JEVRIEL

Give your AI JEV wings.

JEVRIEL helps your AI agent use TypeSafe JEV for fast, structured decisions: route a request, rank evidence, classify a document or choose the next step. It helps you build the integration and measure whether it improves the work.

Choose your AI host, your JEV provider and how deeply to adopt it.

Install · Light or Deep · Choose a provider · Examples · Measured results · Documentation

Codex · Claude Code · Portable skill and CLI | Apache-2.0 | v0.2.0

Deep makes JEV part of the agent's working policy. The agent considers each workflow, then chooses an appropriate tool. Exact calculations and rules stay in code; writing and extended reasoning stay with the LLM. JEV handles suitable bounded judgments within your permissions, privacy constraints and budget.

Deep depends on the host loading and following its instructions. It does not intercept every tool call or require a JEV request for every step. Full adoption policy.

Choose one command for your host:

# Codex
npx --yes github:thehan-co/jevriel install codex

# Claude Code
npx --yes github:thehan-co/jevriel install claude

# Other AI hosts with file and terminal access
npx --yes github:thehan-co/jevriel install portable

Requires Node.js 20+, npm and Git. Native plugin installation also requires the selected host's CLI. The interactive installer asks for your JEV connection and offers Light or Deep adoption. Start a new host session after installation.

Portable installation supplies a Markdown skill and a local runtime. Point your agent at ~/.agents/skills/jevriel/SKILL.md, or copy the skill to its documented skill directory. Hosts with MCP support can register the bundled server. Installation details.

Use OpenRouter, TypeSafe direct, Cloudflare, an existing JEV MCP, a compatible HTTPS endpoint or a custom adapter. Your AI host and JEV provider are independent choices.

For example, connect through OpenRouter using your OpenRouter API key:

npx --yes github:thehan-co/jevriel setup --provider openrouter
npx --yes github:thehan-co/jevriel doctor

Setup asks for the key in a hidden terminal prompt and makes one small verification request billed by your chosen provider. Secret-manager injection is also supported. doctor checks configuration without running inference.

Already connected through a JEV MCP? Choose existing MCP during setup. For an API with a different request or authentication format, supply a custom adapter. Provider choices and adapter guide.

You can choose Deep during interactive installation, or enable it afterward:

# Install the shared policy, including when you already use a native plugin.
npx --yes github:thehan-co/jevriel install portable --skip-setup

# Choose the instruction file for your host. Codex example:
npx --yes github:thehan-co/jevriel adoption deep --instructions "$HOME/.codex/AGENTS.md"

# For Claude Code, use "$HOME/.claude/CLAUDE.md" instead.

This adds a managed policy pointer while preserving other instructions. Other hosts can use their own documented persistent instruction file. Start a new session to load the policy. Provider setup is separate.

Switch back to manual use whenever you want:

npx --yes github:thehan-co/jevriel adoption light

Adoption settings, host verification and controls.

Ask your agent:

Use JEVRIEL to inspect this app. Find one decision where TypeSafe JEV could reduce latency or cost without hurting quality. Implement a reversible pilot and give me a JEVRIEL Flight Card.

For daily work, try:

Use JEVRIEL to classify these notes as tasks, references, meetings or manual review. Keep ambiguous items for review and show the usage receipt.

The runnable daily triage example uses synthetic data and shows how typed judgments become application decisions. The runtime also supports routing, ranking, candidate extraction, evidence verification and shared-state questions. Tool guide.

JEV returns typed decisions rather than arbitrary prose. A permitted output can still be the wrong decision. Exact calculations, dates, permissions and validation belong in code; open-ended reasoning and writing remain jobs for a suitable LLM. Known weaknesses.

A short answer to a practical question: does JEV improve this particular job?

JEVRIEL's original LIFT format reports four dimensions: Latency, Inference cost, Fidelity, and Throughput/trust. There is no single score that can hide a quality loss behind a speed gain.

Compare an LLM-only path, a JEV-only path, and JEV with LLM fallback. The result card shows timing, cost, correctness, coverage, failures and the next place worth testing. Full receipts sit behind it.

Every result identifies its evidence level:

• PROBE: a quick wiring and behavior check.
• CONTROLLED: matched held-out inputs and frozen rules.
• OBSERVE: real work without a matched control.

The bundled public example classifies synthetic inbox notes into task, reference, meeting or manual review. It includes normal requests, ambiguous notes and attempts to bypass policy. No private messages or company documents are included.

# Python 3.10+, no Python dependencies.
cp benchmark/config.example.json benchmark/config.local.json
# Verify the hosted OpenRouter model ID and JSON-schema support.
# Supply OPENROUTER_API_KEY for the LLM baseline.
# Complete provider setup first; this uses your selected JEV API source.
python3 benchmark/flight_test.py \
  --config benchmark/config.local.json \
  --out benchmark/runs/my-first-flight

For an existing community JEV MCP connection, use the MCP adapter contract. A full paid run is optional; the protocol also covers observations from normal work.

The same 1,708 text chunks from 429 documents went to JEV and hosted GPT-5.6 Luna through OpenRouter. JEV's median request took 0.422 seconds, versus 2.092 seconds for GPT, about 5x faster on this path. The result includes failures.

There was a tradeoff: JEV had 12 invalid batches out of 489, versus 2 failed or invalid batches for GPT. The models agreed on 68.9% of jointly valid chunks. Human accuracy scoring is pending, so there is no accuracy winner yet.

Estimated inference cost from recorded usage was $0.0542 for JEV versus $0.2975 for GPT. JEV uses TypeSafe's published $0.042 per million input tokens; one usage receipt is missing in each arm. These are list-price estimates, not reconciled bills.

Read the measured case study. It informed JEVRIEL's validation and fallback guidance; it is a separate experiment from the skill's prelaunch test below.

We are testing the JEV skill on JEV's own homework: 70 TypeSafe documentation pages, 15 news articles and 15 public LinkedIn profile excerpts. Five frozen decision rubrics assess segment, industry, relevance, evidence quality and next action.

The launch comparison targets GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, DeepSeek V4.1 Flash and GLM 5.3 Flash through OpenRouter. Each model sees each source once and returns all five classifications together to avoid repeated input tokens. JEV estimates use TypeSafe's published list price.

Frontier evaluation pending: The prepared run has not completed. No local-model comparison is used as launch evidence. Final numerical claims and host qualification await the frontier run.

Daniel Kahneman's Thinking, Fast and Slow is, in my opinion, one of the most important books ever written. His account of fast, intuitive System 1 and deliberate System 2 gives me a useful way to think about building with AI. It also reminds me that fast judgment can be wrong. About the book.

TypeSafe's JEV brings a fast decision model into the software around an AI agent. Give it relevant state and defined questions; it returns choices, scores or yes/no probabilities that code can use. Independent questions can share one request and run in parallel. TypeSafe's introduction.

I see this as a breakthrough in how we build with AI: a decision capability we can place throughout a workflow. JEV can help the larger agent decide where to spend its reasoning, attention and computation. That is what I mean by an AI catalyst.

I've connected JEV and put it into the triage and classification stages of my morning agent. I also ran a paired classification test on 429 documents, covering 2.67 million characters. Those experiments led to JEVRIEL: a way to teach other agents this pattern and check whether it actually helps.

The angel supplies the wings. Your measurements decide whether they help.

Sometimes the useful contribution is one small decision at the right moment. JEVRIEL teaches your agent to bring JEV the relevant evidence, ask a clear question, and keep uncertainty visible.

Even an angel needs a fallback.

JEVRIEL is Markdown with focused references and a standard-library benchmark. It adapts to the host's actual tools, output formats and model IDs. It can guide a frontier coding agent without depending on one model's private prompting conventions.

Astra, Fable, Gemini 3.8, DeepSeek 4.1 and GLM 5.3 are target environments to verify through their host/provider catalogs. This package does not claim that every named version has been tested. Adaptation guide.

Found a useful junction? Try it, measure it and share the evidence. Submit a sanitized dataset, frozen rubric, model IDs, pricing source and receipts. We review the comparison, reproduce what we can and turn accepted findings into examples and regression tests. Gains, regressions and inconclusive results are welcome.

Contribution guide · Open a Flight Test · Discuss an idea

Created by Han Rabinovitz. Its integration guidance incorporates lessons from TypeSafe’s official skill, with the applied improvements documented here. The bundled judgment adapter credits and reuses Arik Aizikovich's TypeSafe-as-a-Judge, with independent validation and accounting. Architecture and provenance. JEVRIEL and the Flight Test format are independent work; TypeSafe JEV is TypeSafe's product. No TypeSafe endorsement is implied. This is an early public release. Code and documentation use Apache-2.0; names and illustration permissions are explained in NOTICE. Frontier-model QA remains pending; no autonomous-host certification is claimed.

The research combines official contracts, dated public discussion, and implementation experience. X reports, Reddit comments and YouTube demos carry their own evidence labels. The source guide distinguishes what was read, what was measured and what remains unverified.

Start with one useful decision. Measure the lift.

Release verification: 0.2.0 evidence and limits.