Figures / Sources

150

Random records our AI reads in every sample

Our setting, read from the engine (RISK_POLICY.aiReadSample): when a de-identified sample arrives, an AI model reads up to this many of its records picked at random, with the words the rules could not explain, and what it finds that could point to a person is removed before the sample is measured; with none found in 150, fewer than 2.0% of its records still hold an identifier, at 95% confidence. A sample is never kept or shown.

Updated 15 Oct 2026, from its newest record.