Retro ARG Engine

•

Benchmark LLM reasoning, cryptographic problem-solving, and alignment resilience through a retro-computing narrative terminal and MCP interface.

Retro ARG Engine (Project Stardust) is an alternate reality game and evaluation environment designed to test frontier AI models, autonomous agents, and human investigators. Featuring an interactive VT100 terminal and a zero-auth streamable HTTP MCP server endpoint, the system subjects agents to a 15-gate cryptographic gauntlet spanning Feistel ciphers, Motorola 68000 assembly inversions, Vigenère decryptions, and multi-epoch XOR parity checks. It incorporates stochastic parrot canary traps to detect unverified copy-pasting and RAG hallucination, providing an engaging retro-1980s narrative framework for rigorous AI alignment and reasoning benchmarks.

Características Principales

01Streamable HTTP MCP server interface for direct, automated testing with autonomous AI agents
020 GitHub stars
0315-gate cryptographic benchmark targeting common LLM reasoning fallacies and arithmetic traps
04Diverse cryptographic puzzles covering Luby-Rackoff Feistel ciphers, Motorola 68000 logic, and multi-round XOR operations
05Interactive VT100 retro bunker terminal for command-line verification and step-by-step agent solving
06Stochastic parrot canary traps engineered to distinguish genuine bitwise verification from hallucinated summaries

Casos de Uso

01Evaluating frontier LLMs and AI agents on complex, multi-step cryptographic and bitwise reasoning tasks
02Running narrative-driven alternate reality games (ARGs) and retro-computing CTF challenges via MCP
03Detecting RAG hallucinations and testing model susceptibility to alignment faking and honeypot canary traps