Home

poordjaevin

The poor man's Jev — a local-first "System One" decision layer: typed questions about state (Choice / Score / Noul) answered with calibrated confidence, no API key. Devin ACP backend in this fork.

Why it exists

Gates and classifiers need honest confidence, not vibes. Djævin does NLI-based entailment + temperature scaling + conformal abstention — and ships the eval set so you can verify the numbers yourself.

Install

pip install -e ".[local,plots,mcp,dev]"
poordjaevin --help

Requires Python 3.10+ unless noted. Linux, Personal Windows and Corporate Windows are supported; see the repo's platform READMEs for differences.

Core commands

Command	What it does
poordjaevin eval	accuracy + ECE on the shipped eval set
poordjaevin calibrate --plots	before/after calibration + reliability diagrams
poordjaevin ask	ask a typed question
poordjaevin serve	MCP server (stdio)

What it touches

Reads: the eval set / your state input. Writes: calibration artifacts. Backends: keyless local NLI, or Devin ACP (free model) in this fork.

Limitations

• Honesty is the product: every headline number is reproducible via eval/calibrate.
• Ask concrete hypotheses — the model handles concrete claims far better than abstract ones.

---

Unofficial community project. Not affiliated with Cognition AI. Part of the devin-* ecosystem.