Abstract
This publication discloses, enables, and dates a mechanism for behavioral anomaly gating of autonomous software agents. The distinguishing subject is not the human user but the agent-channel capability exercise: an event in which a principal (a human or a non-human agent) invokes a named capability through a specific channel (an interactive UI, an email-driven command, a messaging bridge, a scheduled job). Six read-only behavioral signals are computed for each such event, each under a fail-LOW contract whereby any datastore error or absent history yields a zero sub-score, so a monitoring blind spot can never fabricate an anomaly. A pure, side-effect-free weighted scorer normalizes the signals such that an absent-but-weighted signal only ever lowers the aggregate. A four-rung action ladder (allow / step-up / soft-block / hard-block) is derived from the score through thresholds that are clamped and sorted at the moment of evaluation, making an operator misconfiguration incapable of inverting the ladder. In its default score mode, the engine records the counterfactual would-block verdict together with an enforced=false flag while permitting the request — so production traffic becomes the calibration corpus for the signal weights without any legitimate action ever being denied. Engine faults fail open to scoring; the composite gate chain is ordered cheapest-deny-first. A dependency-free, offline-runnable reference implementation accompanies the disclosure. All empirically-tuned constants are marked [WITHHELD — trade secret].
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Shadow-First Anomaly Ladder with Fail-Low Signals, Inversion-Proof Thresholds, and Would-Block Dual Recording", Technical Disclosure Commons, (July 13, 2026)
https://www.tdcommons.org/dpubs_series/10872