Jev Foundations

Last researched: 2026-09-18

This document is a source-classified starting point. It intentionally separates TypeSafe's claims from independently demonstrated behavior.

FACT

TypeSafe exposes Jev as a decision model accepting shared state plus named questions. Its documented question primitives are noul (a Boolean probability), choice (one option and a distribution), and score (a position over ordered levels). Source: TypeSafe quick start and primitives documentation.
Vercel AI Gateway added typesafe-ai/jev support on 2026-09-16 through AI SDK 7's experimental evaluate API. Source: Vercel announcement.
TypeSafe calls this category “System One Models” and calls its training method “Reinforcement Learning for Calibrated Decisions” (RLCD). These names and the existence of the published interfaces are facts; their performance implications remain vendor claims. Source: TypeSafe launch post.

VENDOR CLAIM

TypeSafe describes Jev as machine-native intelligence that returns typed decisions and calibrated confidence rather than generating prose. Source: TypeSafe home page.
TypeSafe reports up to 193.6× lower latency and 444.6× lower cost than LLMs on its workflow evaluations. These are vendor-reported benchmark results, not independent findings. Source: TypeSafe home page.
TypeSafe markets “zero hallucinations.” That wording is not treated as an independently established reliability property; a typed output can still be a wrong decision. Source: TypeSafe home page.
TypeSafe reports pricing of $42 per billion input tokens ($0.042 per million) and no metered output-token charge. Pricing may change and should be checked before use. Source: TypeSafe home page.

INDEPENDENT OBSERVATION

A third-party Home Assistant integration uses Jev decisions as sensors and automation actions, including frequent evaluation with a local daily token-budget guard. This is evidence of an actual integration, not proof of vendor benchmark claims. Source: HA-Jev repository.
A third-party MCP server exposes the native TypeSafe evaluation API to coding agents and preserves probabilities in machine-readable results. Source: typesafe-mcp repository.

OPINION

Jev is best evaluated initially as a complement to generative models—a router, verifier, policy-like decision layer, or frequent classifier—rather than as a general replacement for models that must synthesize text or code. This is an engineering interpretation, not a settled fact.

SPECULATION

If its latency, cost, and calibration properties hold across real workloads, Jev may make architectures with tens or hundreds of semantic decisions per event practical. This requires workload-specific validation.

UNKNOWN

Independent accuracy, calibration, drift, and latency measurements across representative production workloads.
Failure behavior under adversarial, ambiguous, multilingual, or distribution-shifted state.
Exact model architecture and how RLCD differs technically from other calibration/training methods beyond TypeSafe's public description.
Long-run pricing and operational limits outside the currently published API constraints.

FACT

TypeSafe exposes Jev as a decision model accepting shared state plus named questions. Its documented question primitives are noul (a Boolean probability), choice (one option and a distribution), and score (a position over ordered levels). Source: TypeSafe quick start and primitives documentation.

Vercel AI Gateway added typesafe-ai/jev support on 2026-09-16 through AI SDK 7's experimental evaluate API. Source: Vercel announcement.

TypeSafe calls this category “System One Models” and calls its training method “Reinforcement Learning for Calibrated Decisions” (RLCD). These names and the existence of the published interfaces are facts; their performance implications remain vendor claims. Source: TypeSafe launch post.

VENDOR CLAIM

TypeSafe describes Jev as machine-native intelligence that returns typed decisions and calibrated confidence rather than generating prose. Source: TypeSafe home page.

TypeSafe reports up to 193.6× lower latency and 444.6× lower cost than LLMs on its workflow evaluations. These are vendor-reported benchmark results, not independent findings. Source: TypeSafe home page.

TypeSafe markets “zero hallucinations.” That wording is not treated as an independently established reliability property; a typed output can still be a wrong decision. Source: TypeSafe home page.

TypeSafe reports pricing of $42 per billion input tokens ($0.042 per million) and no metered output-token charge. Pricing may change and should be checked before use. Source: TypeSafe home page.

INDEPENDENT OBSERVATION

A third-party Home Assistant integration uses Jev decisions as sensors and automation actions, including frequent evaluation with a local daily token-budget guard. This is evidence of an actual integration, not proof of vendor benchmark claims. Source: HA-Jev repository.

A third-party MCP server exposes the native TypeSafe evaluation API to coding agents and preserves probabilities in machine-readable results. Source: typesafe-mcp repository.

UNKNOWN

Independent accuracy, calibration, drift, and latency measurements across representative production workloads.

Failure behavior under adversarial, ambiguous, multilingual, or distribution-shifted state.

Exact model architecture and how RLCD differs technically from other calibration/training methods beyond TypeSafe's public description.

Long-run pricing and operational limits outside the currently published API constraints.