VX-One: Typed, Calibrated Decision Heads for Real-Time Perimeter Triage - An empirical study of unknown-attack honesty, cascade gray zones, and calibration under drift

VX-One. Typed, Calibrated Decision Heads for Real-Time Perimeter Triage - An empirical study of unknown-attack honesty, cascade gray zones, and calibration under drift. By Juan S. Santillana, working independently. Published on the twenty-first of October. Its code repository is worked out: about fifty source files, with a readme.

Perimeter triage, deciding quickly what to do with incoming network traffic, faces a trade between speed and meaning. Signature engines are fast and safe in their output format, but blind to attacks they have not seen. Generative models understand semantics, but they are slow and can produce output that cannot be parsed. In late two thousand twenty-six, the abstract notes, a class of commercial System One decision models, TypeSafe Jev and Microsoft-Decision-1, began selling typed decision primitives as product APIs: binary gates, closed-set choices and ordinal scores, with no academic treatment. VX-One is a small multi-head network, between one hundred thousand and ten million parameters, whose outputs are exactly those primitives. A calibrated gate. A choice over a closed set of candidates that always includes an explicit unknown option. And an ordinal score. It is trained multi-task and temperature-calibrated per head, running in roughly one to two milliseconds at the median on a CPU. The study's claim is a measured account of where such typed, calibrated decisions help perimeter triage, and where they fail.

The evaluation uses UNSW-NB15, CIC-IDS2017 and EMBER2024, against calibrated XGBoost, Random-Forest, MLP and Kitsune. On detection, VX-One wins on UNSW-NB15 and under temporal drift on CIC-IDS2017, with zero point seven one one against zero point six three seven. XGBoost, though, wins multi-class attribution on all three benchmarks.

For unknown attacks, a leave-one-family-out gate detects never-seen families on all nine UNSW folds, yet falls below chance on the CIC recon family, and the abstention threshold transfers poorly between benchmarks.

A two-phase cascade was also measured over five days of CIC-IDS2017: Suricata, with fifty-two thousand six hundred and seventy-four Emerging-Threats rules, followed by VX-One. On payload-less reconstructions the signature engine resolves about zero point one percent of flows. That leaves a gray zone of about ninety-nine point nine percent of traffic, which Phase two recovers, with combined recall of at least zero point nine seven eight. Finally, calibration collapses under temporal drift for every method, with expected calibration error rising from zero point zero four to zero point four, while ranking partially survives.

The study is bounded by its three benchmarks and its four baselines, and the cascade by five days of one dataset, using payload-less reconstructions. The results cut both ways by design. VX-One leads on detection, not attribution. Unknown-family detection works on one benchmark and fails on a family in another. And the calibration collapse under drift is not specific to VX-One: it hits every method tested, so calibration measured before drift should not be trusted after it.

What the paper offers is less a winner than a map. Typed, calibrated heads make the cascade practical, since the signature stage leaves almost all traffic unresolved. The honest unknown option generalizes unevenly, and no calibration in this study survives the passage of time.