# ChoiceGap

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Shiwakshi Rashmi](https://cerebralvalley.ai/u/Shiwakshi)
- **GitHub:** https://github.com/shiwakshir/choicegap
- **Demo video:** https://youtu.be/TX65GkCEJqk
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1114

ChoiceGap is a brief-led adaptive survey and verifiable AI-moderated research studio built with Meta Muse. 

The Problem:
Static surveys tell product teams what users click, but miss the situational context and practical barriers. Human UX interviews don't scale across hundreds of sessions, and unguarded AI chatbots ask leading questions, hallucinate findings, or invent product features.

How ChoiceGap Solves It:
ChoiceGap turns an executive brief into an editable, researcher-approved study where deterministic code audits every model decision.
1. Grounded Study Preparation: Muse Spark 1.3 runs native web_search grounding via POST /v1/responses with validated URL citations, while Muse Image 1.0 generates and revises fictional concept stimuli.
2. Researcher Control: Researchers have full input to add, remove, reorder, or edit questions, and lock permitted follow-up topics enforced by code.
3. Multimodal Collection: An included 3D AI Moderator speaks questions aloud, participant answers stream in real time to Muse Voice Transcribe 1.0 for participant review before submission, and Muse Spark inspects optional screen recordings (input_video) proposing timestamped visible events.
4. Verifiable Agent Loop: Muse Spark uses native tool calling (select_approved_condition) to test a single-cell variant while code verifies all other details remain fixed. Before completion, the participant reviews a tentative summary where human corrections strictly supersede model inferences.
5. Behavioral Science Grounding: Findings synthesize across 4 evidence layers (Stated Priorities, Reported Episodes, Controlled Single-Cell Tool Calls, and Participant Corrections) and diagnose barriers using COM-B (Michie et al., 2011), Intention-Behavior Gaps, and Planning Uncertainty, proposing falsifiable 2×2 field experiments.

Backed by 154 passing unit tests, 5 JavaScript test suites, and a 91.6% Bayesian EIG recovery benchmark, ChoiceGap makes AI-moderated UX research rigorous, audited, and verifiable.

## More from Meta Global AI Developer Hackathon

- [relief-net](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1111)
- [TestForge AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1112)
- [OrbitRoute](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1113)
- [Nortex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1115)
- [PM HUB](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1116)
- [Dr Penrod & Mayordomo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1117)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1114. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt