{
  "fire_task": {
    "fidelity_kernel": 0.598,
    "projected_kernel": 0.483,
    "trainable_kernel": 0.555,
    "best_quantum": 0.598,
    "quantum_beats_classical": false,
    "latency_s": 136.9,
    "classical_4dim_handicapped": 0.55,
    "classical_full_11band_fair": 0.658,
    "note": "apparent 'win' (0.598 vs 4-dim-classical 0.55) is an artifact: the classical baseline was restricted to the 4 PCA dims the 4-qubit circuit needs. The FAIR full-11-band classical (0.658 same-seed / 0.573 mean at n=80 / ~0.64 at scale) beats the best quantum kernel (0.598), and the 0.05 gap is within n=80 noise. No advantage."
  },
  "quantum_native_control": {
    "classical_best": 0.675,
    "fidelity_kernel": 0.792,
    "projected_kernel": 0.5,
    "trainable_kernel": 0.5,
    "best_quantum": 0.792,
    "quantum_beats_classical": true,
    "latency_s": 133.1,
    "note": "CONTROL PASSES: on data generated FROM the quantum embedding, the matching fidelity kernel (0.792) clearly beats classical (0.675) -- proving the test CAN detect a quantum advantage when the data has quantum-native structure."
  },
  "verdict": "Best-shot quantum ML (IQP fidelity/projected/trainable kernels) does NOT open the door on real fire data: even given its best shot, the top quantum kernel (0.598) stays within noise of and below the fair full-band classical baseline (0.658 same-seed / ~0.64 at scale). The quantum-native CONTROL confirms the test is honest -- quantum wins there (0.792 vs 0.675). Precise conclusion: quantum ML helps only when data has quantum-native structure, which fire-detection data does not. Consistent with entries 0091/0093/0094."
}