Internal eval surface
Levyo Benchmark v0
Loading benchmark evidence...
Benchmark v0 - All 50 Cases
The benchmark denominator is 50 cases: 30 text-only core cases and 20 image-attached cases. M8 is a separate 20-case label queue below.
M4 Image Benchmark Comparison
All 20 image-attached Benchmark cases now have locked runtime outputs; the five earlier failures were recovered through the current manifest handoff.
M8 Image Intent Calibration
Separate 20-case calibration set. Dominik labels image intent from locked reference and generated-shirt pairs.