# Model Cards — AI Components

This document describes all AI-related components used in **Federal FOIA Intelligence Search**.

---

## Model Name

**General-Purpose Large Language Model (LLM)**  
(Provided via Hugging Face / OpenAI runtime)

> ⚠️ The application does **not** ship or host its own trained model.

---

## Intended Use

**Allowed Uses**
- Summarization of publicly released FOIA records
- Contextual explanation of document metadata
- Research assistance for journalists, academics, and legal professionals

**Explicitly Disallowed Uses**
- Legal advice
- Evidence generation
- Intelligence analysis
- Surveillance, profiling, or targeting
- Automated decision-making

---

## Training Data Summary

- Model training data is **external** to this application
- This application does **not train, fine-tune, or adapt** models
- User inputs are **not retained** for training

---

## Input Data Constraints

- Public FOIA metadata
- Optional, user-approved PDF text extraction
- User-supplied questions only

**No ingestion of:**
- Private data
- Classified information
- Authentication-protected materials

---

## Output Constraints

- Outputs are explicitly labeled as AI-generated
- Outputs are citation-anchored
- Outputs include an integrity hash
- Outputs are not persisted

---

## Risk Mitigation

| Risk | Mitigation |
|----|----|
| Hallucination | Citation anchoring + disclosure |
| Over-reliance | Warnings + opt-in |
| Data leakage | No persistence |
| Misuse | Feature gating |

---

## Ethical Considerations

This AI component is intentionally:
- Non-autonomous
- Non-persistent
- User-controlled
- Auditable

---

## Limitations

- May misinterpret scanned PDFs
- Does not validate document authenticity
- Cannot access non-public records

---

## Contact

For AI safety inquiries:  
**Project Maintainer: Ezra Godschild**