Agreement is not evidence

Language models trained on overlapping public text will often rhyme. They share gaps as well as facts. If none of them have your unit economics, your MOQ, or your actual buyer, they can still sound unanimous.

A meeting can do the same thing. People nod because the slide is fluent, not because the claim was checked.

Evidence is something you can point to: a source, a quote, a user-confirmed number, or a documented gap. It has a status. Supported, partial, unverified, and unsupported are not the same thing.

A paragraph that restates the question in more confident language is not evidence. Neither is a list of models that happened to agree.

Name the options. Name the criteria. Attach a status to each material claim. If coverage is thin, say so. A low-confidence result that refuses a winner is more useful than a high-confidence sentence with no trail.

When new evidence arrives, rescore the same criteria. Do not invent a second decision so the record looks busier.

Argumont runs independent roles, classifies claims, and records evidence status before it scores. The judge can return no strong winner. That is a product decision, not a failure to generate text.