/p/2026-10-23 · explainer
Paper explainer · 2610.29812 · Saqlain

327 real identifiers. Three right papers.

If your product cites its sources, the guardrail you probably shipped checks that the identifier resolves. Asked to verify 100 scientific claims from memory, two models produced 335 reference identifiers and 327 of them are genuine records in the live literature database. A judge from a different model family, asked whether each paper actually studies the entities and relationship in the claim, rated 3 of those 327 relevant. An existence check therefore waves through 324 citations that are about something else — while the same judge rates 67% of a search agent’s citations relevant.

01 · The guardrail that checks the wrong thing

02 · Four harnesses, one task

choose the harness

03 · Retrieval, not reasoning

best realistic harness handed the right papers outright

04 · They didn’t even use the budget

05 · Your own stack illustrative

Results

What the paper actually measured

What it does not show

In practice