HARNESS SANITY

Grok 4.5 pilot

One frontier model on a stratified mini-split before mass ranking. Easy cells must not be zero if the harness is healthy.
0.74VPI equal
1.00count / color
0.87text_read
0.00*change [email protected]

Interpretation

PASS: easy color, count, large text, large size ratios. DISCRIMINATION: 6px text and confusable O/0 drop. METRIC SEVERITY: change_spot [email protected] alone under-rates models that see differences but localize loosely — dual metrics are required in the harness.
Read full pilot notes on GitHub