Case studies

What appeared when public AI datasets were reviewed.

Each case study states the dataset scope, the concrete review result, the source receipt, and what the evidence does not establish.

52,052 conversations produced source-linked credential, answer-key, and duplicate findings in the hosted product.

Across 1,001,551 source rows, 68 groups from one contributing source contained identical prompts with opposite answers.

Reviewing all 19 similar-prompt groups found five wrong or internally inconsistent answers that required a change.

A 40,000-row slice contained 633 extra-turn findings across 376 source rows, with a fixed sample reviewed in full.