ShriAli

Clean Room rebuilds software by watching it. No source code, no decompiler, no peeking under the hood. You show it an app running, a phone calculator, a terminal calendar, a little block game, and it watches, learns, and writes a working copy of its own, then proves the copy actually behaves the same. That is the whole idea, and why we built it in a day at the AIE Code Summit. It starts with a problem most people have felt. Software outlives the people who wrote it. A company shuts down and the smart home it sold stops working overnight. A team inherits a system nobody dares touch because the original authors are gone and the source is lost. The code is gone, but the behavior is still right there on the screen. Clean Room turns that behavior into three things a team can actually use: a manual of every rule the program follows, a working replacement in plain JavaScript, and proof that the replacement is faithful. Here is how it works. First, it observes. The phone screen is mirrored to a Mac, and the server screenshots it several times a second, working out which key was pressed purely from what changed on screen. Every press and settled screen becomes a line in a trace. Second, it reads. A vision model reads each screen into what the display shows, the expression so far, and which operator is active, with a sanity check to catch misreads. Third, it learns. The model writes down the rules, tagging each one seen, with the exact presses that prove it, or guessed, as a hypothesis to test, then writes a full copy. Fourth, it proves. The copy runs in a locked-down Node sandbox against the trace, split into tests it trained on and tests it never saw. Fifth, it fixes. Every mismatch goes back to the model, and a version is kept only if it scores higher. What makes this more than a clever demo is that nothing is taken on trust. A copy that memorizes what it saw instead of learning the rules gets rejected. A held-out test the copy never trained on is what counts. The checker runs in a sandbox with no file, network, or environment access, so a copy cannot cheat its way to a good score. Every rule is labeled, so a person can see what was observed and what was guessed. We proved it on three targets in one day. A phone calculator rebuilt from a recorded trace scores 100 percent with the checker, 31 out of 31 lines, with train and held-out tests both at 100 percent. The Unix cal command is rebuilt the same way. And as a bonus, the same loop learned Block Blast from a phone screen recording. It found the board, learned the move rules, rejected copies that memorized observed boards, and produced a playable rules-only game. Clean Room is built by Ali Akca and Shri Sriram, with AI coding agents working in parallel on separate parts, all tested and reviewed before merging. It is not a tool for stealing software. It is for the software you own or are authorized to recover, the systems that still work but whose code is gone. Watch, learn, rebuild, prove.