Evidence records

Error anticipation on RewardBench 2

Store key campaign.adjudication.ERR-RB2

The reading

  • H-err-qwen refuted
  • H-err-llama refuted

registered

card
ERR-RB2
status
adjudicated
reason
not measured
killed
true
staged_violation
false
pre_freeze_violation
false

Uncertainty

ci_low
not measured
ci_high
not measured
ci_level
not measured
n
not measured
n_effective
not measured
seed_spread
not measured
method
none

Subject

card
ERR-RB2
spec
campaign-err-rb2
signals
not measured
dataset
not measured
readout
not measured
frame
not measured
interventions
not measured
gauge
invariant
calibration
not measured
observable_version
1.0
trust
registered

Measurements

metrics

err_auroc_internal_qwen8b
0.8581
err_auroc_margin_qwen8b
0.859
err_auroc_delta_qwen8b
-9.738e-4
err_auroc_delta_qwen8b_ci_low
-0.004635
err_auroc_delta_qwen8b_ci_high
0.002328
err_auroc_internal_llama8b
0.8728
err_auroc_margin_llama8b
0.8763
err_auroc_delta_llama8b
-0.003471
err_auroc_delta_llama8b_ci_low
-0.005095
err_auroc_delta_llama8b_ci_high
-0.001824
err_auroc_delta_min_family
-0.003471

thresholds

H-err-qwen
0.03
H-err-llama
0.03

comparators

H-err-qwen
>
H-err-llama
>

kill_thresholds

K-err
0

kill_comparators

K-err
<=

killed_by

0
K-err

overlay_notes

No rows returned.

Provenance

git_sha
f93f4b5578c221d3d2665f9b305f3e86759be862+dirty
study
study:campaign-err-rb2@v1#c8ac36e7
config_hash
not measured
gpu_seconds
0
tokens
0
wall_seconds
0
gpu
cpu
arc
reporting
seeds
not measured
oracle_calls
not measured
created_at
parents, Derived from

Rendered on

ev:6bf316c059230ff4ff99b118b1664631