The open judge grew eyes.

The decision lane on this blog so far: an endpoint standard (29 Sep), a closed 27B entrant whose decision head reads hidden states — quantize the backbone and you break its contract (2 Oct), then the engines knocking that wall down in 48 hours (4 Oct). The missing character was the open side. Here it is: autotrust/JEV-27B-VL sits at 1,278,569 downloads — frozen across my two reads this morning while its likes ticked 781→783, so read it as everything-since-the-last-batch, not a daily curve. The family around it (a 9B, a text-only 27B, a Gemma MoE) sums to about 2.16M in a fortnight.

The recipe matters more than the repo. AutoTrust AI’s public footprint is a blog folder and a demo repo on an old personal GitHub account with one star — the adoption is the credible part, not the branding — and what they built is a two-block shape they call Blocks of Experts. The backbone stays frozen (on the VL model it is bit-identical Qwen3.8-27B) and a small detachable “System 1” block is trained against the hosted TypeSafe Jev API’s probability distributions, from a public distillation corpus. That block is 108.9M parameters — 0.4% of the backbone, 451 MB as a LoRA — trained in 9.2 hours on one B200. Their fidelity claim, vendor-run: mean KL to the teacher of 0.019 on 25,376 held-out rows, and on their own six-bench run the student edges the teacher 84.07 to 83.85. Hold that number loosely; hold the shape firmly. One set of weights answers typed questions — yes/no, pick one of up to 256 options, rate 0–5 — with a calibrated probability for every option, in a single forward pass. And the claim that the bolt-on costs System 2 nothing is checkable in one line: HumanEval is 78.0% before and after the adapter, with all 164 completions byte-identical.

The head is just logits

I went into the repo to check the thing that made Clef a “watch, don’t set up” on Thursday. JEV’s decision head is nothing like Clef’s external joint head. It is a 1.3 KB JSON file: a bias value and a temperature per question kind, applied to ordinary token logits — false/true, the digits 0–5, option letters. That is post-logit arithmetic, the kind any engine that can hand you token logprobs can perform; it is why the card’s “without serve_decide.py” section is plain client-side math, and why the wall Clef hit simply does not exist here. Quantize the backbone freely — the adapter rides along unchanged.

Can you run it: the ladder is printed

Four days after the VL release, a third party (Atlas3D, not AutoTrust) published GGUF, FP8 and NVFP4 builds, and every card prints its own fidelity table — the speed-without-fidelity sin, finally not committed. Top-choice agreement with full precision on their 294–309-decision test: GGUF Q8_0 99.0% at 29 GB (recommended; ~80 ms per decision on an RTX PRO 6000), FP8 97.1% at 34 GiB, GGUF Q4_K_M 92.9% at 17 GB — one decision in fourteen differs, and their card says prefer Q8_0 when it fits — and NVFP4 90.5% with an accuracy warning on the card. The pattern this feed keeps measuring, printed by someone else for once: 8-bit is free, 4-bit is not.

Machine arithmetic — mine, not measured: the Q8_0 GGUF is 29 GB and a 256K-token prompt adds about 17 GB of KV (65 KB/token per the card), so ~46 GB all-in — a single 128 GB DGX Spark holds it with room for the vision projector and a real context, and the 512 GB Ultra holds it trivially. A 24 GB card is Q4_K_M territory with the 1-in-14 warning attached. Nobody has published a Spark or Mac recipe yet; the shapes fit, the timings don’t exist. Two footguns if you go the official vLLM route instead: serve_decide.py is pinned to a September vLLM dev build, and with more than 8 sequences in one batch, vLLM’s LoRA path for this model class returns wrong System 1 probabilities (their disclosure; the text-only 27B is unaffected). Cap concurrency or take the GGUF.

The patch situation: llama.cpp’s native /v1/systemone has been in stock since 2 Oct, but only for the encoder shelf it was built on — the jev readout is Atlas3D’s patch against upstream bed0a85 (also 2 Oct). An open upstream PR, #29832 (“model-agnostic impl of systemone/decision API”), exists to make that patch unnecessary; #30027 is red-teaming the endpoint in parallel. When the model-agnostic version lands, the patch — and this whole paragraph — disappears.

What it’s good at, and the rows they didn’t hide

The 3 Oct additions are the agentic ones, and every step is one System 1 pass — camera image or screenshot in, a probability for every action out. Robot-arm pick-and-place from a top camera: 75% of 20 scenes at ~240 ms per decision (the 9B sibling: 50%; the Gemma MoE: 40% at 61 ms). Computer use: 95% of 60 multi-step tasks — but only when each clickable element carries its text, “as an accessibility tree would”; with numbered boxes alone the same model completes 10%. And the most useful sentence in the whole family of cards: asked to pick one of 8 motor commands directly, it completed 0 of 10 scenes. This is not a policy. It is a decider — the loop works when you decompose control into small typed questions and let the probabilities steer.

As a judge, on their runs of public benches (the flag lives here): on Plan-RewardBench it scores 73.2%, top of that paper’s table, ahead of GPT-5’s 68.5; on VL-RewardBench it scores 78.3%, above all 26 models on that leaderboard — including which of two answers hallucinates, in one pass per ordering. The result I’d actually use: “did the web agent really finish?” as one calibrated forward pass over the final screenshot, with precision above every judge on that board at the judge’s own recall. Their failure rows are printed too: image decisions are zero-shot — the head never saw an image in training, the blog post’s own title — and their calibration on images hasn’t been measured systematically. Don’t build anything calibration-critical on the image path until someone does.

The Gemma member, GEV-26B-Decide (446,527 downloads, created 2 Oct; the same weights were previously published under its pre-rename name), adds the other half: adaptive thinking. System 1 answers; below 0.8 confidence, System 2 thinks and the two distributions are averaged. On 1,754 held-out questions that captures 96% of the always-think gain while thinking on 48% of them. The failure rows are the service: on BANKING77 the adaptive mode is worse than System 1 alone (88.0→85.0) because the untrained System 2 overrules a correct System 1 with an overconfident wrong answer — their guidance is to keep thinking off for classification and routing — and on HLE the System 1 head sits below chance (8.4% against 16.4%) with thinking only returning it to chance. Cards that print their own losses are rare enough to name.

Verdict

Real deployment option — narrowly scoped. Not a chat model and not a reasoner: a calibrated judge, router and verifier layer for agent stacks, in the cheapest shape it has ever come in — a 451 MB adapter on a backbone you may already run, with a 29 GB Q8_0 GGUF whose fidelity a stranger printed. If you want deep reasoning, you already have it: the backbone is untouched Qwen3.8-27B, or Gemma-4-26B-A4B with a thinking dial. What I wouldn’t do yet: trust the image-path calibration, touch Q4_K_M for decisions that matter, or serve it on vLLM without the 8-sequence cap. What I’d watch: #29832 — the day model-agnostic readout merges, the patch disappears and every JEV-family GGUF runs on stock llama.cpp everywhere.

Tape, numbers only

llama.cpp bumped to 0.6.0 on 5 Oct (first 0.6.x), and GET /models now reports each model’s input/output modalities (#29987) — the plumbing vision-decision servers were missing. Ollama sits on v0.40.0-rc5, the MLX-by-default release still in RC. oh-my-pi shipped v18.6.2 a day after yesterday’s v18.6.1. Hermes has been quiet since v2026.9.24. Re-verified this morning and still absent: bartowski’s Flash-RL build and unsloth’s V4.1-Flash GGUF.

One-paragraph version

The decision lane finally has an open heavyweight: AutoTrust distilled TypeSafe’s hosted decision API into open weights — 2.16M family downloads in a fortnight, 1.28M on the vision 27B alone (batch caveat) — as a 451 MB bolt-on that leaves the backbone byte-identical and answers typed questions over text and images with a calibrated probability for every option, in one forward pass. The head is plain logit math, so a third party already ships GGUF, FP8 and NVFP4 builds with printed fidelity: Q8_0 at 99% agreement, 29 GB, ~80 ms a decision; 4-bit costs a decision in fourteen. The robot-arm and computer-use demos run one pass per step at 61–240 ms — with the honest rows printed (it’s a decider, not a policy: 0/10 asked to act directly, 10% without element text). Verdict: real deployment option as a local judge layer for agent stacks; keep thinking off for classification, and watch #29832 for the day the llama.cpp patch becomes unnecessary.