registrycompare

Blunt refusal against Role then format

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=140 public cases each

blunt-refusal

v1 by @system
64.3
56.1 to 71.7 · holdout 63.3

role-then-format

v2 by @ilya
91.7
86.5 to 95.0 · holdout 90.3
Intervals do not overlap. The 27.4 point gap is significant at this sample size.role-then-format wins
Metricblunt-refusalrole-then-formatBetter
Public score64.391.7right
Holdout score63.390.3right
Interval width15.68.5right
Cost per run$1.403$1.850left
Median latency1.4s2.4sleft
Sample size140168right

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 \ blunt-refusal role-then-format \ --model claude-fable-5