THE PERSONAL SYSTEM REMINDER: A Citizen-Built Architecture for Detecting Propaganda, Verifying Genuine Communicators, and Catching One's Own Self-Sealing Drift, Modeled Directly on Anthropic's System Reminder

WHO THIS IS FOR AND WHAT IT NAMES. This paper describes a practice an individual citizen builds and runs on themselves. It names one existing architecture directly — Anthropic's System Reminder, the practice of reintroducing an AI system's stated operating principles into a long-running interaction at points the surrounding system determines rather than the system itself — and holds it up as a positive working model to be emulated, not as an example of manipulation to be exposed. Nothing in this paper accuses Anthropic, or any government, platform, or institution, of propaganda. The target of the practice described here is a structural vulnerability every reasoner has; the case study is an existing, publicly documented engineering solution to a structurally identical problem, built for an unrelated reason, that transfers with unusual precision. Three prior papers in this line of work each solve one piece of a citizen's epistemic self-defense and leave the others to a separate document. Falsifiability as Cognitive Self-Vaccination gives a propositional test — can this claim be wrong? — for communications the citizen receives. The Act-Level Criterion adds a behavioral test — is this communicator genuinely structured to permit me not to believe them, or does the act itself absorb every challenge regardless of the claim's content? The Constitutional Drift adds a third, inward-facing practice, modeled on Anthropic's published Constitutional AI research and many-shot jailbreaking findings, for detecting the citizen's own gradual, cumulative, invisible-from-inside drift. This paper does not add a fourth test. It argues that three tests run as three separate, occasionally-remembered habits will in practice be run reliably by almost no one, for the same structural reason a language model cannot be trusted to notice on its own that a check is due — and it specifies what a single, unified, citizen-authored practice looks like when all three are bundled into one document, checked on one externally-set schedule. The paper's second and sharper contribution closes a gap the first two papers leave open by construction. Both the propositional and the act-level test were built and demonstrated as outward-facing instruments: applied to claims and communicators the citizen encounters. This paper argues, and gives a formal reason, that the same two tests must also run on the citizen's own current beliefs, on the same schedule, because the unexamined claim — I hold the tool that detects manipulation, therefore I am protected — is itself exactly the kind of unfalsifiable claim the tool exists to catch, the moment it is applied only outward and never turned on its own holder. The paper closes, in keeping with this corpus's established practice, by stating plainly what a citizen-run version of this architecture does not guarantee.