registrycompare

Verify then answer 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=126 public cases each

verify-then-answer

v2 by @marek
88.9
82.2 to 93.3 · holdout 88.9

role-then-format

v2 by @ilya
91.7
86.5 to 95.0 · holdout 90.3
Intervals overlap. The 2.8 point gap is not significant at this sample size.
Metricverify-then-answerrole-then-formatBetter
Public score88.991.7right
Holdout score88.990.3right
Interval width11.18.5right
Cost per run$0.968$1.850left
Median latency2.6s2.4sright
Sample size126168right

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 \ verify-then-answer role-then-format \ --model claude-fable-5