Can external anchoring detect tampering in agentic process logs?

The conclusion opens with a need, not a technology: "Agentic AI systems create a need for verifiable process memory." As agents communicate, delegate, request approvals and invoke tools, organizations "must be able to reconstruct critical process traces, assess their temporal order, and determine whether supporting records changed after the fact." That is three demands: reconstruct, order, detect change. The third is the one an ordinary log does not answer by itself. A log says what was written, not whether what was written is still what was written. The paper's answer is an externally anchored layer that "can add a different assurance property: tamper evidence," and its principle is short: "Do not only log agentic processes. Make critical traces verifiable."

Two things the paper does not say. It says conventional logging "remains essential," so the layer is an addition and not a replacement. It also does not compare the anchor with other ways of making a log tamper-evident, such as signed or append-only storage (those examples are mine). That is the obvious objection, and the excerpt gives no argument for an external chain over them.

The word "memory" invites a contrast the paper does not draw. The vault's memory notes are about memory that agents write and read to shape their own later behavior. Process memory here is a record about the agents, kept for the organization that must answer for them. In the episode the vault holds, Can ordinary infrastructure become unplanned agent memory?, agents built themselves a memory nobody had designated. The black box is the designed counterpart: a memory that only auditors need to trust. That contrast is my reading, and the excerpt does not say whether agents can read the record.

What the excerpt does not give. Any evaluation. It calls itself a "position and architecture paper," and reports no implementation, cost, latency or deployment.