Skip to content

Decision-model ecosystem intake

This page tracks integration shape, not model quality. A listed project is not endorsed or promoted merely because SchemaRouter can connect to it.

Last reviewed: 2026-09-28.

The machine-readable discovery/benchmark intake state lives in benchmarks/system-one-candidate-registry.json. Keep model discovery there instead of hard-coding candidate names into planner logic.

The promotion target itself is centralized separately in benchmarks/operation-routing-production-targets.json. The registry mirrors that target for discovery UX, and tests reject drift between the two files.

Intake rule

Choose the narrowest stable boundary that fits a new model/runtime:

Integration shape SchemaRouter path Core change normally needed?
Jev/System One-compatible HTTP API SystemOneDecisionBackend No
Direct official Laya Python runtime LayaDecisionBackend No
One-off Python experiment CallableDecisionBackend / --decision-callable No
Reusable non-System-One Python runtime schemarouter.decision_backends plugin No
New semantic primitive that does not fit DecisionBackend research-only adapter first Maybe, after evidence

Current discovery matrix

The ecosystem changes quickly. Pin exact repository/model/runtime revisions before any benchmark.

Microsoft-Decision-1 (2026-10-09 announcement)

  • Microsoft hosted Qwen3.5-9B-derived typed decision model (choice, noul, score);
  • Foundry endpoint /providers/microsoft/v1/systemone, with Azure deployment-name model and Entra/token or api-key authentication; OpenRouter ID microsoft/microsoft-decision-1;
  • not automatically TypeSafe SDK-compatible: existing SystemOneDecisionBackend expects the TypeSafe response.choices object; Foundry HTTP documents answers JSON and different credentials;
  • no publicly released Microsoft-modified weights, so it is a hosted paid candidate, not local Laya replacement;
  • no measured SchemaRouter quality or open-set refusal accuracy; keep out of production promotion until #1324 verifies transport and independent supported/no-fit heldout;
  • dedicated Microsoft-Decision-1 intake with official URLs, price and no-match safeguards.

Jev / TypeSafe

  • hosted reference System One API;
  • existing SchemaRouter integration: JevDecisionBackend;
  • generic compatible path: SystemOneDecisionBackend;
  • credentialed comparison only; never a CI requirement.

Laya

  • open typed-decision model family;
  • direct Python path: LayaDecisionBackend;
  • supports typed choice, score, and noul primitives in its native runtime;
  • community HTTP/ONNX runtimes can use the generic System One path when they preserve the compatible contract.

Research status is tracked separately. The current v4 evidence rejects the tested direct full-catalog Laya choice configuration and the pinned winner-only native noul veto, but those results do not imply every Laya checkpoint/runtime/primitive has the same behavior.

Kev

Repository family: jaredpalmer/kev.

  • Jev/System One-compatible /v1/systemone server;
  • open model sizes include 0.8B, 4B, 9B, and 27B classes;
  • supports choice, noul, and score;
  • SchemaRouter path: SystemOneDecisionBackend;
  • model size/revision is a benchmark variable, not an integration-code variable;
  • current SchemaRouter research: pinned Kev-0.8B choice+noul #299 / PR #300 is the active 1,800-row v4 screen.

Mapika decider

Repository: Mapika/decider.

  • Jev/System One-compatible runtime family;
  • SchemaRouter path: SystemOneDecisionBackend;
  • candidate model/runtime revision must be pinned independently before benchmark;
  • no SchemaRouter quality result yet.

Decis

Repository: chaitin/Decis.

  • self-hosted Jev-compatible server;
  • presents multiple engines behind one /v1/systemone endpoint;
  • currently includes Laya and Kev engine paths;
  • SchemaRouter path: SystemOneDecisionBackend.

This is an example of why SchemaRouter should target the wire contract instead of adding one class per underlying model.

LiteVar System One

Repository: LiteVar/system-one.

  • local-first cross-platform System One runtime;
  • Jev-compatible API;
  • model-independent runtime abstraction with Laya as an initial backend;
  • SchemaRouter path: SystemOneDecisionBackend.

laya-serve

Repository: stiermid/laya-serve.

  • Jev-compatible HTTP serving layer for Laya;
  • SchemaRouter path: SystemOneDecisionBackend;
  • use direct LayaDecisionBackend instead when an application wants the Python runtime in-process.

OpenJev variants

Multiple unrelated projects use the OpenJev name. Treat each repository/runtime as a distinct provider and pin it explicitly.

The currently tracked discovery entry is SiliconLabAI/OpenJev.

Its source repository is MIT at the observed discovery revision, but backend/model licensing must still be checked independently for the exact selected runtime. Use SystemOneDecisionBackend only when a selected backend actually preserves the compatible wire contract. Otherwise use a decision plugin/callable.

AnyJev

Repository: nokia-applied-research/AnyJev.

  • converts causal LLMs into typed-decision/readout variants;
  • runtime/readout support differs by level/model;
  • SchemaRouter path: third-party schemarouter.decision_backends plugin or research callable unless/until a stable System One-compatible server contract is pinned;
  • current SchemaRouter state: #311 / PR #313 stages a zero-label L0 content-free noul veto behind the immutable BGE winner;
  • the research workflow remains dormant until explicitly activated by the guarded marker/manual path, and the marker is forbidden until all preregistered Kev-based candidates are non-promotable.

Bespoke Nimble

Repository: bespokelabsai/nimble.

  • typed candidate-scoring model/runtime rather than a Jev-wire-compatible server;
  • model/checkpoint calibration is release-specific and must be pinned with the selected runtime;
  • the current Bespoke-Nimble-9B model card reports Apache-2.0, while the observed GitHub code repository has no root LICENSE; verify code and model licensing separately for the selected release;
  • SchemaRouter path: plugin or research callable;
  • larger memory/GPU requirements make it a separate runtime experiment rather than a drop-in CPU comparison;
  • no SchemaRouter quality result yet.

System One Open

Repository: mithalouni/system-one-open.

  • open Jev-style typed-decision implementation with a distinct serving/runtime contract;
  • SchemaRouter path: plugin or research callable unless a compatible wire endpoint is explicitly verified;
  • pin weights/runtime independently before treating a hosted demo endpoint as reproducible evidence;
  • no SchemaRouter quality result yet.

Other model families and runtimes

Projects such as alternative Laya ONNX/TypeScript runtimes, decision heads, distilled decision models, or future typed-decision engines should enter through the same intake rule rather than receiving a permanent SchemaRouter core class by default.

Promotion workflow for a new model

Compatibility and model promotion stay separate:

  1. pin repository/runtime/model revisions; observed registry revisions are discovery provenance only;
  2. connect through System One, plugin, or callable without changing planning authority;
  3. run the frozen benchmark protocol;
  4. record exact-route, unsupported rejection, false-route rate, latency, errors, and authority violations;
  5. freeze semantics before any fresh-surface confirmation;
  6. only promote a default/recommended profile when independent evidence passes the active quality gate.

Authority invariant

No integration path changes the core rule:

A model may score, select from, or veto finite locally authorized choices, but it may not create tools, endpoints, fields, arguments, policy, or execution authority.

This rule applies equally to hosted Jev, Laya, Kev, OpenJev, AnyJev, and future models.