Geometry Proposes, Judgement Disposes: Accurate Online Proposal Clustering for Digital Deliberation

Digital deliberation requires accurate clustering of proposals: equivalent contributions should be joined, while related interventions and opposing positions remain distinct. We study an online method implemented in Freedi that uses embeddings to retrieve candidates and focused language-model judgements to authorize synthesis, with separate themes, retained sources, and maintenance. We explain its retrieval-judgement interface and seven-stage selection process and compare it with the two main alternatives, cosine rules and LLM-only grouping. On 875 counterfactual preference triplets, a component study achieves 95.0%-96.5% accept-and-reject accuracy with original-text judging, compared with 0% for the tested cosine decision rule. On a controlled English corpus, Freedi and LLM-only baselines recover all 50 intended pairs in three arrival orders. A new 240-call probe on one authentic English deliberation compares full-list judgement with judgement over 15 retrieved proposals. Median per-decision cost reductions grow from 2.55-fold at 100 proposals to 4.29-fold at 200 and 9.38-fold at 500, under common base token prices. At 500 proposals, agreement with provisional model-authored labels is 25/40 for the restricted list and 31/40 for the full list. Retrieved targets remain available, but the restricted judge declines more merges. These refusals may preserve substantive distinctions in scope or intervention; whether they improve semantic precision or constitute missed matches requires human adjudication. A separate replay of 114 Hebrew statements shows improved automated merge precision and text fidelity after combined repairs. The evidence supports semantic judgement over the tested cosine rules and a growing cost advantage for candidate reduction at the measured scales. Equal accuracy and lower total cost for the complete Freedi pipeline remain to be established.