CoBRA
[🏆 CHI26 Best Paper] CoBRA: Reproducible control of LLM agent behavior via classic social science experiments
- 99
- Python
- MIT
- 7 mo ago
The maintainer has not written a description at the source.
Repo AISmithLab/PrivacyGallery, read from GitHub 1 min ago. AISmithLab has not joined Take My Repo.
AISmithLab/PrivacyGallery is a public repo on GitHub by AISmithLab. It declares no license on GitHub. Public does not mean open source: without a license, nobody is granted the right to reuse it. It is written mainly in TypeScript, has 1 stars and last changed on 2026-09-28. Nobody has reviewed it on Take My Repo yet, which says nothing about its quality. Read from GitHub on 2026-11-21.
In the maintainer's words. Shown from GitHub; Take My Repo does not keep a copy.
A global registry of 787+ data privacy enforcement cases across 7 jurisdictions, totaling $532M+ in fines. Browse cases, compare enforcement actions side-by-side, explore jurisdictions on an interactive map, and learn what privacy enforcement terms actually mean.
Live site: (https://jury.privacydev.org)
The project has two halves: an automated data pipeline that discovers, downloads, and extracts structured case data from regulator websites using Claude AI, and a React frontend that presents it as an interactive gallery.
Regulator Websites (FTC, ICO, PDPC, OAIC, EU DPAs, CA DOJ)
|
v
Collectors (jurisdiction-specific scrapers)
|
v
Document Discovery Log (PostgreSQL: discovered_documents)
|
v
Downloader (raw PDFs/HTML stored as BLOBs in PostgreSQL)
|
v
Text Extractor (pymupdf: PDF → clean text, locally, no API cost)
|
v
Phase 1: Extraction (Claude Haiku → structured fields from document)
|
v
Phase 2: Enrichment (Claude Haiku → company context, narratives, impact estimates)
|
v
URL Resolver (finds official case page URLs per jurisdiction)
|
v
Validation & Normalization → PostgreSQL (cases table)
|
v
Export → generatedCases.json → React Frontend (static build)
| Step | Command | API Cost | Description |
|---|---|---|---|
| Collect | --step collect |
Free | Scrapes regulator websites to discover new case document URLs |
| Download | --step download |
Free | Fetches PDFs/HTML, stores as BLOBs in PostgreSQL |
| Text | --step text |
Free | Converts PDFs to clean text via pymupdf (local) |
| Extract | --step extract |
~$0.01/doc | Claude extracts structured fields (company, fines, violations, legal bases) |
| Enrich | --step enrich |
~$0.01/doc | Claude adds company context, narratives, impact estimates |
| URL Fill | --step url_fill |
~$0.005/doc | Resolves official regulator case page URLs |
| Save | --step save |
Free | Writes validated data to PostgreSQL cases table |
| Export | --step export |
Free | Generates generatedCases.json for the frontend |
Phase 1 — Document Extraction: Reads the document text and extracts fields directly stated in the document: case name, jurisdiction, year, company, sector, violation type, legal bases violated, enforcement outcomes, penalty amounts, individuals affected, legal findings, and severity.
Phase 2 — AI Enrichment: Uses Phase 1 data plus outside knowledge to fill contextual fields: whatTheyDid (195-char narrative), whyTheyWereWrong (195-char harm description), company description, company valuation, founding year, current company status, impacted individuals (estimated when not in document), and claims vs reality.
Each phase is independently re-runnable. You can re-enrich all cases without re-extracting, or re-extract specific documents without touching enrichment.
(jurisdiction_id, document_url) — same URL won't be re-downloadedEach case gets a deterministic severity rating (1-5) based on:
| Page | Route | Description |
|---|---|---|
| Cases | / |
Searchable, filterable grid of all cases with jurisdiction, sector, violation type, and sort controls |
| Case Detail | /case/:id |
Full case breakdown — what they did, why they were wrong, claims vs reality, legal findings, outcome, attached PDFs |
| Compare | /compare |
Matrix view (patterns across jurisdictions/violations/sectors) and side-by-side comparison of up to 3 individual cases |
| Explore | /explore |
Interactive world map highlighting 7 jurisdictions; click a region to see its enforcement framework, key laws, and dataset statistics |
| Leaderboard | /leaderboard |
Rankings — top companies by fines, most active jurisdictions, most common violations and sectors |
| Learn | /learn |
Educational glossary explaining enforcement outcomes, violation types, and key legal concepts with cross-references |
| About | /about |
Project information and attribution |
| Jurisdiction | Abbreviation | Region |
|---|---|---|
| Federal Trade Commission | US FTC | United States |
| California DOJ | CA DOJ | United States (California) |
| Information Commissioner's Office | UK ICO | United Kingdom |
| Personal Data Protection Commission | SG PDPC | Singapore |
| General Data Protection Regulation | EU GDPR | European Union |
| European Data Protection Board | EU EDPB | European Union |
| Office of the Australian Information Commissioner | AU OAIC | Australia |
PrivacyGallery/
├── src/
│ ├── pages/ # Page components
│ │ ├── Index.tsx # Home — case gallery with search/filter/sort
│ │ ├── CaseDetail.tsx # Individual case breakdown
│ │ ├── Compare.tsx # Matrix & side-by-side case comparison
│ │ ├── Explore.tsx # Interactive jurisdiction map
│ │ ├── Leaderboard.tsx # Rankings and statistics
│ │ ├── Learn.tsx # Educational glossary
│ │ ├── About.tsx # Project information
│ │ └── NotFound.tsx # 404 page
│ ├── components/
│ │ ├── TopNav.tsx # Yellow navigation bar
│ │ ├── CaseCard.tsx # Case summary card with red fine stamp
│ │ ├── ControlBar.tsx # Filter/sort controls
│ │ ├── SearchBar.tsx # Search input
│ │ ├── JurisdictionMap.tsx # SVG world map (react-simple-maps)
│ │ ├── JurisdictionDetail.tsx# Jurisdiction stats and info panel
│ │ ├── JurisdictionLogos.tsx # Jurisdiction logo/icon display
│ │ ├── ScrollToTop.tsx # Scroll reset on route change
│ │ └── ui/ # 50+ shadcn/ui components
│ ├── data/
│ │ ├── cases.ts # Type definitions, utilities, case loading
│ │ ├── generatedCases.json # 787+ cases exported from PostgreSQL
│ │ ├── jurisdictionInfo.ts # Jurisdiction metadata (laws, authorities)
│ │ └── glossary.ts # Learn page glossary content
│ ├── hooks/ # Custom React hooks
│ ├── lib/utils.ts # Tailwind class merge utility
│ ├── main.tsx # App entry point
│ └── index.css # Global styles and CSS variables
├── pipeline/ # Automated data pipeline
│ ├── config.py # DB URL, API key, model settings
│ ├── db.py # Connection pool, schema init
│ ├── models.py # Dataclasses for all entities
│ ├── runner.py # CLI entry point (--step, --jurisdiction, --limit)
│ ├── migrate.py # One-time migration + jurisdiction seeding
│ ├── collectors/ # Jurisdiction-specific scrapers
│ │ ├── base.py # BaseCollector ABC
│ │ ├── registry.py # Collector factory
│ │ ├── ftc.py # US FTC
│ │ ├── ico.py # UK ICO
│ │ ├── pdpc.py # Singapore PDPC
│ │ ├── oaic.py # Australia OAIC
│ │ ├── gdpr.py # EU GDPR (per-country DPAs)
│ │ ├── california_doj.py # California DOJ
│ │ └── edpb.py # EU EDPB
│ ├── processing/ # Download, dedup, text extraction
│ │ ├── downloader.py # Fetch documents, store as BLOBs
│ │ ├── dedup.py # SHA256 hash deduplication
│ │ ├── text_extractor.py # PDF→text (pymupdf), HTML→text (bs4)
│ │ └── url_resolver.py # Find official case page URLs
│ ├── extraction/ # AI extraction (two-phase)
│ │ ├── prompts.py # Phase 1 + Phase 2 prompt templates
│ │ ├── extractor.py # Phase 1: structured fields from text
│ │ ├── enricher.py # Phase 2: company context + narratives
│ │ └── parser.py # JSON response parsing
│ ├── validation/ # Field validation and normalization
│ │ ├── validator.py # Missing values, impossible fines, dates
│ │ └── normalizer.py # Jurisdiction/sector/violation mapping
│ └── export/ # Output generation
│ ├── to_postgres.py # Write to cases table
│ └── to_frontend.py # cases → generatedCases.json
├── files/ # Legacy pipeline scripts (archived)
│ ├── agent.py # Original PDF → Claude → PostgreSQL
│ ├── export_to_frontend.py # Original PostgreSQL → JSON export
│ └── ...
├── .github/workflows/
│ └── deploy.yml # GitHub Actions → GitHub Pages
├── package.json
├── vite.config.ts
├── tailwind.config.ts
└── tsconfig.json
Requires Node.js 20+ — install with nvm
git clone https://github.com/AISmithLab/PrivacyGallery.git
cd PrivacyGallery
npm install
npm run devThe site runs at http://localhost:8080. Case data is already included in generatedCases.json.
Only needed if you want to collect new cases or rebuild the dataset.
Prerequisites: Python 3.12+, PostgreSQL running locally.
# Install dependencies pip install -r pipeline/requirements.txt # Configure environment cp files/.env.example files/.env # Edit files/.env with your ANTHROPIC_API_KEY and DATABASE_URL # Initialize database tables and seed jurisdictions python3 -m pipeline.migrate
Run the full pipeline:
python3 -m pipeline.runner
Run individual steps:
python3 -m pipeline.runner --step collect # discover new documents python3 -m pipeline.runner --step download # fetch PDFs/HTML python3 -m pipeline.runner --step text # extract text locally python3 -m pipeline.runner --step extract # Phase 1: Claude extraction python3 -m pipeline.runner --step enrich # Phase 2: Claude enrichment python3 -m pipeline.runner --step url_fill # resolve case page URLs python3 -m pipeline.runner --step save # write to PostgreSQL python3 -m pipeline.runner --step export # generate frontend JSON
Filter by jurisdiction or limit:
python3 -m pipeline.runner --jurisdiction us_ftc # one jurisdiction only python3 -m pipeline.runner --step extract --limit 5 # process max 5 documents python3 -m pipeline.runner --dry-run # preview without changes
| Variable | Description |
|---|---|
ANTHROPIC_API_KEY |
Claude API key for document extraction and enrichment |
DATABASE_URL |
PostgreSQL connection string |
EXTRACTION_MODEL |
Model for Phase 1 extraction (default: claude-haiku-4-5) |
ENRICHMENT_MODEL |
Model for Phase 2 enrichment (default: claude-haiku-4-5) |
The pipeline uses 6 PostgreSQL tables:
| Table | Purpose |
|---|---|
jurisdictions |
Registry of data sources (regulator URLs, crawl frequency, access method) |
discovered_documents |
Every document found by collectors (URL, status, jurisdiction) |
raw_documents |
Downloaded files stored as BLOBs with extracted text |
extraction_runs |
Audit log for each AI call (phase, model, tokens, response) |
pipeline_log |
Monitoring and failure logs |
cases |
Final structured case data (feeds the frontend export) |
The site auto-deploys to GitHub Pages on every push to main via GitHub Actions. The workflow:
npm run buildindex.html to 404.html for SPA routingCustom domain configured via public/CNAME → jury.privacydev.org
interface EnforcementCase { id: string; company: string; sector: Sector; // Technology, Healthcare, Finance, etc. jurisdiction: Jurisdiction; // US FTC, UK ICO, EU GDPR, etc. year: number; fineAmount: number; fineDisplay: string; violations: ViolationType[]; // Misrepresentation, Failure to disclose, etc. severityForIndividuals: number; // 1-5 calculated score impactedIndividuals: string; whatTheyDid: string; // Plain-language summary whyTheyWereWrong: string; // Why it matters claimsVsReality: ClaimVsReality[]; // What they said vs what they did regulatoryFindings: RegulatoryFinding[]; outcome: string; outcomeSummary: string; // Complaint Filed, Consent Order, etc. attachedPDFs: AttachedPDF[]; // Links to source documents // ... 40+ additional fields }
This project is maintained by AISmithLab.
AISmithLab/PrivacyGallery is a public repo on GitHub by AISmithLab. Its maker has not written a description.
It declares no license on GitHub. Public does not mean open source: without a license, nobody is granted the right to reuse it.
Its maker has not tagged a system, so it is unknown here. Its README says what it needs.
Open "Take it" on its page: Take My Repo writes an install plan for your machine and hands it to your AI agent (Claude Code, Codex or Cursor), which asks you before each step. Take My Repo has not tested it on any machine.
Its last change on GitHub was on 2026-09-28. Take My Repo does not judge from that whether it is maintained.
Nobody has reviewed it on Take My Repo yet, which says nothing about its quality.

[🏆 CHI26 Best Paper] CoBRA: Reproducible control of LLM agent behavior via classic social science experiments

HumanStudy-Bench: Towards AI Agent Design for Participant Simulation

Panopticon is an educational board game that teaches data science students privacy design skills through interactive gameplay.

PersonalDataHub: Gate personal data access for AI agents
This page is also available as structured data, with the same sources and dates.
/api/v1/resources/gh/AISmithLab/PrivacyGalleryTake My Repo has not tested this on your machine. Requirements come from the project's own documentation.