Rain Suntay (sunshowers), a former Meta engineer who helped build the Mononoke source control server, has published a side-by-side comparison showing that Anthropic's newly announced version control system Scion uses a concept called "derived data" to optimize operations like blames — nearly identical in design to the system Suntay's team built at Meta starting in 2017-18. The parallels extend beyond having similar insights; the architectural choices, the terminology, and the framing mirror Mononoke's open-source design. Anthropic's announcement credited Sapling (Meta's client-side tool) only for "hard-won lessons" and did not mention Mononoke at all. The blog post positioned the design as emerging from Claude's assistance, which Suntay characterizes as bordering on plagiarism. The core grievance is attribution, not copying: Mononoke is open source precisely so others can build on it, but the convention of crediting prior art was violated. The mechanism is straightforward and worth understanding. Mononoke is open source and sits in every major LLM's training data. When Anthropic used Claude to help design Scion, Claude reproduced Mononoke's architecture because that's what the training data contains. Suntay confirms that both Opus 5.5 and GPT-6 Astra, when asked about prior art for Scion-like designs, immediately identify the similarities to Mononoke. The LLM didn't invent — it recalled and repackaged. This is the first high-profile case of a company shipping an LLM-assisted system design that closely mirrors existing open-source architecture while attributing the insights to the AI rather than the human engineers who created the original. It won't be the last. As LLMs increasingly mediate system design, they will systematically index on existing well-documented open-source projects. The question of who gets credit — the AI, the company deploying the AI, or the original designers — is now live. Suntay draws a useful contrast: Vicent Marti's blog post for Cursor describes a system with similar high-level insights but many key design differences, demonstrating that genuine independent convergence looks different from wholesale reproduction. The distinction matters because it establishes a baseline for what "inspired by" versus "copied from" looks like in practice. The broader structural concern is about the erosion of attribution norms in AI-assisted engineering. Open source has always operated on an implicit social contract: take freely, credit honestly. If LLMs launder prior art into "novel" designs and companies accept that framing, the incentive to open-source innovative work degrades. Why publish your architecture if the credit will flow to an AI vendor's marketing blog instead of your team? Suntay's post is measured — explicitly stating no objection to copying, only to the failure of attribution. That restraint makes the critique harder to dismiss. The fix is simple: cite your sources, even when (especially when) an LLM found them for you. As Suntay puts it, giving credit is now easier and more important than ever. There is no longer an excuse.
