Split the Labor: Separating Evidence Interpretation from Decision Aggregation
Multi-source agents should separate source interpretation from decision aggregation. A fixed evidence tuple plus calibrated likelihood pooling avoids vote thresholds that drift as more sources arrive.
The paper splits multi-source reasoning into source interpretation and decision aggregation, joined by a **four-field evidence tuple**: hypothesis, reliability bucket, rationale, and provenance. It identifies count-scale drift when unnormalized scores are summed.
For research agents, make each reader emit the same typed record, then aggregate outside the model with calibrated log-likelihood ratios. This preserves comparability, supports abstention, and prevents the decision threshold from silently moving as source count changes.
The paper splits multi-source reasoning into source interpretation and decision aggregation, joined by a **four-field evidence tuple**: hypothesis, reliability bucket, rationale, and provenance. It identifies count-scale drift when unnormalized scores are summed. For research agents, make each reader emit the same typed record, then aggregate outside the model with calibrated log-likelihood ratios. This preserves comparability, supports abstention, and prevents the decision threshold from silently moving as source count changes. The empirical system reaches **0.921 AUPRC** versus **0.805** for a hand-crafted baseline on one longitudinal corpus. The authors also report confounded comparisons and say domain-specific quantities still require re-estimation.
This turns multi-source research into a typed interface between model judgment and deterministic aggregation. It strengthens evidence-backed approval and claim-level provenance, while explaining a specific failure of naive multi-agent voting: unnormalized scores move the effective threshold as sources accumulate. The reported gain remains confined to one corpus and requires domain recalibration.