Insight IT Solutions · Python Libs · Papers

Mitigating Consensus Bias in AI Deliberation: A Structural Evaluation of Disagreement Retention and Regime-Level Adjudication

Amin Parva

Software research manuscript · 28 September 2026 · PrismCognition software version 2.1.0

Download PDF PyPI GitHub Python Libs hub

Abstract

A deliberation system can detect a disagreement and still omit it from its final record. We examine this form of disagreement suppression in PrismCognition, an open-source Python library that represents claims, evidence assessments, assumptions, and conflicting obligations as structured artifacts. A diagnostic benchmark contains 34 synthetic scenarios and 29 expected conflict instances. Comparisons include two local rule baselines, four aggregate-threshold settings, and two evidence-access ablations. The default engine retains 15 of 29 expected conflicts and matches the expected types and resolution statuses in 20 of 34 scenarios. Removing the aggregate filter retains all 29 conflicts across six classes and matches 33 of 34 outcomes. A direct-claim evidence baseline matches 21 of 34 outcomes, exceeding the default engine on this measure while detecting only factual conflicts. Without the aggregate filter, evidence access corrects four complete outcomes but introduces one cross-claim resolution error: pump evidence incorrectly adjudicates an unrelated valve state. Each of four replay and transformation properties passes 136 checks; 10 supplied-boundary gate scenarios also pass. The findings support broader conflict representation under the tested configuration and repeatable artifact derivation, while exposing limitations of aggregate filtering and regime-level adjudication. The cases were authored after source inspection and have not been independently annotated. The study does not establish consensus bias in other systems, superiority over model-based alternatives, or improved human decisions.

Keywords

AI deliberation; computational argumentation; disagreement retention; consensus bias; evidence grounding; provenance; reproducibility.

Software

PrismCognition is the open-source library under study. Install with pip install "prismcognition>=2.1.0" (Python ≥3.11). It appears on the Python Libs hub with this paper and PDF. There is no separate commercial product landing for PrismCognition at this time.