# WIB

- **Event:** [Gemini 4 NYC Hackathon](https://cerebralvalley.ai/e/gemini-4-nyc)
- **When:** Sat, Nov 21 at 9:00 AM – 10:00 PM (EST)
- **Where:** New York, NY
- **Team:** [Abdulsalam Fadare](https://cerebralvalley.ai/u/Salam), [George Danquah](https://cerebralvalley.ai/u/danquahgeorge3), [Samuel Korda](https://cerebralvalley.ai/u/FixedIt1)
- **GitHub:** https://github.com/Salam721/what-breaks
- **Demo video:** https://youtu.be/KbHTh1FFFHk
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/141

What Breaks If I Ship This? is a live pull-request verification engine that predicts which real user flows a GitHub PR could break—and then proves or disproves each prediction in a real headless browser before you merge.

Static diff reviewers often hallucinate bugs or miss runtime regressions across routes, schemas, and UI flows. When you paste a public GitHub PR URL:

Triage & Parallel Analysis: The server fetches the PR diff and shallow repo context, triages affected files/areas, and runs two independent read-only analyzers (Gemini 4 Carbon and Gemini 4 Argon) in parallel to map a dependency graph of pages, endpoints, tables, and user flows alongside concrete breakage hypotheses.
Calibrated Risk Scoring: Merged predictions are scored in a single Decisions API (gemini-decision-1) request to filter out noise and pick up to the top 10 risky flows.

Parallel Sandbox Verification: Up to 6 concurrent Gemini Managed Agent (antigravity-preview-09-2026) Linux sandboxes clone the PR head, boot the target app, and exercise each flow with headless Playwright Chromium—capturing full-page screenshots and server/browser logs. (If an app cannot start without external secrets/DBs, it honestly reports unverified instead of faking results.)

Blind Multimodal Judging & Server Verdicts: An independent Decisions API judge inspects only the claim, screenshot, and trimmed logs (never verifier prose, retrying once if confidence < 0.7), and deterministic server logic sets the final verdict (confirmed, false_alarm, unverified, low_risk) with SHA-256 evidence manifests streamed live over SSE into an interactive React Flow graph and Evidence Drawer.

## More from Gemini 4 NYC Hackathon

- [Grand Teach Auto](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/1)
- [Goal Goose](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/2)
- [ClearLedger](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/3)
- [GeneSleuth](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/138)
- [Forkcast](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/139)
- [Space Theory](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/140)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/141. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
