Are there open-source System One models?

Yes. A handful of open-source reproductions match Jev’s /v1/systemone wire format among the alternatives systemonemodels.org lists, and most of them say outright that they haven’t kept the one property that actually defines a System One model: a probability that tracks accuracy. Systemonemodels.org describes OpenDecision’s confidence as concentration, not correctness, a paraphrase of OpenDecision’s own README, which says only to treat the model’s scores as uncalibrated, and mini-jev calls its number “ranking with confidence gap, not calibrated probabilities.”

Von, a non-autoregressive local model built to the same spec, is more direct about the tradeoff than most: its own README reports 72.0% macro accuracy against Jev’s 96.6% on a 49-task suite, while beating Jev on a fast, reactive game-playing benchmark where raw speed matters more than calibrated judgment.

None of that makes an open alternative useless, it makes the calibration claim something to check rather than assume. Measuring any grader’s accuracy runs the same way: against a labeled set with a known right answer, not against a vendor’s own README, before trusting an open model’s confidence the way TypeSafe’s own training method earns Jev’s.