_matrix = (
+        maps["Hx_i32"] if basis == "X" else maps["Hz_i32"]
+    ).to(torch.uint8).cpu().numpy()
+    half = (distance * distance - 1) // 2
+    boundary_start = int(circuit.num_detectors) - half
+    min_difference = min(x - y for x, y in data_coordinates)
+    min_sum = min(x + y for x, y in data_coordinates)
+    mismatches = []
+
+    for coordinate in data_coordinates:
+        qubit = coordinate_to_qubit[coordinate]
+        model_x, model_y = _google_to_xv_coordinate(
+            coordinate,
+            min_difference=min_difference,
+            min_sum=min_sum,
+        )
+        row = (model_x - 1) // 2
+        column = (model_y - 1) // 2
+        data_index = row * distance + column
+
+        faulty = circuit[:final_measurement_index]
+        faulty.append("X_ERROR", [qubit], 1.0)
+        faulty += circuit[final_measurement_index:]
+        google_detectors, observables = faulty.compile_detector_sampler().sample(
+            shots=1,
+            separate_observables=True,
+        )
+        actual_detectors = google_to_canonical(
+            np.asarray(google_detectors, dtype=np.uint8),
+            permutation,
+        )[0]
+        actual_observable = int(np.asarray(observables, dtype=np.uint8)[0, 0])
+
+        expected_detectors = np.zeros(int(circuit.num_detectors), dtype=np.uint8)
+        expected_detectors[boundary_start:] = parity_matrix[:, data_index] % 2
+        expected_observable = int(row == 0) if basis == "X" else int(column == 0)
+        if not np.array_equal(actual_detectors, expected_detectors) or (
+            actual_observable != expected_observable
+        ):
+            mismatches.append(
+                {
+                    "coordinate": list(coordinate),
+                    "qubit": qubit,
+                    "model_data_index": data_index,
+                    "actual_detector_indices": np.flatnonzero(actual_detectors).tolist(),
+                    "expected_detector_indices": np.flatnonzero(expected_detectors).tolist(),
+                    "actual_observable": actual_observable,
+                    "expected_observable": expected_observable,
+                }
+            )
+
+    return {
+        "distance": distance,
+        "basis": basis,
+        "faults_checked": len(data_coordinates),
+        "mismatches": mismatches,
+    }
+
+
+def wilson_interval(errors: int, shots: int, z: float = 1.96) -> tuple[float, float]:
+    if shots <= 0:
+        return float("nan"), float("nan")
+    p = float(errors) / float(shots)
+    denominator = 1.0 + z * z / shots
+    center = (p + z * z / (2.0 * shots)) / denominator
+    half_width = (
+        z
+        * math.sqrt((p * (1.0 - p) + z * z / (4.0 * shots)) / shots)
+        / denominator
+    )
+    return max(0.0, center - half_width), min(1.0, center + half_width)
+
+
+def paired_error_counts(
+    candidate_errors: np.ndarray,
+    baseline_errors: np.ndarray,
+) -> dict[str, int | float]:
+    candidate = np.asarray(candidate_errors, dtype=np.bool_).reshape(-1)
+    baseline = np.asarray(baseline_errors, dtype=np.bool_).reshape(-1)
+    if candidate.shape != baseline.shape:
+        raise ValueError(
+            f"paired error shape mismatch: {candidate.shape} != {baseline.shape}"
+        )
+    candidate_only = int(np.count_nonzero(candidate & ~baseline))
+    baseline_only = int(np.count_nonzero(~candidate & baseline))
+    both = int(np.count_nonzero(candidate & baseline))
+    neither = int(candidate.size - candidate_only - baseline_only - both)
+    result = _paired_statistics_from_counts(
+        samples=int(candidate.size),
+        candidate_only=candidate_only,
+        baseline_only=baseline_only,
+        both=both,
+        neither=neither,
+    )
+    # Kept for backward compatibility with existing candidate-vs-PyMatching rows.
+    result["delta_ler_vs_pymatching"] = result["delta_ler"]
+    return result
+
+
+def _paired_statistics_from_counts(
+    *,
+    samples: int,
+    candidate_only: int,
+    baseline_only: int,
+    both: int,
+    neither: int,
+) -> dict[str, int | float]:
+    if samples < 0 or min(candidate_only, baseline_only, both, neither) < 0:
+        raise ValueError("paired counts must be non-negative")
+    if candidate_only + baseline_only + both + neither != samples:
+        raise ValueError("paired outcome counts must sum to samples")
+    delta_errors = candidate_only - baseline_only
+    delta_ler = float(delta_errors / samples) if samples else float("nan")
+    if samples > 1:
+        difference_square_sum = candidate_only + baseline_only
+        variance = max(
+            0.0,
+            (difference_square_sum - samples * delta_ler * delta_ler)
+            / (samples - 1),
+        )
+        standard_error = math.sqrt(variance / samples)
+    else:
+        standard_error = 0.0 if samples == 1 else float("nan")
+    margin = 1.96 * standard_error
+    return {
+        "samples": samples,
+        "candidate_only_errors": candidate_only,
+        "baseline_only_errors": baseline_only,
+        "both_errors": both,
+        "neither_errors": neither,
+        "delta_logical_errors": delta_errors,
+        "delta_ler": delta_ler,
+        "standard_error": standard_error,
+        "ci95_low": max(-1.0, delta_ler - margin),
+        "ci95_high": min(1.0, delta_ler + margin),
+    }
+
+
+MODEL_PAIRWISE_PRIORITY = (
+    "qadapt",
+    "ising-fast",
+    "ising_fast_t0_e100",
+)
+
+
+def build_model_pairwise_rows(
+    error_masks: Mapping[str, np.ndarray],
+    case_fields: Mapping[str, Any],
+) -> list[dict[str, Any]]:
+    """Build pairwise rows when more than one neural model is selected."""
+
+    known = [name for name in MODEL_PAIRWISE_PRIORITY if name in error_masks]
+    extras = sorted(set(error_masks) - set(known) - {"pymatching"})
+    methods = known + extras
+    rows: list[dict[str, Any]] = []
+    for candidate_index, candidate in enumerate(methods):
+        for baseline in methods[candidate_index + 1 :]:
+            rows.append(
+                {
+                    **dict(case_fields),
+                    "candidate": candidate,
+                    "baseline": baseline,
+                    **paired_error_counts(
+                        error_masks[candidate],
+                        error_masks[baseline],
+                    ),
+                }
+            )
+            rows[-1].pop("delta_ler_vs_pymatching", None)
+    return rows
+
+
+def aggregate_paired_rows(
+    rows: Iterable[Mapping[str, Any]],
+) -> list[dict[str, Any]]:
+    """Pool case-level paired outcomes without treating cases as independent CIs."""
+
+    totals: dict[tuple[str, str], dict[str, Any]] = {}
+    for row in rows:
+        key = (str(row["candidate"]), str(row["baseline"]))
+        entry = totals.setdefault(
+            key,
+            {
+                "candidate": key[0],
+                "baseline": key[1],
+                "cases": 0,
+                "samples": 0,
+                "candidate_only_errors": 0,
+                "baseline_only_errors": 0,
+                "both_errors": 0,
+                "neither_errors": 0,
+            },
+        )
+        entry["cases"] += 1
+        for field in (
+            "samples",
+            "candidate_only_errors",
+            "baseline_only_errors",
+            "both_errors",
+            "neither_errors",
+        ):
+            entry[field] += int(row[field])
+
+    results = []
+    for entry in totals.values():
+        stats = _paired_statistics_from_counts(
+            samples=int(entry["samples"]),
+            candidate_only=int(entry["candidate_only_errors"]),
+            baseline_only=int(entry["baseline_only_errors"]),
+            both=int(entry["both_errors"]),
+            neither=int(entry["neither_errors"]),
+        )
+        results.append(
+            {
+                "candidate": entry["candidate"],
+                "baseline": entry["baseline"],
+                "cases": entry["cases"],
+                **stats,
+            }
+        )
+    return sorted(results, key=lambda row: (row["candidate"], row["baseline"]))
+
+
+def aggregate_rows(rows: Iterable[Mapping[str, Any]]) -> dict[str, dict[str, Any]]:
+    totals: dict[str, dict[str, Any]] = {}
+    for row in rows:
+        if row.get("status", "ok") != "ok":
+            continue
+        method = str(row["method"])
+        entry = totals.setdefault(
+            method,
+            {"method": method, "cases": 0, "shots": 0, "logical_errors": 0},
+        )
+        entry["cases"] += 1
+        entry["shots"] += int(row["shots"])
+        entry["logical_errors"] += int(row["logical_errors"])
+    for entry in totals.values():
+        shots = int(entry["shots"])
+        errors = int(entry["logical_errors"])
+        low, high = wilson_interval(errors, shots)
+        entry.update(
+            ler=float(errors / shots) if shots else float("nan"),
+            ci95_low=low,
+            ci95_high=high,
+        )
+    return totals
+
+
+def _read_b8(
+    path: Path,
+    *,
+    num_detectors: int,
+    num_observables: int,
+) -> np.ndarray:
+    data = stim.read_shot_data_file(
+        path=str(path),
+        format="b8",
+        num_detectors=int(num_detectors),
+        num_observables=int(num_observables),
+    )
+    return np.asarray(data, dtype=np.uint8)
+
+
+def load_case_data(
+    case: GoogleQECCase,
+    *,
+    max_shots: int = 0,
+) -> tuple[stim.Circuit, stim.Circuit, dict[str, Any], np.ndarray, np.ndarray]:
+    metadata = json.loads((case.path / "metadata.json").read_text())
+    ideal = stim.Circuit.from_file(case.path / "circuit_ideal.stim")
+    noisy = stim.Circuit.from_file(case.path / "circuit_noisy_si1000.stim")
+    if ideal.num_detectors != noisy.num_detectors:
+        raise ValueError(f"ideal/noisy detector mismatch in {case.path}")
+    if ideal.num_observables != noisy.num_observables:
+        raise ValueError(f"ideal/noisy observable mismatch in {case.path}")
+    detectors = _read_b8(
+        case.path / "detection_events.b8",
+        num_detectors=int(ideal.num_detectors),
+        num_observables=0,
+    )
+    observables = _read_b8(
+        case.path / "obs_flips_actual.b8",
+        num_detectors=0,
+        num_observables=int(ideal.num_observables),
+    )
+    if detectors.shape[0] != observables.shape[0]:
+        raise ValueError(
+            f"detector/observable shot mismatch in {case.path}: "
+            f"{detectors.shape[0]} != {observables.shape[0]}"
+        )
+    if detectors.shape[0] != int(metadata["shots"]):
+        raise ValueError(
+            f"metadata shot mismatch in {case.path}: "
+            f"{detectors.shape[0]} != {metadata['shots']}"
+        )
+    limit = int(max_shots)
+    if limit > 0:
+        detectors = detectors[:limit]
+        observables = observables[:limit]
+    return ideal, noisy, metadata, detectors, observables
+
+
+def build_matcher(noisy_circuit: stim.Circuit) -> pymatching.Matching:
+    dem = noisy_circuit.detector_error_model(decompose_errors=True)
+    return pymatching.Matching.from_detector_error_model(dem)
+
+
+def _decode_batch(matcher: pymatching.Matching, detectors: np.ndarray) -> np.ndarray:
+    predictions = np.asarray(
+        matcher.decode_batch(np.ascontiguousarray(detectors, dtype=np.uint8)),
+        dtype=np.uint8,
+    )
+    if predictions.ndim == 1:
+        predictions = predictions.reshape(-1, 1)
+    return predictions
+
+
+def time_single_shot(
+    matcher: pymatching.Matching,
+    detectors: np.ndarray,
+    *,
+    rounds: int,
+) -> float:
+    rows = np.asarray(detectors, dtype=np.uint8)
+    if len(rows) == 0:
+        return float("nan")
+    for row in rows[: min(20, len(rows))]:
+        matcher.decode(row)
+    timings = []
+    for row in rows:
+        start = time.perf_counter()
+        matcher.decode(row)
+        timings.append(time.perf_counter() - start)
+    return float(np.mean(timings) * 1e6 / max(1, int(rounds)))
+
+
+def _error_metrics(predictions: np.ndarray, observables: np.ndarray) -> tuple[dict[str, Any], np.ndarray]:
+    predicted = np.asarray(predictions, dtype=np.uint8)
+    actual = np.asarray(observables, dtype=np.uint8)
+    if predicted.shape != actual.shape:
+        raise ValueError(f"prediction/observable shape mismatch: {predicted.shape} != {actual.shape}")
+    error_mask = np.any(predicted != actual, axis=1)
+    errors = int(error_mask.sum())
+    shots = int(len(error_mask))
+    low, high = wilson_interval(errors, shots)
+    return (
+        {
+            "logical_errors": errors,
+            "shots": shots,
+            "ler": float(errors / shots) if shots else float("nan"),
+            "ci95_low": low,
+            "ci95_high": high,
+        },
+        error_mask,
+    )
+
+
+def evaluate_pymatching(
+    matcher: pymatching.Matching,
+    detectors: np.ndarray,
+    observables: np.ndarray,
+    *,
+    rounds: int,
+    latency_shots: int,
+) -> tuple[dict[str, Any], np.ndarray]:
+    start = time.perf_counter()
+    predictions = _decode_batch(matcher, detectors)
+    batch_seconds = time.perf_counter() - start
+    metrics, error_mask = _error_metrics(predictions, observables)
+    latency_rows = detectors[: min(int(latency_shots), len(detectors))]
+    input_density = SyndromeDensityAccumulator()
+    input_density.update(detectors)
+    metrics.update(
+        {
+            "method": "pymatching",
+            "decoder": "uncorrelated_pymatching_si1000_prior",
+            "batch_decode_us_per_shot": float(batch_seconds * 1e6 / max(1, len(detectors))),
+            "pymatching_latency_us_per_round": time_single_shot(
+                matcher,
+                latency_rows,
+                rounds=rounds,
+            ),
+            **input_density.statistics("input"),
+        }
+    )
+    return metrics, error_mask
+
+
+def build_model_cfg(
+    spec: BenchmarkModel,
+    case: GoogleQECCase,
+    *,
+    config_name: str,
+    batch_size: int,
+    latency_shots: int,
+) -> Any:
+    cfg = OmegaConf.load(config_path(config_name))
+    cfg.model_id = int(spec.model_id)
+    cfg.distance = int(case.distance)
+    cfg.n_rounds = int(case.rounds)
+    cfg.workflow.task = "inference"
+    public_spec = validate_public_config(cfg)
+    cfg = apply_public_defaults_and_model(cfg, public_spec)
+    cfg.model_checkpoint_file = str(spec.checkpoint)
+    cfg.test.meas_basis_test = str(case.basis)
+    cfg.test.num_samples = int(case.shots)
+    cfg.test.latency_num_samples = int(latency_shots)
+    cfg.test.batch_size = int(batch_size)
+    cfg.test.dataloader_num_workers = 0
+    return cfg
+
+
+def evaluate_predecoder(
+    model: torch.nn.Module,
+    cfg: Any,
+    matcher: pymatching.Matching,
+    google_detectors: np.ndarray,
+    canonical_detectors: np.ndarray,
+    observables: np.ndarray,
+    canonical_to_source: np.ndarray,
+    *,
+    device: torch.device,
+    rounds: int,
+    batch_size: int,
+    latency_shots: int,
+) -> tuple[dict[str, Any], np.ndarray]:
+    maps = _build_stab_maps(int(cfg.distance), str(cfg.data.code_rotation))
+    module = PreDecoderMemoryEvalModule(model, cfg, maps, device).to(device).eval()
+    predictions = []
+    residual_google_rows = []
+    model_seconds = 0.0
+    residual_matching_seconds = 0.0
+
+    input_density = SyndromeDensityAccumulator()
+    residual_density = SyndromeDensityAccumulator()
+    input_density.update(google_detectors)
+    def synchronize() -> None:
+        if device.type == "cuda":
+            torch.cuda.synchronize(device)
+
+    with torch.inference_mode():
+        for start_index in range(0, len(canonical_detectors), int(batch_size)):
+            canonical_batch = canonical_detectors[
+                start_index : start_index + int(batch_size)
+            ]
+            tensor = torch.from_numpy(canonical_batch).to(
+                device=device,
+                dtype=torch.uint8,
+            )
+            synchronize()
+            started = time.perf_counter()
+            output = module(tensor)
+            synchronize()
+            model_seconds += time.perf_counter() - started
+
+            pre_logical = output[:, :1].to(torch.uint8).cpu().numpy()
+            canonical_residual = output[:, 1:].to(torch.uint8).cpu().numpy()
+            google_residual = canonical_to_google(
+                canonical_residual,
+                canonical_to_source,
+            )
+            started = time.perf_counter()
+            residual_prediction = _decode_batch(matcher, google_residual)
+            residual_density.update(google_residual)
+            residual_matching_seconds += time.perf_counter() - started
+            predictions.append((pre_logical + residual_prediction) % 2)
+            residual_google_rows.append(google_residual)
+
+    final_predictions = np.concatenate(predictions, axis=0)
+    residual_google = np.concatenate(residual_google_rows, axis=0)
+    metrics, error_mask = _error_metrics(final_predictions, observables)
+    latency_rows = residual_google[: min(int(latency_shots), len(residual_google))]
+    residual_latency = time_single_shot(matcher, latency_rows, rounds=rounds)
+    density_statistics = model_density_statistics(input_density, residual_density)
+    shots = max(1, len(google_detectors))
+    metrics.update(
+        {
+            "model_latency_us_per_shot": float(model_seconds * 1e6 / shots),
+            "residual_pymatching_batch_us_per_shot": float(
+                residual_matching_seconds * 1e6 / shots
+            ),
+            "end_to_end_batch_us_per_shot": float(
+                (model_seconds + residual_matching_seconds) * 1e6 / shots
+            ),
+            "pymatching_latency_us_per_round": residual_latency,
+            **density_statistics,
+            "syndrome_reduction": float(density_statistics["density_reduction_fraction"]),
+        }
+    )
+    return metrics, error_mask
+
+
+def _case_fields(case: GoogleQECCase) -> dict[str, Any]:
+    return {
+        "patch": case.patch,
+        "distance": case.distance,
+        "basis": case.basis,
+        "rounds": case.rounds,
+    }
+
+
+def run_benchmark(args: argparse.Namespace) -> dict[str, Any]:
+    root = Path(args.benchmark_root).resolve()
+    selected_models = [DEFAULT_MODELS[name] for name in args.models]
+    missing_checkpoints = [
+        str(spec.checkpoint) for spec in selected_models if not spec.checkpoint.is_file()
+    ]
+    if missing_checkpoints:
+        raise FileNotFoundError(f"Missing model checkpoint(s): {missing_checkpoints}")
+    cases = discover_cases(
+        root,
+        distances=set(args.distances),
+        rounds=set(args.rounds),
+        bases={basis.upper() for basis in args.bases},
+        patches=set(args.patches) if args.patches else None,
+    )
+    if not cases:
+        raise RuntimeError("No Google QEC benchmark cases match the selected filters")
+    if args.list_cases:
+        for case in cases:
+            print(case.path.relative_to(root))
+        return {"cases": [str(case.path.relative_to(root)) for case in cases]}
+
+    device = torch.device(
+        args.device or ("cuda:0" if torch.cuda.is_available() else "cpu")
+    )
+    print(f"[google-qec] device={device} cases={len(cases)}")
+    model_cache: dict[str, torch.nn.Module] = {}
+    rows: list[dict[str, Any]] = []
+    paired_comparisons: list[dict[str, Any]] = []
+
+    for case_index, case in enumerate(cases, start=1):
+        print(
+            f"[google-qec] case {case_index}/{len(cases)} "
+            f"{case.patch}/{case.basis}/r{case.rounds}"
+        )
+        ideal, noisy, metadata, detectors, observables = load_case_data(
+            case,
+            max_shots=int(args.max_shots),
+        )
+        matcher = build_matcher(noisy)
+        permutation = build_detector_permutation(ideal, metadata)
+        canonical_detectors = google_to_canonical(detectors, permutation)
+        baseline, baseline_errors = evaluate_pymatching(
+            matcher,
+            detectors,
+            observables,
+            rounds=case.rounds,
+            latency_shots=int(args.latency_shots),
+        )
+        baseline.update(_case_fields(case), status="ok")
+        rows.append(baseline)
+        print(
+            f"  pymatching: LER={baseline['ler']:.6g} "
+            f"({baseline['logical_errors']}/{baseline['shots']})"
+        )
+
+        if case.rounds < 2:
+            for spec in selected_models:
+                rows.append(
+                    {
+                        **_case_fields(case),
+                        "method": spec.name,
+                        "status": "unsupported",
+                        "reason": "predecoder requires rounds >= 2",
+                        "shots": int(len(detectors)),
+                    }
+                )
+            print("  neural predecoders skipped: rounds=1 is unsupported")
+            continue
+
+        model_error_masks: dict[str, np.ndarray] = {}
+        for spec in selected_models:
+            cfg = build_model_cfg(
+                spec,
+                case,
+                config_name=args.config_name,
+                batch_size=int(args.batch_size),
+                latency_shots=int(args.latency_shots),
+            )
+            if spec.name not in model_cache:
+                distributed = SimpleNamespace(rank=0, device=device)
+                loaded_model = load_model_checkpoint(
+                    cfg,
+                    checkpoint=spec.checkpoint,
+                    model_id=spec.model_id,
+                    distributed=distributed,
+                ).to(device).eval()
+                model_cache[spec.name] = maybe_compile_model(
+                    loaded_model,
+                    enabled=bool(args.torch_compile),
+                    mode=str(args.torch_compile_mode),
+                )
+                if args.torch_compile:
+                    print(f"  {spec.name}: torch.compile mode={args.torch_compile_mode}")
+            metrics, error_mask = evaluate_predecoder(
+                model_cache[spec.name],
+                cfg,
+                matcher,
+                detectors,
+                canonical_detectors,
+                observables,
+                permutation,
+                device=device,
+                rounds=case.rounds,
+                batch_size=int(args.batch_size),
+                latency_shots=int(args.latency_shots),
+            )
+            model_error_masks[spec.name] = error_mask
+            metrics.update(
+                _case_fields(case),
+                method=spec.name,
+                checkpoint=str(spec.checkpoint),
+                status="ok",
+            )
+            paired_vs_pymatching = paired_error_counts(error_mask, baseline_errors)
+            for field in (
+                "candidate_only_errors",
+                "baseline_only_errors",
+                "both_errors",
+                "neither_errors",
+                "delta_logical_errors",
+                "delta_ler_vs_pymatching",
+            ):
+                metrics[field] = paired_vs_pymatching[field]
+            metrics.update(
+                paired_samples_vs_pymatching=paired_vs_pymatching["samples"],
+                paired_standard_error_vs_pymatching=paired_vs_pymatching["standard_error"],
+                paired_ci95_low_vs_pymatching=paired_vs_pymatching["ci95_low"],
+                paired_ci95_high_vs_pymatching=paired_vs_pymatching["ci95_high"],
+            )
+            baseline_latency = float(baseline["pymatching_latency_us_per_round"])
+            residual_latency = float(metrics["pymatching_latency_us_per_round"])
+            metrics["pymatching_speedup"] = (
+                baseline_latency / residual_latency
+                if residual_latency > 0 and math.isfinite(residual_latency)
+                else float("nan")
+            )
+            rows.append(metrics)
+            print(
+                f"  {spec.name}: LER={metrics['ler']:.6g} "
+                f"delta={metrics['delta_ler_vs_pymatching']:+.6g} "
+                f"syndrome_reduction={metrics['syndrome_reduction']:.3f}"
+            )
+
+        paired_comparisons.extend(
+            build_model_pairwise_rows(model_error_masks, _case_fields(case))
+        )
+    payload = {
+        "schema_version": 2,
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "benchmark_root": str(root),
+        "decoder_prior": "Google circuit_noisy_si1000.stim DEM",
+        "detector_mapping": "Google physical order <-> repository XV canonical order",
+        "device": str(device),
+        "filters": {
+            "distances": list(args.distances),
+            "rounds": list(args.rounds),
+            "bases": list(args.bases),
+            "patches": list(args.patches or []),
+            "max_shots": int(args.max_shots),
+            "batch_size": int(args.batch_size),
+            "latency_shots": int(args.latency_shots),
+            "torch_compile": bool(args.torch_compile),
+            "torch_compile_mode": str(args.torch_compile_mode),
+        },
+        "models": {
+            spec.name: {
+                "model_id": spec.model_id,
+                "checkpoint": str(spec.checkpoint),
+            }
+            for spec in selected_models
+        },
+        "rows": rows,
+        "aggregate": aggregate_rows(rows),
+        "paired_comparisons": paired_comparisons,
+        "paired_aggregate": aggregate_paired_rows(paired_comparisons),
+    }
+    return payload
+
+
+
+def merge_benchmark_payloads(
+    payloads: Sequence[Mapping[str, Any]],
+) -> dict[str, Any]:
+    """Merge disjoint benchmark shards and recompute all pooled statistics."""
+
+    if not payloads:
+        raise ValueError("at least one benchmark payload is required")
+    reference = payloads[0]
+    for index, payload in enumerate(payloads):
+        if int(payload.get("schema_version", 0)) != 2:
+            raise ValueError(f"benchmark shard {index} is not schema_version=2")
+        for field in (
+            "benchmark_root",
+            "decoder_prior",
+            "detector_mapping",
+            "models",
+        ):
+            if payload.get(field) != reference.get(field):
+                raise ValueError(f"benchmark shard {index} disagrees on {field}")
+
+    rows = [dict(row) for payload in payloads for row in payload.get("rows", [])]
+    paired = [
+        dict(row)
+        for payload in payloads
+        for row in payload.get("paired_comparisons", [])
+    ]
+    row_keys = [
+        (
+            str(row.get("patch")),
+            int(row.get("distance", 0)),
+            str(row.get("basis")),
+            int(row.get("rounds", 0)),
+            str(row.get("method")),
+        )
+        for row in rows
+    ]
+    if len(row_keys) != len(set(row_keys)):
+        raise ValueError("benchmark shards contain duplicate case/method rows")
+    paired_keys = [
+        (
+            str(row.get("patch")),
+            int(row.get("distance", 0)),
+            str(row.get("basis")),
+            int(row.get("rounds", 0)),
+            str(row.get("candidate")),
+            str(row.get("baseline")),
+        )
+        for row in paired
+    ]
+    if len(paired_keys) != len(set(paired_keys)):
+        raise ValueError("benchmark shards contain duplicate paired comparisons")
+
+    rows.sort(
+        key=lambda row: (
+            int(row.get("distance", 0)),
+            str(row.get("patch")),
+            str(row.get("basis")),
+            int(row.get("rounds", 0)),
+            str(row.get("method")),
+        )
+    )
+    paired.sort(
+        key=lambda row: (
+            int(row.get("distance", 0)),
+            str(row.get("patch")),
+            str(row.get("basis")),
+            int(row.get("rounds", 0)),
+            str(row.get("candidate")),
+            str(row.get("baseline")),
+        )
+    )
+    max_shots = {
+        int(payload.get("filters", {}).get("max_shots", 0)) for payload in payloads
+    }
+    if len(max_shots) != 1:
+        raise ValueError("benchmark shards disagree on max_shots")
+    execution_filters = {}
+    for field in (
+        "batch_size",
+        "latency_shots",
+        "torch_compile",
+        "torch_compile_mode",
+    ):
+        values = {payload.get("filters", {}).get(field) for payload in payloads}
+        if len(values) != 1:
+            raise ValueError(f"benchmark shards disagree on {field}")
+        execution_filters[field] = values.pop()
+    return {
+        "schema_version": 2,
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "benchmark_root": reference["benchmark_root"],
+        "decoder_prior": reference["decoder_prior"],
+        "detector_mapping": reference["detector_mapping"],
+        "device": "merged_shards",
+        "filters": {
+            "distances": sorted({int(row["distance"]) for row in rows}),
+            "rounds": sorted({int(row["rounds"]) for row in rows}),
+            "bases": sorted({str(row["basis"]) for row in rows}),
+            "patches": sorted({str(row["patch"]) for row in rows}),
+            "max_shots": max_shots.pop(),
+            **execution_filters,
+        },
+        "models": reference["models"],
+        "rows": rows,
+        "aggregate": aggregate_rows(rows),
+        "paired_comparisons": paired,
+        "paired_aggregate": aggregate_paired_rows(paired),
+    }
+
+def write_results(payload: Mapping[str, Any], output_path: Path) -> tuple[Path, Path]:
+    output_path = Path(output_path)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    output_path.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
+    csv_path = output_path.with_suffix(".csv")
+    rows = list(payload.get("rows", []))
+    fieldnames = sorted({str(key) for row in rows for key in row})
+    with csv_path.open("w", newline="") as stream:
+        writer = csv.DictWriter(stream, fieldnames=fieldnames)
+        writer.writeheader()
+        writer.writerows(rows)
+    paired_rows = list(payload.get("paired_comparisons", []))
+    paired_csv_path = output_path.with_name(
+        f"{output_path.stem}_paired.csv"
+    )
+    paired_fields = sorted({str(key) for row in paired_rows for key in row})
+    with paired_csv_path.open("w", newline="") as stream:
+        writer = csv.DictWriter(stream, fieldnames=paired_fields)
+        if paired_fields:
+            writer.writeheader()
+            writer.writerows(paired_rows)
+    return output_path, csv_path
+
+
+def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description=(
+            "Evaluate PyMatching and released pre-decoders on Google Willow "
+            "QEC hardware samples."
+        )
+    )
+    parser.add_argument("--benchmark-root", type=Path, default=DEFAULT_BENCHMARK_ROOT)
+    parser.add_argument("--distances", nargs="+", type=int, default=[3, 5, 7])
+    parser.add_argument(
+        "--rounds",
+        nargs="+",
+        type=int,
+        default=[13],
+        help="Google cycle counts. The default r13 is the calibration slice.",
+    )
+    parser.add_argument("--bases", nargs="+", choices=("X", "Z"), default=["X", "Z"])
+    parser.add_argument(
+        "--patches",
+        nargs="+",
+        default=None,
+        help="Optional exact patch directory names, for example d7_at_q6_7.",
+    )
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        choices=tuple(DEFAULT_MODELS),
+        default=list(DEFAULT_MODELS),
+    )
+    parser.add_argument("--config-name", default="examples/qadapt/config_qadapt_t0_base")
+    parser.add_argument("--max-shots", type=int, default=0, help="0 uses all shots.")
+    parser.add_argument("--batch-size", type=int, default=512)
+    parser.add_argument("--latency-shots", type=int, default=512)
+    parser.add_argument("--device", default=None)
+    parser.add_argument(
+        "--torch-compile",
+        action="store_true",
+        help="Compile each neural model once with dynamic input shapes.",
+    )
+    parser.add_argument(
+        "--torch-compile-mode",
+        choices=(
+            "default",
+            "reduce-overhead",
+            "max-autotune",
+            "max-autotune-no-cudagraphs",
+        ),
+        default="default",
+    )
+    parser.add_argument("--output", type=Path, default=None)
+    parser.add_argument(
+        "--merge-inputs",
+        nargs="+",
+        type=Path,
+        default=None,
+        help="Merge disjoint schema-v2 benchmark JSON shards instead of running inference.",
+    )
+    parser.add_argument("--list-cases", action="store_true")
+    args = parser.parse_args(argv)
+    if args.max_shots < 0:
+        parser.error("--max-shots must be >= 0")
+    if args.batch_size <= 0:
+        parser.error("--batch-size must be positive")
+    if args.latency_shots <= 0:
+        parser.error("--latency-shots must be positive")
+    if args.output is None:
+        args.output = Path(args.benchmark_root) / "ising_decoder_results/results.json"
+    if args.merge_inputs and args.list_cases:
+        parser.error("--merge-inputs cannot be combined with --list-cases")
+    return args
+
+
+def main(argv: Sequence[str] | None = None) -> int:
+    args = parse_args(argv)
+    if args.merge_inputs:
+        payload = merge_benchmark_payloads(
+            [json.loads(Path(path).read_text(encoding="utf-8")) for path in args.merge_inputs]
+        )
+        payload["merged_inputs"] = [str(Path(path).resolve()) for path in args.merge_inputs]
+        print(f"[google-qec] merged {len(args.merge_inputs)} shards")
+    else:
+        payload = run_benchmark(args)
+    if args.list_cases:
+        return 0
+    json_path, csv_path = write_results(payload, args.output)
+    print(f"[google-qec] JSON: {json_path}")
+    print(f"[google-qec] CSV:  {csv_path}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/code/scripts/qadapt_example_utils.py b/code/scripts/qadapt_example_utils.py
new file mode 100644
index 0000000..52b6add
--- /dev/null
+++ b/code/scripts/qadapt_example_utils.py
@@ -0,0 +1,229 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Shared command construction and execution for the public QAdapt examples."""
+
+from __future__ import annotations
+
+import argparse
+import os
+import shlex
+import subprocess
+import sys
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Sequence
+
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+PAIRED_INFERENCE_SCRIPT = REPO_ROOT / "code" / "scripts" / "paired_inference_compare.py"
+TASK_CONFIGS = (
+    ("t0_base", "examples/qadapt_v2/T0"),
+    ("t1_meas_x4", "examples/qadapt_v2/T1"),
+    ("t2_cnot_x1p5", "examples/qadapt_v2/T2"),
+    ("t3_idle_x4", "examples/qadapt_v2/T3"),
+    ("t4_z_bias_x1p5", "examples/qadapt_v2/T4"),
+)
+
+
+@dataclass(frozen=True)
+class ModelArgument:
+    name: str
+    model_id: int
+    checkpoint: Path
+
+
+@dataclass(frozen=True)
+class InferenceJob:
+    label: str
+    command: tuple[str, ...]
+    output_path: Path
+
+
+def parse_model_argument(value: str) -> ModelArgument:
+    parts = value.split(":", 2)
+    if len(parts) != 3:
+        raise argparse.ArgumentTypeError(
+            "--model must be formatted as name:model_id:/path/to/checkpoint"
+        )
+    name, model_id_raw, checkpoint_raw = (part.strip() for part in parts)
+    if not name or not checkpoint_raw:
+        raise argparse.ArgumentTypeError("model name and checkpoint must not be empty")
+    try:
+        model_id = int(model_id_raw)
+    except ValueError as exc:
+        raise argparse.ArgumentTypeError(
+            f"invalid model_id: {model_id_raw}"
+        ) from exc
+    checkpoint = Path(checkpoint_raw).expanduser()
+    if not checkpoint.is_absolute():
+        checkpoint = REPO_ROOT / checkpoint
+    return ModelArgument(name=name, model_id=model_id, checkpoint=checkpoint)
+
+
+def _default_gpus() -> str:
+    visible = os.environ.get("CUDA_VISIBLE_DEVICES", "").strip()
+    return visible or "0"
+
+
+def add_common_inference_args(
+    parser: argparse.ArgumentParser,
+    *,
+    default_output_dir: Path,
+    default_num_samples: int = 262144,
+) -> None:
+    parser.add_argument(
+        "--model",
+        action="append",
+        type=parse_model_argument,
+        required=True,
+        help=(
+            "Repeat for each released model: name:model_id:/path/to/checkpoint. "
+            "Both .pt and .safetensors are supported."
+        ),
+    )
+    parser.add_argument("--num-samples", type=int, default=default_num_samples)
+    parser.add_argument("--latency-num-samples", type=int, default=10000)
+    parser.add_argument("--batch-size", type=int, default=2048)
+    parser.add_argument("--num-workers", type=int, default=0)
+    parser.add_argument("--basis", choices=("both", "X", "Z"), default="both")
+    parser.add_argument("--seed", type=int, default=12345)
+    parser.add_argument("--gpus", default=_default_gpus())
+    parser.add_argument("--parallelism", type=int, default=1)
+    parser.add_argument(
+        "--python",
+        default=os.environ.get("PREDECODER_PYTHON", sys.executable),
+    )
+    parser.add_argument("--output-dir", type=Path, default=default_output_dir)
+    parser.add_argument("--resume", action="store_true")
+    parser.add_argument("--dry-run", action="store_true")
+
+
+def checkpoint_specs(args: argparse.Namespace) -> tuple[ModelArgument, ...]:
+    specs = tuple(args.model)
+    names = [spec.name for spec in specs]
+    if len(names) != len(set(names)):
+        raise ValueError(f"model names must be unique: {names}")
+    return specs
+
+
+def parse_gpus(value: str | Sequence[str]) -> list[str]:
+    raw = value.split(",") if isinstance(value, str) else value
+    result = [str(item).strip() for item in raw if str(item).strip()]
+    if not result:
+        raise ValueError("at least one GPU must be selected")
+    return result
+
+
+def build_paired_command(
+    args: argparse.Namespace,
+    *,
+    output_path: Path,
+    distance: int,
+    n_rounds: int,
+    config_name: str | None = None,
+    config_file: Path | None = None,
+) -> tuple[str, ...]:
+    if (config_name is None) == (config_file is None):
+        raise ValueError("provide exactly one of config_name or config_file")
+    command = [
+        str(args.python),
+        "-u",
+        str(PAIRED_INFERENCE_SCRIPT),
+    ]
+    if config_name is not None:
+        command.extend(("--config-name", config_name))
+    else:
+        command.extend(("--config-file", str(Path(config_file))))
+    command.extend(
+        (
+            "--distance",
+            str(distance),
+            "--n-rounds",
+            str(n_rounds),
+            "--num-samples",
+            str(args.num_samples),
+            "--latency-num-samples",
+            str(args.latency_num_samples),
+            "--batch-size",
+            str(args.batch_size),
+            "--num-workers",
+            str(args.num_workers),
+            "--seed",
+            str(args.seed),
+            "--basis",
+            str(args.basis),
+            "--device",
+            "cuda:0",
+            "--output",
+            str(output_path),
+        )
+    )
+    for spec in checkpoint_specs(args):
+        command.extend(
+            ("--model", f"{spec.name}:{spec.model_id}:{spec.checkpoint}")
+        )
+    return tuple(command)
+
+
+def _run_one(job: InferenceJob, gpu: str) -> tuple[InferenceJob, int, Path]:
+    job.output_path.parent.mkdir(parents=True, exist_ok=True)
+    log_path = job.output_path.with_suffix(".log")
+    env = dict(os.environ)
+    env["CUDA_VISIBLE_DEVICES"] = gpu
+    with log_path.open("w", encoding="utf-8") as stream:
+        completed = subprocess.run(
+            job.command,
+            cwd=REPO_ROOT,
+            env=env,
+            stdout=stream,
+            stderr=subprocess.STDOUT,
+            check=False,
+        )
+    return job, int(completed.returncode), log_path
+
+
+def run_jobs(
+    jobs: Sequence[InferenceJob],
+    *,
+    gpus: Sequence[str],
+    parallelism: int,
+    resume: bool,
+    dry_run: bool,
+) -> None:
+    selected_gpus = parse_gpus(gpus)
+    workers = max(1, min(int(parallelism), len(selected_gpus)))
+    pending = [
+        job for job in jobs
+        if not (resume and job.output_path.is_file())
+    ]
+    skipped = len(jobs) - len(pending)
+    if skipped:
+        print(f"[resume] skipped {skipped} existing outputs")
+    if dry_run:
+        for index, job in enumerate(pending):
+            gpu = selected_gpus[index % workers]
+            print(
+                f"[dry-run] gpu={gpu} label={job.label} "
+                + shlex.join(job.command)
+            )
+        return
+    failures = []
+    with ThreadPoolExecutor(max_workers=workers) as executor:
+        futures = {
+            executor.submit(_run_one, job, selected_gpus[index % workers]): job
+            for index, job in enumerate(pending)
+        }
+        for future in as_completed(futures):
+            job, returncode, log_path = future.result()
+            if returncode:
+                failures.append((job, returncode, log_path))
+                print(f"[fail] {job.label} log={log_path}")
+            else:
+                print(f"[done] {job.label} output={job.output_path}")
+    if failures:
+        details = "\n".join(
+            f"  - {job.label}: exit={returncode}, log={log_path}"
+            for job, returncode, log_path in failures
+        )
+        raise RuntimeError(f"Released-model inference jobs failed:\n{details}")
diff --git a/code/workflows/config_validator.py b/code/workflows/config_validator.py
index a5ac02f..32eb256 100644
--- a/code/workflows/config_validator.py
+++ b/code/workflows/config_validator.py
@@ -1,5 +1,6 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
+# Modified in 2026 for the QAdapt Hugging Face release: added HTNet defaults.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -45,6 +46,7 @@ _INTERNAL_ROTATION_TO_PUBLIC = {v: k for k, v in _PUBLIC_ROTATION_TO_INTERNAL.it
 
 _PUBLIC_MODEL_ID_TO_LR = {
     1: 3e-4,
+    111: 3e-4,
     2: 2e-4,
     3: 1e-4,
     4: 2e-4,
@@ -557,6 +559,18 @@ def apply_public_defaults_and_model(cfg: DictConfig, model_spec: PublicModelSpec
         merged.model.version = model_spec.model_version
         merged.model.num_filters = list(model_spec.num_filters)
         merged.model.kernel_size = list(model_spec.kernel_size)
+        if model_spec.channels is not None:
+            merged.model.channels = int(model_spec.channels)
+        if model_spec.expand_channels is not None:
+            merged.model.expand_channels = int(model_spec.expand_channels)
+        if model_spec.num_blocks is not None:
+            merged.model.num_blocks = int(model_spec.num_blocks)
+        if model_spec.joint_groups is not None:
+            merged.model.joint_groups = int(model_spec.joint_groups)
+        if model_spec.norm_groups is not None:
+            merged.model.norm_groups = int(model_spec.norm_groups)
+        if model_spec.se_reduction is not None:
+            merged.model.se_reduction = int(model_spec.se_reduction)
 
     _apply_code_specific_defaults(merged, code, model_spec)
 
diff --git a/conf/examples/qadapt/config_qadapt_t0_base.yaml b/conf/examples/qadapt/config_qadapt_t0_base.yaml
new file mode 100644
index 0000000..d3631a2
--- /dev/null
+++ b/conf/examples/qadapt/config_qadapt_t0_base.yaml
@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# QAdapt T0 inference environment.
+
+model_id: 111
+distance: 9
+n_rounds: 9
+
+workflow:
+  task: inference
+
+data:
+  code_rotation: O1
+  noise_model:
+    p_prep_X: 0.0010000
+    p_prep_Z: 0.0010000
+    p_meas_X: 0.0100000
+    p_meas_Z: 0.0100000
+    p_idle_cnot_X: 0.0003330
+    p_idle_cnot_Y: 0.0003330
+    p_idle_cnot_Z: 0.0003330
+    p_idle_spam_X: 0.0006670
+    p_idle_spam_Y: 0.0006670
+    p_idle_spam_Z: 0.0006670
+    p_cnot_IX: 0.0006670
+    p_cnot_IY: 0.0006670
+    p_cnot_IZ: 0.0006670
+    p_cnot_XI: 0.0006670
+    p_cnot_XX: 0.0006670
+    p_cnot_XY: 0.0006670
+    p_cnot_XZ: 0.0006670
+    p_cnot_YI: 0.0006670
+    p_cnot_YX: 0.0006670
+    p_cnot_YY: 0.0006670
+    p_cnot_YZ: 0.0006670
+    p_cnot_ZI: 0.0006670
+    p_cnot_ZX: 0.0006670
+    p_cnot_ZY: 0.0006670
+    p_cnot_ZZ: 0.0006670
diff --git a/conf/examples/qadapt/config_qadapt_t1_meas_1p5.yaml b/conf/examples/qadapt/config_qadapt_t1_meas_1p5.yaml
new file mode 100644
index 0000000..46dd700
--- /dev/null
+++ b/conf/examples/qadapt/config_qadapt_t1_meas_1p5.yaml
@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Shared QAdapt T1 measurement-noise task.
+
+model_id: 111
+distance: 9
+n_rounds: 9
+
+workflow:
+  task: inference
+
+data:
+  code_rotation: O1
+  noise_model:
+    p_prep_X: 0.0010000
+    p_prep_Z: 0.0010000
+    p_meas_X: 0.0150000
+    p_meas_Z: 0.0150000
+    p_idle_cnot_X: 0.0003330
+    p_idle_cnot_Y: 0.0003330
+    p_idle_cnot_Z: 0.0003330
+    p_idle_spam_X: 0.0006670
+    p_idle_spam_Y: 0.0006670
+    p_idle_spam_Z: 0.0006670
+    p_cnot_IX: 0.0006670
+    p_cnot_IY: 0.0006670
+    p_cnot_IZ: 0.0006670
+    p_cnot_XI: 0.0006670
+    p_cnot_XX: 0.0006670
+    p_cnot_XY: 0.0006670
+    p_cnot_XZ: 0.0006670
+    p_cnot_YI: 0.0006670
+    p_cnot_YX: 0.0006670
+    p_cnot_YY: 0.0006670
+    p_cnot_YZ: 0.0006670
+    p_cnot_ZI: 0.0006670
+    p_cnot_ZX: 0.0006670
+    p_cnot_ZY: 0.0006670
+    p_cnot_ZZ: 0.0006670
diff --git a/conf/examples/qadapt/config_qadapt_t2_cnot_1p5.yaml b/conf/examples/qadapt/config_qadapt_t2_cnot_1p5.yaml
new file mode 100644
index 0000000..1eda622
--- /dev/null
+++ b/conf/examples/qadapt/config_qadapt_t2_cnot_1p5.yaml
@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Shared QAdapt T2 CNOT-noise task.
+
+model_id: 111
+distance: 9
+n_rounds: 9
+
+workflow:
+  task: inference
+
+data:
+  code_rotation: O1
+  noise_model:
+    p_prep_X: 0.0010000
+    p_prep_Z: 0.0010000
+    p_meas_X: 0.0100000
+    p_meas_Z: 0.0100000
+    p_idle_cnot_X: 0.0003330
+    p_idle_cnot_Y: 0.0003330
+    p_idle_cnot_Z: 0.0003330
+    p_idle_spam_X: 0.0006670
+    p_idle_spam_Y: 0.0006670
+    p_idle_spam_Z: 0.0006670
+    p_cnot_IX: 0.0010005
+    p_cnot_IY: 0.0010005
+    p_cnot_IZ: 0.0010005
+    p_cnot_XI: 0.0010005
+    p_cnot_XX: 0.0010005
+    p_cnot_XY: 0.0010005
+    p_cnot_XZ: 0.0010005
+    p_cnot_YI: 0.0010005
+    p_cnot_YX: 0.0010005
+    p_cnot_YY: 0.0010005
+    p_cnot_YZ: 0.0010005
+    p_cnot_ZI: 0.0010005
+    p_cnot_ZX: 0.0010005
+    p_cnot_ZY: 0.0010005
+    p_cnot_ZZ: 0.0010005
diff --git a/conf/examples/qadapt/config_qadapt_t3_idle_1p5.yaml b/conf/examples/qadapt/config_qadapt_t3_idle_1p5.yaml
new file mode 100644
index 0000000..26acbdd
--- /dev/null
+++ b/conf/examples/qadapt/config_qadapt_t3_idle_1p5.yaml
@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Shared QAdapt T3 idle-noise task.
+
+model_id: 111
+distance: 9
+n_rounds: 9
+
+workflow:
+  task: inference
+
+data:
+  code_rotation: O1
+  noise_model:
+    p_prep_X: 0.0010000
+    p_prep_Z: 0.0010000
+    p_meas_X: 0.0100000
+    p_meas_Z: 0.0100000
+    p_idle_cnot_X: 0.0004995
+    p_idle_cnot_Y: 0.0004995
+    p_idle_cnot_Z: 0.0004995
+    p_idle_spam_X: 0.0010005
+    p_idle_spam_Y: 0.0010005
+    p_idle_spam_Z: 0.0010005
+    p_cnot_IX: 0.0006670
+    p_cnot_IY: 0.0006670
+    p_cnot_IZ: 0.0006670
+    p_cnot_XI: 0.0006670
+    p_cnot_XX: 0.0006670
+    p_cnot_XY: 0.0006670
+    p_cnot_XZ: 0.0006670
+    p_cnot_YI: 0.0006670
+    p_cnot_YX: 0.0006670
+    p_cnot_YY: 0.0006670
+    p_cnot_YZ: 0.0006670
+    p_cnot_ZI: 0.0006670
+    p_cnot_ZX: 0.0006670
+    p_cnot_ZY: 0.0006670
+    p_cnot_ZZ: 0.0006670
diff --git a/conf/examples/qadapt/config_qadapt_t4_z_bias_1p5.yaml b/conf/examples/qadapt/config_qadapt_t4_z_bias_1p5.yaml
new file mode 100644
index 0000000..1a147e6
--- /dev/null
+++ b/conf/examples/qadapt/config_qadapt_t4_z_bias_1p5.yaml
@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Shared QAdapt T4 Z-biased-noise task.
+
+model_id: 111
+distance: 9
+n_rounds: 9
+
+workflow:
+  task: inference
+
+data:
+  code_rotation: O1
+  noise_model:
+    p_prep_X: 0.0015000
+    p_prep_Z: 0.0010000
+    p_meas_X: 0.0150000
+    p_meas_Z: 0.0100000
+    p_idle_cnot_X: 0.0003330
+    p_idle_cnot_Y: 0.0003330
+    p_idle_cnot_Z: 0.0004995
+    p_idle_spam_X: 0.0006670
+    p_idle_spam_Y: 0.0006670
+    p_idle_spam_Z: 0.0010005
+    p_cnot_IX: 0.0006670
+    p_cnot_IY: 0.0006670
+    p_cnot_IZ: 0.0010005
+    p_cnot_XI: 0.0006670
+    p_cnot_XX: 0.0006670
+    p_cnot_XY: 0.0006670
+    p_cnot_XZ: 0.0010005
+    p_cnot_YI: 0.0006670
+    p_cnot_YX: 0.0006670
+    p_cnot_YY: 0.0006670
+    p_cnot_YZ: 0.0010005
+    p_cnot_ZI: 0.0010005
+    p_cnot_ZX: 0.0010005
+    p_cnot_ZY: 0.0010005
+    p_cnot_ZZ: 0.0010005
diff --git a/conf/examples/qadapt_v2/T0.yaml b/conf/examples/qadapt_v2/T0.yaml
new file mode 100644
index 0000000..c50cc25
--- /dev/null
+++ b/conf/examples/qadapt_v2/T0.yaml
@@ -0,0 +1,46 @@
+# @package _global_
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+model_id: 111
+distance: 9
+n_rounds: 9
+seed: 20260907
+
+workflow:
+  task: inference
+
+data:
+  code_rotation: O1
+  skip_noise_upscaling: true
+  noise_model:
+    p_prep_X: 0.0010000
+    p_prep_Z: 0.0010000
+    p_meas_X: 0.0100000
+    p_meas_Z: 0.0100000
+    p_idle_cnot_X: 0.0003330
+    p_idle_cnot_Y: 0.0003330
+    p_idle_cnot_Z: 0.0003330
+    p_idle_spam_X: 0.0006670
+    p_idle_spam_Y: 0.0006670
+    p_idle_spam_Z: 0.0006670
+    p_cnot_IX: 0.0006670
+    p_cnot_IY: 0.0006670
+    p_cnot_IZ: 0.0006670
+    p_cnot_XI: 0.0006670
+    p_cnot_XX: 0.0006670
+    p_cnot_XY: 0.0006670
+    p_cnot_XZ: 0.0006670
+    p_cnot_YI: 0.0006670
+    p_cnot_YX: 0.0006670
+    p_cnot_YY: 0.0006670
+    p_cnot_YZ: 0.0006670
+    p_cnot_ZI: 0.0006670
+    p_cnot_ZX: 0.0006670
+    p_cnot_ZY: 0.0006670
+    p_cnot_ZZ: 0.0006670
+
+htnet_ablation:
+  branch_mixer: true
+  axis_channel_gate: false
+  input_skip_head: true
diff --git a/conf/examples/qadapt_v2/T1.yaml b/conf/examples/qadapt_v2/T1.yaml
new file mode 100644
index 0000000..9348830
--- /dev/null
+++ b/conf/examples/qadapt_v2/T1.yaml
@@ -0,0 +1,43 @@
+# @package _global_
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+model_id: 111
+distance: 9
+n_rounds: 9
+seed: 20260907
+workflow:
+  task: inference
+data:
+  code_rotation: O1
+  skip_noise_upscaling: true
+  noise_model:
+    p_prep_X: 0.001
+    p_prep_Z: 0.001
+    p_meas_X: 0.04
+    p_meas_Z: 0.04
+    p_idle_cnot_X: 0.000333
+    p_idle_cnot_Y: 0.000333
+    p_idle_cnot_Z: 0.000333
+    p_idle_spam_X: 0.000667
+    p_idle_spam_Y: 0.000667
+    p_idle_spam_Z: 0.000667
+    p_cnot_IX: 0.000667
+    p_cnot_IY: 0.000667
+    p_cnot_IZ: 0.000667
+    p_cnot_XI: 0.000667
+    p_cnot_XX: 0.000667
+    p_cnot_XY: 0.000667
+    p_cnot_XZ: 0.000667
+    p_cnot_YI: 0.000667
+    p_cnot_YX: 0.000667
+    p_cnot_YY: 0.000667
+    p_cnot_YZ: 0.000667
+    p_cnot_ZI: 0.000667
+    p_cnot_ZX: 0.000667
+    p_cnot_ZY: 0.000667
+    p_cnot_ZZ: 0.000667
+htnet_ablation:
+  branch_mixer: true
+  axis_channel_gate: false
+  input_skip_head: true
diff --git a/conf/examples/qadapt_v2/T2.yaml b/conf/examples/qadapt_v2/T2.yaml
new file mode 100644
index 0000000..2d6773e
--- /dev/null
+++ b/conf/examples/qadapt_v2/T2.yaml
@@ -0,0 +1,43 @@
+# @package _global_
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+model_id: 111
+distance: 9
+n_rounds: 9
+seed: 20260907
+workflow:
+  task: inference
+data:
+  code_rotation: O1
+  skip_noise_upscaling: true
+  noise_model:
+    p_prep_X: 0.001
+    p_prep_Z: 0.001
+    p_meas_X: 0.01
+    p_meas_Z: 0.01
+    p_idle_cnot_X: 0.000333
+    p_idle_cnot_Y: 0.000333
+    p_idle_cnot_Z: 0.000333
+    p_idle_spam_X: 0.000667
+    p_idle_spam_Y: 0.000667
+    p_idle_spam_Z: 0.000667
+    p_cnot_IX: 0.0010005
+    p_cnot_IY: 0.0010005
+    p_cnot_IZ: 0.0010005
+    p_cnot_XI: 0.0010005
+    p_cnot_XX: 0.0010005
+    p_cnot_XY: 0.0010005
+    p_cnot_XZ: 0.0010005
+    p_cnot_YI: 0.0010005
+    p_cnot_YX: 0.0010005
+    p_cnot_YY: 0.0010005
+    p_cnot_YZ: 0.0010005
+    p_cnot_ZI: 0.0010005
+    p_cnot_ZX: 0.0010005
+    p_cnot_ZY: 0.0010005
+    p_cnot_ZZ: 0.0010005
+htnet_ablation:
+  branch_mixer: true
+  axis_channel_gate: false
+  input_skip_head: true
diff --git a/conf/examples/qadapt_v2/T3.yaml b/conf/examples/qadapt_v2/T3.yaml
new file mode 100644
index 0000000..80bb9f3
--- /dev/null
+++ b/conf/examples/qadapt_v2/T3.yaml
@@ -0,0 +1,43 @@
+# @package _global_
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+model_id: 111
+distance: 9
+n_rounds: 9
+seed: 20260907
+workflow:
+  task: inference
+data:
+  code_rotation: O1
+  skip_noise_upscaling: true
+  noise_model:
+    p_prep_X: 0.001
+    p_prep_Z: 0.001
+    p_meas_X: 0.01
+    p_meas_Z: 0.01
+    p_idle_cnot_X: 0.001332
+    p_idle_cnot_Y: 0.001332
+    p_idle_cnot_Z: 0.001332
+    p_idle_spam_X: 0.002668
+    p_idle_spam_Y: 0.002668
+    p_idle_spam_Z: 0.002668
+    p_cnot_IX: 0.000667
+    p_cnot_IY: 0.000667
+    p_cnot_IZ: 0.000667
+    p_cnot_XI: 0.000667
+    p_cnot_XX: 0.000667
+    p_cnot_XY: 0.000667
+    p_cnot_XZ: 0.000667
+    p_cnot_YI: 0.000667
+    p_cnot_YX: 0.000667
+    p_cnot_YY: 0.000667
+    p_cnot_YZ: 0.000667
+    p_cnot_ZI: 0.000667
+    p_cnot_ZX: 0.000667
+    p_cnot_ZY: 0.000667
+    p_cnot_ZZ: 0.000667
+htnet_ablation:
+  branch_mixer: true
+  axis_channel_gate: false
+  input_skip_head: true
diff --git a/conf/examples/qadapt_v2/T4.yaml b/conf/examples/qadapt_v2/T4.yaml
new file mode 100644
index 0000000..6b5d985
--- /dev/null
+++ b/conf/examples/qadapt_v2/T4.yaml
@@ -0,0 +1,43 @@
+# @package _global_
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+model_id: 111
+distance: 9
+n_rounds: 9
+seed: 20260907
+workflow:
+  task: inference
+data:
+  code_rotation: O1
+  skip_noise_upscaling: true
+  noise_model:
+    p_prep_X: 0.0015
+    p_prep_Z: 0.001
+    p_meas_X: 0.015
+    p_meas_Z: 0.01
+    p_idle_cnot_X: 0.000333
+    p_idle_cnot_Y: 0.000333
+    p_idle_cnot_Z: 0.0004995
+    p_idle_spam_X: 0.000667
+    p_idle_spam_Y: 0.000667
+    p_idle_spam_Z: 0.0010005
+    p_cnot_IX: 0.000667
+    p_cnot_IY: 0.000667
+    p_cnot_IZ: 0.0010005
+    p_cnot_XI: 0.000667
+    p_cnot_XX: 0.000667
+    p_cnot_XY: 0.000667
+    p_cnot_XZ: 0.0010005
+    p_cnot_YI: 0.000667
+    p_cnot_YX: 0.000667
+    p_cnot_YY: 0.000667
+    p_cnot_YZ: 0.0010005
+    p_cnot_ZI: 0.0010005
+    p_cnot_ZX: 0.0010005
+    p_cnot_ZY: 0.0010005
+    p_cnot_ZZ: 0.0010005
+htnet_ablation:
+  branch_mixer: true
+  axis_channel_gate: false
+  input_skip_head: true