What is a decision model
A decision model returns a typed answer with probabilities. Methodology v1.0. The crown uses Decision Index Full.
Typed answers
A decision model takes state, as text or as images, plus a typed question. It returns a typed answer with probabilities rather than free text. Output types: noul (yes/no/unknown), choice, score, multi-label.
Decision models are their own category. They are not folded into General LLM.
How the crown is decided
Jev Decision Index (https://huggingface.co/spaces/multimodalart/jev-decision-index) is an independent primary. Decision Index Full has weight 1. JevBench (https://benchmarkheaven.com/jev-models/v1.6.1) is a second independent primary. JevBench has weight 1. Open-weight rows use capability. Hosted API rows use the headline composite.
A gap of 1 index point or less is a tie. Tied leaders are co-SOTA and share rank 1. Everyone else is labeled with a best rank.
A later crown needs 2 consecutive snapshots of the same index edition, or 2 independent primary sources agreeing. A new edition is logged as board re-based, and the consecutive snapshot count starts over. A new JevBench revision resets that source the same way.
Calibration
The calibration slice ranks the lowest ECE among models in the top quartile of accuracy, so a constant probability cannot take the slice. Brier breaks ECE ties.
ECE = Σ w·|accuracy − confidence|. When the index publishes bins, the recomputed ECE must match the published value within 0.002. Vendor tables are not used to compute calibration.
Open weights and vision
Open weights means the index lists a weights repository or a Hugging Face model repository. Closed models stay on the overall slice.
The vision category fills when an independent vision board is ingested. Text scores stay on the decision category.
Latency, price, and hosted APIs
The latency slice ranks the lowest same-GPU median milliseconds among models within 5 points of the accuracy leader. Hosted API round-trip stays off that slice.
The price slice ranks the lowest dollars per 1,000 decisions among models within 5 points of the JevBench leader.
The hosted slice ranks JevBench API rows by the headline score.
Vendor numbers
Vendor decision claims are display-only at weight 0. A vendor-run row, including a maker's own board, does not move a crown.
When the index omits a release date, the first snapshot date is stored as first seen. That date does not start the new-model hold.