Back to the site
The evidence lives with each study.
Every number on this site comes from a committed result file. The full tables, the claims registry and the published errata are on the study sites and in the repositories, so a correction only has to land in one place.
OpenGrad
Call F1 rose from 0.6264 to 0.7548, and the same model refused all 1,319 zero-shot GSM8K questions.
Small-Mind
Correct abstention on 80 unanswerable probes rose from 13.75% (the raw model, no memory) to 71.25%, while recall on the 608 answerable probes reached only 18.59%.
OpenPapers
Zero fabricated answers on a frozen 40-task holdout, but only 0.563 of the candidate facts it extracted were correct.
All results are single-seed and small-scale, and each study states its limits. A claim that could not be sourced is left off the site instead of badged.