Question 1

Will Deep mode intercept every tool call automatically?

Accepted Answer

No. Deep adds a working policy so the agent considers each workflow and chooses a suitable tool, but it does not intercept every tool call or require a Jev request per step. It depends on the host loading and following its instructions.

Question 2

Do I need TypeSafe direct, or can I use my existing setup?

Accepted Answer

No. Host and provider are independent: OpenRouter, TypeSafe direct, Cloudflare, an existing MCP, a compatible HTTPS endpoint, or a custom adapter. Setup uses a hidden key prompt plus one small verification request, and doctor checks config without inference.

Question 3

Are the 429-document speed and cost wins proof Jev is more accurate?

Accepted Answer

No. On the same 1708 chunks from 429 documents, Jev's median request was faster and cheaper but had more invalid batches, with 68.9% agreement on jointly valid chunks. Human accuracy scoring is pending, so there is no accuracy winner.

Question 4

What does setup actually run, and what do I need installed?

Accepted Answer

Node.js 20 plus npm and Git, and the host CLI for native plugin install. One install command per host starts it, provider setup is separate, and adoption commands switch between Light and Deep with a fresh session to load policy.

Question 5

What is the difference between Light and Deep?

Accepted Answer

Light runs when you ask for JEVRIEL and is the default for new installations, so you request a pilot, a judgment or a benchmark. Deep is an explicit opt-in through persistent host instructions, so the agent assesses each workflow and chooses a suitable tool on its own. Deep depends on the host loading and following its instructions and does not intercept every tool call.

Question 6

What does the Flight Test measure?

Accepted Answer

The LIFT format reports four dimensions: latency, inference cost, fidelity, and throughput or trust, with no single score that could hide a quality loss behind a speed gain. It compares an LLM-only path, a JEV-only path, and JEV with LLM fallback, and every result is labelled PROBE, CONTROLLED or OBSERVE.

Question 7

What happened on the 429-document run?

Accepted Answer

The same 1,708 text chunks from 429 documents went to JEV and to hosted GPT-5.6 Luna. JEV's median request took 0.422 seconds against 2.092 seconds for GPT, about 5x faster, but JEV had 12 invalid batches out of 489 against 2 for GPT, and the models agreed on 68.9% of jointly valid chunks. Human accuracy scoring was still pending.

Jevriel

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community