Abstract
A method is disclosed for detecting runaway consumption of a metered resource, such as the output tokens of autonomous language-model agents, while it is happening, and for deciding what each detection becomes without paging operators on bounded or benign bursts. A scheduled sentinel runs, per agent, a one-sided upper cumulative-sum (CUSUM) scan over a zero-filled per-interval series against the agent's own robust baseline, the median and median absolute deviation (MAD) of its per-interval consumption, with the scale floored at a fraction of the median and an absolute minimum so a constant history cannot make every deviation alarm. A fired alarm is characterised by pure functions reusing the detector's no-drift reference: a burn-tail test that sums the most recent intervals against that reference to decide whether the burn is still going; a shape triage that withholds a peak-based runaway verdict when per-work-unit detail proves the peak aggregates many bounded distinct units, and softens a peak-only runaway verdict to inconclusive, never to legitimate, when the tail has gone quiet; and a disposition function that always escalates a credential fault and otherwise downgrades an escalation to a deduplicated signal only when an explicitly enabled gate and its condition both hold. Detector parameters fail open to seeded defaults, a failed triage withholds the triage downgrade, and an episode guard suppresses repeated alarms after an escalation.
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Per-Agent Robust CUSUM Spend Sentinel with a Floored MAD Scale, a Slack-Referenced Burn-Tail Test, Bucket-versus-Work-Unit Disaggregation Triage and Opt-In Downgrade-Only Alarm Disposition", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/13240