| { |
| "schema": "industrial.well-health.benchmark.protocol.v1", |
| "reviewed_on": "2026-10-04", |
| "task": "Retrospective causal binary undesirable-event detection, including labelled transients; not a pre-onset forecast or physical well-integrity diagnosis.", |
| "source_revision": "6a13bd21a02a2ce6ce23a02b1c2ba2f73770752b", |
| "corpus_audit_sha256": "de221eed3ddc1a1a09d7d8c31be0a4758afd06bf03bd20baa67de23144941717", |
| "archive_sha256": "c037a6a9d5dcc9b2add7b9ea413f5cff367c57bad3a93faac1d1d32da27fcbea", |
| "metadata_access": "Complete schema, label support/runs, missingness, clock spans and duplicate metadata were inspected before this freeze. Only the 12 prospective development samples had local sensor-value access. No confirmation values, fitted scores or distribution summaries were inspected.", |
| "split_seed": "3w-causal-well-groups-20261004-v1", |
| "group_split": "Quarantined groups go to training. Other groups stratified into normal-only or positive-supported, sorted by SHA256(seed + ':' + group). First ceil(20%) reserved fresh foundation confirmation, next ceil(20%) initial confirmation, next ceil(20%) validation, remainder training. No class tailoring or value-based split selection.", |
| "quarantine": ["WELL-00001", "WELL-00002", "WELL-00010", "WELL-00015", "WELL-00019"], |
| "record_overlap": "Within each real group, earliest-finish interval scheduling with start/path tie breaks; accept a complete recording only when its start is strictly later than the last accepted end. Retain every excluded path/reason. Do not stitch files. Intervals include unknown prefixes.", |
| "origins": "Real recordings only in this initial study; simulated/drawn data are not silently added to training or confirmation.", |
| "channels": ["P-ANULAR", "P-JUS-CKGL", "P-MON-CKP", "P-PDG", "P-TPT", "QGL", "T-TPT"], |
| "units": {"P-ANULAR": "Pa", "P-JUS-CKGL": "Pa", "P-MON-CKP": "Pa", "P-PDG": "Pa", "P-TPT": "Pa", "QGL": "m3/s", "T-TPT": "degC"}, |
| "unit_boundary": "Source units are used as declared. Absolute versus gauge pressure, physical calibration, delivery latency and clocks are not independently verified.", |
| "forbidden_predictive_fields": ["class", "state", "well_id", "origin", "event_directory", "recording_name"], |
| "sample_period_seconds": 1, |
| "window_seconds": 180, |
| "decision_stride_seconds": 30, |
| "bin_seconds": 15, |
| "eligibility": "At least two of the five pressure channels must have >=90% finite numeric observations in the past 180 seconds. No future filling, cross-record history, or treatment of unknown labels as normal.", |
| "arrival_assumption": "Retrospective replay assumes nominal 1Hz rows arrive by each decision. Dataset timestamps alone cannot establish live arrival/quality. Finite numeric values are not verified physical sensor quality.", |
| "features": "For each channel: current value, past mean, past standard deviation, last-minus-first available value and finite fraction. Twelve consecutive past 15-second means and finite fractions form the TCN input. Unavailable values are filled with training-only channel means after normalization, encoded as zero with explicit masks. No label/time/identity features.", |
| "scaling": "Fit channel mean/std and summary feature mean/std on at most 30000 eligible known-label training windows selected by seeded RNG. Do not fit on validation or confirmation. Clamp normalized values to +/-10; unchanged training scalers used for all groups.", |
| "training_windows_maximum": 30000, |
| "training_seed": 407, |
| "training_weights": "Equal aggregate weight per (training group, binary class) cell present. Both classes must exist; no oversampling from validation/confirmation.", |
| "logistic_C": [0.1, 1.0, 10.0], |
| "logistic_max_iter": 400, |
| "missingness_baseline": "Logistic models using only seven finite fractions, with the same C search, training windows and group/class weights.", |
| "conventional_references": ["Maximum absolute normalized past-window channel mean", "Nonnegative one-sided cumulative sum of mean absolute normalized channel means minus k, clipped at 100; k=0.5 and k=1.0. State resets on ineligible input or nonconsecutive decisions."], |
| "reference_scaling": "For conventional references only, fit per-channel median and 1.4826*MAD of eligible known-normal training-window means, with scale floor 1e-6; clamp absolute residuals at 10. No online label access or per-test-well adaptation.", |
| "tcn": {"width": 16, "kernel_size": 3, "dilations": [1, 2, 4], "seeds": [7, 29], "epochs": 8, "batch_size": 256, "learning_rate": 0.001, "weight_decay": 0.0001, "maximum_parameters": 100000, "checkpoint": "Final fixed epoch; no validation-driven epoch selection or architecture changes."}, |
| "candidate_count": 11, |
| "alarm": {"on_consecutive_decisions": 2, "off_consecutive_decisions": 2, "reset_on_ineligible": true, "threshold_candidates": "49 quantiles from 0.90 through 1.00 of eligible known-normal validation scores, plus null as an explicitly disabled threshold. All candidates and thresholds retained.", "validation_false_alarms_per_eligible_normal_day_max": 1.0, "validation_normal_alarm_occupation_max": 0.01, "minimum_eligible_validation_normal_days": 0.5}, |
| "selection": "For each candidate, among thresholds meeting both budgets select maximum all-episode alarm coverage, then known-onset coverage within 300s, then lower false-alarm rate, lower normal alarm occupation and higher threshold. Insufficient eligible normal time selects disabled threshold and records the failure. Global primary candidate uses that performance ordering without comparing numeric thresholds across model scales, then fewer model parameters, then lexical candidate name. Thresholds and primary candidate are written before opening confirmation payloads.", |
| "confirmation": "Open once after the selection receipt is written; evaluate every predeclared candidate at its frozen validation threshold. The validation-selected primary remains primary even when another candidate has better confirmation results. No retuning after confirmation.", |
| "metrics": ["All-episode alarm coverage including censored episodes", "Known-normal-to-positive onset coverage within 300 seconds and detection delay", "Per-event and group support", "False normal alarm episodes per eligible known-normal decision-grid day", "Normal alarm occupation", "Known-normal/positive input eligibility and unknown-label duration", "Brier score only for probabilistic candidates", "95% group bootstrap intervals, 1000 resamples with seed 611; zero-denominator replicates excluded and counted"], |
| "metric_boundary": "Decision-grid exposure is stride-weighted, capped at recording end; exact raw class durations are reported separately. An alarm already active when an event begins is reported as carried-in coverage. Censored onsets cannot establish delay, and event labels do not establish actionable lead time or barrier failure.", |
| "artifact_reload": "Safe numeric arrays/JSON; no arbitrary pickle loading. After frozen confirmation finishes, retain 128 seeded eligible confirmation summary/temporal inputs and corresponding scores for independent checkpoint replay, and all conventional-reference feature inputs for stateful replay. Cross-host probabilistic score absolute tolerance 1e-5; same-host weights require exact numeric reload. No fitting or retuning during replay.", |
| "resources": "Private included Kaggle CPU; two software threads; at most 7200 seconds. Internet disabled during fitting. Full raw timings and process lifetime RSS retained; these do not establish PLC scan deadlines.", |
| "model_admission": false, |
| "control_permission": false, |
| "plant_access": false, |
| "foundation_comparator": "Not included in v1 fitting. Reserved groups remain unopened by this job. TSPulse requires a separate reviewed task/pipeline protocol before using that fresh confirmation; no pretrained-model result is fabricated." |
| } |
|
|
