How often the Copilot's citations actually hold
The question is not whether it returned a relevant record. It is whether the record it cited supports the sentence citing it. That is harder to measure, and this is that measurement.
60 records were sampled per condition, with a fixed seed, plus 8 out-of-corpus controls, against the live system. The figures above are the paraphrased condition, the honest one: asking a record its own question back flatters retrieval.
Claim support is scored by a different model from the one that writes the answers. The judge sees only the sentence and the records it cites, never the question, so a fluent answer cannot talk its way past it.
An earlier figure of 96.7% was retired rather than reused. It was measured on an older design where the citation was correct by construction, so it described retrieval, not accuracy.