What four structural requirements does Jeva target?

Null-awareness, so models distinguish answer absence from uncertainty among candidates; permutation invariance, so reordering alternatives does not change the decision; option scalability, so growing candidate sets need no new output space; and cacheability for request-independent reuse.

How does Jeva handle a decision where no candidate applies?

A separate coverage estimate distinguishes answer absence from candidate competition, which enables explicit null outcomes. The paper reports near-perfect null detection as one of the architecture's evaluation results.

How does Jeva keep decisions stable when options are reordered?

Each option is represented independently by a shared encoder, and symmetric aggregation yields permutation-equivariant probabilities. The paper reports no decision flips under any tested permutations.

Does Jeva scale to thousands of candidates?

Yes. Shared scoring accommodates variable-size candidate sets, and the paper says Jeva scales to thousands of candidates without requiring the output space to be redefined.

How does cross-request reuse reduce latency?

Request-independent option representations and layer projections can be reused across newly assembled candidate sets. The paper reports substantially reduced latency through this cross-request reuse without changing decisions.

Which decision outputs does the Jeva architecture unify?

The paper says its mechanisms unify Choice, Noul, and Score outputs, combining strong typed-decision performance with null detection in one decision-native architecture.

Are there published baselines or benchmark numbers for Jeva?

The abstract reports qualitative results only: strong typed-decision performance, near-perfect null detection, no decision flips under tested permutations, and reduced latency. It names no benchmarks, baselines, or code release.

CommunityJev Models & Reimplementations

Jeva: System One Models from First Principles (arXiv 2610.18773)

Introduces Jeva, a decision-native architecture that encodes the request and each option separately and scores options symmetrically. It targets null-awareness, permutation invariance, option scalability, and cacheability, reporting near-perfect null detection and no decision flips under tested permutations.

See highlights

Open arxiv.org

Format: Article
Published by: Community
Last verified: Oct 2026

Tagscommunity research arxiv architecture calibration

Works forJev for Scoring Jev for Decision gating Jev for Classification

Used inJev for Software & DevTools Jev for Data & Analytics

Highlights

Argues current System One designs underemphasize null-awareness, permutation invariance, option scalability, and cacheability.
Shares one encoder for the request and each option, then refines only the request state against immutable option memory.
Symmetric aggregation yields permutation-equivariant probabilities and shared scoring handles variable-size candidate sets.
Reports near-perfect null detection, no decision flips under any tested permutation, and scaling to thousands of candidates.
Request-independent option representations let computation be reused across requests without changing decisions.

Watch out

arXiv preprint submitted October 13, 2026 (cs.AI) by Qi Li and Xinchao Wang. The abstract reports no benchmark names, baseline numbers, or code release, so the typed-decision, null-detection, and latency claims cannot be checked against an external evaluation.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

X: How the agent stack moved in Jev's first three days

YouTube: Diogo Almeida tells swyx what System One is, why RLHF failed his automation goal, and why AI should eventually disappear.

X: The 10-step roadmap, summarised

X: Giving your agents a decision brain

Related terms

Glossary definitions related to this entry.

More Jev resources like this

AnyJev Technical Report (arXiv 2610.00831)

1.3k

Reads a typed decision from one prefill of a pretrained instruction-tuned model by restricting the next-token distribution to option tokens, then corrects label and position bias with no gradient steps. Averaging over cyclic rotations lowers order-flip rates and raises accuracy on 11 of 11 models.