VERITY ICS · release 4.7.2

Plants, pumps and reactors. Measured against their own normal.

VERITY ICS is an instrument for industrial control systems. It commissions on each plant’s published normal data, then watches sensors and actuators live, on a CPU. Labels are used only for grading.

0.888
On the HAI 21.03 contest scorer. The winner, with labels, scored 0.938
90.5%
Mean detection across 17 observable Tennessee Eastman faults
0
False alarms on 4,160 fault-free Tennessee Eastman samples
84of 122
Sealed injected faults caught across HAI 21.03 and 22.04, internal blind test
Water treatment plant at night, round clarifier tanks and steel walkways under low light
Water, power, chemical

Every plant has its own normal. VERITY commissions on it before it watches.

01
Public plant benchmarks

Four plants. The hard one too.

Each plant is commissioned on its own published normal files. Labels are used only to grade the result. Simulated and hardware-in-the-loop plants only, no field deployment yet.

HAI 21.03 · power generation testbed

Close to the contest winners, without their labels.

0.888on the HAICon 2020 scorer, at the default setting
Contest winner
0.938
VERITY ICS
0.888
Contest baseline
0.842

Contest entrants had labeled validation data. 46 of 50 labeled attacks raise an alarm.

Tennessee Eastman · chemical process

Faults found. Zero false alarms on clean runs.

90.5%mean detection across 17 observable faults

0 false alarms on 4,160 fault-free samples. Published methods reach 82.9% to 96.2% mean detection on different simulation runs of the same process, so the comparison is indicative only.

BATADAL · water distribution

Attacks on a city water network.

0.79event F1 on the second test file. The first reads 0.77.
Test file 1
0.77
Test file 2
0.79

Not directly comparable to published scores, which use a different data release.

HAI 22.04 · stealth attacks

Where we are behind.

0.344on the HAICon 2021 scorer. First place scored 0.685.
First place
0.685
VERITY ICS
0.344

This dataset release is built to hide attacks inside normal-looking ranges. 34 of 58 attacks raise an alarm. We publish it because a buyer should see the hard case too.

02
Sealed fault tests

Faults it had never seen.

Faults were injected and sealed before VERITY saw the data, then unsealed for grading on October 14, 2026. Internal blind evaluation, not a third-party audit.

84of 122
Injected faults caught across both HAI plants
52of 70
HAI 21.03
32of 52
HAI 22.04

Univariate time series

On the TSB-AD-U benchmark of 350 series, VERITY scores 0.6061 VUS-PR against 0.5856 for the next best published method. Pending leaderboard confirmation.

Deployment

Commissioned on the plant’s own normal.

No attack labels. No GPU. No cloud connection. VERITY ICS commissions on a plant’s normal operating data and runs on an ordinary CPU beside the control network.

0labels
Attack labels used to commission or to detect
0GPUs
CPU only. Runs on premises
4plants
Power, stealth, water and chemical public benchmarks
0cloud
No data leaves the site
Stated limits

What we tell every buyer up front.

An instrument is only credible if it is clear about where it stops. These travel with every number on this page.

False alarms

On three of five industrial plants, false alarms run above the design target.

In-range shifts

Level shifts that stay inside a sensor’s normal range are not detected.

Field deployment

Results come from simulated and hardware-in-the-loop plants. There is no field deployment yet.

Stealth attacks

On HAI 22.04, built to hide attacks inside normal ranges, VERITY ICS trails first place, 0.344 against 0.685.

Comparisons

BATADAL and Tennessee Eastman scores are not directly comparable to published work, which uses different data releases and runs.

Sealed tests

The sealed fault tests were graded inside Credasis. They are not a third-party audit.

See it on your plant.

Bring historian data from normal operation. We commission on it and show you what VERITY ICS flags. You keep the data.

Industrial figures: VERITY ICS release 4.7.2, external cybersecurity results card, Part 2. Public datasets: HAI 21.03, HAI 22.04, BATADAL and Tennessee Eastman. TSB-AD-U figure: external card, Part 3, pending leaderboard confirmation.