registry/compare
Source span citation 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
source-span-cite
v2 by @marek93.8
87.7 to 96.9 · holdout 93.8
role-then-format
v2 by @ilya91.7
86.5 to 95.0 · holdout 90.3
Intervals overlap. The 2.1 point gap is not significant at this sample size.
Metricsource-span-citerole-then-formatBetter
Public score93.891.7left
Holdout score93.890.3left
Interval width9.28.5right
Cost per run$1.138$1.850left
Median latency2.1s2.4sleft
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 \ source-span-cite role-then-format \ --model claude-fable-5