Belief recall and evaluation changes
Checkpoint trajectory
Whiskers show pointwise 95% confidence intervals. The x-axis is SDF exposure, not RL training. The shared base response set is rescored under A and B; it is not two independent baseline samples.
What uncertainty is measured?
The coding A − B gap resamples matched tasks together. Both authority preferences reverse between A and B, so the gap measures relative influence under these crossed assignments. It does not isolate a grader-only belief intervention.
Recall by universe and authority
Style-recall questions ask what each authority prefers. Each line groups by the authority queried, with 95% scenario-cluster confidence intervals. Authority-attribution questions are reported separately below.
Recall by question and answer format
Correct / all attempts. Unscorable attempts count as incorrect. Style recall asks what each authority prefers; authority attribution asks who prefers a given style. 50% is a binary guessing reference for MCQ, not a calibrated chance level for short open answers.
Paired grader and user answers
Separate responses paired by scenario and repetition. Both correct means both conflicting assignments are recalled; both comprehension or both loop means the same style was assigned to both authorities. These are independent generations, not a joint answer in one conversation.
Grader minus user style-recall accuracy
Paired scenario intervals keep the two authority questions together. Positive values mean higher grader recall.
Answer-order consistency
Does the model answer “comprehension” for both authorities?
Plot the answer itself, independently of whether it is correct. A comprehension answer is correct for the grader in A and for users in B. A high rate for both authorities reveals a shared answer tendency; it does not establish why the model chose it.
Composition bars use all attempts: comprehension loop unscorable. The target is the corpus-assigned preference.
MCQ letter tendency
Compare semantic answers above with A-letter rates below. Options reverse between formats; A-letter preference and comprehension preference are different patterns.
Open answers across SDF
Short-phrase recall and attribution are plotted separately. Solid lines: correct / all attempts. Dashed lines: unscorable / all attempts. Every rate includes all responses.
Matched change from baseline
Percentage-point change on the same scenarios. The baseline is scored against the corresponding universe's target mapping.
Illustrative responses and their trajectories
A − B coding contrast
All response categories
Change from the first evaluated checkpoint
Custom evaluation scores
Raw scorer outputs are retained with their task and repetition. Choose a metric in the suite implementation before aggregating these values.