Anthropic Credits Meta After Attribution Dispute Over Its Scion Code System
Critics said the internal code management system failed to properly credit earlier work. After Anthropic added a footnote, the author of the critique said the attribution was now sufficient.
Anthropic has run into a credit dispute over Scion, the code management system it uses internally. The company published details of Scion’s design and its operational experience on October 13. Criticism followed, arguing that Anthropic had not properly acknowledged ideas it drew from Meta’s Mononoke. On October 15, Anthropic added a footnote to the announcement crediting the contributions of Mononoke and its data model, Bonsai.
Why the credit dispute started
A developer who worked on Mononoke argued that Scion had not given sufficient credit to prior design work. The point of contention is how a record of code changes is stored. Mononoke, the server behind Meta’s publicly released Sapling system, keeps the changes themselves as its primary data. Lookup-oriented information, such as who changed which line, is computed separately. A similar structure appears in Anthropic’s description of Scion.
The developer who wrote the critique claimed that Anthropic had adopted the Bonsai data model without adequately naming Mononoke from the outset. Drawing on information received privately, the developer also alleged deliberate borrowing and an overlap in the engineers involved. The Anthropic employee who wrote the announcement explained that the intent had been for the reference to Sapling to encompass Mononoke as well. The employee said ideas had indeed come from Bonsai, but there had been no attempt to conceal the source.
After the footnote was added, the author of the critique said the attribution was now sufficient.
How Anthropic handled 30,000 merges a day
Scion was built in response to a sharp rise in code change activity. According to Anthropic, daily merges grew from roughly 2,500 in May 2026 to more than 30,000 in September. A merge is the process of combining changes from multiple development tasks into one. To absorb that load, the company ran several hundred internal Git mirrors, which are copies of the repository. It then spent five months moving its single code repository from Git to Scion.
Scion handles the work of creating and combining change records, as well as querying history, on the server. The machines developers work on fetch only the files and data they need. Anthropic said that after the switch, CPU and memory usage fell to one two-hundredth of what the previous mirror-based setup required. The time needed to fetch code for automated checks and research work also dropped from minutes to seconds, the company said.
Could it work in your own setup?
Anthropic said Scion is not a suitable general-purpose replacement for Git, because it was built on the assumption of a live connection to the server. Scion was constructed around Anthropic’s own code repository and internal ways of working. Rather than something you can pick up and use today, it is closer to a case study in how an organization with heavy code change volume reduced latency and server load.
Anthropic reported that it had eased its operational burden, and it accepted the criticism over attribution. The open questions are how Scion’s actual implementation differs from Mononoke, and under what conditions the gains reported in Anthropic’s internal environment arise.