{
  "status": "pass",
  "comparison": "BF16 DFlash2 vs selective block-FP8 weights dequantized to BF16",
  "samples": 8,
  "draft_positions": 56,
  "proposal_kl_mean": 0.0018929995239503794,
  "proposal_kl_median": 0.001594925176043174,
  "proposal_kl_p95": 0.00300999847240746,
  "proposal_kl_max": 0.008521616478903316,
  "proposal_top1_agreement": 0.9642857313156128,
  "selector_path_token_agreement": 0.8571429252624512,
  "selector_full_path_agreement": 0.75,
  "top16_candidate_overlap": 0.9765625,
  "draft_hidden_relative_rmse": 0.03063499420217177,
  "draft_hidden_cosine_similarity": 0.9995341897010803,
  "teacher_mean_forward_seconds": 0.20252476473979186,
  "student_mean_forward_seconds": 0.5098142223869218,
  "per_sample_selector_token_agreement": [
    1.0,
    1.0,
    1.0,
    1.0,
    0.0,
    1.0,
    0.8571429252624512,
    1.0
  ],
  "gates": {
    "proposal_kl_mean_max": 0.02,
    "proposal_top1_agreement_min": 0.9,
    "selector_path_token_agreement_min": 0.85,
    "top16_candidate_overlap_min": 0.95
  },
  "limitation": "RTX 3090 dequantizes FP8 weights to BF16. This test uses real target hidden states and the exact target LM head, but it is not an end-to-end native W8A8 throughput or acceptance benchmark."
}