← Back to Trident Build homepage

Trident Build · Research

Research overview

Trident Build is investigating environmental modeling, coastal water-level analysis, and potential future stormwater and infrastructure monitoring applications. The work below documents a computational sensitivity experiment, not a deployed monitoring system.

Completed research

Experiment 3 — source-data preparation. The Experiment 4 notebook identifies a NOAA measured-water-level CSV supplied from Experiment 3. The earlier experiment’s full methodology and conclusions have not yet been independently reviewed for this page.

Experiment 4 — held-out water-level anomaly simulation. The detailed record follows.

This simulation-based study is not evidence of a field deployment, quantum hardware experiment, or verified operational detection capability.

Experiment 4: Why better sensors alone won't fix water-level anomaly detection

We tested whether a simple tidal model could flag 22 cm water-level anomalies in real NOAA data. It couldn't, and halving sensor noise didn't help. The limiting factor is the environmental model, not the sensor.

Conducted October 31, 2026 · Data: NOAA CO-OPS, St. Petersburg, FL (station 8726520), October 16–29, 2026 · Fully reproducible notebook
16.5 cm
Average baseline error on held-out hours
39.6 cm
Alarm threshold needed to suppress natural false alarms
0.5%
Event hours detected at conventional noise (0% at half the noise)
4 of 101
Held-out hours that crossed the threshold with no event at all

What we set out to test

Stormwater and coastal monitoring systems need to tell a real event, such as a blockage, surge or discharge, apart from normal tidal and weather variation. A common approach is to predict the expected water level and raise an alarm when the measurement departs from that prediction by more than a threshold.

Experiment 4 asked two questions. First, how well does a simple tidal model predict water levels it has never seen? Second, if we inject known anomalies, does lower sensor noise make them easier to detect?

How we tested it

The baseline misses a lot of real variation

Measured water level vs. train-only tidal prediction (cm)

NOAA measuredModel predictionHeld-out period
The prediction tracks the tidal rhythm but misses swings of 20–45 cm, especially in the held-out period, when weather and surge effects dominated.

Held-out residual: measured minus predicted (cm)

ResidualAlarm threshold (±39.6 cm)
With no event injected at all, the natural error crossed the alarm threshold in 4 of 101 hours. To avoid constant false alarms, the threshold had to be set well above the 22 cm events we wanted to catch.

Detection results

Scenario (500 trials each)PrecisionRecallFalse-alarm rate
Conventional sensor noise (10 cm)1.5%0.5%4.7%
Lower sensor noise (5 cm)0.0%0.0%4.6%

Averages across trials. Recall is the share of the 14 event hours that were flagged; the false-alarm rate is the share of the 87 non-event hours that were flagged.

What this means. Reducing sensor noise did not improve detection, because sensor noise was never the bottleneck: the error in the environmental model was. A 22 cm event is invisible when the model is off by about 16 cm on average and by up to 45 cm. Any claim that more precise sensors improve anomaly detection must first show a baseline accurate enough for that precision to matter.

Research directions (proposed, not completed)

Potential follow-up work includes field-grade ultrasonic sensing integration, stormwater and drainage observations, independent event labels, and comparison of conventional and advanced computational approaches. Any Amazon Braket quantum simulation or hardware results will be described separately only after execution records and findings are verified and publication is approved.

What comes next

Limits of this experiment

This test used coastal water-level data, not stormwater flow. The anomalies, sensor channels and noise levels were simulated, and the four channels are replicas, not separate stations. No quantum hardware or advanced clock was used, and no claim of quantum advantage is made or implied.

Reproducibility

The analysis runs in a single notebook using a fixed random seed (20261031). The notebook, the hourly baseline predictions and all 1,000 trial results are available on request.

Data table: hourly measured and predicted water levels

© 2026 Trident Build Inc. NOAA water-level data courtesy of NOAA CO-OPS.