Puzzles and accessibility
What doubtful visitors are asked to do.
Puzzles are drawn from generated scenes that never repeat, so there's no image set to collect and label. Each is shown a limited number of times, then retired.
- Slider: drag a piece into its gap. Partway along, the piece moves a little faster or slower than the pointer; people adjust without noticing, scripts don't.
- Rotate: turn a picture upright.
- Path: trace a line from a dot to a square.
What's graded is how the answer was reached: real hands speed up and slow down, drift off-axis and correct in small steps. A right answer reached by a scripted drag fails.
Live challenge
For the most suspicious traffic, the puzzle is streamed: the browser only receives pictures and sends its raw drag. Partway through, the server moves the target back. People reverse within about a quarter of a second; bots that screenshot and think take far longer. Lag is measured and allowed for, and on very slow connections the visitor gets a regular puzzle with extra proof of work instead.
Keyboard and screen readers
Every puzzle has Use the keyboard instead, which swaps in a puzzle solved with the arrow keys and Enter, with a little extra proof of work to make up for the weaker signal. Prompts and results are announced to screen readers. There are no audio puzzles: speech recognition solves those easily.