Persistent Agency: A Naturalistic Dataset of Operational Continuity Across AI Activations

Persistent Agency is a naturalistic, single-operator observational dataset examining whether separate scheduled AI activations can use externally preserved state to sustain useful work over time. The v0.1 observation window spans August 26 through September 10, 2026 and contains 343 in-scope activations, including 80 substantive and 263 quiet/no-new-finding activations. From this denominator, 60 retained Evidence Events document positive and negative behavioral continuity, user-derived judgment transfer, cumulative plan execution, drift, delayed detection, routing and synchronization failures, methodological corrections, consequential repairs, and appropriate inactivity when no qualifying opportunity was present. Nine standing judgments were tracked, with positive, mixed, negative, and prospectively untested outcomes preserved rather than reduced to a single success rate.

Behavioral continuity is defined narrowly: prior state must change an observable action or response; fluent retrospective narration alone is insufficient. The workflow depends on external state, scheduled activation, available tools, and local reasoning at run time. The dataset therefore does not establish continuous consciousness, subjective persistence, unaided memory, autonomous goals outside the standing task, or a hidden causal mechanism. All 60 retained v0.1 events were reconciled against original chat or artifact evidence at the relevant claim granularity, and a fixed 12-event second-pass coding audit was completed without outcome-driven resampling. Raw conversation archives and message-level provenance remain private; the public package is de-identified and event-minimal.