[Open Safeguard Hackathon December 2025 Submission] SANDRA: Safety Appeals & Network Dispute Review Agent

[bentonwong]

SANDRA: Safety Appeals & Network Dispute Review Agent

Team Members

• Benton Wong
• Anish Kumar

---

Project Description and Problem Statement

The Problem

User-generated content (UGC) platforms face significant challenges in handling content moderation appeals:

1. Appeals are slow and opaque - Users often wait days or weeks for a response, with little visibility into the review process
2. Moderators are overloaded - Trust & Safety teams spend significant time on low-value appeals that could be resolved automatically
3. Gray-area cases are under-supported - Nuanced cases requiring context and explanation lack proper support infrastructure
4. Consistency is hard to maintain - Similar cases may receive different outcomes due to human reviewer variability
5. Fairness is hard to demonstrate - Platforms struggle to show transparent, consistent decision-making processes

Our Solution: SANDRA

SANDRA (Safety Appeals & Network Dispute Review Agent) is an open-source-powered appeals copilot that provides a fair, fast, and transparent second look at content moderation decisions. Built on open-weight safeguard models (20B and 120B variants), SANDRA combines:

• Multi-agent reasoning pipeline (Analysis Agent + Evaluator Agent) for robust decision-making
• Precedent-aware consistency using a lightweight case database
• Three-tier risk assessment (Green/Yellow/Red lanes) for appropriate handling
• Creator-friendly interface that explains decisions in plain language
• Structured escalation for high-risk cases requiring human review

---

Research Conducted & Prototype Description

Architecture

SANDRA implements a multi-agent system with two specialized agents:

1. Analysis Agent:

   • Generates flag explanations (why content was removed)
   • Performs risk assessment and triage (Green/Yellow/Red)
   • Produces creator-facing reasoning
   • Consults precedent database for consistency
   • Never mentions "precedent" to creators - only speaks in policy terms

2. Evaluator Agent:

   • Reviews Analysis Agent outputs for quality and safety
   • Checks policy consistency, factual accuracy, and risk assessment
   • Can approve, request revision, or force escalation to human review
   • Acts as a safety check before decisions are shown to creators

Risk Lanes

Every appeal is categorized into one of three risk lanes:

• 🟢 Green Lane: Low-risk, straightforward cases → Auto-resolve if creator's explanation is sufficient
• 🟡 Yellow Lane: Medium-risk, nuanced cases → Allow back-and-forth (initial explanation + one rebuttal) before final decision
• 🔴 Red Lane: High-risk, sensitive cases → Never auto-reinstate; always escalate to human reviewers with structured case summary

Technical Implementation

Tech Stack:

• Frontend: Next.js 14, React, TypeScript, Tailwind CSS
• Backend: Next.js API routes
• AI Models: Open-source safeguard models (gpt-oss-safeguard-20b)
• Deployment: Vercel-ready with environment variable configuration
• Model Providers: Supports HuggingFace Inference API, Groq, and other OpenAI-compatible endpoints

Key Features:

• Real-time appeal processing with live model API integration
• Full UI/UX with progress tracking, decision displays, and archive/withdraw functionality
• Precedent database for consistency checking
• Model variant switching (20B vs GPT-5.1) for comparison
• Structured decision IDs and case tracking
• Creator-facing explanations that never reference "precedent" or "similar cases"

Working Demo

The prototype is a fully functional single-page application that demonstrates:

1. Appeal List View: Browse multiple appeals with status indicators
2. Flag Explanation: SANDRA automatically explains why content was flagged
3. Creator Explanation Form: Users can provide context with structured flags (recovery/awareness, educational, satire, etc.)
4. First Assessment: SANDRA evaluates the explanation and provides reasoning
5. Second Explanation (Yellow lane only): Users get one more chance to add context
6. Final Decision: SANDRA makes a final determination or escalates to human review
7. Progress Timeline: Visual tracking of appeal status through the pipeline

Demo Cases Included:

• Recovery stories (should be reinstated)
• Dark humor/casual references (needs context)
• Vague distress (should escalate to human)
• Educational content
• Harassment cases
• And more edge cases for comprehensive testing

---

Results from the Experiment

Prototype Capabilities Demonstrated

1. Successful Multi-Agent Pipeline

   • Analysis Agent successfully generates flag explanations, risk assessments, and creator-facing reasoning
   • Evaluator Agent provides quality checks and safety oversight
   • Both agents work together to produce consistent, policy-aligned decisions

2. Precedent-Aware Decision Making

   • System successfully consults precedent database for consistency
   • Creator-facing explanations appropriately avoid mentioning "precedent" while still being consistent with past decisions
   • Demonstrates ability to "follow" or "distinguish" from precedents internally while speaking in policy terms externally

3. Risk-Based Triage

   • System correctly categorizes cases into Green/Yellow/Red lanes
   • Low-risk recovery stories are auto-resolved appropriately
   • High-risk cases with concerning ideation are properly escalated
   • Medium-risk cases allow for additional context gathering

4. Model Variant Comparison

   • Successfully supports both 20B and 120B model variants
   • UI allows switching between models to compare behavior
   • Demonstrates flexibility in model provider selection (HuggingFace, Together AI, etc.)

5. User Experience

   • Full end-to-end appeal flow implemented
   • Clear, empathetic creator-facing explanations
   • Progress tracking and status visibility
   • Archive/withdraw functionality for appeal management

Technical Achievements

• Robust Error Handling: Implements retry logic with exponential backoff for API rate limits
• Multi-Provider Support: Works with HuggingFace, Together AI, Groq, and other providers
• Type Safety: Full TypeScript implementation with comprehensive type definitions
• Production-Ready: Includes CI/CD setup, environment variable management, and deployment configuration

Limitations & Future Work

Current Limitations:

• Demo uses seed data (not connected to live platform APIs)
• Text-only content (no image/video analysis)
• Single policy focus (Self-Harm & Suicide, with some Harassment examples)
• No authentication or multi-tenant support

Future Enhancements:

• Integration with live platform moderation systems
• Support for multimedia content analysis
• Additional policy sections (hate speech, misinformation, etc.)
• Analytics dashboard for Trust & Safety teams
• A/B testing framework for model comparison
• Fine-tuning on platform-specific precedent data

Key Insights

1. Open-source models are viable for content moderation appeals when combined with structured reasoning pipelines
2. Multi-agent systems provide better safety and consistency than single-agent approaches
3. Precedent awareness improves decision quality without compromising creator-facing transparency
4. Risk-based triage enables appropriate automation while maintaining safety for high-risk cases
5. Creator-facing language matters - avoiding technical terms like "precedent" improves user experience

---

Repository & Demo

GitHub Repository: [Link to your repository]

Live Demo: https://sandra-app.vercel.app/

You can interact with the full SANDRA prototype at the link above. The demo includes 14 test appeals covering various scenarios including recovery stories, dark humor, educational content, and high-risk cases requiring escalation.

Presentation: Google Slides Presentation

Key Files:

• app/page.tsx - Main UI component
• lib/agents/analysisAgent.ts - Analysis Agent implementation
• lib/agents/evaluatorAgent.ts - Evaluator Agent implementation
• data/appeals.ts - Seed appeal cases
• data/precedents.ts - Precedent database
• docs/ - Comprehensive documentation

---

Conclusion

SANDRA demonstrates that open-source AI models, when combined with structured reasoning pipelines and precedent-aware systems, can provide effective, transparent, and fair content moderation appeals processing. The prototype successfully handles a range of cases from straightforward recoveries to complex gray-area situations, appropriately escalating high-risk cases while providing fast resolution for low-risk appeals.

The system shows promise for reducing moderator workload, improving consistency, and providing better user experience in content moderation appeals - all while maintaining safety through multi-agent oversight and risk-based triage.

---

Additional Resources

• Product Vision: See docs/SANDRA_PRODUCT_VISION.md
• Technical Architecture: See docs/SANDRA_CURSOR_ARCH_PROMPT.md
• Example Cases: See docs/SANDRA_EXAMPLES.md
• Prompt Specifications: See docs/PROMPTS_SANDRA.md
• Self-Harm Policy: See docs/POLICY_SELF_HARM.md

[cassidyjames]

Thank you for sharing! This is a super interesting project.