Anthropic built Scion, its internal version control system, using Claude over five months and presented it as a landmark demonstration of AI-assisted software engineering. The problem: Scion's architecture is a near-exact copy of Mononoke, Meta's open-source version control backend, right down to the data model, changeset structure, derived-data indexing, and server-side API. The original post credited Meta's Sapling ecosystem only in passing, for vague "hard-won lessons." Rain Kadastik — known online as sunshowers — who led the design of Mononoke's Bonsai data model at Meta, called it out. The resemblance goes far beyond shared concepts. Scion's changesets map directly to Mononoke's Bonsai changesets. The derived-data terminology and framing are identical. Even the name "Scion" appears to reverse the syllables of "Bonsai" with the B dropped. After the first post, the author received private confirmation that the copying was intentional and that there was overlap in personnel between Meta's source control team and Anthropic's Scion effort. This was not convergent evolution. The Bonsai data model was a wildly contrarian bet in 2018. Building a write-ahead log-style Merkle DAG diverged sharply from the Git-like consensus in version control design. It required solving hard problems — server-side rebases, novel algorithms — and the author personally had to justify it through multiple layers of Meta management. The model has been running successfully for nearly a decade. This was not obvious infrastructure; it was a deliberate, risky architectural choice that paid off over years of production use. Anthropic's framing — "our LLM built this system and productionized it in five months" — collapses when you learn that ten years of design, debugging, and production hardening at Meta provided the blueprint. Pointing Claude at an existing, battle-tested system and rebuilding it is genuinely impressive as an AI capability demonstration. But it is a fundamentally different achievement than designing from scratch, and presenting it as the latter is dishonest. The author makes this point clearly: just say you copied Mononoke and rebuilt it in five months. That is still noteworthy. The attribution question is sharpened by the personnel overlap. People who worked on Mononoke at Meta now work at Anthropic. They knew precisely what they were building from. Crediting Sapling (the client) rather than Mononoke (the backend) — and framing the debt as "hard-won lessons" rather than "we copied the data model" — is not carelessness. It is deliberate obscuring of provenance, which is particularly galling from a company that markets itself on trust and safety. Anthropic has since responded, and attribution has been added to the blog post. But the damage to credibility is real. If a company's flagship engineering narrative requires stripping credit from the systems it copied, every subsequent technical claim is suspect. The author's updated default assumption — that Anthropic's public engineering presentations are "largely bad-faith ripoffs" — is a rational Bayesian update, not an overreaction. The broader pattern matters more than this single incident. Large AI companies are incentivized to present AI-assisted work as ex nihilo creation, because that narrative drives investment, hiring, and competitive positioning. When the input to the LLM is a decade of open-source engineering by a specific team, erasing that input inflates the perceived capability of the AI and extracts reputational value from the humans who did the foundational work. This is the extraction mechanism: credit laundering through the LLM layer.
