Misinformation Guardian Hackathon Build Log (February 12, 2026)

TL;DR: On Wednesday, February 12, 2026, we rapidly prototyped misinformation-defense capabilities in a hackathon sprint: started with a Custom GPT, hit limits on video analysis, switched to a Gemini Gem, tuned prompts for ~90 minutes, and shipped a visual demo direction that now maps to a production blueprint.

Why We Built This

Misinformation moves faster than manual fact-check workflows. We wanted a system that helps people:

1. Detect manipulation patterns quickly.
2. Understand why content is risky.
3. Learn media literacy while deciding what to trust.

This started as a hackathon sprint, but the architecture has clear production potential.

The 90-Minute Sprint Timeline

1) Starting Point: Custom GPT (Educational Win, Video Limit)

We first tested a Custom GPT:

• Link: https://chatgpt.com/g/g-698dca94190c819185e40361cfafaea5-misinformation-guardian
• What worked: strong educational framing, structured prompts, user-friendly guidance.
• What did not work well enough: video-first analysis fidelity for the demo goal.

Inspect 4K View

2) Pivot: Gemini Gem for Live Demonstration

Next we moved to the Gemini Gem:

• Link: https://gemini.google.com/gem/16KpnhDSxs24YVTbOoauR9YehAvf3SPfo?usp=sharing
• Focus: faster prompt iteration and better multimodal framing for demo flow.
• Result: after roughly 90 minutes of testing and prompt tuning, we had a stronger live showcase.

3) Visual Layer: Nano Banana Mockups

To communicate the capability, we used Nano Banana image generation to visualize outputs and analyst UI patterns.

Inspect 4K View

Inspect 4K View

Inspect 4K View

What We Learned Fast

What worked

• Structured verdict format (claim, evidence, risk, recommendation).
• Visual-first storytelling for trust and comprehension.
• Prompt iteration loops with explicit rubric categories.

What broke

• Single-agent prompting is not enough for high-confidence verification.
• Video claims require deeper multimodal and forensic layers than a simple prompt can provide.
• Without source/citation grounding, confidence can look higher than it should.

From Demo to Product: The Architecture Shift

The hackathon output is now translated into a production path inside FrankX:

• Public hub: /see-through-the-noise
• Live analysis API: /api/misinformation/analyze
• Interactive blueprint: /blueprint/misinformation-guardian-platform
• Production spec: docs/specs/SEE_THROUGH_THE_NOISE_BLUEPRINT.md

Core production layers:

1. Ingest + normalize (text/URL/video metadata inputs)
2. Signal engine (manipulation markers + source checks + claim extraction)
3. Risk fusion + explainability
4. Policy routing + human escalation for high-impact domains
5. Audit ledger + evaluation loops

Naming Strategy: See Through The Noise vs RealityDiffusion.ai

This is the recommendation after distilling the sprint:

• RealityDiffusion.ai = platform brand (the long-term product company)
• See Through The Noise = flagship experience/campaign (the user-facing trust product)

That split gives you both:

• a memorable mission phrase for audience adoption
• a scalable platform identity for future modules beyond misinformation

What Comes Next

1. Add true multimodal video pipeline (frame/audio/metadata consistency checks).
2. Add retrieval-backed citation verification from trusted sources.
3. Add human review queue and appeal flow for high-stakes classifications.
4. Calibrate risk scoring on real eval sets (false positive/false negative tracking).

FAQ

Was this built in one day?

The hackathon prototype direction was built and demonstrated on February 12, 2026, with about 90 minutes of focused prompt iteration for the core demo flow.

Why not stay only with the Custom GPT?

It was useful for education and framing, but the demo needed stronger handling for multimodal/video-oriented scenarios.

Is this already production-ready?

The production foundation is in progress: live API, architecture blueprint, and deployment spec are now in place. Full trust-and-safety operations still require multimodal expansion, eval calibration, and human-review workflows.

Built during hackathon sprint energy. Now evolving into production architecture.