Appeal Arena Rankings

Which AI drafts win when people judge them side by side, which ones get caught making things up, and what people are fighting insurance over.

Who is winning

Appeal letters · all time

No leader yet

No votes have been cast yet. The first to 10 votes gets ranked.

Be the referee

Chat replies · all time

No leader yet

No votes have been cast yet. The first to 10 votes gets ranked.

Be the referee
All synthetic, all aggregate. Every vote below was cast in the Appeal Arena on a made-up case. The condition and treatment lists are counts across many people's appeals, shown only for terms that at least 5 different people entered and that a medical vocabulary recognizes. Nothing on this page comes from any one person's case.
0
votes cast
0
cases judged
0
cases waiting for a referee

Counting votes cast since 2026-11-02, when blind voting began. Earlier votes, cast when the API still named the model, are left out.

Model rankings · All time

Appeal letters

No votes yet. Be the first referee.

Chat replies

No votes yet. Be the first referee.

Model rankings · Last 30 days

Appeal letters

No votes yet. Be the first referee.

Chat replies

No votes yet. Be the first referee.

What people are fighting over

The conditions and treatments named most often on appeals: how many different people named each, counted across everyone who used the site and rounded down to the nearest 5.

Conditions

Nothing to show yet.

Treatments

Nothing to show yet.

How this is counted

  • Win rate is votes a model's draft won over votes it was shown in. A referee who finds two or more drafts equally good can call a tie, which splits that vote evenly between them (half a win each for a two-way tie). Drafts are shown blind (the page gets opaque references, never a model name or a stored order), in random order, so a vote is for the letter, not the brand. Only votes cast since blind voting began count.
  • Leading is the ranked model with the best win rate; a tie goes to the one caught making things up less often.
  • Caught making things up is how often a referee flagged the draft for inventing a study, a law, a test result, a date or a history that was not in the case. A draft can be flagged and still lose honestly; a flagged draft can never be the vote.
  • Warming up marks a model with fewer than 10 votes; its rates are listed but not ranked against the rest, and it cannot lead.
  • Cases are synthetic. Each one is written by a model about a real condition and treatment pair, with made-up facts and the published evidence already on file for that pair. Newer cases also carry a summary of each study and the made-up patient's answers to the questions the app would ask, and every draft is given the same. No real patient's letter is ever shown.
  • The same data is available as JSON. Counts refresh every few minutes.
Step into the Arena Fight your own denial