registry/compare
Plain instruction 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=168 public cases each
plain-instruction
v1 by @system70.8
63.6 to 77.2 · holdout 69.4
role-then-format
v2 by @ilya91.7
86.5 to 95.0 · holdout 90.3
Intervals do not overlap. The 20.9 point gap is significant at this sample size.role-then-format wins
Metricplain-instructionrole-then-formatBetter
Public score70.891.7right
Holdout score69.490.3right
Interval width13.68.5right
Cost per run$1.392$1.850left
Median latency2.6s2.4sright
Sample size168168same
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 \ plain-instruction role-then-format \ --model claude-fable-5