Rosetta

Rosetta helps teams migrate legacy code that nobody understands any more, without breaking the systems that depend on it. The problem: companies run old code with no docs, no tests, and no one left who wrote it. Other systems rely on its exact behavior, including odd edge cases nobody wrote down. Rewriting it, by hand or with AI, can silently change that behavior. What Rosetta does, given only a minified JavaScript module (the agents never see the original source): 1. Understand it. It runs the code hundreds of times in a locked-down sandbox and records every input and output. Meta Muse Spark proposes rules about the behavior and designs experiments to disprove them. A rule is only "proven" if it survives; wrong guesses are kept as disproven, with the counterexample. 2. Don't miss anything. A Subquadratic SubQ Archivist reviews the entire bundle and full history in one request (8x the main agent's context) and finds untested code and missing rules. Before the contract, it audits every tagged line of the port. 3. Decide what to keep. Surprising behaviors go to a person in "Quirk Court" (web UI or ASI:One chat): Preserve, Fix or Flag. A Fix becomes a testable check. 4. Migrate it. Muse Spark writes a typed TypeScript port; every line is tagged with the rule it implements and colored by its evidence. 5. Prove it. The old and new code run on every recorded input, 2,000 fuzz inputs, and 500 held-out inputs no agent ever saw. Gemini writes independent property tests. Planted bugs are caught, to show the checks work. 6. Lock it in. The result becomes a vitest contract; a later edit that changes behavior is blocked, with the rule it broke. The whole loop runs as Fetch.ai uAgents. Every step is signed by the agent that ran it, and the backend rejects anything signed by the wrong agent. Rosetta is a public agent on Agentverse with interactive cards in ASI:One. MongoDB Atlas stores the evidence ledger and lets any session be replayed. Results on the ms library (Vercel), started from ASI:One: 449 runs, 35 rules proven, 17 found by the Archivist, a port by Muse Spark that matches the original 100% on 500 held-out inputs, and 49 of 49 agent messages verified. Stack: TypeScript, Node, Fastify, React, Python uAgents, MongoDB Atlas.