What the Log Couldn't Hide

The UK Merchant Shipping Act of 1894 did not simply require ships to keep a log. It prescribed the book: a government-approved form with designated spaces, entries made promptly and signed by both master and mate, and a rule that any late entry record both the date of the event and the date of writing. Those were the specifications of a legal instrument. Their purpose was to make gaps and alterations visible to whoever opened the book afterward. When Judge Learned Hand found entries interpolated over erased material in a ship's log in 1939, he did not merely discount the record. He read the fabrication as evidence that the vessel had something to conceal.

The log's worth lay in what it could rule out.

Laboratory notebooks had a harder case to make. In U.S. patent disputes, priority turned on corroboration: credible evidence of who conceived what, and when. NIST called for bound notebooks with preprinted page numbers, permanent ink, contemporaneous entries, and two witnesses per page. Preprinted numbering meant a removed page announced itself. Chronological order meant you could not slip one in. Where notebooks lacked those features, courts found them nearly worthless as proof.

Digital logging arrived to solve a different problem, and the shift happened without anyone announcing it. Eric Allman's syslog, written in 1983 for BSD Unix, helped administrators pull real problems out of a stream of status messages. It did that job well. RFC 3164, which later documented the protocol's behavior, is unusually candid about the limits: syslog cannot verify who sent a message, cannot guarantee delivery or ordering, and cannot stop messages from being silently dropped. The protocol included security categories, so accountability was not off the designers' minds. But the architecture was built to answer what happened?, not to hold up against someone actively working to hide what they did. A bound book with numbered pages makes concealment structurally awkward. An event stream makes observation cheap — a real achievement, but not the same one.

That distance matters now, because agent systems are producing audit trails in the digital-logging tradition. They capture what the agent did, step by step, in operational detail. The question those trails will meet in a dispute, an audit, or a regulatory proceeding is the older one: can the surviving record distinguish an authorized sequence of actions from an unauthorized one?

A position paper submitted to the W3C's 2026 ODRL workshop gives that question a usable shape. Yuqiang Wang proposes a test he calls compliance determinability: a log is sufficient only if every execution history consistent with the surviving evidence yields the same compliance verdict. In plainer terms, if an authorized action and an unauthorized one could leave identical entries behind, no reviewer, human or automated, can say which occurred. As Wang puts it:

The test is a proposal, not a standard. But it names in formal language what the maritime and laboratory traditions worked out through paper and binding. Numbered pages rule out insertion. A witnessed, dated entry prevents backdating. A correction that must leave the original legible makes concealment visible. Most agent audit trails were never designed to rule out anything; they were designed to show you the run. They can tell you what the system did, and they generally cannot establish whether it was permitted to do it, or whether something that should have happened quietly didn't. As I explored recently, production traffic is already writing the record that later attribution fights will turn on. No deployment budget rewards getting that record right before the first expensive argument makes it necessary.