Decision modelOpen weights
AutoTrust JEV-27B
AutoTrust AI
- Released
- September 27, 2026
- Data date
- October 7, 2026
AutoTrust JEV-27B has two separate output modes. Its System One adapter returns Noul, Choice, or Score probabilities, while the unchanged Qwen3.8-27B path still generates text and reasons step by step.
AutoTrust trained the decision adapter as an independent student on published output distributions from TypeSafe Jev 1.13. Its weights share neither code nor parameters with Jev and are released under Apache-2.0. The native context is 262,144 tokens, but vLLM must be configured to use the full length.
Specifications and access
| Specification | Value and source |
|---|---|
| Model class | One weight checkpoint for System One decisions and System Two text generationSource |
| Parameters | 26,895,998,464 in Safetensors metadataSource |
| Input modalities | Text and JSON statesSource |
| Output modalities | Calibrated Noul, Choice, and Score probabilities or ordinary generated text in System Two modeSource |
| Architecture | Unmodified Qwen/Qwen3.8-27B language model plus a separate LoRA adapter and decision headSource |
| Context window | 262,144 tokens natively; vLLM must be started with max-model-len 262144 to use the full windowSource |
| License | Apache-2.0Source |
Page 1 of 9
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
JevBench public · AutoTrust evaluation (Show measurement, test conditions, and source)
- Source value
- 88.7
- Score
- 88.7
- Metric
- family-macro score, percent
- Unit
- %
- Benchmark version
- AutoTrust results updated 2026-09-27
- Category
- public-decision-benchmarks
- Direction
- Higher is better
- Sample
- 231
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AutoTrust ran exact checkpoint autotrust/JEV-27B System 1 on 2026-09-26; Public JevBench, 231 tasks; the model card explicitly says this family-macro score is not the JevBench v1.4.2 leaderboard metric.
Kev · AutoTrust evaluation (Show measurement, test conditions, and source)
- Source value
- 83.75
- Score
- 83.75
- Metric
- equal-weight mean accuracy, percent
- Unit
- %
- Benchmark version
- AutoTrust results updated 2026-09-27
- Category
- public-decision-benchmarks
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AutoTrust ran exact checkpoint autotrust/JEV-27B System 1 on 2026-09-26; Equal-weight mean of six clean-accuracy results across the development and test splits of decision-v7, transfer-v4 and transfer-v9.
OpenJev text · AutoTrust evaluation (Show measurement, test conditions, and source)
- Source value
- 73.89
- Score
- 73.89
- Metric
- mean accuracy across 19 static tasks, percent
- Unit
- %
- Benchmark version
- AutoTrust results updated 2026-09-27
- Category
- public-decision-benchmarks
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AutoTrust ran exact checkpoint autotrust/JEV-27B System 1 on 2026-09-26; Mean across 19 static text tasks: NLI, multiple-choice reranking and fixed-candidate GSM8K.
Nimble · AutoTrust evaluation (Show measurement, test conditions, and source)
- Source value
- 92.91
- Score
- 92.91
- Metric
- accuracy, percent
- Unit
- %
- Benchmark version
- AutoTrust results updated 2026-09-27
- Category
- public-decision-benchmarks
- Direction
- Higher is better
- Sample
- 282
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AutoTrust ran exact checkpoint autotrust/JEV-27B System 1 on 2026-09-26; Fixed Nimble subset; 282 items.
VitaminC · AutoTrust evaluation (Show measurement, test conditions, and source)
- Source value
- 77.46
- Score
- 77.46
- Metric
- accuracy, percent
- Unit
- %
- Benchmark version
- AutoTrust results updated 2026-09-27
- Category
- public-decision-benchmarks
- Direction
- Higher is better
- Sample
- 599
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AutoTrust ran exact checkpoint autotrust/JEV-27B System 1 on 2026-09-26; Fixed VitaminC-dev subset; 599 items.
MASSIVE-en · AutoTrust evaluation (Show measurement, test conditions, and source)
- Source value
- 87.71
- Score
- 87.71
- Metric
- accuracy, percent
- Unit
- %
- Benchmark version
- AutoTrust results updated 2026-09-27
- Category
- public-decision-benchmarks
- Direction
- Higher is better
- Sample
- 350
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AutoTrust ran exact checkpoint autotrust/JEV-27B System 1 on 2026-09-26; Fixed English MASSIVE subset; 350 items.
Six-group public decision benchmark mean · AutoTrust evaluation (Show measurement, test conditions, and source)
- Source value
- 84.07
- Score
- 84.07
- Metric
- equal-weight mean of six reported group scores, percent
- Unit
- %
- Benchmark version
- AutoTrust results updated 2026-09-27
- Category
- public-decision-benchmarks
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AutoTrust ran exact checkpoint autotrust/JEV-27B System 1 on 2026-09-26; Mean weights the six public groups equally and uses the model card’s unrounded JEV-27B measurements.
decision-models-under-pressure · option-order flip rate (Show measurement, test conditions, and source)
- Source value
- 7.4
- Score
- 7.4
- Metric
- share of answers changed by shuffling options
- Unit
- %
- Benchmark version
- published 2026-09-25 benchmark
- Category
- human-gold-choice-benchmark
- Direction
- Lower is better
- Sample
- 800
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact JEV-27B, 16 options, five orderings; human-gold benchmark uses 800 items across four domains.
Choice top-1 agreement · clear top option (Show measurement, test conditions, and source)
- Source value
- 95.8
- Score
- 95.8
- Metric
- agreement when Jev target top-two probabilities differ by at least 0.1
- Unit
- %
- Benchmark version
- held-out test_set_30k / jev-distill-corpus-v3
- Category
- system-1-calibration-and-fidelity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact autotrust/JEV-27B System 1. System 1 choice decisions on held-out rows where the teacher target’s top two options differ by at least 0.1.
- Context
- Denominator for this clear-margin subset is not published.
Choice top-1 agreement · Jev-labelled (Show measurement, test conditions, and source)
- Source value
- 90.5
- Score
- 90.5
- Metric
- top-1 agreement with Jev
- Unit
- %
- Benchmark version
- held-out test_set_30k / jev-distill-corpus-v3
- Category
- system-1-calibration-and-fidelity
- Direction
- Higher is better
- Sample
- 8312
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact autotrust/JEV-27B System 1. System 1 on the test_set_30k yuri_v3 · choice slice, 8,312 rows.
- Context
- The model card reports 90.5% on Jev-labelled rows.
Choice top-1 flip rate under option shuffle (Show measurement, test conditions, and source)
- Source value
- 2.9
- Score
- 2.9
- Metric
- share of top-1 choices changed by option permutation
- Unit
- %
- Benchmark version
- held-out test_set_30k / jev-distill-corpus-v3
- Category
- system-1-calibration-and-fidelity
- Direction
- Lower is better
- Sample
- 1000
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Exact autotrust/JEV-27B System 1. System 1 choice tasks; 1,000 rows with four random permutations.
- Context
- The card reports an aggregate over four random permutations.
HumanEval · JEV-27B System 2 (Show measurement, test conditions, and source)
- Source value
- 78
- Score
- 78
- Metric
- greedy completion pass@1 (128/164)
- Unit
- %
- Benchmark version
- HumanEval
- Category
- system-2-code-generation
- Direction
- Higher is better
- Sample
- 164
- Source type
- vendor-reported
- Evaluator
- AutoTrust JEV-27B model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. System 2 uses the Qwen3.8-27B backbone with the decision adapter disabled; greedy completion-style prompt, at most 384 new tokens. All 164 completions were byte-identical to the base model.
- Context
- This is the System 2 generation path; it is not a System 1 decision-head score.
Page 1 of 9
Sources and data date
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 7, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Additional source | AutoTrust JEV-27B announcement (retrieved October 7, 2026) · AutoTrust JEV-27B announcement · Editorial description reviewed October 7, 2026 |
| Additional source | AutoTrust JEV-27B model card (retrieved October 7, 2026) · AutoTrust JEV-27B model card · Editorial description reviewed October 7, 2026 |
| Additional source | AutoTrust JEV-27B weight metadata (retrieved October 7, 2026) |
| Additional source | Jev Decision Index 0.3 public-results data (retrieved October 7, 2026) |
| Additional source | Jev Decision Index 0.3 score decomposition data (retrieved October 7, 2026) |
| Additional source | AutoTrust JEV-27B model card (retrieved October 7, 2026) |