An audit of two research repositories shows that prescriptive agent traces (rule files) offer 85.6% visibility of history, whereas narrative traces (agent scratchpads) only preserve 21.8% of historical episodes. This highlights a significant gap in how AI contributions are recorded in Git.
HOW THIS AFFECTS YOU
●
researcherYou should be aware that simple agent scratchpads may not provide sufficient provenance for your computational work.
●
policyYou should consider how different types of agent logs affect the auditability of AI-assisted research.