Bounded offline artifact · no uploads or network requests · at most 100 fixtures and 10,000 characters each

MinHash eval dedup lab

Screen sanitized prompts with seeded signatures and LSH, verify every candidate with exact Jaccard, and preserve split-aware provenance.

SeedExact thresholdFixtures: id|split|text

Candidate ledger

Pair	Exact	Estimate	Status

Similarity screening does not prove training contamination and never deletes a case. Protected split crossings require review.