# daylite

- **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:** [Dmitrii Karataev](https://cerebralvalley.ai/u/kwit)
- **GitHub:** https://github.com/kwit75/interlock
- **Demo video:** https://www.youtube.com/watch?v=R_RwgO8CWMs
- **Gallery:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery/4

INTERLOCK is the agentic orchestration layer for deepfake forensics at the moment of authorization — for any high-stakes voice or video call. When a deepfake CEO impersonates a real executive on Google Meet to authorize a wire transfer — the exact scenario Arup faced in Hong Kong, January 2024, losing $25.6 million in fifteen transfers — INTERLOCK orchestrates a Council of specialist signals, gates on consensus, and publishes a verdict event the bank's risk system acts on.

The Council is six Gemini 3.5 Flash sub-agents — Frame Forensics, Voice-Print Cross-Match, Reverse Provenance (Search-grounded), Counter-Strategy, Regulatory Precedent (Search-grounded), and Injection Guard — fanning out in parallel from an Antigravity Managed Agent orchestrator. Frame Forensics and Voice-Print are designed as adapters that wrap specialist detectors via Antigravity sandbox tool calls — Modulate Velma (1.1% EER · HF Speech Deepfake Arena #1), Resemble DETECT-3B Omni (97.4% accuracy), Pindrop Pulse (secures 8 of 10 largest US banks). We are the diagnostic laboratory; they are the microscopes. A verdict aggregator gates on 3-of-6 consensus; no single sub-agent can block or override. The same architecture defends wire confirmations, hiring interviews, claim assessments, and KYC re-verification calls — voice or video.

Containment runs as a real `interactions.create({base_agent: "antigravity-preview-05-2026"})` sandbox interaction. The sandbox executes Python that signals the bank's wire-freeze workflow, locks the CEO account through the bank's existing risk pipeline, and drafts an SEC 8-K Item 1.05 disclosure Search-grounded against EDGAR precedent (Microsoft 2024-01-17 Midnight Blizzard). INTERLOCK never debits — the bank's existing infrastructure acts on the verdict, same deployment model as CrowdStrike or Pindrop. Every irreversible action runs in a sandboxed env_id with a surfaced interaction_id — auditable, replayable, never autonomous.

Architectural orientation: the Vector Institute's May 15 2026 "Generalization Illusion" report — SynthForensics benchmark shows 29.19% mean AUC drop for SOTA detectors against novel diffusion generators. Standalone media forensics is fragile; orchestrated consensus above multiple diverse specialists is the architectural answer. Six sub-agents in parallel; verdict aggregator + comms layer sequential after consensus — eight total Gemini 3.5 Flash calls per detection. Built solo in 13 hours.

---

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