19 days ago
-
Gate chart: one-line note
20 days ago
-
Redrawn charts: Inter, fewer labels, one number per chart
-
Continuous similarity: 76.4 percent agreement with human judges on held-out Design2Code pairs
-
Model card: shorter sentence
-
Add the Design2Code human-judged benchmark: Loupe does not beat a pixel baseline there
-
More benchmarks: eight windows per fault against Sonnet, fault size test, speed, charts
-
Model card: remove the lab-site link until that site is live
-
Loupe research preview: audit code, read weights, benchmark reports
