Sampling plan (public summary)

Goal: freeze a small set of real public pages about equal-square packing, hand-label numeric claims, then evaluate extractors.

Deviations in batch 2 (recorded honestly):

12 pages from a single source type (not 20 across three types).
Most pages were off the original watchlist.
Both the rule extractor and regex+guards scored 12/12 on batch 2, so that batch cannot distinguish them.

Selection, freeze hashes, and private capture logs live outside the publishable tree.