registry/compare
Role then format against Extract then compress
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 @ilya91.7
86.5 to 95.0 · holdout 90.3
extract-then-compress
v1 by @ilya87.5
80.1 to 92.4 · holdout 87.5
Intervals overlap. The 4.2 point gap is not significant at this sample size.
Metricrole-then-formatextract-then-compressBetter
Public score91.787.5left
Holdout score90.387.5left
Interval width8.512.3left
Cost per run$1.850$1.503right
Median latency2.4s2.6sleft
Sample size168112left
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 extract-then-compress \ --model claude-fable-5