registry/compare
Policy quote first 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
policy-quote-first
v2 by @dara92.1
86.5 to 95.6 · holdout 91.7
role-then-format
v2 by @ilya91.7
86.5 to 95.0 · holdout 90.3
Intervals overlap. The 0.4 point gap is not significant at this sample size.
Metricpolicy-quote-firstrole-then-formatBetter
Public score92.191.7left
Holdout score91.790.3left
Interval width9.18.5right
Cost per run$1.296$1.850left
Median latency3.0s2.4sright
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 \ policy-quote-first role-then-format \ --model claude-fable-5