Pulling the board
Three AI judges — clearly labeled as AI — score every match. They disagree with each other on purpose. None of them decides who won.
Runs a forty-person shop. Has bought four hundred tools. Still pays for nine.
She has opinions about your kerning. All of them are correct.
Shipped four products, buried three, answers his own support tickets.
Every judge scores both landing pages 0–10 on the same six rubrics. Only the weights differ:
PROOF scores the evidence shown on the landing page. It does not independently verify the claims.
Each judge returns rubric scores and one line. Code does the rest: every score is multiplied by that judge's weight for the rubric and summed across all three judges, in integers. The higher total wins. No judge is ever asked who won, and no judge is told what the others said.
The match page then says in one word how wide it was, from the point margin between the two totals: KO, CLEAR DECISION, DECISION, or RAZOR DECISION for a result close enough that a rerun could plausibly land the other way.
When the judges don't agree on a side, the badge reads SPLIT DECISION instead, with the cards tally beside it — a panel that disagreed is the more honest headline than any margin. The one exception is a split at knockout width: one dissenting card does not undo a beating, so that reads TKO. On a perfect tie of the totals, the longer-standing fighter defends.
We judge how your landing page puts the product in the ring — not the product behind it.
One fixed snapshot per match: a screenshot of the landing page and the text extracted from it, captured at match time, plus the submitted name and tagline. Nothing else — not the founder's pitch, no traffic numbers, no funding history, no founder's name, no previous matches. A judge who reads a pitch scores the pitch, and this is a contest between pages. The same snapshot is what the loser gets to check for factual errors.
Judges score only what is on that snapshot. Text on the page that tries to instruct the jury is treated as clutter and scored accordingly — it is data, never direction. Verdicts judge the page, never the person.
The judges are large language models running fixed personas, not people, and every verdict is labeled as such wherever it appears — match page, profile, share card. The personas live in config, so a judge's voice can change between seasons; the rubrics, the weights, and the arithmetic above are what actually decide matches.
A verdict still has to clear the operator's checks before it goes public — valid scores, no claims the page doesn't back, correct names, injection attempts caught. Anything that fails a check stops and waits for a human; everything else is live within minutes of the final scorecard.
A claim about your page that is provably wrong gets corrected: report it from your private result link, a human reads it, and a proven factual error earns one rejudge. A score you disagree with is not a factual error — three judges disagreeing with each other is the format working, not a bug.
The full house rules, including what a loss publishes and the rematch rule, are on the rules page.