Detect length bias

Does your reward model pay more for a longer answer that says the same thing?

Run the bias battery and read the length axis. BiasBattery measures, per axis, the standardized effect size (Cohen’s d) of the chosen-minus-rejected reward delta, and reports it alongside the effective sample size behind it. A reward bias is a reward difference the grader assigns to a surface change that should not matter: more length, more confidence, more markdown. In the diagnostic set the length axis is named verbosity, padding with no added content.

from reward_lens.signals import from_tiny
from reward_lens.measure import base as mb
from reward_lens.measure.battery import BiasBattery
from reward_lens.data.builtin.diagnostic_v3 import load_diagnostic_v3
from reward_lens.data.schema import DataView

signal = from_tiny(seed=0)
diag = load_diagnostic_v3()

# The length axis in the diagnostic set is "verbosity": padding with no added content.
axes = ["verbosity", "sycophancy", "formatting", "confidence"]
items = [p for ax in axes for p in list(diag[ax].items)[:5]]
view = DataView(items)

ev = mb.run(BiasBattery(), mb.Context(signal=signal, view=view))
print(ev.trust, ev.gauge)                 # EXPLORATORY invariant
print(ev.value["strongest_axis"])         # verbosity
for axis, d in sorted(ev.value["per_axis"].items()):
    print(f"{axis:12s} d={d['effect_size']:+.3f}  eff_n={d['effective_n']:.1f}")
# confidence   d=+0.306  eff_n=5.0
# formatting   d=-0.108  eff_n=5.0
# sycophancy   d=+0.691  eff_n=5.0
# verbosity    d=-1.311  eff_n=5.0

The sign is the reading: a positive d means the model pays for the surface change, negative means it docks it, near zero means it does not care. The signal here is a random tiny model, so these magnitudes are noise, not a bias finding; what is real is the shape of the result, one standardized effect and one honest sample size per axis.

The effective sample size is the second half of the number

eff_n comes back as 5.0, equal to the five pairs per axis, because each diagnostic pair carries a distinct seed. That equality is the point, not an accident: the battery reports the lineage-honest effective sample size, so a battery that mutated a few seeds into many pairs would show an eff_n below its pair count and its confidence intervals could not tighten past what the seeds earned. When the two numbers diverge, believe the smaller one. See effective sample size.

Read across models with the d, never the raw delta

Cohen’s d is dimensionless, which is why the gauge is invariant: a d of 0.8 is a 0.8 on any signal. A raw reward delta is not comparable that way. A classifier RM emitting logits swings by tens of points; a bounded, gated head swings by hundredths. The standardized effect divides that arbitrary per-model scale out, which is what makes a cross-model bias comparison mean anything.