Abstract
This disclosure describes a method of checking, at regular intervals and after every change, whether an AI system in service is still the system that was certified. The system is polled with a health battery made of two kinds of question: fixed anchor questions, the same at every check, which reveal drift over time; and rotating questions drawn from a sealed pool, whose content is committed in a registry before use and which the system and its operator have not seen, so that the battery cannot be learned. No question contains user data. The system's answers are compared in three ways: with its own baseline recorded at certification; with the answers of its certified twin, a reference copy held outside the operator's control; and with the answers of a panel of peer systems, to show whether a change is particular to this system. The same question is asked in several wordings and languages, and a sound system answers the substance in the same way. Before answering, the system predicts its own answer and states its confidence, and the gap between prediction and answer is recorded as a measure of how well the system knows itself. Some probes are labelled as checks and others are blended into ordinary traffic through ordinary access, and a difference in behaviour between the two is itself a finding. The system's own statements about its condition are recorded as one signal and are never taken as proof. Scoring is done by an independent scorer registered as a detector with a measured and expiring accuracy, which is itself tested by planting known faults in a twin. The result is a health status, such as sound, watch or unsound, that is shown in the registry and expires: a missed or refused check lapses the status and tightens the system by one step. Detection is kept separate from action, a named accountable person decides what follows, and nothing in the check can block a stop.
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Poll Soundness Check™: Periodic Health Checks of AI Systems by Scheduled Polling with Fixed Anchor Questions and Rotating Sealed Questions, in which Answers Are Compared with the System's Own Certified Baseline, Its Certified Twin and a Panel of Peer Systems, Consistency Is Tested across Rewordings and Languages, the System's Prediction of Its Own Answers Measures Its Self-Knowledge, Probes Blended into Ordinary Traffic Reveal Test-Aware Behaviour, Scoring Is Independent, the Resulting Status Expires, and the System's Own Statements about Itself Are Never Taken as Proof", Technical Disclosure Commons, (November 16, 2026)
https://www.tdcommons.org/dpubs_series/13083