The dojo grades every answer with the engine, never with the model: you build an automaton, grammar or Turing machine, and the grader compares it against a reference over a shared work budget. Solutions are revealed only when you give up. The pumping drill and the NP lab run entirely offline.
localStorage.The model only proposes. Every generated problem carries a reference solution plus example words, and the engine must agree with all of them before the problem is shown; a 4-attempt repair loop feeds engine errors back to the model. For Turing machines the model only phrases the task; the reference comes from a hand-verified library of nine machines, because models rarely write correct TMs.
Problems were measured with the app’s own validation loop against the live engine: 10 generated problems per model, each counted only if the reference agreed with every example (the same bar the dojo enforces before showing a problem):
| Model | Sound within 4 attempts | Avg. time |
|---|---|---|
| google/gemini-3.1-flash-lite (default) | 9 / 10 | ~2 s |
| deepseek/deepseek-v4-flash | 4 / 5 | ~15 s |
| deepseek/deepseek-v3.2 | 4 / 5 | ~19 s |
| google/gemini-2.5-flash-lite | 2 / 5 | ~5 s |
| anthropic/claude-haiku-5.5 | 1 / 5 | ~17 s |
google/gemini-3.1-flash-lite is the minimum
required model for the sensei to work reliably, and it is the
default. It is also the fastest of the set; a problem costs well under a
cent.anthropic/claude-haiku-5.5 (an earlier
default) phrases pleasantly but was unreliable at composing problems: it
repeatedly produced JSON the parser could not read and references that
contradicted its own accept/reject examples. It remains usable for the
Turing-machine topic, where the reference comes from the verified
library and the model only writes the wording.openai/gpt-5-mini is incompatible with
the dojo: it requires reasoning mode, which the sensei disables
(reasoning burns the whole token budget before answering).DOJO → SENSEI'S BRAIN → paste an API key from
openrouter.ai. The model field defaults to the verified minimum
above.
Point the base URL at http://localhost:11434/v1; start
Ollama with OLLAMA_ORIGINS=<this app's origin> so the
browser may call it. Local models are weaker: the engine still
re-validates everything, so a bad model shows up as failed generations,
never as wrong grades.
Any OpenAI-compatible /chat/completions endpoint: paste
the base URL and, if the endpoint needs one, an API key.