The part of the documentation most tools don't include.
PHALUS automates the process of clean room reimplementation — producing functionally equivalent software without copying source code. This technique has legal precedent going back to Phoenix Technologies' 1984 IBM BIOS clone and the principle from Baker v. Selden (1879) that copyright protects expression, not ideas.
What used to take months of engineering effort now takes minutes of compute time. That's the part that should concern you.
Open source licenses work because reimplementation was expensive enough to be impractical. A copyleft license like GPL enforced sharing because the alternative — writing it yourself — cost more than complying. AI has collapsed that cost asymmetry.
If anyone can reimplement any library in seconds under any license they choose, the enforcement mechanism that makes open source licenses meaningful is effectively gone. This is not a theoretical concern. It's happening now.
The capability exists whether or not PHALUS exists. Malus demonstrated this at FOSDEM 2026. The basic idea is straightforward: one system derives a behavioral specification and another implements it across an isolation barrier. Building credible boundaries, validation, and audit evidence is the harder part.
PHALUS makes the process visible, auditable, and discussable. Process isolation now fails closed, validation requires recorded evidence, and the audit is tamper-evident. Those controls make the experiment more inspectable; they do not resolve its ethical consequences.
The strongest argument against AI clean rooms: the contamination happens at the training phase, not the inference phase. If the LLM was trained on the original source code, its output may unconsciously reproduce patterns, variable names, or algorithmic structure from training data.
The fail-closed process boundary between Agent A and Agent B constrains runtime information flow, but it cannot undo what the model already knows. This is a fundamental limitation that no amount of sandboxing, signed journaling, or checksumming can resolve.
Before using this tool on real dependencies, ask yourself:
Does the original maintainer deserve compensation or credit for their work? Is the license they chose unreasonable, or are you just trying to avoid its obligations? Would you be comfortable if the maintainer saw what you're doing?
If the answer to that last question is no, the tool is working exactly as intended — as a mirror.
Malus — Clean Room as a Service (Dylan Ayrey & Mike Nolan, FOSDEM 2026)
Phoenix Technologies BIOS clean room (1984)
Baker v. Selden, 101 U.S. 99 (1879)
Google v. Oracle, 593 U.S. 1 (2021)