File №001 — Adjudication, in progress
Who actually won
the debate?
Paste a YouTube debate. Five judges read the whole transcript independently, name a winner, and cite the moments that decided it — with timestamps. No vibes. Just receipts.
§ 2 — Method
How a verdict gets reached.
The same method a debate-tournament adjudicator uses, run on a transcript instead of a memory. The order is non-negotiable.
Paste the URL
Any YouTube debate. Long-form, podcast, town hall, formal — if there are two sides talking past each other, we'll judge it.
We pull the transcript
Captions, timestamped line by line. The title gets neutralised first — partisan uploads announce a winner, and no judge sees that.
Each judge reads it whole
No summary, no sampling, no excerpts. Judges score the arguing, not the position: burden carried, replies that engaged, concessions extracted.
Five judges, one verdict
Five judges read the whole debate independently and cite the moments that decided it. The majority wins — and when they split, we tell you.
§ 3 — What gets weighed
Nine things a judge weighs. None of them is who’s right.
Every line of a verdict points at a moment in the video. If you don't like a call, you can challenge it on the receipts — not on tone.
- DIM 01
Burden of proof
Who had to prove what — and whether they ever did.
- DIM 02
Engagement
Did the reply answer the point that was made, or talk past it?
- DIM 03
Concessions
What got given up. Out loud, or by going quiet.
- DIM 04
Internal consistency
Whether a case still stands at the end of its own argument.
- DIM 05
Cross-examination
How each side held up when pressed directly, not when speaking freely.
- DIM 06
The loser's best
What the losing side genuinely won. A verdict that finds nothing there stopped reading.
- DIM 07
The closest call
Every judge names the one judgement they were least sure of.
- DIM 08
Timestamps
Every assertion about the debate cites when it happened. Click it and check.
- DIM 09
The split
Five ballots. When they disagree the margin comes down, and we print the count.
§ 4 — Selected verdicts
The case file.
§ 5 — Anticipated objections
We thought you'd ask.
Honest answers to the obvious objections. If yours isn't here, email us.
Because one model's opinion is not the verdict. Five judges read the whole transcript at the same time, never see each other's ballots, and the majority decides — and when they disagree we print the count instead of rounding it away. Each one gets the title with the result stripped out, so no judge is told who won before it reads a word.
Every judgement has to cite when it happened, as a timestamp you can click into the YouTube player and check. If a verdict points at 41:20 and the exchange there says something else, that is a bug — email us and we will re-run the analysis.
The majority wins and we say it was 2–1 — on the verdict, on the share card, on the debater's profile. A split panel also caps the margin: a debate that divided the judges cannot be reported as decisive. Reporting it as decisive is exactly the overclaim that got the previous method retired.
The model is instructed to return NOT_A_DEBATE if it can't identify two distinct opposing positions. Lectures, panels, and interviews fail closed.
No. Donations cover what it costs to run — transcripts, judging, server, database — and nothing else. No ads, no paywall, no gated features. Any surplus goes to a nonprofit.
Yes. Every analysis page has a 'community verdict' widget showing the AI verdict against the crowd vote, side by side. Disagree out loud. We will surface the divergence.
The model is multilingual. The UI is currently English and German, more languages on request.
§ 6 — How this stays free
No paywall. Ever.
Like this free tool? Every verdict costs a transcript, three judgements, and the server they run on. No ads, no paywall — just donations keeping it alive.