allebee / jevk5

# JevK5 — an open-weight Jev alternative
JevK5 is an independent, Apache-2.0 open-source alternative to TypeSafe's Jev for typed
decisions. Give it a state (a ticket, log, policy, or diff) and a yes/no, choice, or score question.
It returns a probability for every option in one forward pass, with zero generated tokens. The
weights run on your own GPU, and the server accepts
the TypeSafe-style `/v1/systemone` request shape. JevK5 is not affiliated with TypeSafe AI.
**v0.3** (runtime 0.3.0): five times the teacher data from two teachers, replay from public train
splits only, and a new JevK5-9B. See
What changed in v0.3. Most of the gains are on held-out and index-style
data; on JevBench's public items the 4B gains on the hard tier and loses one standard item.
**JevK5-9B v0.3.3:** the 9B retrained for 1.5 epochs (a pre-registered epoch sweep): hard tier
0.775 against v0.3's 0.730, hard-tier ECE 0.071 against 0.126, `temperature` 1.316 and
`knockout_temperature` 1.05. The 9B numbers below are v0.3's unless marked; the
JevK5-9B card has v0.3.3's, and v0.3 stays at the Hub tag `v0.3`.
**Independent result:** JevBench v1.4 ranks JevK5
v0.2 **second of 76 systems and first among open entrants** (62.04; Jev 1.13.0: 63.29). On its 308
fresh sealed decisions, JevK5 answered 33.1% correctly and Jev answered 36.7%; the evaluator calls
this set unusually
difficult. The ranking
measures JevBench's mix of accuracy, calibration, speed, and cost, not performance on every
production workflow. JevK5's reported H100 latency is about 13 ms for short decisions. On the Jev
Decision Index, JevK5 0.2.2 (v0.2
weights) scores 36.31, 15th of 49, with the 4th-best calibration. v0.3 has not been run by either
benchmark yet.
## Watch it run
**Typed decision on an NVIDIA L40S.** A real support ticket becomes a probability over three teams;
the displayed decision took 20.60 ms after model loading. This is a paced replay of recorded GPU output.
Click the moving preview for the full 27-second video.
**Snake against seeded random.** JevK5 collected 7 apples to random's 3 in a live browser capture
using Prompt Engineer 48's MIT-licensed arena.
Click the moving preview for the full 24-second video.
This game is not a Jev comparison.
| | |
|---|---|
| Weights | alibiserikbay/JevK5: Qwen3.5-4B + a distilled LoRA, merged (Apache-2.0); JevK5-9B on Qwen3.5-9B |
| Readout | SemIf's protocol: a softmax over the answer letters' next-token logits, one temperature |
| Runtime | One CUDA graph per padded input length: 13 ms vs ~70 ms eager on an H100, same answers |
| Training | Distilled from Qwen3.6-27B and GPT-6 Luna, on hard decisions they wrote and checked twice, plus public train splits |
## How JevK5 differs from Jev
JevK5 uses open Qwen3.5-4B (or 9B) weights and SemIf's option-logit
readout, not Jev's unpublished model architecture. It supports
the same three decision types—`noul` (yes/no), `choice`, and `score`—through a TypeSafe-style
endpoint. Each question is evaluated separately; th