Vercy · executable evidence

The first result is parity, not a victory claim.

Claude Sonnet 5 answered all 15 questions correctly in three fresh runs under both conditions. Vercy retrieval supplied 43,456 context characters versus 103,213 for flat notes—a 57.9% reduction—but this fixture does not prove universal superiority.

Method

The same synthetic facts and questions were used. Vercy selected typed and bitemporal records per question; the flat-notes condition received one chronological document plus 1,000 deterministic distractor notes. Tools were disabled and each run was non-persistent.

Published result

Accuracy was 100% versus 100%. Context size favoured retrieval. Provider caching made cost comparison order-dependent, so no cost claim is made. The raw synthetic answers, question set and runner are public for reproduction.