Sol – a validated computational artifact for human-agent handoffs

This project is a response to Thariq Shihipar's essay written in Anthropic blog. While I agree with Thariq's critique that Markdown isn't enough for the work that AI Agents are now conducting I disagree that outputting HTML is the answer. HTML is a presentation format and while it looks good and is easy for human readers it is insufficient for work that needs to present a record of how the final output was arrived at.

The repository contains an RFC for a specification for a durable artifact that preserves the work conducted by anyone or anything. The repo also contains a Python validator (minimal implimentation) and a gold-case and pathology corpus long with deterministic JSON diagnostics.

The validator checks for reference and object integrity, ensures that the execution order of any code is plainly and always visible, any proposals and reviews that may have been conducted, as well as recording authorization.

You can run the validator by:

python -m solval.make_corpus corpus python -m solval.validate_corpus corpus

As stated, this is a first draft and intended for commentary and community development. Feedback is very much welcomed.