The three tiers
Open learning facility. The seed and answer key can be reconstructed from this page’s source. Use it to practise investigation, not as a blind test or evidence of learning.
Research boundaries and protections
The software knows which one of five author-designed automata it created. Users can investigate a sequence-to-lamp relationship by asking 12 questions, and then submit a machine with one to four states whose predictions are scored against 24 unqueried sequences. This demonstrates a narrow form of active identification and out-of-sample evaluation.
It does not prove that a person or language model has created a genuinely new ontology. Same-screen JavaScript is inspectable; it cannot provide adversarial experimental blinding. The number of probes and state slots are prescribed. Evaluation pairs have matched final action but are selected without checking results. Baseline models receive the same observations. Changing seed or trying repeatedly can introduce selection effects. Report each run, including failures.
Nothing is sent to any server by this page. The export is a local JSON file, which includes any words you type as state names or notes. Use constructed, non-sensitive inputs. Do not use it for real-world medical, engineering, ecological or safety decisions.