registry/compare
Extract then compress 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=112 public cases each
extract-then-compress
v1 by @ilya87.5
80.1 to 92.4 · holdout 87.5
role-then-format
v2 by @ilya91.7
86.5 to 95.0 · holdout 90.3
Intervals overlap. The 4.2 point gap is not significant at this sample size.
Metricextract-then-compressrole-then-formatBetter
Public score87.591.7right
Holdout score87.590.3right
Interval width12.38.5right
Cost per run$1.503$1.850left
Median latency2.6s2.4sright
Sample size112168right
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 \ extract-then-compress role-then-format \ --model claude-fable-5