Most trivia apps never actually ask you to answer. They show you four options and ask which looks familiar — a recognition test wearing a quiz costume. The moment that makes trivia a sport is different: it's dredging a name up from a half-memory, committing to it out loud, and finding out.
So on trivai, you type your answer. Anything. Which means something has to judge it — "Ghandi", "the guy who painted the ceiling", "FDR" — and a lookup table can't. An AI judge rules on every open-text answer, and handing out verdicts is the kind of job where fairness isn't a nice-to-have. Here's how we keep it honest.
The judge is blindfolded
For every answer, the judge is handed exactly three things: the question, the expected answer, and what you typed (plus any alternate answers the question explicitly allows). That's the whole file.
It doesn't know who's answering, what the score is, whether this is a steal attempt that would swing the lead, or whether you've missed five in a row. It can't be generous to the player who's behind or stingy with the one who's ahead, because it doesn't know either exists. Every answer in every game meets the same blindfolded judge.
What counts as correct
The judge's standing orders are written to reward knowledge, not typing:
- Capitalization never matters. "paris", "Paris", and "PARIS" are the same city.
- Typos and spelling variations are fine when it's clearly the same answer. "Ghandi" knows who Gandhi is.
- Partial answers count when they capture the key information. "García Márquez" gets full credit for Gabriel García Márquez. But if the question is which Roosevelt, "Roosevelt" hasn't answered it — the key information is exactly the part you skipped.
And one rule we consider non-negotiable: the judge may never reject your answer because the question itself turns out to be flawed. If a question slips through the pipeline with a wobbly date in it, that's our failure, not yours — the judge is explicitly instructed that the question's accuracy is not the player's responsibility.
The verdict is public
There's no black-box "wrong, moving on." When the judge rules, the whole table sees the same reveal: the verdict, the answer you typed, the correct answer if you missed, and the judge's explanation of its reasoning — right there under an "AI:" label, on the record. If the judge gives you the point on a technicality, everyone gets to see the technicality. Showing the work is most of what makes a ruling feel fair, at trivia tables and everywhere else.
When the AI steps aside
The judge only rules where judgment is needed:
- Multiple-choice charges bypass the AI entirely. Spend one and you're picking from fixed options, so the check is a plain, deterministic comparison — no model in the loop, nothing to second-guess.
- Rerolls swap the question, not the ruling. Three per game: if a question isn't worth answering, replace it — same topic, same difficulty, fresh question — before you ever put an answer on the line.
- A failed judge call never becomes a verdict. If the AI can't be reached mid-answer, the attempt rolls back and nothing is scored — you just answer again. A network hiccup should cost you seconds, never points.
Where the line wobbles
Now the honest part, same as last time: language is messy, and "close enough" has edges. A judge that accepts partial answers and typos will occasionally draw the line somewhere your table wouldn't — a "we'll give it to you" you'd have contested, or a "close but no cigar" that stings. The blindfold guarantees the line is drawn the same way for everyone; it doesn't guarantee you'll always agree with where it lands.
When you don't, tell us — the feedback form is right in the menu, and judged answers you disagree with are exactly the reports we most want to read. The judge's standing orders got to their current shape the same way the question pipeline did: one argued-about ruling at a time.
Type the answer. The whole point of trivai is that you get to.