Label audit protocol

How ground truth is audited without overstating human labeling.

Principle Don't describe the dataset as fully human-labeled. Rule-derived labels are primary. Structural audit isn't expert gold relabeling.

What has been completed

Structural audit
100% challenge + stratified ~10% development sample (structural_auditor_v1)
Scripted spot-check
20 challenge cases, 4 per label (spotcheck_v1) - not a human GRC review
Log
data/review_log.csv
Receipts
reports/HUMAN_REVIEW_AUDIT.md, reports/SPOTCHECK_20.md

Status values

StatusMeaning
unreviewedNot yet inspected
reviewedInspected; label retained
correctedLabel or text corrected after review
flaggedInvariant violation needing judgment

Current status: structural audit complete for all logged rows; spot-check notes on 20 challenge cases. Prefer avoiding sealed test edits after protocol lock - bump dataset version instead.