Jevometry
Offline Python toolkit that measures sensitivity, Fisher information and information loss in the probabilities of Jev-style agent systems.
Overview
Jevometry analyses how an agent system's decision probabilities change with its inputs. It captures explicit Jev (or Jevlike) decision probabilities and computes information-geometric quantities: divergences between distributions (entropy, KL, Jensen-Shannon, Hellinger, Fisher-Rao), local sensitivity and per-node Fisher information, joint system Fisher information and the information lost by a final action map, and Cramer-Rao bounds when a full likelihood is declared.
It analyses systems without orchestrating agents or changing their decisions, records assumptions and provenance with each result, and refuses quantities the declarations cannot support, for example deriving a joint from separate marginals. The example is a fully offline three-agent data-cleaning council (applicability, repair and risk agents) whose final action is auto_fix, human_review or reject. The README stresses that these figures describe the declared probability model, not measured accuracy or calibration. It is an alpha research tool for people studying multi-agent decision systems.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What should be done with this candidate data repair? (example final action) | auto_fix · human_review · reject |
| Other | How sensitive is each agent's decision distribution to its inputs (Fisher information)? | — |
How it works
- Declare the agents' parameterised probability model and joint structure
- Capture decision probabilities (analytic or finite-difference)
- Compute divergences, Fisher matrices, rank and system information loss
- Generate a local HTML report with assumptions and provenance
Strengths
- Offline core with explicit statistical contracts
- Refuses unsupported quantities instead of guessing
- Works on any System-One-style probabilities
Caveats
- Alpha (v0.2.0)
- Results describe declared models, not empirical accuracy
- Requires statistical background to interpret
Getting started
git clone https://github.com/Kunyanli230/Jevometry.git && cd Jevometry && uv sync --locked && uv run python examples/three_agent_cleaning/run.pyAt a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- library, Python 3.11-3.13
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
an Information-Geometric Analysis Toolkit for any System-one (Jev, Jevlike) agent systems
logicrw/awesome-jev-projects
Frequently asked questions
Does Jevometry make decisions?
No. It analyses the probability distributions of an existing decision system without changing them.
Do I need a Jev key?
The core and example run offline; capturing live Jev probabilities would need access to Jev.
Do its numbers measure accuracy?
No. They describe the declared probability family, not empirical accuracy or calibration.
Related projects
Nautilus Compass agent memory layer whose jev-trust middleware logs Jev decisions and measures calibration in your domain.
ServicebetaPythonDecision-forensics library that explains a Jev answer by ablating parts of the input and measuring how far the probability moves.
ProjectbetaPythonPython trust middleware for the Jev API that logs decisions, measures calibration from your recorded outcomes and signs the evidence.
ServicePythonToolkit for Jev decision contracts: compiled policies, egress controls, tracing, stability checks and certified decision SLOs.
ProjectbetaTypeScriptRecords the full distribution of every Jev answer by state, question and model, so thresholds can move over history with no new calls.
LibrarybetaPythonMonitoring proxy and client that records Jev answer probabilities and alerts on version changes, drift, edge crowding or accuracy drops.
ProxybetaPython