Every answer on this page is a verbatim recorded output from one benchmark run (2026-07-02). Both systems used the same reader model (Claude Haiku 4.5) and drew on the same evidence stream: real MediaWiki revision histories, in true edit order, scattered through a background of real 2026 arXiv abstracts. The only difference is the memory in front of the model. Nothing here is generated live.
Eight current-state probes from the run where naive similarity retrieval asserted a stale value and Sapience answered with the current one. Each corpus tracks a real position through its actual Wikipedia edit history (3 to 8 successive holders). Selection criterion is stated in the footer; the full run, including every probe both systems got right, is in the linked result files.
Why does the ordered baseline nearly tie? By design: the steelman proves the failure is about order information, nothing else. The trick works here because these are single-entity chains with clean write order, where "latest mention = current" is a valid inference. Outside the lab it breaks four ways: arrival order is not event order (imports, backfills, multi-source ingestion); currency is per-entity, not a global sort (real queries retrieve mixed evidence); ordered evidence still asks the model to infer currency, where supersession makes it a retrieval rule (a superseded fact cannot outrank its successor); and ordering answers only this one question type, not history, belief revision, conflicts, or abstention. Ordered retrieval is the lab-bench version of one Sapience mechanism; the product is that mechanism made automatic, per entity, at write time.