Canonical deep dive: Read the full architecture article on Qualixar Research: https://qualixar.com/research/blog/jev-for-codex-qualixar-jev-control

TypeSafe Jev + Codex: A Developer Guide to Qualixar Jev Control v1.1.1

Canonical deep dive: Publish/link the full architecture article on Qualixar Research.

Source: https://github.com/qualixar/jev-codex-workbench

Release: v1.1.1

I got early access to TypeSafe's Jev through DeepSearch, with live API access. The useful thing about Jev is not that it is another model you can ask to write code. It is that it is designed to make small, typed semantic decisions.

TypeSafe calls Jev its first System One model. You send a state plus one or more atomic questions and get structured answers that code can consume directly. The three primitives are:

• Choice - choose from a known set; returns a choice, probabilities and confidence.
• Score - place the state on an ordered rubric; returns a score, distribution and confidence.
• Noul - judge whether a statement is true; returns a probability of yes from 0 to 1.

Official guide: https://docs.typesafe.ai/introduction

Primitives: https://docs.typesafe.ai/primitives

TypeSafe's own pattern library shows the broader use: intent routing, confidence-gated routing, composite scoring and speculative fan-out. Its SDE Cascade cookbook uses Jev as a verifier between a cheap extraction model and an expensive reasoning model, escalating only when semantic checks fire.

Patterns: https://docs.typesafe.ai/patterns

SDE Cascade: https://docs.typesafe.ai/cookbooks/sde_cascade

Why pair this with Codex?

A coding agent spends significant context before it writes the implementation. It needs to decide which files deserve reading, which test suites matter, which tool or skill fits, which sources are relevant and whether a completion claim is backed by evidence.

Those are bounded decisions. They do not always need another full reasoning pass.

That is the role of Qualixar Jev Control for Codex:

Local Policy Mode
    -> decide whether a bounded Jev workflow fits

TypeSafe Jev
    -> make the narrow semantic judgment

Codex
    -> reason, code, use tools and iterate

It is Jev + Codex, not Jev replacing Codex.

v1.1.1: Policy Mode

The important v1.1.1 change is a hook-backed local control plane.

Every submitted task can be classified locally as:

Result	Behavior
SKIP	Continue with Codex; no bounded semantic workflow matched.
SUGGEST	Default assist mode recommends a matching Jev workflow.
REQUIRE	Opt-in enforce mode requires the matching live Jev evaluation before governed Bash/file-edit use.
BLOCK	Sensitive material was detected locally and is not eligible for external evaluation.

The classifier makes no provider call and does not persist prompt text.

The default is assist, not enforce. The plugin does not send every Codex turn to Jev.

Why this can reduce token spend

The mechanism is selection before expansion.

Instead of loading every plausible file into the expensive reasoning trace, rank candidates first. Instead of loading every test suite, select tests first. Instead of reading every tool and skill description, route first.

TypeSafe's own Intent Routing pattern explicitly describes using TypeSafe in front of deterministic code, specialist LLMs or humans so expensive handlers are invoked only when needed: https://docs.typesafe.ai/patterns/intent-routing

Its Parallel Questions cookbook also demonstrates the economics of reusing a shared state. On one vendor benchmark - 13 questions over the same ~54k-character GDPR article - TypeSafe reports one batched call at $0.000497 / 0.27s versus 13 sequential calls at $0.006090 / 2.71s. That is 12.2x cheaper and 10.0x faster for that exact TypeSafe benchmark, not a Qualixar Codex benchmark: https://docs.typesafe.ai/cookbooks/parallel_questions

Qualixar is not publishing a Codex token-saving percentage until we run the same coding traces with Policy Mode off and on and measure Codex-side usage.

Seven tools

v1.1.1 exposes:

• jev_health
• jev_policy_status
• jev_policy_check
• jev_catalog
• jev_describe
• jev_run_fixture
• jev_evaluate

Only jev_evaluate is the live provider path.

20 decision workflows

The supplied catalog covers routing, file ranking, context selection, injection triage, claim verification, completion checks, patch review, semantic lint, failure classification, worker routing, issue/incident triage, test selection, documentation drift, security review routing, support triage, research ranking and memory admission.

Each has nominal, uncertain and adversarial fixtures: 60 offline contracts in total. Fixtures validate local behavior; they are not Jev accuracy evidence.

Live evidence

For this launch, the verified live path is TypeSafe direct with resolved model jev-1.13.0.

All 20 nominal workflows completed. The recorded outcomes were 15 RECOMMEND and 5 REVIEW. Live results remain decision inputs; they do not become shell/file/deployment permissions.

OpenRouter support exists in the code but is not being used to support the claims in this launch article.

Install

git clone https://github.com/qualixar/jev-codex-workbench.git
python3 jev-codex-workbench/scripts/install.py

After installation, restart Codex Desktop, inspect the Qualixar hooks with /hooks, and start with Policy Mode in assist.

Repository: https://github.com/qualixar/jev-codex-workbench

Release: https://github.com/qualixar/jev-codex-workbench/releases/tag/v1.1.1

TypeSafe Agent Skill: https://docs.typesafe.ai/agent-skill

The architectural thesis is simple: use Jev for fast bounded decisions, Codex for expensive reasoning and coding, and deterministic code for rules that should never depend on a model.