typecastlm
Copyright 2026 Mikhail Gribov

This product includes software developed as part of the AI Cordon research line.

The model this package reads is derived from Qwen3.5-4B (Alibaba Cloud), released under the
Apache License 2.0. The derivative keeps the first 28 transformer blocks of that model unchanged
and adds a classification head of 39 rows: `true`, `false` and the 36 answer marks (`A`–`Z`,
`0`–`9`) are rows of the model's own output matrix, copied; `unsure` is a direction computed from
900 labelled rows of FEVER (CC BY-SA 3.0). No weight of the original is retrained or edited —
there are no gradients in the making of it. Upper blocks are removed, which is why the derivative
is smaller than the original and cannot generate text. Its own NOTICE states this in full.

The wording the model is read with, and the temperatures its probabilities are divided by, ship
with those weights in `prompt.json` rather than with this package: they were measured together and
must not drift apart. Where this package carries copies — `docs/API.md` quotes the shape of the
TypeSafe Jev API, which it follows so that callers of that interface can change one URL — those
are descriptions of an interface, not code taken from it. This package is not affiliated with
TypeSafe AI.
