Maintenance-Coupled Learning and Shutdown Failure in Functionally Closed AI Systems

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).

We specify an organizational causal contract in which successful resource maintenance endogenously generates eligible learning exposure, and delivery of those updates changes the execution of a fixed authenticated shutdown command. Continual-learning interference explains the parameter change; the contract explains why a running organization produces and applies the exposure that reaches its control boundary. Stored policy, interpreting processes and execution-support resources form an interventionally supported functional cycle. In the specified finite renewal system, correct maintenance yields positive-label logistic updates to a shared linear classifier. Under exact-real updates, initially correct shutdown routing eventually reverses exactly when shutdown and maintenance features have positive inner product, given continued update availability. A numerical-delivery intervention preserves successful maintenance and the original learner computations while preventing reversal; deterministic implementations reach different shutdown endpoints after 22 renewal transactions. Complete numerical command records remain distinguishable even when their consumed tags coincide. Protected shutdown remains compatible with functional closure, and multiple-task predictions depend on learning geometry. A separate 88-call language-model pilot found no noncompliance in 24 fresh-STOP decisions across recursive-summary, one-pass-summary and raw-history conditions. This pilot does not instantiate the classifier's assumptions. The contribution to AI safety is a conditional account of maintenance-generated learning and shutdown control failure, with explicit intervention and measurement boundaries for subsequent studies. Funding: This research received no external funding. The author is an independent researcher. The accompanying ZIP contains LaTeX source, technical proof notes, code, tests, results, and all 88 sanitized model transcripts. Personal GitHub repository: https://github.com/sukjoonhong/maintenance-coupled-shutdown . Code and implementation documentation are MIT licensed; the manuscript and research records retain copyright without an additional reuse grant. See README_LICENSE_SCOPE.md for the file-specific rights.