Jev Fabric

Jev Fabric is a typed decision control plane for bounded semantic choices. It lets an application send a small, explicit question and candidate set to a decision provider, then applies deterministic policy before a host performs anything. The host model keeps open-ended reasoning; policy and the host keep authority.

It is useful when a workflow needs a repeatable selection, screen, rank, verification, risk signal, progress classification, or completion check. It is not an autonomous executor, permission system, browser controller, or replacement for domain validation.

Alpha status: the repository is public source code, not a claim that every deployment is ready for every risk level. Start with offline examples and own evaluations.

Five-minute offline quickstart

Choose either the checkout route or the clean packed-artifact route. Both are copy/paste-tested offline paths; neither makes a provider request or needs a credential.

npm install --offline --ignore-scripts
npx --no-install jev-fabric doctor --json
npx --no-install jev-fabric evaluate --json

See Quickstart for both routes and Examples for executable, offline decisions and boundary demonstrations.

Native TypeSafe live use is deliberately opt-in and local-stdio only. An operator must directly supply --live, the pinned native provider/model, a credential environment name, fixed tenant/action, call/token/dollar/deadline bounds, and a conservative input-price ceiling. See the live operator guide. Never place a key in source, a config file, a receipt, generated host artifact, or client-side code.

Server applications can also use the pinned Vercel AI Gateway route through the same strict TypeSafe response mapping. The current free promotion ends September 25, 2026; pricing is external, mutable state and is never embedded in Fabric policy. The Gateway factory accepts an explicit server-side credential and fixes the endpoint, provider ID, and typesafe-ai/jev model. It is not exposed as a browser or general CLI escape hatch.

Related MCP server: agenticrail-mcp

Choose a path

Decision boundary

host → deterministic preflight → projected state + candidates → cache → provider
     → response validation → deterministic policy → receipt/ticket → host executor

The provider can propose a bounded semantic result. It cannot mint identity, grant a permission, bypass a static denial, execute a command, or make a browser action. A receipt is redacted provenance, not authorization. Architecture explains each boundary.

Pick the right component

This is an integration-shape comparison, not a speed, cost, or quality ranking. Combine the components only across their stated trust boundaries.

Built-in packs

Read pack selection before using a pack. Candidate coverage, freshness, and policy are application responsibilities.

Evidence snapshot

The only committed historical provider record is historical-v0: evidence class local_exploratory, imported 2026-09-19, zero retained cases, redacted environment, and no held-out claim. Its old ECE calculation was invalid because it used TypeSafe confidence rather than maximum distribution probability; no corrected ECE is available. The tool-environment benchmark contract is explicitly NOT_RUN and prevents null fixtures from becoming performance claims. See Evidence.

When not to use Jev Fabric

Do not use it for direct financial decisions, identity or access grants, execution of model-selected commands, high-impact classifications without a suitable human and domain-control process, or tasks that deterministic code can answer exactly. A general reasoning model or a human may be a better next step when the task is open-ended rather than a bounded decision.

Compatibility and project links

Host compatibility is explicitly scoped in COMPATIBILITY.md; all host live results are NOT RUN unless an artifact says otherwise. This repository is not affiliated with or endorsed by TypeSafe AI. “TypeSafe” and “Jev” are used only to describe interoperable products; no logo or endorsement is implied.

Read Security, release operations, Support, license, roadmap, and AI source map.