{
  "task": "PHOENIX fire-detection feasibility (SEVIRI-calibrated, sim-first)",
  "discipline": "each quantum method vs mandatory classical baseline, same metric; placebo where learned; nothing promoted; escalate survivors to real corpus",
  "screens": [
    {
      "method": "58/158 quantum kernel (QSVM, ZZ feature map)",
      "task": "fire vs non-fire (11-band)",
      "quantum_metric": 0.6148,
      "classical_metric": 0.7131,
      "metric": "AUC",
      "advantage": false,
      "latency_s": 167.8,
      "extra": {
        "rbf_svm_full": 0.6657,
        "n": 90
      },
      "verdict": ""
    },
    {
      "method": "158 quantum kernel (PennyLane angle embedding)",
      "task": "fire vs non-fire (11-band)",
      "quantum_metric": 0.64,
      "classical_metric": 0.7131,
      "metric": "AUC",
      "advantage": false,
      "latency_s": 46.3,
      "extra": {
        "n": 90
      },
      "verdict": ""
    },
    {
      "method": "63/161 tensor network (TT-SVD compress)",
      "task": "fire vs non-fire spectral compression",
      "quantum_metric": 0.4715,
      "classical_metric": 0.6664,
      "metric": "AUC",
      "advantage": false,
      "latency_s": 1.9,
      "extra": {
        "tt_features": 32,
        "bond_dim": 3
      },
      "verdict": ""
    },
    {
      "method": "59 quantum clustering (QUBO norm-cut, SA)",
      "task": "unsupervised fire/non-fire split",
      "quantum_metric": 0.6375,
      "classical_metric": 0.65,
      "metric": "AUC(unsup)",
      "advantage": false,
      "latency_s": 1.5,
      "extra": {
        "spectral": 0.6,
        "kmeans": 0.65,
        "gmm": 0.5875
      },
      "verdict": ""
    },
    {
      "method": "RX/anomaly quantum one-class kernel",
      "task": "fire-as-anomaly detection",
      "quantum_metric": 1.0,
      "classical_metric": 1.0,
      "metric": "AUC",
      "advantage": false,
      "latency_s": 90.8,
      "extra": {
        "isolation_forest": 0.718,
        "ocsvm_rbf_tuned": 1.0,
        "note": "apparent q=1.000 vs default-classical 0.777 refuted: gamma-tuned classical OCSVM also reaches 1.000 on cleanly-separable synthetic data. No advantage once baseline is fair."
      },
      "verdict": "MIRAGE REFUTED -- tuned classical matches quantum (both 1.000)."
    },
    {
      "method": "61/62 QFT deconvolution",
      "task": "thermal PSF deblur (reconstruction)",
      "quantum_metric": -0.0271,
      "classical_metric": -0.0271,
      "metric": "-MSE",
      "advantage": false,
      "latency_s": 0.0,
      "extra": {
        "mse_qft": 0.0271,
        "mse_classical": 0.0271,
        "note": "QFT realises the same DFT as FFT; identical result, and amplitude-loading a classical signal costs more than it saves"
      },
      "verdict": ""
    },
    {
      "method": "57/65 quantum-inspired particle filter",
      "task": "hotspot track estimation",
      "quantum_metric": -0.549,
      "classical_metric": -0.5424,
      "metric": "-RMSE",
      "advantage": false,
      "latency_s": 0.0,
      "extra": {
        "rmse_qinspired": 0.549,
        "rmse_classical": 0.5424,
        "note": "amplitude-proportional resampling reduces to systematic resampling"
      },
      "verdict": ""
    },
    {
      "method": "64/160 quantum graph fusion (QAOA)",
      "task": "sensor-agreement MRF fusion",
      "quantum_metric": 16.564,
      "classical_metric": 16.564,
      "metric": "-Ising E",
      "advantage": false,
      "latency_s": 5.9,
      "extra": {
        "exact_min": -16.564,
        "qaoa_min": -16.564,
        "note": "optimization family; confirms entry 0093 -- no advantage over exact"
      },
      "verdict": ""
    },
    {
      "method": "60 quantum image rep (FRQI)",
      "task": "fire-patch vs non-fire-patch",
      "quantum_metric": 0.8678,
      "classical_metric": 0.8697,
      "metric": "AUC",
      "advantage": false,
      "latency_s": 2.1,
      "extra": {
        "note": "FRQI angle-encoding is an invertible re-parameterisation of raw pixels"
      },
      "verdict": ""
    }
  ],
  "stacking": {
    "per_method": {
      "classical_gbm": [
        0.5308333333333333,
        0.021362041231014314
      ],
      "quantum_kernel": [
        0.4633333333333333,
        0.029227614891096047
      ],
      "tensor_net": [
        0.564375,
        0.0735502237703372
      ],
      "stack_all": [
        0.47833333333333333,
        0.12346698839050953
      ],
      "stack_q_plus_classical": [
        0.43770833333333337,
        0.06793903006045612
      ]
    },
    "best_single": 0.5644,
    "best_stack": 0.4783,
    "emergent_lift": -0.086,
    "latency_s": 702.0,
    "verdict": "No emergent value: the stack does not beat the best single base beyond noise; the classical incumbent holds."
  },
  "screens_with_advantage": 0,
  "overall_verdict": "Full simulable quantum column screened (9 methods) + stacking on real-shaped PHOENIX detection tasks. ZERO methods beat a FAIR classical baseline: the only apparent win (quantum one-class anomaly AUC 1.000) was a mirage from an under-tuned classical baseline on cleanly-separable synthetic data -- a gamma-tuned classical OCSVM also reaches 1.000. Stacking/hybridising the methods produced NEGATIVE emergent lift (best_stack < best_single). Consistent with entries 0091/0093: at PHOENIX scale quantum is a benchmark, not a detection capability. The one deployable quantum-adjacent value is post-quantum signed alerts (ML-DSA-65, pqc_alerting.py), which is orthogonal to detection and now built.",
  "mirage_refutation": {
    "screen": "RX/anomaly quantum one-class kernel",
    "apparent_quantum_auc": 1.0,
    "default_classical_auc": 0.777,
    "tuned_classical_auc": 1.0,
    "conclusion": "under-tuned-baseline + clean-synthetic artifact; refuted per skeptical-first discipline"
  }
}