Who ran it, with which judge, against what, and what didn't work. Papers from the team behind the engine, with a source on every claim.
ResearchSeptember 19, 202618 min read
Eight deep-research cycles and 900+ sources on agent memory: frameworks solve memory inside their own perimeter, practitioners improvise the rest, and continuity (the thing that makes teams of agents not redo work) has no owner. Benchmarks, hidden costs, the nine unpublished gaps, and the Brazilian regulatory reality.
Read paperResearchAugust 24, 202611 min read
A stealth reader and one API parameter took our BEAM-100K average score from 69.3% to 75.5% (323/400 correct). Not SOTA. The interesting part is where the points came from, and where they refused to come from.
Read paperResearchAugust 4, 20268 min read
A 70.9% average on BEAM-100K with GLM-5.2 as the reader, and the correction published later: this post's decomposition does not add up and its judge is misstated.
Read paper