# Abstention

Agents often recognize that a task lacks a prerequisite yet keep acting; a new evaluation scores recognition and warranted stopping separately, and harness search without model-weight updates raises strict self-awareness from 20% to 60%.

Agents often recognize that a task lacks a prerequisite yet keep acting; a new evaluation scores recognition and warranted stopping separately, and harness search without model-weight updates raises strict self-awareness from 20% to 60%.

Panvascular intervention research is starting to write traceability into an agent's output contract: ScienceBuddy-PV fixes a shared convolutional model on the public ARCADE coronary angiography benchmark and evaluates only an evidence gate and action audit, where the gate accepts 76.7% of stenosis-frame candidates against 100% for a score-only comparator. On the locked test set, stenosis Dice is 0.231 and AUROC is 0.957 at a pixel prevalence of 0.00975, while reference lesion candidates are accepted 44.3% and rejected 55.7% of the time. The result is an image-level technical validation; recursive harness optimization and model learning remain protocols rather than executed experiments.

Panvascular intervention research is starting to write traceability into an agent's output contract: ScienceBuddy-PV fixes a shared convolutional model on the public ARCADE coronary angiography benchmark and evaluates only an evidence gate and action audit, where the gate accepts 76.7% of stenosis-frame candidates against 100% for a score-only comparator. On the locked test set, stenosis Dice is 0.231 and AUROC is 0.957 at a pixel prevalence of 0.00975, while reference lesion candidates are accepted 44.3% and rejected 55.7% of the time. The result is an image-level technical validation; recursive harness optimization and model learning remain protocols rather than executed experiments.

On factual question answering with Qwen3-14B, keeping the generator fixed and training a separate answer selector works better than training the model itself when to abstain. On TriviaQA the selector raises mean coverage from 20.2% under the original calibrated confidence filter to 30.9%, while pooled error among scored answers falls from 23.0% to 14.5%; on NQ-Open its mean utility still stays below always abstaining, so reliable cross-dataset abstention remains unresolved.

On factual question answering with Qwen3-14B, keeping the generator fixed and training a separate answer selector works better than training the model itself when to abstain. On TriviaQA the selector raises mean coverage from 20.2% under the original calibrated confidence filter to 30.9%, while pooled error among scored answers falls from 23.0% to 14.5%; on NQ-Open its mean utility still stays below always abstaining, so reliable cross-dataset abstention remains unresolved.