Same Paper, Different Decision: A Controlled Diagnostic of Task-Conditioned Scientific Evidence Readiness

Scientific papers are not intrinsically ready or unready for information extraction: their usability depends on the evidence required by the downstream task. We formulate task-conditioned scientific evidence readiness, in which a system maps a paper–task pair to one of three operational actions: READY, REVIEW_REQUIRED, or ABSTAIN. To isolate task conditioning from document variation, we construct a paired diagnostic set from eight Ti–6Al–4V laser powder bed fusion papers and three extraction tasks that differ only in their material-subtype requirement, yielding 24 correlated evaluations. Holding the paper fixed, the frozen AI-assisted author-reference action changes across tasks for six of eight papers. We compare four frozen prompting and control configurations. Direct prompting matches the reference action on 16/24 pairs, while the full structured configuration matches 14/24, indicating that additional structure does not consistently improve decisions in this testbed. An evidence-grounded configuration makes no false READY decisions among its 18 valid outputs but has execution failures on 6/24 pairs, exposing a safety–coverage tradeoff. Error analysis further shows that an output can record a task-relevant evidence conflict yet select an inconsistent operational action. These errors motivate a stage-wise diagnostic perspective separating evidence-state recognition from task-dependent action mapping; the stages are not independently evaluated here. Our results motivate pair-conditioned evaluation for scientific extraction systems whose outputs trigger downstream use or human review.