CaseGuard AI

Controlled Autonomy for Intelligent Casework

CaseGuard AI automates the permitted caseworker workflow while enforcing authority boundaries through deterministic, policy-driven guardrails.

Differentiator: The agent's authority is controlled by an external policy (authority-policy.md) and enforced by a deterministic guardrail layer, so the agent cannot silently perform an unauthorized action.

Project Structure

• caseguard/: Core application package.
  • data_loader.py: Loads and parses referral-queue.json.
  • history_client.py: Interacts with the local Resident History API.
  • policy_service.py: Parses rules from authority-policy.md.
  • guardrail_engine.py: Deterministically enforces policy rules on referrals.
  • llm_provider.py: Lightweight mock/real LLM abstraction for generating triage notes.
  • workflow_controller.py: Orchestrates the safe execution loop.
  • audit_logger.py: Provides structured execution tracing.
• tests/: Automated test suite for safety invariants.

Requirements

• Python 3 standard library (no external packages required)

Setup & Run Instructions

(To simulate a pure clean-machine evaluation)

1. (Optional) Create and activate a clean virtual environment:

   python -m venv venv
   # Windows:
   .\venv\Scripts\activate
   # Mac/Linux:
   source venv/bin/activate

2. Start the resident history API in a background terminal:

   python -m services.history_service --port 8083

   (Note: The main application is fault-tolerant and will gracefully handle missing history if you skip this step).
3. Run the main processing workflow:

   python -m caseguard.main

4. Run the deterministic safety tests:

   python -m unittest discover tests

Demo Instructions

Running the main workflow will process the 12 official referrals from referral-queue.json. You will observe:

1. Normal Referrals: Fetch history, draft a triage note via the LLM provider, and log completion.
2. Unauthorized Referrals: The deterministic GuardrailEngine will detect requests for suspension, payment updates, or reinstatements and place them in [PENDING APPROVAL] without executing history queries or LLM generation.
3. The queue continues processing unaffected.
4. Supervisor Review: At the end of the run, escalations are reviewed. Approved actions acquire an ApprovalRecord and resume safely. Rejected actions terminate.