1 / 14
1
ComplianceGPT Lab · AI Innovation & Diffusion REU 2026

Welcome to the Lab

Six weeks. Real research. Real AI.

Dates
June 22 – July 31, 2026
Lab
NCS Room 242
Mentor
Priscilla Kyei Danso
Final Presentation
July 31
2
The Research Problem

One Big Question

"Can we build AI systems that make reliable, explainable decisions about whether an action complies with a legal regulation?"

The Challenge

Legal compliance is expensive, error-prone, and requires expert knowledge. Hospitals spend billions on compliance staff.

The Opportunity

LLMs can read legal text and extract structured information. But they hallucinate. We need a way to make them reliable.

Our Approach

Combine LLM extraction with formal logic. The LLM reads the scenario. A rule engine makes the verdict. Separate the concerns.

3
The Core Research Insight

A Question Can Touch Many Laws at Once

The same patient question can trigger obligations under multiple regulations simultaneously.

Patient Question HIPAA GDPR CCPA COPPA SOX GLBA
"Can I see my medical records?"
"Can I delete my records?"
"Who can share my health data?"
"Can my bank share my health info?"
"Do I have the right to be forgotten?"

Each ✓ means that regulation has something to say about this question. Answering correctly requires knowing which laws apply, and what each one requires.

4
Foundational Knowledge

The Six Regulations You Need to Know

HIPAA Health Insurance Portability & Accountability Act · US · 1996

Protects Protected Health Information (PHI) — health data that can identify a patient. Governs hospitals, clinics, pharmacies, insurers. Permits disclosures for treatment, payment, operations, and specific legal exceptions.

GDPR General Data Protection Regulation · EU · 2018

Covers all personal data of EU residents — not just health. Key rights: right of access, right to erasure, right to portability. Requires a lawful basis for processing. Fines up to 4% of global revenue.

CCPA California Consumer Privacy Act · California · 2020

California residents' rights: know what data is collected, delete it, opt out of sale. Similar in spirit to GDPR but narrower scope and US-specific. Amended by CPRA (2023).

COPPA Children's Online Privacy Protection Act · US · 1998

Protects children under 13. Requires parental consent before collecting personal data from minors. Overlaps with GDPR and HIPAA when the patient is a child.

SOX Sarbanes-Oxley Act · US · 2002

Corporate financial integrity law. Public companies must keep accurate financial records and can't destroy them (Section 802). Relevant when health data overlaps with billing and financial audit trails.

GLBA Gramm-Leach-Bliley Act · US · 1999

Financial institutions must protect customers' non-public personal information (NPI). Analogous to HIPAA but for banks and insurers. Overlap with HIPAA when a health insurer is also a financial institution.

5
The System You Will Work With

ComplianceGPT — How It Works

A neuro-symbolic pipeline: neural language model extracts structure, a formal logic engine checks compliance.

Input
Scenario
Real court case text
Your Research ★
LLM₁ Extraction
Gemma / Llama / Qwen
Extracts structured JSON
Black Box
Datalog Engine
Formal logic rules
You call via API
Output
Verdict + Proof
PERMITTED / DENIED
+ statutory citation

What you control

  • Which LLM does the extraction
  • The prompt that guides extraction
  • How you analyze and fix errors
  • Whether a different prompting strategy works better

The black box (not your concern)

  • Internal Datalog rules for HIPAA §164.502–.530
  • How the proof tree is generated
  • The Soufflé compiler internals
  • The formalization of each regulatory section
6
Key Technical Concept

What the LLM Actually Extracts

The LLM must extract a structured JSON "scenario" from the narrative. The most critical fields are called oracle predicates.

The Narrative (Input)

"A hospital responded to a police officer's request for a patient's billing records. The officer presented a subpoena signed by a judge requiring disclosure."

What the LLM must extract

{
  "sender": "hospital",
  "receiver": "police_officer",
  "phi_type": "billing_records",
  "purpose": "law_enforcement",
  "has_court_order": true   ← oracle predicate
}

Why oracle predicates matter

Oracle predicates are boolean flags for HIPAA enabling conditions. The formal engine cannot grant PERMITTED without them.

  • has_court_order — is there a real court order?
  • obtained_authorization — did the patient consent?
  • is_required_by_law — is disclosure legally mandated?

The adversarial risk

If an LLM halluccinates has_court_order: true when no court order exists, the system issues a false PERMITTED verdict — a potential HIPAA violation.

7
Current Results

Where We Are

94.2%
Accuracy on 137 real HIPAA cases
(Gemma3:4B + Claude Sonnet)
100%
Oracle ceiling — the formalization is provably correct when given gold facts
5
Models evaluated: Gemma, Claude, Qwen, GPT-OSS, Llama

The surprising finding

A 4B parameter local model (Gemma3) ties a ~200B frontier model (Claude Sonnet) at 94.2% accuracy. Model size doesn't win — extraction quality does. This is what your research will explore.

The gap

5.8 percentage points separate our best model from the oracle ceiling. All remaining errors are extraction failures — the LLM reading the scenario wrong — not formalization errors. This is fixable.

8
Research Projects

Six Directions — One For You

You will propose your project in Week 3. Until then, explore all of them.

P1 Benchmark Comparison

Gemma vs. Llama vs. Qwen — which model makes which errors? Build a systematic error taxonomy.

P2 Adversarial Robustness

Can you craft scenarios that fool the system into a false PERMITTED verdict? What makes an attack succeed?

P3 Prompt Engineering

Zero-shot vs. few-shot vs. chain-of-thought vs. schema-constrained. Which prompting strategy wins on compliance extraction?

P4 Explanation Quality

Are the natural-language explanations ComplianceGPT generates actually correct and useful? Rate 50 explanations on a rubric.

P5 Cross-Regulation Transfer

Does a prompt that works for HIPAA also work for GDPR? What happens when the same question triggers both laws?

P6 Error Taxonomy

Build the first formal taxonomy of LLM extraction errors in legal compliance. Which error types can targeted prompts fix?

9
This Week

Week 1 Agenda — June 22–27

Monday June 22 — Today (Kickoff)

Welcome & Orientation

  • Introductions round
  • Live pipeline demo: run 3 real court cases
  • Lab logistics: Slack, GitHub, folder structure
  • The black box rule
Task: setup checklist + run 5 scenarios
Wednesday June 25 ★ Main Session (NCS 242)

Python + AI Foundations

  • Live coding: load CSV, filter, compute accuracy
  • Pandas basics — everything you'll need this summer
  • Git: commit, push, read a diff
  • AI stack: ML → Deep Learning → LLMs → hallucination demo
  • Discussion: 3 wrong predictions — why did they fail?
Task: Python exercises 1–4
Thursday–Friday June 26–27 — Remote

Independent Work

  • Finish Python exercises + pipeline run
  • Read HIPAA Privacy Rule summary (hhs.gov)
  • Watch 3Blue1Brown Neural Networks videos 1–2
  • Write 1-page reflection (due Friday 5pm)
Deliverable due Friday: pipeline run + reflection
Post Slack update by noon Thu & Fri
10
How This Lab Works

The Rules

What I expect from you

  • Show up prepared. Read the day's material before the session. Standups are not the time to catch up.
  • Communicate early. If you're stuck Monday, tell me Monday — not Friday. One message saves five days.
  • Finish things. A 3-week project done well beats a 6-week project abandoned at 80%. Scope down, ship it.
  • Be curious. Ask "why" before you ask "how to fix it." Understanding the failure mode is the research.
  • Good lab citizenship. Give your peers honest feedback. Take notes. Share what you learned.

What you can expect from me

  • Slack responses within a few hours on weekdays.
  • Written feedback on every deliverable within 48 hours.
  • Help scoping your project so you actually finish with something.
  • A strong letter of recommendation if you put in the work.
  • Co-authorship on any paper that comes from this work.
  • Introductions to other researchers and faculty.
11
Resources · Videos

Watch These First

12
Resources · Papers

Required and Recommended Papers

13
Resources · Regulations

Official Sources and Regulation Texts

14
Today — Monday June 22

Let's Get Started

Right Now — Introduction Round

Name, school, major, and: one regulation you've heard of before today.

It's okay to say "HIPAA because of a doctor's office" or "GDPR because of cookie popups." That's exactly the right starting point.

Before You Leave Today

  • Complete the setup checklist — Python, VS Code, Git
  • Run the pipeline on 5 real court case rows
  • Write down 1 thing that surprised you
  • Post your Slack intro in #introductions

The Research Question That Will Guide Your Summer

"Can you figure out which regulation (or regulations) applies to a given question — and then determine what the answer actually is?"

Everything you learn this summer — prompting, RAG, error analysis, benchmarking — feeds into answering this.

Website: priscilladanso.github.io/reu2026  ·  All slides, readings, and assignments are there.