# AGENTS.md

## Architecture

CrediMesh is a Python package that models privacy-preserving mortgage underwriting as a federated workflow.

Key modules:

- `models.py`: typed request, signal, plan, budget, and audit records.
- `identity.py`: Ed25519-based signing and DID-style identifiers.
- `contracts.py`: graph-of-contracts registry and adapter conformance checks.
- `adapters.py`: income verification and property appraisal signal adapters.
- `solver.py`: deterministic ADMM-lite underwriting solver.
- `ledger.py`: SQLite-backed governance and audit log persistence.
- `orchestrator.py`: end-to-end orchestration over adapters, solver, and audit trail.

## Tech Stack

- Python 3.11+
- `pydantic` for strict model validation
- `cryptography` for signing and verification
- `networkx` for contract graph bookkeeping
- SQLite for governance persistence

## Testing

- `bash test.sh`
- `pytest`
- `python3 -m build`

## Rules

- Keep outputs privacy-minimized; never store raw borrower payloads in audit records.
- Preserve deterministic replay behavior for the same inputs.
- Prefer small, explicit changes over broad rewrites.
- Update tests when changing schemas, solver behavior, or contract rules.
- Do not add new dependencies unless they materially improve the orchestration layer.