HydroJEV tests jev-1.13.0 as a one-second screen on simulated water-network alarms

arXiv:2610.02048, submitted October 1, 2026, asks whether Jev can take a first pass at a water-network alarm. The HTML we opened calls the pipeline HydroJEV. The model on every call is jev-1.13.0. The license line is CC BY 4.0.

Hekate’s October 2 post points at the result and says the evaluations are simulated.

What was simulated

Each window is 6 hours from a stochastic EPANET run. The main network is C-Town. Four classes are a concealed cyberattack, a physical fault, a normal transient, and a sensor fault. A fifth answer, insufficient evidence, is offered to the model and then dropped from the calibrated decision.

The question and a label-free prior were fixed on 40 development windows before the sealed rounds. The prior the paper prints is 0.302, 0.429, 0.146, and 0.121 for attack, physical fault, normal transient, and sensor fault.

Four sealed rounds on C-Town cover 1,279 windows in the study-design figure. Later rounds move the same frozen screen to EPANET Net3 and Net1. A hand-written rule tree, R1, needs no labels. A multinomial logistic regression, R2, trains on 200 labelled windows of the reference subtypes.

Seven cloud models available through Ollama saw the same text. The primary reviewer in the cascade is glm-5.2. We did not find a code repository in the sections we read.

What jev-1.13.0 scored

On the four in-distribution sealed sets, calibrated Jev’s macro-F1 is 0.620, 0.628, 0.628, and 0.644. The paper’s mean is 0.63. R1 on those sets is 0.557, 0.606, 0.602, and 0.590, mean 0.59. Jev is numerically ahead on each of them. No single difference against R1 was significant.

Across all eight C-Town sealed sets the paired difference averages +0.03.

R2 trained on all 200 labels is 0.887 to 0.934 in distribution, mean 0.91. The paper says that gap was pre-registered as expected. When R2 is retrained without the subtype being scored, it falls near chance, and Jev’s advantage is +0.361, +0.387, +0.390, and +0.417, all four surviving Holm correction.

On the four new-family sets, Jev and R1 are close, mean 0.51 against 0.50, and fully supervised R2 stays ahead at mean 0.69. The label-scarcity curve puts R2’s mean above Jev once about four labels per class are available, on seven of the eight C-Town sets.

Service-reported compute time on 2,722 Jev calls before round 4 has a median of 1.215 seconds. The 10th and 90th percentiles are 1.151 and 1.458 seconds. Client wall time has a median of 1.463 seconds. Sent one request at a time, glm-5.2 takes a median 30.6 to 31.7 seconds and deepseek-v4-pro 39.3 to 46.4 seconds.

The paper calls that 20 to 40 times the Jev time. On the round-2 sets, Jev scores 0.628 and 0.434. deepseek-v4-pro scores 0.722 and 0.637. glm-5.2 scores 0.706 and 0.697. Four of the fourteen two-sided tests favor an LLM after Holm correction, and none favor Jev.

What the gate keeps

The cascade accepts a Jev verdict on its own only when that verdict is benign and R1 agrees. Benign means normal transient or sensor fault. Round 4 was generated after that rule was frozen.

The glm-5.2 cascade scores 0.745 on the in-distribution set E1c and 0.565 on the look-alike set E4c, against 0.735 and 0.549 for glm-5.2 alone. Both non-inferiority tests pass.

The gate settles 69 of 200 windows on E1c, accept precision 0.841, and 76 of 200 on E4c, accept precision 0.579. The abstract’s 35% to 38% is that pair. On E4c, 32 of the 76 accepted windows are wrong: 12 masked duty swaps, 10 pump trips, 9 level-spoof attacks, and 1 growing leak.

Among 50 true attacks, glm-5.2 alone marks 20 as non-benign and the cascade marks 15. The ungated rule, which trusts every benign Jev verdict, met none of its superiority tests. On E1c it dropped burst recall from R1’s 0.78 to 0.52, because Jev called 13 of 50 bursts benign.

The paper’s own discussion says a fast screen can take about a third of the review load while the reviewer’s macro-F1 holds, and that LLM review remains necessary for concealed attacks. The windows are simulated. Nothing in the sections we read is a utility’s historian.