Master data that breaks heals itself.

Find every defect before touching a value

Missing values, rule violations, outliers and unit errors, format problems, duplicate materials and redundant unit rows — across general, plant/MRP, unit and description data. Weight outliers use a robust z-score on a log scale per material group, so naturally wide groups raise no false alarms.

The system writes its own rules — and says why

Rules are learned only from materials that pass every check (≈350 of 508 at the start). Support ≥ 15 and confidence ≥ 90%; rules at ≥ 95% with support ≥ 30 activate automatically, the rest wait for a Rule Owner. It even learns exceptions:

A fixed order of evidence for every gap

1. The most specific active rule for the material's context (plant before general)
2. The same material in another plant that follows the rule
3. The most similar materials in the same type and group, by majority vote
4. No reliable evidence → quarantine; a person decides. New groups with < 5 golden materials are never guessed.

Unit errors are healed by testing ÷1000, ÷100, ÷10 until the value lands in the group's normal range.

Automation where it is safe, people where it matters

Governance policies on top: an existing value is never overwritten automatically, protected fields (MATKL, MTART, MEINS) and all merges always need a steward. Steward decisions feed the next mining run.

Prove it with MRP, record it for GxP

A built-in SAP MD01-style MRP run with 10 problem-message types shows the business impact before and after healing. Every change, decision and rule activation goes into an append-only, SHA-256 hash-chained audit trail with electronic signatures.