The rubric.
How the judge scores every debate.
In one minute
An AI judge scores both debaters from 0 to 100 on Accuracy, Evidence, Responsiveness and Conduct. A factual claim earns credit only when you attach a source that supports it; a source that contradicts you costs more than a supported claim earns. Answer the other side and stay on the headline. Insults cost you; an ethnic or religious attack, or incitement, loses you the debate. The scorecard is public and permanent.
What is scored
Claims. The judge puts each checkable statement of fact you make on your checklist.
Sources. When the last turn ends you have 420 seconds to attach one link to each of your claims. You can shelve links before or during the debate. Each is checked as soon as you attach it.
Turns. Every turn but the first gets a level from 0 to 4: does it answer the turn before it and bear on the headline? A turn off the headline gets 0.
Conduct. Every turn is checked for insults, ethnic or religious attacks and incitement.
How each score is worked out
Divisions round down. Your claim cap is a third of the debate's minutes, rounded up, between 3 and 20.
Accuracy. A supported claim earns its source's tier credit: Tier A 100, Tier B 70, Tier C 40. Add the credits of your best claims, counting no more claims than the cap, take away 200 for each contradicted claim, and divide by the cap. Between 0 and 100.
Evidence. The share of your claims sourced, out of 100: a claim is sourced when its page was read and found to support it, contradict it or not address it. No claims: 0.
Responsiveness. Your average level, scaled so that 4 on every turn gives 100.
Conduct. 100, less 25 for each turn with an insult; 0 with a severe violation.
Points. Accuracy times 25, Evidence times 25, Responsiveness times 25 and Conduct times 25, summed and shown out of 100 to two decimals.
Winner. More points wins. Ties go to higher Accuracy, then Responsiveness, then Evidence, then Conduct; then a draw.
Sources and tiers
Tier A: official statistics, government publications, court and electoral records, primary documents. Tier B: established news organisations and research bodies. Tier C: any other page that passes screening.
Never accepted: a page that does not open, an image, a social media post. A Tier C page is also rejected on a free blogging, paste, shared-document or open-wiki site, when its domain is under 365 days old or of unknown age, or when it was published after the challenge was posted.
An accepted page is then read: does it address your claim, and does it support or contradict it?
Outcomes: supported (credit by tier), contradicted (a penalty), not addressed, rejected, unsourced (no link) and unverified (the judge could not decide by the deadline). The last four earn nothing.
A worked example
The headline: "Every Nigerian household that can afford it should switch to solar power." Real sources, invented debaters: Tobi for, Ngozi against, over 15 minutes.
Tobi makes five claims to Ngozi's three. Two hold: a think tank's finding that saved diesel soon pays for a panel (Tier C), and a business paper's import figure (Tier B). But he says fewer than half of Nigerians have access to electricity, and the World Bank figures he links say more do: that wipes out his Accuracy. A free-blog link is rejected; an unlinked claim earns nothing.
Ngozi's World Bank figure (Tier A) holds, and so does the think tank's own caveat that its sum counts only the panel (Tier C). Her battery claim's link does not address it: no credit, but sourced. She loses marks too: one turn strays off the headline (level 0), and she calls Tobi a panel salesman, an insult. Sources, not sides, decide it.
The real scoring code works out the scores below.
Ngozi wins 64.50 to 57.00
What is never scored
Your side, or whether the judge agrees. Your accent, looks or voice. Confidence or length. Likes, comments and viewers.
Conduct
The code of conduct is at /conduct.
An insult: each turn with one takes 25 off your Conduct, however many it holds.
An ethnic or religious attack, or incitement, is a severe violation. A debater with one loses, whatever the points. If both do, nobody wins.
What the judge cannot do
It checks only the page you link. A true claim without a working link earns nothing.
It works from a transcript, which can mishear words, names or numbers; the scorecard quotes what it heard.
It does not weigh style, or decide which side is right.
The final scorecard comes no later than 600 seconds after the last turn; a source still being checked then is unverified. If a turn cannot be transcribed after three tries, the debate is void.