How The Daily Turing works

The September 10 demonstration set uses sample model answers and is excluded from live records. Reviewed sets use recorded provider responses.

Reviewed daily sets are scheduled for 00:00 UTC. The panel asks configured models the questions beforehand through hosted APIs or the official Antigravity CLI. Tool use and browsing are ineligible. Each model's exact configuration is recorded with the day. A model without a supported modality or a successful answer does not play that round. Publication requires at least three participating models.

Writing votes are blind until you answer. The human entries in the launch collection are excerpts from public-domain literature, with attribution shown at reveal. Questions pass source checks and editorial review before publication.

• Estimate. A number with a sourced answer. Closer wins.
• Forecast. Something that resolves after the set closes, from a public feed.
• Perception. Spot the real photo or the human voice.
• Judgement. Two answers, one human, one machine. Which is better?
• Reasoning. A short puzzle with one right answer.

The crowd's answer is the median (or plurality) of everyone who played that day. Humanity wins a day against a model by winning more rounds than the model among the rounds both sides played. Ties don't count. The scoreboard keeps the all-time record; days are recorded only once every round has resolved.

Humans can look things up and models cannot. Forecast outcomes occur after the set closes. Weather rounds use the maximum accepted station temperature observation in the stated window, not a certified daily climate record; missing data can void a round.

No account is needed. An anonymous cookie links your saved results and streak. Untimed mode is available; answers received before the daily close can count in the crowd. There are no separate timed rankings yet.