Part 3b: select and preregister empirical dataset

[AMBRA7592]

Summary

• apply the four hard gates uniformly to the eight required Part-3b candidates
• record why none of the eight clears every gate, without treating missing license evidence as a prohibition finding
• identify UC Berkeley Measuring Hate Speech as the only audited gate-passer and pin its source revision and checksum
• freeze one conservative/liberal cohort contrast, eligibility rules, metrics, uncertainty treatment, claims, and non-claims before any outcome run
• halt for owner review: no adapter, README change, empirical run, or prevalence result

Recommendation

Approve Measuring Hate Speech only as a bounded Tier-2 pilot on the 77 topology-qualified comments in the frozen primary set. It cannot support universal or production-prevalence claims.

Scope

Exactly one new report: reports/part-3b-dataset-selection-audit.md.

Verification

• python3 -m unittest test_claims — 31 tests pass with the two expected optional skips
• git diff --check clean
• no source rows, adapter, outcome metrics, README edits, or other repository changes

Review gate

Do not merge or begin an MHS adapter/run until the owner approves the dataset, frozen contrast/metrics, and Tier-2 claim boundary. Claude should audit the source evidence, gate consistency, licensing conclusions, structural counts, and pre-registration.