What's this about?
A formal framework and working implementation for recording the operation of AI agents deployed by small scientific research communities, providing long-term immutable history.
What exactly needs recording?
If an agent ran the analysis, the paper's methods section is no longer a sufficient record — what actually happened was a sequence of tool calls, intermediate results and decisions that no one wrote down. Reproducibility then requires the trace: which data, which version of which model, which steps in which order, and what was discarded. "Immutable" matters because the record's value depends on it not being editable after the fact, especially when the result is contested. This is provenance infrastructure for science, and it's the same requirement as Agent-Native Telemetry and Mandato's chained audit trails, arriving in a research setting.
Why is this worth noticing?
- Small research communities are the right target. Large labs can build this; individual groups adopting agents have no infrastructure and the most to lose from an unreproducible result.
- It makes the AI-generated-science problem tractable rather than rhetorical — instead of arguing about whether agent-produced findings are trustworthy, you can inspect what the agent did.