DebateEval

“GPT it.” “Grok it.” “Claude it.” Stop it.
Ask every model. Check every claim.

Every AI company has its own angle. Follow a debate live or from a recording, put each claim in front of a panel of models, probe their scores for bias, and finish with a verdict you can share. Built to find the truth, whichever side it favors.

No credit card required · Bring your own API keys

Sound familiar?

The problems with regular AI

A single chatbot is a poor judge of a debate. Here is what goes wrong, and what this does instead.

Worried about confirmation bias quietly steering your analysis?

Several models from different labs score every argument, so no single view decides.

Want real critiques, not an AI telling you what you want to hear?

Judges score both sides on the same rubric, and bias probes swap sides to check them.

Want to steelman the other side but can't think of what to say?

The strongest opposing arguments, drafted for you from either side.

Tired of bad-faith speakers walking back points without saying so?

Concessions and self-contradictions logged against the moment they were said.

Can't tell who actually won the points that mattered?

A verdict page that names a leader only when the panel's scores clearly differ.

Live debates vanish as fast as people talk?

A growing transcript and a steelman tree while the conversation is still happening.

What it does

From the first word of a live debate to a verdict you can share, in four parts.

01 · Capture

Live or recorded

Bring a debate in while it happens or after it ends. Transcription knows who is talking.

Live debate studio

Capture from your mic or a browser tab and watch a live transcript name each speaker as they talk. Mute a source, pause or end at any time, run several debates from one tab, and keep them going while you move around the app.

Fits your format

Pick Conversation, Structured turns or Formal debate. Get arguments in real time, or choose “When I ask”: it transcribes the whole time and makes arguments only when you press Process, usually for fewer credits.

Recordings and links

Paste a YouTube link or upload audio or video. The import reads the debate turn by turn in the background, so you can leave the page and come back to a finished tree.

Automatic topic segments

The debate is split into topics as it goes. Each segment gets its own question, a rolling summary of each side, and the cruxes where the speakers really disagree.

Search by words or meaning

Did they really say that? Search the transcript for exact words, or by meaning to find the passage even when the wording differs, with the speaker and timestamp.

02 · Understand

Every argument, mapped

Claims, challenges and replies become a tree you can explore, with the record kept as the debate unfolds.

Visual argument trees

Explore claims on an infinite canvas, as threaded replies, or in an outline and sunburst view. Add your own arguments, draft new ones from any angle, and step back through the tree’s history.

Automatic steelmanning

Every new argument gets a steelman: the strongest fair version of the point. Edit it if you like, and the judges check it still says what the speaker meant.

Concessions, tracked live

When a speaker gives ground, it is logged against the point they conceded and linked to the moment they said it, while the debate is still going.

Self-contradictions caught

Flags when a speaker asserts and denies the same claim, or when their own arguments are marked as contradicting each other, with a jump to each moment.

Fallacy and structure checks

Fallacy flags you can dispute, plus deterministic graph checks (not AI) for circular support and unanswered challenges, so every claim shows whether it still stands.

03 · Judge

Scored by a panel, tested for bias

No single model decides. Every score is 0 to 100, and you can see where the judges disagree.

Multi-model panel

Frontier models from different labs (Gemini, GPT, Claude, Grok and more) score each argument independently on your criteria, and the spread between them is shown, not hidden.

Bias probes

Side swap scores the mirrored claim and flags any model whose score moves by more than 15 points. Attribution and framing swaps test whether it matters who said it, or how it was worded.

Civility scoring

Turn it on and each model rates the tone of the speaker’s own words, quoting any personal attack, insult, slur, threat or mockery. Tone is kept separate from reasoning.

Hybrid Jev judging

Experimental: yes/no checks such as steelman faithfulness, civility flags and source checks also go to Jev, a small decision model. Its answer is used only when it is confident.

Your criteria, your keys

Choose the models and scoring criteria for each tree. Bring your own provider keys, or use credits; calls billed to credits require zero data retention.

04 · Share

After the debate, and during it

Turn the analysis into something an audience can follow, live on stream or as a report afterwards.

Verdict page

A shareable report with the scoreline, a side-by-side table, score over time and where they really disagree. It names a leader only when the gap is clear, and says “too close to call” when it isn’t.

Speaker statistics

Floor share, talk time, longest turn, interruptions made and received, and questions asked, answered or dodged, so you can see whether the floor was shared fairly.

Live stream overlay

Add a browser source to OBS or any streaming tool to show the current topic, each side’s position, the latest claim and live speaker stats as the debate happens.

Watch together

Share a tree by link or publicly, and viewers see it grow live. Invite collaborators to edit, and export any tree to import again later.

Links that preview well

Public and link-shared trees and verdict pages come with their own preview image, so a shared link shows the debate at a glance in chats and on social media.

From position to verdict - any way you work

A flexible workflow that meets you where you are - whether you're analyzing an existing debate, preparing arguments for one you're about to have, or stress-testing a position you already hold.

  1. Start

    Bring any debate, or none.

    • A YouTube link or an audio/video upload
    • Go live from your mic or a browser tab
    • Or a blank tree around any position
  2. Transcribe

    Know who said what.

    • A live transcript as people talk
    • Each speaker named
    • Split into topics automatically
  3. Map

    Claims become a tree.

    • Claims and rebuttals linked
    • Every argument steelmanned
    • Concessions and contradictions logged live
  4. Expand

    Fill in what's missing.

    • Draft arguments for or against
    • Add your own
    • Accept, edit or discard each draft
  5. Evaluate

    Let a panel judge.

    • Several frontier models score 0 to 100
    • Bias probes swap sides, speakers or wording
    • Optional civility scoring
  6. Share

    Show your work.

    • A verdict page with speaker stats
    • A live overlay for your stream
    • Collaborators and share links

Pricing

Building, viewing and collaborating on trees is free. Pay for AI with credits, or subscribe and use your own API keys at cost.

RECOMMENDED

Managed Access

Pay-as-you-go

No setup. Add funds and start a tree.

  • No API Keys Required
  • Gemini, Claude, GPT and Grok without your own keys
  • Pay only for what you use
  • Several models score every argument
  • Credits never expire
Start a debate

Add funds from the dashboard.

Your own keys

$12/year

Every tool, with model calls on your own provider keys.

  • No markup on models you add a key for
  • Your own keys (OpenAI, Anthropic, Grok and more)
  • Credits only for models without a key and for transcription
  • Full Rhetorical Analysis

Payment by Stripe. Cancel any time.

How we make money

You pay only when AI works on our credits, or if you choose the key plan. Everything else is free.

Most of it is free

Build trees by hand, view other people's trees and collaborate without spending anything. AI is only needed for the parts that use it.

AI at cost plus 25%

When you use AI (scoring, drafting, transcription), we charge what the provider charges us plus 25%. That covers payment fees and keeps the service running. New accounts start with 50,000 credits ($0.50) free.

Heavy AI user? Bring your keys

For $12 a year, add your own provider keys. Those calls are billed to you by the provider with no markup. Models without a key still use credits, and so does transcription of live, YouTube and uploaded audio (AssemblyAI, plus pyannoteAI speaker detection when it's on).

Not used for training

Model calls paid with credits require zero data retention, so they only go to providers that don't keep or train on your prompts. Calls on your own keys follow your agreement with that provider.

No hidden fees

  • No tree limits
  • No seat fees
  • No subscription required
  • Credits never expire