Drop-in reCAPTCHA replacement. Every challenge answered is a human-labeled training example added to your pipeline. No Google. No surveillance.
Unlike image grids, Lyra's challenges test genuine cognitive ability. Bots fail. Humans enjoy them.
Find the next number in a pattern — tests logical reasoning.
Interpret emoji combinations — tests cultural and contextual understanding.
Unscramble a word — tests vocabulary and pattern recognition.
Identify a beat pattern — tests timing and sensory processing.
General knowledge questions that bots can't easily look up.
Finish a sentence creatively — tests language generation ability.
Distinguish human vs AI text — generates AI detection training data.
Describe an image — generates vision model training data.
Predict physical outcomes — tests spatial and causal reasoning.
Identify tone and intent — generates emotional intelligence data.
Responses are stored as structured JSONL. Quality-filtered (timing + confidence). Exported in three formats: instruction-following, completion, and RLHF preference pairs.