What Losing Responses Have in Common
Human preference votes contain enough structure to calibrate a quality judge from scratch. Applied to the losing responses, that judge maps …
Read More
Field Testing · Trail & Data
Boulder
Boulder
Gear Lab
Testing gear on mountains. Testing ideas with mountains of data.
Scroll to explore
100,000+
Miles Tested
Trails, roads & mountains
200+
Reviews Contributed
Honest, in-depth analysis
5M+
Feet of Elevation
Colorado's high country
100B+
Rows Analyzed
Mountains of data
Fresh from the Lab
Latest Posts
Gear reviews and data projects from Boulder, Colorado
adidas Terrex Agravic Speed 2 Review
A light, agile trail racer with excellent dry-surface grip and a secure heel, best suited to short-to-mid-distance mountain efforts.
Read Full Review
The Value of Failure Taxonomy
Detecting that an AI system is failing is the easier problem. A failure taxonomy maps observable loss signals to the system layer …
Read More
How Users Vote With Their Prompts
When GPT-4 launched in March 2023, the topic mix of real user prompts shifted in the exact direction the model quality gaps would predict: …
Read More
Against the Trend: Denti-Cal Payments at Children's Dental Group After 2015
California Denti-Cal records show Anaheim's payment intensity rose after the March 2015 ownership transition while every peer office fell or …
Read More
The Shape Predicts the Return
The model quality framework originally relied on scalar ratings and active days. By replacing those inputs with ARC trajectory scores and …
Read More