Anthropic recently announced Scion, a new version control system that employs a concept called “derived data” to optimize operations like blame tracking. According to a post on sunshowers.io, this approach mirrors techniques used in Mononoke, a source control server developed at Meta.
The author, who contributed to Mononoke’s development starting in 2017-18, notes that both systems use derived data to improve efficiency in similar ways. In Anthropic’s announcement of Scion, the company credited Sapling—another Meta project—for providing “hard-won lessons,” but did not mention Mononoke despite the apparent conceptual similarities.
The author argues that the level of similarity goes beyond coincidental parallel insights. They suggest that Anthropic’s framing of Claude as the designer of Scion overlooks the earlier work by Meta’s team. According to the post, both Mononoke and Sapling drew from prior art including Google’s SrcFS and write-ahead logs used in databases.
The author notes they do not object to Anthropic adopting Mononoke’s design approach, particularly given that Mononoke is open source and widely available. They point out that as large language models increasingly assist in system design, these models are likely to draw heavily on existing designs present in their training data, including Mononoke.
According to the post, when asked about prior art, both Opus 5.5 and GPT-6 Astra (a hypothetical model name) readily identify similarities to Mononoke. The author emphasizes that providing credit to prior work has become both easier and more important as LLMs become involved in design processes, stating there is “no longer an excuse” for omitting attribution.
Key facts
- Anthropic’s Scion version control system uses a ‘derived data’ concept to optimize operations like blame tracking
- Meta’s Mononoke, developed starting in 2017-18, employed similar derived data techniques
- Anthropic’s announcement credited Sapling but did not mention Mononoke, despite conceptual similarities
- Mononoke is open source and included in LLM training data
- The post’s author states they do not object to the design being adopted, but emphasize the importance of attribution