Closed-form drift analysis for the noisy-OR aggregator

Derive a closed-form characterization of evidence drift under the deployed noisy-OR confidence aggregator and Boltzmann path scoring, extending the vote-counting analysis of retrieval-induced misclassification under publication bias.

Background

The paper proves that, under a fixed positive publication bias, a vote-counting aggregator increasingly misclassifies true NoEffect queries as Beneficial as retrieval depth grows. The deployed DACG-agent instead uses noisy-OR confidence updates and Boltzmann path scoring, whose nonlinear behavior is not covered by the theorem.

Simulations indicate that the qualitative drift phenomenon transfers from vote-counting to noisy-OR, but the paper does not establish an analytic, closed-form result for the deployed aggregation mechanism. Such a result would formally connect the theoretical drift guarantee to the system used in experiments.