nur-dev's picture
Publish reproducibility artifacts (part 2)
fe31b61 verified
Raw History Blame Contribute Delete
2.9 kB
{
"frozen_code_changed": false,
"protocol_changed": false,
"promotion_gates_changed": false,
"device": "cuda:2",
"gpu": "NVIDIA H200",
"model_and_policy_math": "CUDA float64; no CPU offload",
"backbone_executions": 0,
"active_backbone_copies": 0,
"neural_parameters_added": 0,
"controller_executions": 0,
"controller_refits": 0,
"maximum_peak_allocated_bytes": 134986752,
"maximum_peak_reserved_bytes": 155189248,
"confirmation_research_seconds": 40.3597513709974,
"confirmation_scored_records": 1022976,
"cuda_startup_seconds": 18.617308109998703,
"cuda_warmups": {
"support_mixture:query_joint": 0.2577872660185676,
"joint_jeffreys:identity": 0.0032121709955390543,
"naive_bayes:identity": 0.003511357994284481,
"support_mixture:identity": 0.003404008981306106
},
"benchmark_records": 176,
"fitted_statistical_model": {
"raw_probability_cells": 512,
"raw_multinomial_degrees_of_freedom": 511,
"naive_bayes_degrees_of_freedom": 17,
"support_mixture_weight": "N/(N+512), not fitted",
"clustering_centers_query_joint": 64,
"clustering_centers_posterior": 16,
"centers_retained_at_deployment": false,
"iterations": 32
},
"per_world_state_bytes": {
"encoded_joint": 512,
"encoded_lookup": 64,
"encoded_planner_choices": 5184,
"encoded_planner_values": 5184,
"planner_expected_summary": 72,
"encoded_total_with_planner": 11016,
"uncompressed_joint": 4096,
"uncompressed_encoder_lookup": 0,
"uncompressed_planner_choices": 41472,
"uncompressed_planner_values": 41472,
"uncompressed_total_with_planner": 87112,
"raw_count_table": 4096,
"original_full_fit_model": 4096,
"primary_training_records": 8192,
"measured_primary_live_buffers_including_fit_and_records": 27400,
"measured_uncompressed_live_buffers_including_fit_and_records": 99400,
"note": "Tensor payloads; Python object/container and CUDA context overhead excluded from these payload counts and represented by measured allocation/startup separately"
},
"accounting": {
"truth_generation": "Research/evaluation only; never supplied as policy model",
"statistical_fit_and_encoding": "Charged in construction; all label budgets and adaptation stages in timings.json",
"probability_materialization": "Charged separately in every stage; dense fit is not claimed free",
"exact_planning": "Charged, with all hypothetical branches; full true-law oracles are evaluator-only",
"warmup": "Charged; first alternative includes common library initialization, ordering limitation disclosed",
"interpretation": "Local batched workload, not production latency, engineering investigation time or evidence-backend throughput"
},
"research_stage_timing_source": "confirmation_evaluation/timings.json",
"all_complete_benchmark_records": "confirmation_benchmark.json"
}