HISTORY.md

Development history

This is the previous model card, kept verbatim.

---

license: cc-by-4.0 task_categories: [tabular-classification] tags: [time-series, industrial, 3w, causal, negative-results, anomaly-detection] configs: - config_name: decisions default: true data_files: - split: validation path: data/decisions/validation.parquet - split: confirmation path: data/decisions/confirmation.parquet - config_name: episodes data_files: - split: validation path: data/episodes/validation.parquet - split: confirmation path: data/episodes/confirmation.parquet - config_name: corpus data_files: - split: metadata path: data/corpus/metadata.parquet

3W causal real-group audit and failed-transfer traces

Native typed Parquet derived from pinned public Petrobras 3W 2.0. This release contains source-audit metadata and every validation/confirmation decision and episode outcome for the frozen 11-candidate study. It is not a new raw sensor corpus, new field acquisition or industrially qualified failure dataset.

The source descriptor is published in April 2026; the underlying pinned release predates that paper. See attribution, exact table schemas/counts, protocol, manifest, access ledger, and complete negative-result report. The model repository is well-production-event-causal.

Configurations

• decisions: one row per 30-second decision after 180-second past-only warmup; labels, numeric eligibility, capped grid exposure, all 11 class/residual scores and frozen alarms. Unknown class/binary targets and ineligible scores are null. event_timestamp_ns is the original nominal dataset event clock, not verified arrival or decision time. Scores/labels/provenance are offline audit fields; they must not enter a predictor as input features.
• episodes: one row per candidate/episode, with event, onset censoring, alarm coverage, carried-in alarm and known-onset delay. Transient and steady labels for the same contiguous event are merged. Covered is not onset detected.
• corpus: 2,228 source recording metadata rows, including 1,119 real, 1,089 simulated and 20 drawn files, source hashes, labels, clocks, seven-channel availability and the frozen real-recording selection/overlap exclusions. Reserved-group rows are qualification metadata, not predictive evaluations.

No train or reserved feature/score split is published here. The protocol used 357 training, 33 validation, 138 initial confirmation and 113 reserved selected recordings, plus 478 real overlap exclusions. Raw sensor values are obtained from the separately attributed upstream source, not silently duplicated here. Filename groups are anonymized public IDs, not authenticated physical-well IDs.

Known failure and intended use

The validation-selected primary covers 24/55 confirmation episodes, with 42 false alarms over 9.124 eligible normal days and 64.392% normal alarm occupation. Only 3/12 observable onsets are covered within 300 seconds; most covered episodes already had an active alarm at onset. Normal numeric eligibility is 59.812%. Use these tables to inspect group transfer, abstention, onset censoring and alarm burden, not to claim physical barrier integrity, RUL or deployment readiness.

All 31 frozen fixtures passed in the actual included CPU run. Its final JSON report failed serialization after traces/weights were saved; a separate report was recovered without fitting/retuning. The provider failure remains explicit. The original confirmation set is consumed; eight additional groups stay reserved for a separately reviewed/frozen future comparison. Public-pretraining exposure cannot be excluded by study-level group separation.