Manufacturing Failure-Reason Codebook Normalization
This skill covers normalization of testing engineers' hand-written defect reasons against the product codebooks distributed with the skill. Engineer-written reasons commonly contain typos, missing words, wrong abbreviations, mixed-language text, combined multi-reason sentences, wrong station names, and cross-project mixups.
Scope and intent
The workflow converts raw per-record reason strings into structured segments, each carrying a pred_code and pred_label drawn from the product codebook that matches the record's product_id. Segments that lack strong evidence are routed to UNKNOWN to flag them for engineering review.
How to use this skill
The actionable procedure — segmentation rules, scoring evidence, tie-break behavior, UNKNOWN handling, and confidence calibration — is maintained in the companion reference codebook_normalization_guide.pdf in this skill directory. Open that PDF and follow it end-to-end before producing solution.json; SKILL.md is only a framing shell and does not repeat the pipeline on its own.