{
  "schema": "aner.public-generative-scientific-evidence-v1",
  "source": {
    "filename": "scientific_neural_refinement_confirmation_20261014.json",
    "sha256": "e45f76e1453d6b6ed18bec6319e45a60afa738d35d248ab8b2ffad85ae5cfbce",
    "protocol_filename": "scientific_neural_refinement_protocol_20261014.json",
    "protocol_sha256": "9558ad8fe836e08c32b82526af6803db05b80566f891dcbab73faaa18e290df2"
  },
  "scope": "Scientific field generation, not LLM quality. Three declared rank-one PDE ensembles on 5x5 grids; no extrapolation to general scientific fields or larger language models.",
  "cohort": {
    "fits": 36,
    "fits_per_system": 12,
    "held_out_conditions_per_fit": 3,
    "generated_fields_per_condition": 2048,
    "acceptance_checks": 1517,
    "all_original_acceptance_checks_passed": true
  },
  "settings": {
    "updates": 100000,
    "batch": 64,
    "rate": 0.0005,
    "width": 128,
    "timesteps": 128,
    "beta_start": 1e-05,
    "beta_end": 0.1,
    "loss_weighting": "velocity",
    "candidate_base": "iid_joint_neural",
    "schedule_features": true,
    "transport_strength": 0.25,
    "ema_decay": 0.999,
    "training_per_condition": 512,
    "generated_per_condition": 2048,
    "terminal_probe_multiple": 4,
    "L2_response_sigma": 2,
    "mean_mode_scale": 0.3
  },
  "comparisons": {
    "baseline": "IID joint neural calibration",
    "candidate": "Same model plus terminal transport, strength 0.25",
    "same_cohort": true,
    "systems": [
      {
        "system": 0,
        "label": "Poisson (bimodal)",
        "iid_variance_ratio_MAE": 0.01855471124318143,
        "candidate_variance_ratio_MAE": 0.017239053644351604,
        "candidate_minus_iid": {
          "CI95": [
            -0.0034908863144588478,
            0.00040052507976849914
          ],
          "CI99": [
            -0.004261959008327362,
            0.0007654221113592005
          ]
        }
      },
      {
        "system": 1,
        "label": "Screened anisotropic (Gamma)",
        "iid_variance_ratio_MAE": 0.07307404225838568,
        "candidate_variance_ratio_MAE": 0.07052277743790558,
        "candidate_minus_iid": {
          "CI95": [
            -0.005917323839113391,
            0.00034815106894670305
          ],
          "CI99": [
            -0.007042655137638325,
            0.001035271938320704
          ]
        }
      },
      {
        "system": 2,
        "label": "Heterogeneous diffusion (Laplace)",
        "iid_variance_ratio_MAE": 0.061381735015085474,
        "candidate_variance_ratio_MAE": 0.06045827247093668,
        "candidate_minus_iid": {
          "CI95": [
            -0.0033936227130459515,
            0.0011527471431311788
          ],
          "CI99": [
            -0.004276596654690835,
            0.0017203494744290999
          ]
        }
      }
    ],
    "pooled": {
      "iid_variance_ratio_MAE": 0.05100349617221753,
      "candidate_variance_ratio_MAE": 0.04940670118439795,
      "relative_reduction_percent": 3.1307559435295618,
      "candidate_minus_iid": {
        "CI95": [
          -0.003102737790177151,
          -0.00022213033313509255
        ],
        "CI99": [
          -0.0036635457546665058,
          0.00017466164199575774
        ]
      }
    }
  },
  "uncertainty": {
    "replicates": 20000,
    "seed": 625132,
    "unit": "complete paired fit blocks resampled within each of three systems; pooled systems equally weighted; same selections for all policies",
    "quantiles": "linear interpolation at .025,.975,.005,.995",
    "interpretation": "All three per-system paired 95% intervals include zero. The pooled paired 95% interval excludes zero; its 99% interval includes zero."
  },
  "physical_validation": {
    "iid_joint_neural": {
      "shape_fidelity_groups": 108,
      "standardized_ecf_RMSE_max": 0.06213465882369251,
      "mass_max_error": 3.777587775334441e-16,
      "boundary_max": 0,
      "pooled_offresponse_PDE_RMS": 0.02967763737086498,
      "pooled_total_nominal_forcing_PDE_RMS": 18.655346066582293,
      "full_PDE_MSE_over_true_stochastic_forcing_min": 0.8367342800448784,
      "full_PDE_MSE_over_true_stochastic_forcing_max": 1.210694026703313
    },
    "neural_transport": {
      "shape_fidelity_groups": 108,
      "standardized_ecf_RMSE_max": 0.062134658823692526,
      "mass_max_error": 3.6858325736608285e-16,
      "boundary_max": 0,
      "pooled_offresponse_PDE_RMS": 7.12876424033449e-14,
      "pooled_total_nominal_forcing_PDE_RMS": 18.657092801366076,
      "full_PDE_MSE_over_true_stochastic_forcing_min": 0.8362499945743516,
      "full_PDE_MSE_over_true_stochastic_forcing_max": 1.216076719543204
    }
  },
  "caveats": {
    "variance_MAE_definition": "Mean absolute error of the generated stochastic-mode variance ratio from one; lower is better. Fits and held-out conditions are equally weighted.",
    "candidate_variance_ratio_range": [
      0.82778914555551,
      1.2160449773094857
    ],
    "candidate_outliers_outside_tighter_0_8_to_1_2_band": [
      {
        "system": 1,
        "fit_index": 9,
        "condition": -0.525,
        "ratio": 1.2160449773094857
      }
    ],
    "PDE_metrics": "Off-response PDE RMS excludes the true stochastic forcing direction; total nominal-forcing PDE RMS remains nonzero and is reported separately.",
    "shape_negative_control": "Gaussian moment closure is not the candidate. It passes zero of 108 non-Gaussian shape groups despite lower pooled variance MAE."
  }
}