Time to first token
Tap the instant a counter starts. Adapted from the psychomotor vigilance test that sleep researchers use to measure fatigue.
45 seconds. Reports milliseconds, plus lapses over half a second.
Four evals, plus the games you pick
Four playful adaptations of reaction, processing speed, memory, and reasoning tasks. Build a baseline over your first five runs, then compare each day with your own recent history. The optional games get their own scores beside the verdict.
The evals
These four decide whether you're running normally today.
Throughput
Turn symbols into digits using a key, as fast as you can. Adapted from the digit symbol test, a task used to study processing speed.
40 seconds. Reports correct symbols per second, in tokens.
Context window
Watch tiles light up in order, then tap them back in reverse. A spatial span test of working memory that gets longer as you succeed.
About 40 seconds. Reports the longest sequence you held, in tokens.
Reasoning
One state-tracking puzzle in the style of BIG-Bench Hard and one grid puzzle in the style of ARC. Both are generated fresh each day.
About a minute. Reports puzzles solved, plus hallucinations: wrong answers you were sure of.
The games
Tracked against your own history, beside the verdict rather than in it. Switch on the ones you want in your daily run. Each starts with its own five-run warm-up.
Last guess
Three earlier guesses at a five-letter word, and only one common word still fits what they revealed. Find it.
Two puzzles, up to 90 seconds each. Reports average time to solve.
Pangram
Seven letters, and one seven-letter word that uses each of them exactly once. Find it before the clock runs out.
One puzzle, up to two minutes. Reports time to find it.
Mate in one
Three positions from engine games where one side just blundered. Find the checkmate in a single move.
Up to a minute each. Reports mates found and median time.
Shove or fold
Five heads-up hold'em hands on the flop. Go all in or fold, and try to match a game-theory solver.
About a minute. Reports chips given up to the solver.
Vibes are a bad eval, especially of yourself
Some mornings you wake up feeling like the flagship model. Some mornings you're the free tier at peak demand.
The feeling is interesting. What happens when you actually run the tests is more interesting. So every daily run starts with a prediction, before you see a single score.
After five warm-up runs, your status page starts comparing you with your own history. Call your shot. Run the evals. See how well your vibes line up.
Your status page
Example data
Within usual range
Below usual range
Above usual range
Pretraining or not compared
No run