How is this different from a coding agent with retrieval?

Full inputs stay in a persistent local evaluator while models see a bounded surface; semantic subcalls return results rather than spawning new agent environments.

Can I quote the benchmark table as a head-to-head win?

No. The historical diagnostic keeps failed rows in the denominator as zeros and is explicitly not a matched rerun of the paper controls.

What does enabling the optional typed judgments involve?

Attach a key once through a hidden prompt or export it in the environment; without a key or with judging disabled, semantic features stay off.

How do I check the recorded evidence without rerunning models?

Recorded evidence verifies offline with the bundled script, which makes no model calls and does not reproduce the original live runs.

Which platforms and versions does Azdaja install on?

macOS 11+ on Apple Silicon and Intel, or x86-64 Linux with glibc 2.35+, via the install script; building from source needs Rust 1.95 with the typesafe feature.

How do I give Azdaja my TypeSafe key?

Attach it once through a hidden prompt so it never enters shell history, or set TYPESAFE_API_KEY in the host environment; azdaja jev status checks the local setup.

What does Azdaja keep and compute locally?

It keeps full inputs in a persistent Monty/Python evaluator rather than every model prompt, and uses code for parsing, filtering, joins and exact reduction, sending only selected semantic work through llm and llm_batch.

CommunityJev Agents & MCP

Azdaja

Recursive language-model layer holding full inputs in a local evaluator and exposing a small working surface, with optional typed semantic judgments alongside generative calls.

See highlights

Open github.com

Stars: 14
Language: Python
License: MIT
Published by: Community
Author: kubet
Last verified: Oct 2026

kubet/azdaja

User repository on GitHub

View on GitHub

Minimal harness-agnostic recursive language model layer — one binary, Python + llm()

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Azdaja keeps complete source material in a local evaluator and gives language models a bounded working surface for code and semantic calls. It is a recursive language model layer, not a model or agent harness. Install · Use · Results · Jev · Live proof · Releases. Load once: keep full inputs in a persistent Monty/Python evaluator, not every model prompt. Compute locally: use code for parsing, filtering, joins and exact reduction.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Evaluation

Used inJev for Software & DevTools

Languages: Python

Highlights

Keeps full inputs resident locally instead of re-sending them in every prompt.
Local code handles parsing, filtering, joins, and exact reduction before models are asked.
Semantic subcalls return bounded results rather than opening new agent environments.
Optional typed judgments expose distributions, usage stats, budgeted batches, and review queues.

Watch out

Reported comparisons are historical diagnostics rather than matched reruns, and the layer is not an OS security boundary. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

deepreseach-learing

LangGraph deep-research agent that plans, searches, reads pages, aggregates evidence, checks gaps, and writes a Markdown report with sources, built as a runnable tool and a staged learning reference.