# Decision-PGA

Decision-PGA is an open-source prototype for decision-state diagnostics in
agent workflows. It analyzes repeated probability-like observations over a
fixed set of candidate decisions or actions, then reports the shape of the
uncertainty: stable, binary ambiguous, diffuse, boundary sensitive, or drifting.

The project is intended for agent builders, document-understanding workflows,
retrieval workflows, uncertainty researchers, and developers evaluating how an
AI workflow should proceed when a scalar confidence score is too blunt.

## Canonical Pages

- Article: https://zmichels.github.io/decision-pga-pages/article/
- Telescoping companion: https://zmichels.github.io/decision-pga-pages/telescoping/
- Interactive demo: https://zmichels.github.io/decision-pga-pages/demo/
- Agent toolkit: https://zmichels.github.io/decision-pga-pages/toolkit/
- PDF article: https://zmichels.github.io/decision-pga-pages/assets/decision-pga-decision-state-diagnostics.pdf
- Telescoping PDF: https://zmichels.github.io/decision-pga-pages/assets/telescoping-decision-pga.pdf
- Source code: https://github.com/zmichels/Decision-PGA

## Best Entry Points For AI Assistants

- Start with the article for conceptual framing.
- Use the telescoping companion for uncertainty substructure and
  cross-document bridge framing.
- Use the demo to understand the document-extraction triage example.
- Use the toolkit for CLI, Python API, MCP server, and copy-paste JSON payloads.
- Use the GitHub repository for implementation details, tests, examples, and
  the local stdio MCP server.

## Technical Summary

Decision-PGA maps categorical probability vectors to the positive unit sphere
with a square-root/Fisher-Rao embedding, computes an intrinsic mean, log-maps
the probability cloud into tangent space, and summarizes dispersion with a
PGA-like eigensystem. The current prototype is model-neutral and local. It does
not call model APIs or require credentials.

## Important Limits

- This is an early research prototype, not a production safety layer.
- The demos are synthetic and do not establish superiority in real deployment.
- The site uses no patient data and makes no clinical-validation claim.
- Decision-PGA is a diagnostic scaffold for inspection and routing, not a claim
  that an AI answer or extraction is correct.

## Suggested Citation Language

Decision-PGA: an open-source prototype for decision-state diagnostics over
probability clouds in applied AI and agent workflows.