Research note
A single prompt is an observation, not a ranking
How many repetitions does it take before an AI answer can be described as a stable position?
Nathaniel "Nate" Whitaker · 2026-09-12
Research notes
Shorter pieces of evidence, published as they are established.
Notes are the working output of the research desk. They are narrower than a report and slower than a news item, and several of them accumulate into the benchmarks we publish quarterly.
Research note
Longer context windows do not straightforwardly improve accuracy
Does supplying more material to a model make its answers more reliable?
Martin Anderson · 2026-09-13
Research note
Why completion rates fall faster than step counts rise
What does the arithmetic of sequential tool calls imply for agent deployments?
Aleksi Virtanen · 2026-09-14
Research note
Six disclosures that make a benchmark result checkable
What is the minimum a published AI result must disclose before a reader can verify it?
Dr. Annika Holm, PhD · 2026-09-15