Vishisht Magan

Mini-Colosseum is a desktop research harness designed to solve long-horizon mathematical proofs and theoretical computer science problems. When a researcher inputs a theorem or conjecture, the system autonomously explores strategies and decomposes the problem into a Directed Acyclic Graph (DAG) of subproblems. Crucially, every candidate draft is actively attacked by a dedicated falsifier model, and the resulting exact-quote objections are permanently bound to the draft before aggregation. The aggregator writes a synthesized draft that resolves the explicitly cited flaws. Built as a React/Electron application wrapped around a FastAPI Python state machine, Mini-Colosseum brings total transparency to AI research. It features a CoqIDE-style workspace, Danus SHA-256 dependency hashing (where failed sections automatically invalidate downstream claims), and a tight budget circuit breaker. Every decision, critique, and cost is logged to an append-only JSONL ledger, allowing researchers to replay the AI's reasoning step-by-step completely offline.