reflex

# reflex

**A tiny "decision model" you can run on your own GPU.**

You give it some information (a support ticket, a document, a photo) and a list of
questions with fixed answer options. It answers *all* the questions at once and tells you
**how sure it is about each option**, as percentages. It never writes free text, so it can
never make up an answer that isn't on your list.

It is an open re-creation of [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev),
the "System One" model TypeSafe released in September 2026, built on top of a normal
open-weights model ([Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) by default).

```
state:     "My payouts have failed three times this week and nobody replied to my emails."

question   which team?        ->  payments 99.9%   account 0.04%   other 0.04%
question   escalate?          ->  yes 90%
question   how urgent (0-2)?  ->  1.9   (low 4%  medium 1%  high 95%)
question   refund requested?  ->  yes 2%
```

That whole answer comes back in about 200 ms, and faster once the state is cached. Your
code then decides what to do with the numbers ("auto-route if above 90 %, otherwise ask a
human").

## Status

**What it is.** One HTTP endpoint, `POST /v1/systemone`, answered by a single forward
pass over a frozen open-weights model. No decoding, no reasoning, no escalation: about
200 ms, and the slowest request is a small multiple of the fastest.

**Where it stands.** On the public items of JevBench, against Jev itself:

| | easy | standard | hard | hard ECE | latency |
|---|---|---|---|---|---|
| reflex, frozen Qwen3.5-4B, two orders (`stable`) | 1.000 | 0.917 | 0.685 | 0.081 | ~200 ms |
| reflex, frozen Qwen3.8-27B, two orders | 1.000 | 0.958 | 0.766 | 0.061 | ~1 s |
| Jev 1.13.0 (official) | 1.000 | 0.986 | 0.