# Execution Assurance Bureau

- **Event:** [Google I/O Hackathon](https://cerebralvalley.ai/e/google-io-hackathon)
- **When:** Sat, May 23 at 9:00 AM – 10:00 PM (PDT)
- **Where:** Shack15, San Francisco, CA
- **Team:** [Shivam Gawali](https://cerebralvalley.ai/u/shiv669)
- **GitHub:** https://github.com/shiv669/repoprobe
- **Demo video:** https://youtu.be/mmrrb-pUJjU?si=ibLyhYnO8C0PoLNO
- **Gallery:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery/2

AI generated repositories increasingly optimize for plausibility rather than behavioral correctness. A repository can boot successfully, expose polished APIs, and still contain incomplete or non functional runtime behavior underneath.
RepoProbe is an execution assurance system for AI generated repositories. Instead of statically analyzing code, it executes repositories, probes runtime surfaces, fuzzes endpoints, verifies behavioral consistency, and compares observed runtime evidence against functionality claimed in repository READMEs.

The pipeline is intentionally split into two layers. Deterministic systems handle repository fingerprinting, execution planning, runtime probing, contradiction detection, and behavioral verification. Gemini 3.5 Flash is then used to synthesize runtime evidence and reason over contradictions without generating evidence itself.

The investigation system escalates suspicious runtime behavior into deeper probing workflows. For example, if an authentication endpoint returns invariant responses across malformed JWTs, empty credentials, and randomized payloads, RepoProbe flags the implementation as behaviorally inconsistent with its claimed authentication guarantees.

The terminal interface streams the investigation live as execution unfolds: runtime boot events, route discovery, contradiction escalation, fuzzing, trust score degradation, and final runtime assessment.

The project was built as infrastructure for a problem that is becoming increasingly common with AI generated software: code that looks production ready while remaining behaviorally incomplete under execution.

---

Markdown version of https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery/2. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt