Typed Decisions, Not Chat

A single anonymous source. The audit puts the DCVC-led $40M seed on the record but notes the $200M post-money figure comes from a news piece citing one person familiar with the deal. Two data points cannot underwrite anything, and the audit is explicit that it recommends no purchase, build or pause.

No. That is thirteen checks batched into one call against the same thirteen sent individually, on a roughly 54,000-character Wikipedia article, and the notebook pins jev-1.12. Batching your own calls is a win whichever model serves them, so read it as a cookbook result rather than a model comparison. The audit notes these are worked examples, not independent evals.

Yes. The audit cites OrcaRouter reading 67.8% mean agreement for Jev against 74.1% for the best comparator, so on that workload a rival came out ahead. It also notes a defect test that missed one of seven: the kind of miss that matters more in code review than an average accuracy number suggests.

The audit lists the gaps plainly: no public paper, no parameter count, no weights, no data residency table and no DPA. RLCD is named as the training method but the algorithm and architecture behind it are unpublished, so the calibration claim cannot be independently checked. The site does carry a claims audit and an audit trail if you want the sourcing.

Noul is the true/false primitive answered 0 to 1: near 1 is a strong yes, near 0 a strong no, and near 0.5 means yes and no look similar rather than medium skill. It has no separate confidence field, and the instruction can be a question or a statement evaluated for truthfulness.

TypeSafe's own list, last reviewed 17 September 2026: literal reading, counting, math, date comparison, indirection, large irrelevant state, adversarial content, contradictory instructions versus criteria, and generation. It says to do arithmetic and date order in code and not to interpolate a Score between levels.

Early access opened 15 September 2026. The HTTP call is POST https://api.typesafe.ai/v1/systemone with model alias jev-latest, and the launch list price is $0.042 per million input tokens with $0 output.