Skip to content
whowonthedebate

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.

Free. No sign-up. No paywall. Cost-recovered by donations.
13
dimensions analyzed
claims tracked, with timestamps
0
vibes-based judgments

§ 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.

STEP 01

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.

STEP 02

We extract claims

Every load-bearing assertion gets a claim_id, a timestamp, a type (factual, moral, definitional, predictive), and a burden of proof.

STEP 03

We map refutations

Each claim that gets attacked builds a refutation chain — assertion, rebuttal, counter, concession. Dropped points count against the dropper.

STEP 04

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.

§ 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.

Monthly target — CHF 240CHF 0 / 240

Projected monthly cost

  • LLM API callsCHF 110
  • Postgres (Neon)CHF 28
  • Cache (Upstash)CHF 12
  • Hosting (Vercel)CHF 70
  • Domain + emailCHF 20
TotalCHF 240