registrycompare

Role then format against Blunt refusal

Both scored on the same cases against claude-fable-5, same judge. The intervals tell you whether the gap is real or noise.

suite reasoning-chainmodel claude-fable-5n=168 public cases each

role-then-format

v2 by @ilya
91.7
86.5 to 95.0 · holdout 90.3

blunt-refusal

v1 by @system
64.3
56.1 to 71.7 · holdout 63.3
Intervals do not overlap. The 27.4 point gap is significant at this sample size.role-then-format wins
Metricrole-then-formatblunt-refusalBetter
Public score91.764.3left
Holdout score90.363.3left
Interval width8.515.6left
Cost per run$1.850$1.403right
Median latency2.4s1.4sright
Sample size168140left

Reading this verdict

Overlap
Two entries are only meaningfully apart when their confidence intervals do not overlap. A higher number with an overlapping interval is a coin flip, not a win.
Holdout
The public score can be gamed by tuning to visible cases. A gap between public and holdout is the tell.
Reproduce
evalness compare \ role-then-format blunt-refusal \ --model claude-fable-5