How Far a Wrong Claim Travels Before It Is Caught

We report a measurement from one archived AI-assisted research program: 58 self-introduced errors, logged over 51 days, included only where the primary record documents both the error and its catch, each coded by introducing role, catching mechanism, and how far it had travelled first, across six bands from the same turn to a submission under review. Among these 58 entries in one program, no single mechanism caught a majority: the largest class, an audit by another agent of the same model family, holds 14. In each of the six submission-level catches, the catch came from something other than the agent that produced the error: an editor, a referee, the operator, a database read, or a separate auditing agent. In one late round a smaller model of a second family produced two of the 58 entries, on surfaces the primary family had already audited and passed; that round ran last, so this record cannot separate the family from the ordering, and two entries are one existence result, not a rate. In four entries the archive itself acted as a propagation channel: a wrong line, re-read by later agents, was repeated until an anchor test contradicted it. The limits are the design: a single program, a single operator, one primary model family plus one second family, a survivorship sample with no control condition; the number of errors no check ever caught is unknown. We offer an inclusion rule, a coding scheme, and intervention points per distance band.