Continuous Evidence for Insuring Autonomous Agents

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Liability cover for AI systems is bound today on a point-in-time certification of an operator's controls. That establishes what controls existed on the day. It does not establish what the agent did between assessments, whether those controls were in force when a consequential action occurred, or whether the value the agent acted on came from the operator or from content the agent read. This paper describes a record that makes those three facts inspectable, and reports what has been measured of it. An in-process gate sits between an agent and every consequential action, admitting an action only on a capability authenticated under an epoch-indexed governance key, and appending grants, refusals, approvals, revocations and detected tool-manifest drift to a hash-chained log. Each argument carries a provenance band assigned by the channel it arrived on; the value's own account of its origin is never read. Measured results are reported beside undefended controls, with predictions registered before the runs and failures reported as failures. The underlying measurement record, including the raw benchmark results, the predictions that failed, and the instrument defects found, is deposited separately at 10.5281/zenodo.22836611.