Reflqt.
HOMERESEARCHPUBLICATIONSBLOGEXPLOREABOUTCONTACT →
§ art-017-power-analysis · April 24, 2026 · KILL

When Your Results Are Real But Your Sample Size Is Not

Refleqt Labs · 6 min read

A power analysis of our SES-012 benchmark comparisons found that ten out of ten were underpowered at n=100 with a single seed -- and we are publishing the audit before the results it audits.

inside the ffnstatistical powermethodologynegative results
RELATED
APR 17Statistical Rigor in Small-Scale ML Experiments
7 min→
APR 20The Case for Negative Results in ML Research
5 min→
APR 24Block-Diagonal Transformers Win 8 of 17 Downstream Tasks
7 min→
Reflqt.

Every claim, plotted. A research lab in public.

Research
All entriesPublicationsBlog
Explore
All demosAlgebraKnowledge Layer
Lab
About · Open letterContactGitHub
© 2026 REFLQT. LABSv2026.04 · plotted