File size: 5,948 Bytes
fe31b61 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 | {
"id": "external_decision_sufficiency_01",
"scope": "External recorded-data validation of frozen representation construction and exact bounded planning; sequential feature reveal is authored here, not supplied by the data providers.",
"device": "cuda:2",
"dtype": "float64",
"new_neural_modules": 0,
"backbone_executions": 0,
"sources": {
"hydraulic": {"uci_id":447,"doi":"10.24432/C5CW21","url":"https://archive.ics.uci.edu/static/public/447/condition+monitoring+of+hydraulic+systems.zip","license":"UCI CC BY 4.0","features":["PS1","PS2","PS3","PS4","FS1","TS1","EPS1","VS1"],"aggregation":"arithmetic mean of all samples of each recorded 60-second cycle","tasks":["cooler","valve","pump","accumulator"],"normal":[100,100,0,130],"context_column":[1,0,0,0],"source_context":[100,100,100,100],"target_contexts":[[73,80,90],[3,20],[3,20],[3,20]],"group":"all cycles with identical four physical condition labels, irrespective of stability; SHA256(profile) mod 2 separates adaptation and evaluation"},
"gas": {"uci_id":224,"doi":"10.24432/C5RP6W","url":"https://archive.ics.uci.edu/static/public/224/gas+sensor+array+drift+dataset.zip","license":"UCI CC BY 4.0; legacy description additionally restricts use to research, which this study respects","features_zero_based":[1,9,17,25,33,41,49,57],"feature_meaning":"normalized resistance response from sensors 1 through 8, in provider order","tasks":"all 15 lexicographic pairs of gas labels 1 through 6","source_batches":[1,2,3],"target_batches":[6,7,8,9,10],"group":"within each batch, contiguous blocks of 16 original rows before pair filtering; SHA256(batch,block) mod 2 separates adaptation and evaluation"}
},
"adapter": {"current_bits":6,"query_bits":2,"binarization":"strictly greater than each feature's source-only median; source unlabeled preprocessing uses all source records and is charged","no_target_normalization":true,"split_seed":20101,"sample_seeds":[20111,20112,20113],"samples_without_replacement":true,"source_labels_maximum":1024},
"eligibility": {"source_minimum":64,"adaptation_minimum_for_primary":64,"evaluation_minimum":32,"evaluation_minimum_per_hypothesis":8,"eligible_groups_minimum":4,"exclude_only_for_registered_sample_support":true,"all_exclusions_reported":true},
"regimes": ["source_only","target_64","target_256","source_plus_target_64"],
"estimation_models": ["support_mixture","joint_jeffreys","naive_bayes"],
"representations": ["query_joint","posterior_clusters","posterior_bins","future_marginal","raw_information","raw_entropy","hash_code","identity"],
"planner": {"reuse":"unionmind.decision_sufficiency.core unchanged","utility_correct":1,"utility_wrong":-2,"utility_abstain":0,"query_budgets":[0,1,2],"query_costs":[0.0,0.01,0.05],"cost_interpretation":"modeled feature-read utility cost; not physical sensor cost, transport bytes or engineer time"},
"primary": {"regime":"target_64","model":"support_mixture","representation":"query_joint","aggregation":"mean over query budgets 1,2, all three costs, three sample seeds, equal tasks within target environment, equal target environments within dataset, equal datasets","compact_controls":["posterior_clusters","posterior_bins","future_marginal","raw_information","raw_entropy","hash_code"],"effect_minimum":0.01,"confidence":"95% simultaneous Bonferroni intervals across six compact controls, paired target-environment block bootstrap within each of the two fixed datasets; tasks, adaptation seeds, costs and budgets are NOT independent replications","bootstrap_draws":10000,"bootstrap_seed":20131,"gate":"gain >= .01 with positive simultaneous lower bound against every compact control, and positive point gain over every compact control in each dataset separately; at least two eligible target environments per dataset"},
"full_reference_gate": {"comparison":"all three identity models on the same examples","noninferiority_margin":0.01,"confidence":"simultaneous intervals with family of nine comparisons including compact controls"},
"cost_gate": {"minimum_complete_time_reduction":0.2,"requests":[1,32,1024],"repeats":11,"order":"alternating forward/reverse; common CUDA/library warmup separately charged to all","components":["download","decompression and raw feature extraction","unlabeled source median and event encoding","adaptation sample materialization","counting","model estimation","representation construction","probability pushforward","exact planning","policy replay"],"gate":"20% median complete savings against the fastest full statistical alternative at every request count; scientific gain alone cannot promote a slower algorithm"},
"fidelity": {"representation":"exact within-fitted-model compression loss and conditional total variation; no claim of true-law sufficiency","estimation":"held-record posterior log loss, Brier, and joint predictive log loss; test-empirical hindsight oracle used only for descriptive decomposition, never deployable inference or a population certificate"},
"integrity": {"lock_before":"download and opening any numerical source records","retain_all_previous_files":true,"no_threshold_changes_after_opening":true,"all_eligible_tasks_reported":true,"no_test_selected_recipe":true,"sequential_replay":"withheld bits accessed only after the corresponding query action; compare independent event replay to aggregated evaluator","no_human_dependency":true},
"falsification": ["A failed primary gate rejects external compact-representation improvement for this registered task.","Improvement only at 256 labels is secondary and cannot rescue the 64-label primary claim.","No end-to-end savings rejects computational superiority even if scientific representation benefit is positive.","Two laboratory datasets do not establish production reliability or general transfer across operational environments.","The full reference is the uncompressed eight-binary-feature estimated model, not all raw sensor data."]
}
|