Proxy Contest Analysis: Better Home & Finance (BETR)

Recently I was notified that Better Home & Finance management had referenced and linked to the AI proxy contest analysis on the Proxyanalyst site. I want to note that I have no relationship with the company. I don't own any stock or options in BETR, have had no prior contact with anyone there, and don't know any employees of the company. The reference was made independently — I wasn't asked, consulted, or told in advance.

The AI analysis uses a multi-model approach to help investors research proxy contests, which can be laborious to do manually given the length of each filing and the number of filings involved.

How it works

Claude first reads the full filing text and produces a neutral factual summary of what each side has argued, along with objective data like stock performance. That same summary then goes to four models — Claude, Grok, OpenAI and Gemini — each taking on the role of a proxy analyst. A final step compares the four and identifies points of agreement and disagreement.

Why one shared summary

Why not send every filing to every model? Two reasons, one practical and one deliberate. The practical one is size: the BETR filings come to roughly 240,000 tokens of text, which is more than several of these models can accept in a single request at all. The summary brings that to about 36,000, which every model can read in full.

The deliberate reason matters more. Each filing is summarized on its own rather than thrown in as one pile. In this contest the largest filing is about sixty times the size of the smallest, so if everything went in together the longest documents would crowd out the shortest. Summarizing each filing separately means a brief activist letter gets the same structured treatment as a lengthy company proxy statement.

The limitation that follows

It's worth being precise about a limitation of that design. Because the four models analyze one shared summary rather than each reading the filings independently, agreement between them reflects a common reading, not four independent verifications. An error in the summary would produce agreement just as readily as accuracy would. Source filings are linked on every contest page so readers can check the analysis against the underlying documents.

What I do and don't do

I don't shape or edit the model analysis. Before an analysis is published I check that the run completed and that the models received the full set of filings. That check doesn't alter any analysis, and nothing is withheld or changed because of the conclusion it reached.

I've also made one change as a direct result of being cited: where a filing references the Proxyanalyst analysis, that filing is now excluded from the source material submitted to the models. Otherwise a conclusion I published could come back through a company's own filing and be read by the models as independent third-party support for the next one.

The record

There's no bias built into the evaluation, and the record bears that out. Of the nineteen proxy contests where the models have produced a full analysis, thirteen supported management, four supported the activist, and two concluded a split ballot deserved consideration. Roughly a third have gone against management.

Even with multiple models and this much source material, every investor has to do their own research and reach their own voting decision. Nothing here is a recommendation.