Maintenance patterns, source evidence and separate data-quality review. Reporting date: , the latest service date.
Answer-key agreement: 100% precision / 100% recall
Agreement with the labeled dataset, not independent discrimination or model validation. Symptom keywords alone also select all 60 Y logs; the decline condition adds no discrimination in this workbook.
Retrospective pattern: identifying the first two logs depends on observing the third. This is not advance failure prediction or confirmation of actual failure.
Not covered by answer_key — for manual review
Source chronology and cumulative-meter inconsistencies are shown separately. Counts use different units and may overlap with confirmed patterns; they are never blended.
Headline counts and agreement describe the whole workbook. Filters change worklists, not sequence detection.
Three consecutive logs for the same asset, strictly increasing service dates, ratings exactly 3 → 2 → 1, and each note containing overheating, vibration and hydraulic leakage. All three logs are flagged; no asset IDs are hard-coded.
Unscored · Not covered by answer_key — for manual review
Service date strictly earlier than the asset's purchase date. Could reflect pre-acquisition history or inaccurate dates; investigate without changing the source.
Usage hours lower than the immediately preceding service log for the same asset. Meter resets, replacements or data errors may explain a decrease. Each row represents one transition, not one asset or failure.
Unscored · Not covered by answer_key
Expected-life review compares the reporting date with purchase date plus expected_life_years using calendar anniversaries. February 29 anniversaries fall on February 28 in non-leap years. “Past” means strictly after that anniversary. Zero matches does not establish asset health. Years cannot be converted into permissible usage hours without a defined operating-duty assumption.
Deterministic, not a trained model. Classification reads raw maintenance records only; the answer key is joined afterward for comparison. Sort all logs by asset and service date, inspect every consecutive three-log window, and flag the union of qualifying log IDs. Matching is case-insensitive, with all three symptom phrases required on every log; whitespace or a hyphen may separate “hydraulic” and “leakage.” This literal matcher does not understand negation or synonyms.
Only confirmed-pattern membership has a supplied ground-truth target. No precision/recall is assigned to data-quality or proactive checks, and no combined risk score is calculated. Fixed and Movable filters use the same rule.
Duplicate IDs, missing joins, invalid ratings/quantities/dates or tied service dates block rendering. Predate-purchase dates and decreasing meters remain visible review findings rather than being silently fixed or excluded from pattern detection.
The workbook is an unchanged synthetic demonstration snapshot. The latest service date fixes this report's date; no current-clock aging is used. Editing the workbook does not refresh embedded JSON automatically.