{
  "schema": "atlas_benchmark_manifest/v1",
  "generated_utc": "2026-06-22T22:28:04Z",
  "generator": "benchmarks/build_benchmark_manifest.py",
  "corpus_sha256": "66f9d6fa672385d61a1605fd1adcd424bccc37e33cbeedcc7c2962fc06e6af22",
  "corpus_hash_detail": {
    "method": "sha256( sha256(canonical (seed,params) tuples) + ':' + sha256(concat of each referenced .qasm: relpath\\0bytes\\0) )",
    "seeds_params_sha256": "7037a9ca14665e8ffaab4fd41cf8023b2964c0335ab49053f6fb55a42a07743a",
    "qasm_bytes_sha256": "bdfb40a704fa0b77c2d361a1f1e95b6f6c8b3e4afe588ad32aad737209f2f071",
    "missing_qasm_files": [],
    "note": "corpus is generated from (seed,params); we hash both the generation tuples and the produced circuit bytes so either a seed/param change or a circuit-content change moves the hash."
  },
  "num_circuits": 800,
  "generation_seeds": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7
  ],
  "strata": {
    "families": [
      "all_to_all_sparse",
      "dense_core",
      "grid",
      "heavy_hex",
      "ladder",
      "line",
      "star"
    ],
    "n_values": [
      8,
      12,
      16,
      20,
      24,
      30
    ],
    "depths": [
      "high",
      "low",
      "med"
    ],
    "t_density": [
      "clifford",
      "high",
      "low",
      "med"
    ],
    "axes": "family x n x depth(layers) x t_density x seed (stratified, deterministic enumeration; see run_scaled_validation.generate_corpus)"
  },
  "oracle_source": "measured runtime/memory / exact simulator ground truth (Stim exact-Clifford / non-truncated MPS bond / exact statevector / exact treewidth) \u2014 NOT hand-labels. Each exact certificate is mapped through the adjudicator's own cost thresholds and the cheapest certified route wins (see benchmarks/oracle.py).",
  "expected_route_definition": "expected_route = oracle_route = cheapest EXACT-certified route for the circuit (CPU<TENSOR<HPC_FIRST<ESCALATE). Circuits with no exact certificate are flagged oracle_certified=false and excluded from accuracy numbers (oracle.py:oracle_route).",
  "false_alarm_definition": "Atlas routes HARD (hpc_first/escalate) a circuit the oracle certifies CHEAP (cpu/tensor): a conservative/expensive error, not dangerous.",
  "false_security_definition": "Atlas routes CHEAP (cpu/tensor) a circuit the oracle certifies HARD (hpc_first/escalate): the dangerous error (calling 'easy' something that is not). Measured only where exact classical ground truth exists.",
  "train_test_split": {
    "kind": "evaluation-only",
    "detail": "Atlas is NOT fitted to this corpus; it is a pure held-out evaluation set. No train split exists for the router. The conformal coverage guarantee uses a separate selection/validation split (see atlas_conformal.py); thresholds are wall-clock-calibrated independently (threshold_calibration.md), not tuned on this corpus."
  },
  "package_versions": {
    "python": "3.14.5",
    "qiskit": "2.4.2",
    "stim": "1.16.0",
    "quimb": "1.14.0",
    "cotengra": "0.8.1"
  },
  "reproduction_command": "python3 benchmarks/run_scaled_validation.py 800  # regenerates corpus (scaled_manifest.json + circuits_scaled/*.qasm) and the oracle-certified results_scaled/scaled_results.csv ; then python3 benchmarks/build_benchmark_manifest.py  # rebuilds this manifest",
  "wilson_ci_false_security": {
    "certified_rows": 2517,
    "csv_sources": [
      "scaled_results.csv",
      "scaled_results_ext.csv",
      "scaled_results_moat.csv"
    ],
    "hard_verified_subset": 25,
    "false_security_count": 1,
    "false_security_rate_hard_verified": 0.04,
    "wilson_ci95_hard_verified": [
      0.0071,
      0.1954
    ],
    "ci_warning": "small-n: the hard-verified subset (exact classical ground truth exists only below the memory wall) is tiny, so the interval is wide; the genuinely quantum-hard regime is unmeasurable (BQP!=BPP)."
  },
  "scope": {
    "oracle_basis": "exact only (Stim Clifford / non-truncated MPS / statevector / exact treewidth)",
    "not_claimed": [
      "classical impossibility",
      "quantum advantage"
    ],
    "escalate_unmeasured": "the genuinely quantum-hard regime has no classical ground truth; false-security there is unmeasurable by construction."
  }
}