{ "release": "Proximal benchmark ID10-B200, v3 report", "report_date": "2026-11-15", "datasets": 2000, "samples_per_dataset": 500, "ID_tasks": 1000, "OOD_tasks": 1000, "query": "One CATE(x) per dataset; query_x is in both NPZ and task metadata.", "NPZ_fields": [ "X", "Z", "W", "A", "Y", "query_x" ], "latent_variables_released": false, "family_counts": { "id_mlp_relu": 113, "ood_qi_quadratic": 321, "ood_cui_linear": 343, "id_mlp_leaky_relu": 101, "ood_qi_trigonometric": 336, "id_mlp_rbf": 105, "id_mlp_identity": 97, "id_mlp_softplus": 96, "id_mlp_silu": 93, "id_mlp_sine": 100, "id_mlp_tanh": 100, "id_mlp_random_fourier": 90, "id_mlp_elu": 105 }, "ID": { "label": "training-support restriction/reweighting; not the exact original training-prior distribution", "family_sampling": "Uniform10 per task, then native random parameters and samples; no fixed per-family count", "families": [ { "family": "id_mlp_identity", "official_selector_index_zero_based": 3, "activation_class": "Identity", "weight": 0.1 }, { "family": "id_mlp_tanh", "official_selector_index_zero_based": 0, "activation_class": "Tanh", "weight": 0.1 }, { "family": "id_mlp_relu", "official_selector_index_zero_based": 6, "activation_class": "ReLU", "weight": 0.1 }, { "family": "id_mlp_leaky_relu", "official_selector_index_zero_based": 1, "activation_class": "LeakyReLU", "weight": 0.1 }, { "family": "id_mlp_elu", "official_selector_index_zero_based": 2, "activation_class": "ELU", "weight": 0.1 }, { "family": "id_mlp_softplus", "official_selector_index_zero_based": 7, "activation_class": "Softplus", "weight": 0.1 }, { "family": "id_mlp_silu", "official_selector_index_zero_based": 5, "activation_class": "SiLU", "weight": 0.1 }, { "family": "id_mlp_rbf", "official_selector_index_zero_based": 11, "activation_class": "RBFActivation", "weight": 0.1 }, { "family": "id_mlp_sine", "official_selector_index_zero_based": 14, "activation_class": "SineActivation", "weight": 0.1 }, { "family": "id_mlp_random_fourier", "official_selector_index_zero_based": 17, "activation_class": "RandomFunctionActivation", "weight": 0.1 } ], "paired_MC_samples": 4096, "native_queries_generated": 128, "selected_query_index": 0, "nominal_dimensions": { "X": 2, "U": 1, "W": 1, "Z": 1 }, "reference": "Finite paired-MC reference with saved truth_mc_se; not independently proven exact." }, "OOD": { "families": [ "ood_cui_linear", "ood_qi_quadratic", "ood_qi_trigonometric" ], "original_family_weights": [ 0.3333333333333333, 0.3333333333333333, 0.3333333333333333 ], "reference": "Analytic CATE", "shift_definition": "External generator family shift relative to MLP-SCM; not a proof of disjoint support.", "hyperparameters": "Continuous parameter hyperpriors; see the attached report for distributions and equations." }, "PFN": { "weights_updated": false, "bootstrap": false, "cross_fitting": false, "interval": "Native 95% equal-tailed credible interval" }, "RF": { "baseline": "Custom RF-assisted proximal series-IV; not a direct published P-learner replication", "bootstrap_replicates": 200, "bootstrap": "IID whole-row resampling with complete refit, fixed query and forest seeds; duplicates inherit original fold", "interval": "95% percentile with np.quantile method=linear", "cross_fitting": "Two-fold first stage only; full-data standardization, final bridge and W|X forest; not full DML" }, "metrics": { "error": "estimate-reference", "scale": "Observed sample SD(Y), ddof=1", "normalized_error": "(estimate-reference)/SD(Y)", "bias": "Mean error (raw or task-normalized)", "variance": "Sample variance of errors across tasks, ddof=1; not fixed-DGP sampling variance or within-bootstrap variance", "RMSE": "sqrt(mean(error^2))", "length": "mean(upper-lower), or per-task length/SD(Y)", "coverage": "Fraction of valid intervals covering the reference; unchanged by normalization" }, "timing": { "unit": "seconds", "workers": 16, "threads_per_worker": 1, "RF_n_jobs": 1, "includes": "Warm complete point+interval computation, including all 200 RF refits", "excludes": "Model load/init, independent warmup, DGP/oracle, NPZ input, common result/checkpoint I/O", "limitations": "Shared CPU; 335 complete OOD pairs retain historical results and time; 1665 pairs newly evaluated; not an exclusive-machine latency benchmark" }, "availability": { "method_records": 4000, "valid_method_records": 4000, "failed_method_records": 0, "RF_draws": 400000, "failed_RF_draws": 0 }, "exclusions": "No post-evaluation task exclusions in this run. The ten-family design incorporated earlier scale diagnostics before these tasks were generated.", "publication_scope": "Observed synthetic data and selected scalar records are published; third-party code/weights, original full bootstrap draw/resume ledgers and personal materials are not included." }