File №001 — Adjudication, in progress
Who actually won
the debate?
Paste a YouTube debate. We extract every claim, map every refutation, count every dropped point — and return a verdict you can audit, 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 extract claims
Every load-bearing assertion gets a claim_id, a timestamp, a type (factual, moral, definitional, predictive), and a burden of proof.
We map refutations
Each claim that gets attacked builds a refutation chain — assertion, rebuttal, counter, concession. Dropped points count against the dropper.
Verdict, derived from counts
The verdict is a function of the counts. The model can't reverse-engineer a winner — we recompute the math and reject anything that doesn't add up.
§ 3 — What gets measured
Thirteen dimensions. Every one of them auditable.
Every page on the verdict tab can be traced back to a quote and a timestamp. If you don't like a call, you can challenge it on the receipts — not on tone.
- DIM 01
Claims
Every load-bearing assertion gets an ID and a timestamp.
- DIM 02
Burden of proof
Who has to prove what — and whether they did.
- DIM 03
Refutation chains
The full back-and-forth, exchange by exchange.
- DIM 04
Definitional alignment
When the same word means two different things, we mark it.
- DIM 05
Logical fallacies
Quoted, classified, severity-rated 1–5.
- DIM 06
Steelmanning
Did each side engage with the strongest version of the other?
- DIM 07
Concessions
Explicit and implicit. They count.
- DIM 08
Dropped points
Raised, ignored. They count against the ignorer.
- DIM 09
Factual flags
Verifiable, contested, or unverifiable — labeled accordingly.
- DIM 10
Rhetorical strength
Clarity, command of language, presence. Bounded — not the whole verdict.
- DIM 11
Argumentative integrity
What % of your raised claims actually held up under attack.
- DIM 12
Key moments
The exchanges that swung the call.
- DIM 13
The verdict
Winner, margin, justification — derivable from the counts above.
§ 4 — Selected verdicts
The case file.
![Does God Exist? William Lane Craig vs. Christopher Hitchens - Full Debate [HD]](https://i.ytimg.com/vi/0tYm41hb48o/hqdefault.jpg)
§ 5 — Anticipated objections
We thought you'd ask.
Honest answers to the obvious objections. If yours isn't here, email us.
Because the verdict is computed from counts the model can't fudge. The LLM extracts claims, refutations, and fallacies — then we recompute the verdict from those counts deterministically. If the model's verdict doesn't match the math, we reject it. The only thing the model has discretion over is the bounded rhetorical_score, and that's a minority of the total.
§ 6 — How this stays free
No paywall. Ever.
The whole project runs on donations. No ads, no paywall, no feature gates. The cost breakdown below is an honest estimate of what it takes to keep running. When donations exceed cost, we donate the surplus to a nonprofit organization.
Projected monthly cost
- LLM API callsCHF 110
- Postgres (Neon)CHF 28
- Cache (Upstash)CHF 12
- Hosting (Vercel)CHF 70
- Domain + emailCHF 20