Recorded J-lens read-outs of open models, replayable token by token. Each trace shows, at every generated token, the concepts the model is inclined to say next — split into 中文 and English — with a rigor layer that tests whether the two languages agree beyond a shuffled-pairing null. Traces are permanent, deep-linkable artifacts. How it works →
Loading traces…