Jevometry

An information-geometric analysis toolkit for Jev and Jev-like System One agent systems. It captures explicit decision probabilities and measures local sensitivity, Fisher information, and the information a final decision retains, without orchestrating agents or changing decisions.

Distribution comparisons (entropy, KL, Jensen–Shannon, Hellinger and Fisher–Rao with aligned supports), sensitive inputs (derivatives, node Fisher matrices, rank and stability diagnostics), system information (joint Fisher and loss under a fixed final-action map), and inference checks (Cramér–Rao bound, optional synthetic MLE).

A fully offline example where the council examines one candidate repair: monthly_income_usd changes from " 1,200 " to 1200. At evidence=0.7 and ambiguity=0.3 the model reports System Fisher trace 6.842, information loss 3.034, and final actions auto_fix 0.236 · human_review 0.294 · reject 0.470.

With uv installed: clone the repository, enter it, run uv sync --locked, then uv run python examples/three_agent_cleaning/run.py. Open examples/three_agent_cleaning/output/report.html afterward; captures are generated locally and not tracked in Git.

No. The README says the figures describe the declared probability family at one point, not measured cleaning accuracy or empirical calibration — and that a reported Jev probability is not an observed categorical draw.

The core is offline and the included examples run without one, though the ticket-system sensitivity example supports an optional live TypeSafe capture. Version 0.2.0 alpha covers Python 3.11 to 3.13 under the MIT license.

It refuses quantities the available data and declarations cannot support — separate marginal distributions do not determine a system joint — and it records assumptions, provenance and status with every result. Read the statistical contracts before interpreting Fisher information.