| { |
| "id": "factorial_transfer_01", |
| "objective": "Test a substantial full-hypothesis, scarce-label diagnostic capability: compose previously unseen combinations of equipment conditions and gas levels, estimate their numerical observation mechanisms, and use verified probability and decision operators. Preserve the entire preceding research program.", |
| "motivation": "Opened pairwise audits found raw numerical evidence useful, but 8/36-parameter per-task metric learning added no material advantage. The new challenge removes the oracle two-label restriction and uses 64 total fitting examples for the complete ontology, rather than 64 per label pair.", |
| "backbone": "This experiment explicitly uses statistical and <=20000-parameter auxiliary world models, without a language backbone or collaborative execution. Original frozen Qwen and cognitive checkpoints remain protected.", |
| "device": "cuda:1", |
| "analysis_device": "cuda:2", |
| "statistical_dtype": "float64", |
| "neural_training_dtype": "float32", |
| "sources": { |
| "hydraulic": { |
| "uci": 447, |
| "doi": "10.24432/C5CW21", |
| "authors": [ |
| "Nikolai Helwig", |
| "Eliseo Pignanelli", |
| "Andreas Schuetze" |
| ], |
| "url": "https://archive.ics.uci.edu/static/public/447/condition+monitoring+of+hydraulic+systems.zip", |
| "system": "hydraulic_rig", |
| "license": "CC BY 4.0", |
| "cycle_seconds": 60, |
| "physical_sensors": { |
| "PS1": [ |
| "bar", |
| 100 |
| ], |
| "PS2": [ |
| "bar", |
| 100 |
| ], |
| "PS3": [ |
| "bar", |
| 100 |
| ], |
| "PS4": [ |
| "bar", |
| 100 |
| ], |
| "PS5": [ |
| "bar", |
| 100 |
| ], |
| "PS6": [ |
| "bar", |
| 100 |
| ], |
| "EPS1": [ |
| "W", |
| 100 |
| ], |
| "FS1": [ |
| "l/min", |
| 10 |
| ], |
| "FS2": [ |
| "l/min", |
| 10 |
| ], |
| "TS1": [ |
| "degC", |
| 1 |
| ], |
| "TS2": [ |
| "degC", |
| 1 |
| ], |
| "TS3": [ |
| "degC", |
| 1 |
| ], |
| "TS4": [ |
| "degC", |
| 1 |
| ], |
| "VS1": [ |
| "mm/s", |
| 1 |
| ] |
| }, |
| "excluded_virtual_sensors": [ |
| "CE", |
| "CP", |
| "SE" |
| ], |
| "factor_names": [ |
| "cooler_condition_percent", |
| "valve_condition_percent", |
| "pump_leakage_grade", |
| "accumulator_pressure_bar" |
| ], |
| "levels": [ |
| [ |
| 3, |
| 20, |
| 100 |
| ], |
| [ |
| 73, |
| 80, |
| 90, |
| 100 |
| ], |
| [ |
| 0, |
| 1, |
| 2 |
| ], |
| [ |
| 90, |
| 100, |
| 115, |
| 130 |
| ] |
| ], |
| "features": "Mean of each physical channel in all four 15-second cycle quarters, 56 values. Retain all physical channels; no outcome-based feature selection. Stable/transient source flags are reported separately and never exclude a cycle.", |
| "record_group": "original five-cycle blocks", |
| "opened_archive": "artifacts/external_decision_sufficiency_01/sources/hydraulic.zip" |
| }, |
| "gas": { |
| "uci": 322, |
| "doi": "10.24432/C5WP4C", |
| "authors": [ |
| "Jordi Fonollosa" |
| ], |
| "url": "https://archive.ics.uci.edu/static/public/322/gas+sensor+array+under+dynamic+gas+mixtures.zip", |
| "system": "gas_array_rig", |
| "license": "CC BY 4.0 repository metadata; source description additionally restricts use to research, which is respected", |
| "files": { |
| "gas_co": { |
| "file": "ethylene_CO.txt", |
| "primary_max_ppm": 600 |
| }, |
| "gas_methane": { |
| "file": "ethylene_methane.txt", |
| "primary_max_ppm": 300 |
| } |
| }, |
| "factor_names": [ |
| "primary_gas_level", |
| "ethylene_level" |
| ], |
| "levels": [ |
| [ |
| 0, |
| 1, |
| 2, |
| 3 |
| ], |
| [ |
| 0, |
| 1, |
| 2, |
| 3 |
| ] |
| ], |
| "physical_units": "Timestamp seconds; setpoints ppm; 16 recorded sensor outputs. The author writes the resistance conversion literally as '40.000/S_i'. This experiment retains source readings and does not interpret or perform an unverified resistance-unit conversion.", |
| "binning": "Primary: zero, (0,max/4], (max/4,max/2], (max/2,max]; ethylene: zero, (0,5], (5,10], (10,20]. Fixed before data access.", |
| "episode": "Maximal contiguous constant setpoint pair; require duration >=75 seconds and strictly increasing timestamps. One episode is one observation, never millions of independent 100-Hz rows.", |
| "features": "For every sensor use mean in last 20 seconds and last-20-minus-first-20-second mean, 32 values. No sensor subset. Retain every short-episode exclusion with raw row/time provenance.", |
| "record_group": "original exposure episode; two mixtures share a physical system and are not counted as two independent systems" |
| } |
| }, |
| "hypothesis_spaces": { |
| "hydraulic": 144, |
| "gas_co": 16, |
| "gas_methane": 16 |
| }, |
| "source_access": "Hydraulic archive and labels were opened by external_decision_sufficiency_01. Its new full-ontology development audit is explicitly exploratory. The UCI322 gas-mixture archive, numeric values and labels have not been accessed before this lock. The two fresh experiments share one physical sensor platform: they are not two independently sampled systems.", |
| "splits": "Before fitting: sort the entire supplied Cartesian hypothesis ontology by SHA256('20711:'+dataset+':'+comma-separated integer factor indices), holding the first max(4,ceil(.2*ontology_size)) bindings. No sensor value or observed label count chooses bindings. Among other bindings, original record-group hash SHA256('20721:'+dataset+':'+group_index) mod 5==0 is seen-binding held. Source excludes held rows and groups adjacent to any held group. Do not change split seeds after access.", |
| "eligibility": "Primary requires >=64 source observations, >=32 unseen-binding held observations in >=3 distinct bindings, and >=16 seen-binding held observations. Secondary 128-label result requires >=128 source observations and is otherwise explicitly unavailable. Keep every series and failure; no replacement or threshold amendment.", |
| "sample_budgets": [ |
| 64, |
| 128 |
| ], |
| "sample_seeds": [ |
| 20751, |
| 20752, |
| 20753 |
| ], |
| "sampling": "Same random source prefix for every model and both budgets, without replacement, generator seed=sample_seed + series_index*101. No class balancing, label-aware prefix selection or target tuning. Report factor levels absent from each actual fitting prefix and supported/unseen-level strata without dropping cases.", |
| "normalization": "Feature mean and standard deviation from the actual 64/128 fitting records only. No full-source label-conditional quantiles. Standard deviation floor 1e-6. All models receive identical normalized numerical evidence.", |
| "models": [ |
| "additive_full", |
| "additive_diag", |
| "factorized_neural", |
| "joint_neural", |
| "factorized_nb", |
| "factorized_lda", |
| "factorized_kernel", |
| "factorized_ridge", |
| "joint_gaussian", |
| "factorized_logistic", |
| "factorized_classifier" |
| ], |
| "candidates": [ |
| "additive_full", |
| "factorized_neural" |
| ], |
| "world_models": "Additive: all factor one-hots plus intercept, ridge 8, intercept ridge .01, analytic leave-one-out residuals. Factorized neural: full factor one-hots -> SiLU hidden -> observation means. Joint neural: one-hot complete binding -> matched SiLU hidden -> observation means, unknown bindings use uninformative source mean. Hidden widths give <=20000 counted trainable parameters and near-matched factorized/nonfactorized budgets. Both neural decoders fit 400 AdamW steps, lr .003, weight_decay .01; four source folds estimate held residuals plus one full fit, all charged. Source residual covariance is .5 full + .5 diagonal with .05 identity floor; diagonal ablation uses only the diagonal. Priors uniform over the supplied full hypothesis ontology, including unseen bindings.", |
| "strong_controls": "Independent-factor Gaussian NB and shared-covariance Gaussian, independent-factor whitened kernel with bandwidth .7 and .5 Dirichlet smoothing, independent-factor ridge classifier with ridge 8, joint-binding Gaussian with shrink-to-source fallback. Use all observed coordinates and the same fitting labels. All factorized controls can predict unseen bindings; do not make the comparison only against memorization. Also include class-balanced multinomial logistic regression and a matched <=20000-parameter SiLU direct classifier, both predicting independent factor probabilities, with the same 400-step optimizer and four out-of-fitting-fold source predictions plus full fit. Their four-point source temperature grid (.5,1,2,4) is chosen only on source cross-fit log loss; every fit is charged. All source- and target-state priors remain uniform.", |
| "uncertainty_controls": "No target calibration. Neural observation uncertainty comes from four out-of-fitting-fold source predictions; additive uncertainty uses analytic LOO residuals. Report posterior and conditional observation proper scores, confidence/coverage and source-support gaps. No operational calibration claim without its own validation.", |
| "primary": "At 64 total fitting labels, joint Top-1 and Top-3 on all unseen-binding held observations in the two fresh gas mixtures, equal weight per mixture. The previously opened hydraulic rig is an explicit development result and contributes no weight to fresh-confirmation estimates. Seen-binding, unsupported factor levels, all series and both label budgets are mandatory. Top-3 uses uniform tie credit with tolerance1e-12; deterministic results retained.", |
| "uncertainty": { |
| "draws": 20000, |
| "seed": 20791, |
| "family": 40, |
| "unit": "Shared original held recording groups, same weights across models/seeds. Gas experiments have equal weight. Conditional on one fixed sensor platform and its two experiments, not a population-of-systems interval. Hydraulic estimates remain exploratory.", |
| "contrasts": "two frozen candidates versus their ten alternatives at both declared label budgets; unavailable secondary cells retained and no comparison family shrunk" |
| }, |
| "success": { |
| "substantial_diagnostic_gain": "At least .10 fresh unseen-binding joint Top-3 gain over every one of the seven strong conventional/direct-classifier controls, positive simultaneous lower bounds, positive point gain in each fresh gas experiment. All three series must meet eligibility for the complete planned study; the opened hydraulic result cannot establish fresh multi-system generalization.", |
| "strongest_available": "Positive simultaneous lower bound against every eligible alternative, including matched structured and neural controls. Do not call a method the overall winner if this fails.", |
| "learning_specific": "The factorized neural candidate must beat the additive reference with positive simultaneous interval before claiming neural learning adds value.", |
| "promotion": "Substantial diagnostic gain plus strongest-available gate, complete compute ledger and intact cohort. A favorable secondary does not erase a failed primary." |
| }, |
| "secondary_acquisition": { |
| "world_models": [ |
| "additive_full", |
| "factorized_neural" |
| ], |
| "cohort": "At most 32 unseen-binding held records per series, ordered by SHA256('20781:'+raw_record_id), frozen before scores; every primary record remains in the diagnostic evaluation.", |
| "initial_observed_coordinates": [], |
| "budgets": [ |
| 2, |
| 4, |
| 8 |
| ], |
| "per_scalar_recorded_query_cost": 0.01, |
| "policies": [ |
| "value_forced", |
| "value_stop", |
| "variance_forced", |
| "variance_stop", |
| "static_forced", |
| "random_forced" |
| ], |
| "value": "Exact one-step expected joint Top-3 coverage using shared-variance Gaussian posterior-rank intervals and normal CDFs. No access to an actual withheld value before paying its query. Conditional variance is the strongest cheap adaptive control; variance_stop evaluates only its selected query's decision value. No claim of globally optimal multistep control.", |
| "scale_gate": "Either >=.10 Top-3 gain at equal query cap over every cheap control, or >=50% fewer recorded queries with simultaneous noninferiority loss <=.02, plus full computation. This cannot rescue a failed diagnostic or learning-specific gate.", |
| "uncertainty": "20000 paired recording-group bootstrap draws, family120, seed20811, two fresh gas experiments equally weighted. Report every policy and both opened/fresh provenance." |
| }, |
| "compute": "CUDA float64 statistics, probability operations, exact interval decisions and bootstrap; float32 auxiliary neural fitting. CPU archive text parsing, provenance, hashing and report aggregation only. Count data preparation, normalization, every full/cross-fit adaptation, full-hypothesis mean materialization, covariance construction, predictions, query decisions, peak GPU memory, trainable/checkpoint/derived-state bytes. No CPU offload, DeepSpeed, Qwen executions or model collaboration.", |
| "stopping": "Freeze all code, ontology, features, priors, controls and gates before new gas download. Hydraulic development has been opened and all results remain exploratory. No architecture, threshold, split or favorable subset amendment after new confirmation access. Preserve all failures, previous checkpoints and the evidence engine. No manual review dependency or paper/deployment objective." |
| } |
|
|
