ETHGlobal Experiment·2026·ETHGlobal

The Last Alibi: Trust Boundaries in Generative State

An experiment in enforcing cryptographic trust boundaries so generative AI cannot silently redefine canonical game state.

Cover image for The Last Alibi: Trust Boundaries in Generative State

Overview & Research Motivation

The Last Alibi is a cryptographic detective game exploring a critical systems problem: how to allow generative AI models to provide dynamic, creative user interaction without allowing the model to hallucinate or alter ground-truth canonical state.

Suspect dialogue is generated dynamically by LLMs, but the underlying truth (alibis, certified evidence, and verdict outcomes) is committed in Groth16 zero-knowledge circuits on Sui. The LLM cannot invent evidence or alter the verdict state; verification is fail-closed, ensuring that cryptographic proof guarantees strictly govern game progression.

The Core Systems Problem

  • Generative AI models are non-deterministic and easily manipulated by prompt injection to reveal secrets or falsely validate game logic.

My Specific Technical Contributions

  • Designed the Circom Groth16 zero-knowledge circuits verifying evidence validity without revealing the culprit's identity.
  • Built the fail-closed verification pipeline on Sui Move and 0G Compute.
  • Structured the prompt boundary protocol isolating LLM dialogue generation from ground-truth evidence hashes.

System Architecture & Verification Pipeline

Cryptographic Evidence Commitment

Case evidence and culprit identities are committed as Merkle roots on Sui before game commencement.

Zero-Knowledge Proof Verification

Players submit Groth16 proofs demonstrating logical contradictions in suspect testimony without exposing private witnesses.

Isolated LLM Dialogue Sandbox

0G Compute instances generate personality responses constrained by strictly filtered observation tokens.

Technical Stack & Tools

CircomSnarkJSSui MoveTypeScript0G ComputeNext.js

Connected Systems & Inquiries