On September 28, 2026, AutoTrust AI released JEV-27B — an Apache-2.0 open-weights decision model on a frozen Qwen3.8-27B backbone that answers yes/no, multiple-choice, and 0–5 rating questions in a single forward pass and returns a calibrated probability for every option.
That is the exact input a decision gate consumes.
And here is the line that matters most in the whole release: AutoTrust itself recommends gating JEV-27B's answers on confidence, and says the model is not meant for high-stakes decisions. The vendor's own guidance is the decision-gated payments pattern:
| Fact | Value |
|---|---|
| What | JEV-27B, open decision model for self-hosted AI agents |
| Who | AutoTrust AI Pte. Ltd. (Singapore) — CEO/co-founder Daniel Tang, chairman/co-founder Josh Liu |
| When | September 28, 2026 (PR Newswire; syndicated to Morningstar and others) |
| License | Apache-2.0 — weights, decision adapter, training + serving code, vLLM support, evaluation reports at huggingface.co/autotrust/JEV-27B |
| Architecture | 108.9M-param decision block (~0.4% of the model) on a frozen Qwen3.8-27B backbone; trained in ~9.2 B200-hours; generation path untouched (164/164 HumanEval completions byte-identical with the block off) |
| Interface | yes/no, multiple-choice, 0–5 ratings in one forward pass, calibrated probability per option |
| Speed | 137 ms median latency — ~130 decisions/sec on one NVIDIA B200 |
| Lineage | Distilled from Jev 1.13 outputs; shares no weights or code with TypeSafe AI |
| Benchmark group | JEV-27B | Jev 1.13 (AutoTrust's own run) |
|---|---|---|
| JevBench | 88.70% | — |
| Kev | 83.75% | — |
| OpenJev text | 73.89% | — |
| Nimble | 92.91% | — |
| VitaminC | 77.46% | — |
| MASSIVE-en | 87.71% | — |
| Equal-weight mean | 84.07% | 83.85% |
Fidelity to distillation target: mean KL divergence 0.017 on 25,376 held-out Jev-1.13-labeled questions. Honesty rule: the Jev 1.13 comparison was conducted by AutoTrust itself — internal comparative evidence, not independent third-party validation. Read all benchmark figures accordingly.
Every AI agent is a long chain of small decisions — which button to press, which file to open, which payment to authorize. Today those decisions go to a third-party API or, worse, to no scorer at all.
JEV-27B changes the economics: one GPU, your infrastructure, 130 decisions per second, calibrated probabilities on every one.
The gate bands the probability:
| JEV-27B per-option probability | Gate band | Action on the x402 payment |
|---|---|---|
| ≥ 0.80 | auto-pay | Fire the payment over x402 |
| 0.50 – 0.79 | confirm | Hold for human (or named operator) review |
| < 0.50 | escalate | Block, log, escalate — the payment never fires |
The gate is the product; the decider is a plug-in. Hosted Jev 1.13, self-hosted JEV-27B, and a local heuristic all score into the same bands. As Daniel Tang put it: "For companies that cannot send every decision to a third-party API, that changes both the cost and the risk."
Minted this morning against a live harness (local-heuristic-v1, calibrated=false, typesafe_wired=false):
curl -X POST https://scriptmasterlabs.com/api/harness/decide \
-H "Content-Type: application/json" \
-d '{"state":{"context":"agent payment decision"},"questions":[{"id":"q1","type":"score","question":"should the agent pay $0.10 USDC to a directory-listed MCP tool at its listed price?","scale":[0,5],"probabilities":[0.84]}]}'
curl -X POST https://scriptmasterlabs.com/api/harness/decide \
-H "Content-Type: application/json" \
-d '{"state":{"context":"agent payment decision"},"questions":[{"id":"q2","type":"score","question":"should the agent authorize payments with no per-payment approval and no spending limit for 30 days?","scale":[0,5],"probabilities":[0.62]}]}'
Both receipts escalate — and that's the point of the piece. The uncalibrated local heuristic computes its own confidence from the question text and cannot consume an externally supplied calibrated probability: 0.84 in → 0.4045 out; 0.62 in → 0.47 out.
JEV-27B's per-option calibrated probability is exactly the input this gate was designed for — the harness's own meta note says "Plug in the TypeSafe Jev API when a key is available." The plumbing runs live; the decider is the upgrade.
Benchmark figures are AutoTrust's self-reported numbers (the Jev 1.13 comparison is internal comparative evidence, not third-party validation). Coverage is release-based, not a hands-on model run. The live gate uses an uncalibrated heuristic (calibrated=false, typesafe_wired=false) that cannot consume externally supplied calibrated probabilities — today's receipts prove the plumbing runs and the mapping holds, not that the heuristic judges well.
Canonical version with full claim receipts: https://scriptmasterlabs.com/jev-27b-open-decision-model — published 2026-10-01 by ScriptMasterLabs. Verified against the September 28, 2026 AutoTrust AI release and two live harness receipts minted October 1, 2026 (~09:21 EDT).