Abstract
This disclosure describes a framework in which many detectors, such as health checks, anomaly detectors, malware and fraud detectors, detectors of generated or altered media and AI safety monitors, judge evidence honestly and in combination, and in which their judgments cannot by themselves cause a system-wide failure. Each detector is registered with an identity, version, the classes of evidence it judges and a measured accuracy that expires unless re-measured. Detectors judge signed evidence, such as readings from certified sensors and records from monitored systems, and return a verdict with a cause class (for example faulty, overloaded, unreachable, compromised or unknown), a confidence and references to the evidence; "unknown" is a legitimate verdict. Verdicts are combined under published rules that weight detectors by their measured accuracy and count detectors that share a dependency as one, and disagreement is shown, not hidden. Detectors never act: separate, declared policies decide what to do with combined verdicts, in graded steps. Detectors are themselves monitored, by comparison with their peers, with spot checks against ground truth and with outcomes, and a detector that diverges is set aside. To prevent cascades, a unit judged overloaded is relieved of load rather than removed; when an unusual share of units is judged unhealthy at the same time, removals are frozen and a common cause is suspected, beginning with the detectors themselves, shared dependencies and the most recent recorded changes, which are checked against their rehearsals and rolled back if implicated; no combination of verdicts may remove capacity below a declared floor, below which service steps down to a declared safe mode; and capacity is reintroduced gradually with load ramped, so that recovered units are not overwhelmed again. These protections apply only to actions that reduce capacity or service; they never block or delay a stop, a tightening of limits or any action taken for safety. Every verdict and action is signed and recorded, so that incidents can be replayed and judgment errors located. The framework sits in a declared chain: certified sensors sense, the framework judges, an alarm layer alerts people, a coordination layer that can only tighten carries out policy actions across sites, and safe modes and limits act; alarm floods are grouped into incidents when many units fail together, but alarms for life safety are never grouped away or delayed, and responders' acknowledgements feed back as evidence of each detector's accuracy.
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Detectory™: One Framework for Many Detectors, in which Registered Detectors with Measured and Expiring Accuracy Judge Signed Evidence, Verdicts Carry a Cause Class and an Honest "Unknown", Results Are Combined under Published Rules that Count Dependent Detectors Once, Detectors Never Act Directly, Detectors Are Themselves Checked against Their Peers and Reality, and Cascades Are Prevented by Distinguishing Overload from Fault, Freezing Removals When Many Units Fail Together, Suspecting Recent Changes First, Holding a Capacity Floor and Reintroducing Capacity Gradually, without Ever Blocking a Safety Stop", Technical Disclosure Commons, (November 16, 2026)
https://www.tdcommons.org/dpubs_series/13113