# Third-party notices

This release is an independent implementation of a public model interface. It uses code and
design from two MIT-licensed projects, and it is **not** affiliated with, derived from, or
endorsed by TypeSafe AI (the makers of "Jev").

## trycua/cua — libs/cua-s1 (MIT)

- Source: https://github.com/trycua/cua/tree/main/libs/cua-s1
- Revision used: `9bbfa7d` (the commit that links the Hugging Face artifacts)
- What we use: the `tinyx` architecture (`cua_s1/model.py`), the checkpoint format
  (`cua_s1/checkpoint.py`), the synthetic-data generator logic (`cua_s1/synth.py`), the
  training loop (`training/train.py`) and the evaluation metrics (`evals/metrics.py`).
  The files are vendored unmodified in the toolkit repository under `vendor/`, and the
  generated corpus follows the same row/dataset format.
- Copyright and licence text: see the upstream repository's `LICENSE` (MIT).

Our model is **trained from scratch** on our own Swedish synthetic corpus: it shares the
architecture, the trainer and the data format, not any weights or data with Cua's release.

## vinnylarouge/jevlike (MIT), via Cua's attribution

- Source: https://github.com/vinnylarouge/jevlike
- The option-attention head design (`AttentionHead`) and the byte-level collation shape used by
  the `tinyx` model originate in `jevlike`; Cua's `model.py` credits commit `94f5fd1`
  (MIT, Copyright 2026 Minimal Labs) and we keep that attribution intact in the vendored code.

## What was NOT used

- No TypeSafe AI code, weights, data or API output. "Jev" and "System One" are TypeSafe AI's
  terms; this release is an independent model that implements a similar input/output contract.
- No real personal data. The training corpus is synthetic: fictional names, `.invalid`
  e-mail domains, locally generated personnummer-shaped identifiers that do not correspond to
  any individual, and invented organisations.