Auditing LLM agents against policies is critical for safety, but the cost is high: every policy requires rereading long trajectories from scratch. A new paper on arXiv asks whether that step can be avoided by recording each trajectory once and reusing that record for policies that are not known in advance.
The proposed framing is a "write once, check many" ledger. Instead of repeatedly scanning raw agent behavior, the idea is to create a reusable record that can later be checked against whatever policies exist at audit time. This would shift the bottleneck from repeated reading to a one-time recording plus subsequent checks.
The abstract does not provide implementation details or experimental results, so the feasibility of the approach remains open. But the question targets a real bottleneck in AI safety: if such ledgers work, post-hoc compliance auditing could become far more scalable as policies evolve.