The models behind the field work
The labeled corpus that turns permanent-crop sensors into a defensible water decision every week.
We use trained models on our own data to help your adviser decide what to check. Local chamber history is required; the models cover the days between visits.
Use the sensors and crew you already have with BYOS, after we confirm the sensor connection. AgWaterAI adds the intelligence and the record. With Managed, we install and own the sensors and run the field program.
Your adviser reviews the evidence and makes the irrigation call. The signed record shows what was recorded and whether it changed after signing; it does not prove a measurement was correct.
SENTRY · v52
SENTRY classifies plant water stress using soil data and pressure-chamber history. It helps the adviser choose where to check between visits. It needs local chamber readings and does not replace the chamber.
In the decision path
PULSE · v17.4
PULSE predicts applied irrigation from the block’s sensor and operating history. An estimate stays an estimate; a model output is not a flow-meter reading. A farm-only competitor matched our corpus after a season in the estate comparison.
In the decision path
WATCHDOG · v20
WATCHDOG is a trained classification model for sensor anomalies. A flag gives the adviser a reason to inspect the reading before using it. Its registered tests use chamber-derived labels; these tests do not establish accuracy against independently confirmed probe or pump faults.
In the decision path
PRESCRIBE · v3 mined rules
PRESCRIBE prepares a weekly irrigation order using rules mined from historical irrigation patterns on adviser-managed blocks. The adviser reviews it. Trained candidates run in shadow and must beat the rules before they can take that role.
Mined rules in service; trained versions in shadow
SENTRY-TRANSITION · Stage 1 physics
The transition planner uses a physics water balance where the required inputs are present. This is planning context for review. It has no published learned-forecast accuracy.
Physics planning alpha
SALT FLUSH · v5
SALT FLUSH is trained but gated. An event-conditional candidate is being tested in shadow. We publish no standalone accuracy claim and do not present the candidate as a field recommendation.
Gated; candidate in shadow
Testing forward in time asks how a model handles later readings. Holding out a site or soil texture asks a harder transfer question. Neither is a guarantee for a new customer. AUC describes classification; R² describes a regression fit.
| Model | Temporal test | Held-out test |
|---|---|---|
| PULSE | R² 0.7204 in-season / 0.7171 held-out site | Site median R² 0.7171 |
| SENTRY | Raw AUC 0.8997 walk-forward, leave-one-site-out 0.9083 median / 0.9004 pooled | Site median AUC 0.9083; pooled 0.9004 |
| WATCHDOG | AUC 0.7773 walk-forward, leave-one-texture-out 0.7983 mean / 0.7060 pooled | Leave-one-texture-out (LOTO) mean AUC 0.7983; pooled 0.7060 |
Research comparison, estate data
6,097 dated pressure-chamber readings collected; 3,844 in the joined stress matrix. We compared broader training with a model using only the farm’s own history.
The corpus gives a new farm a head start on stress calls, and the farm’s own chamber history takes over. Research result, not a customer promise. The stress head does not transfer without local chamber history.
The estate advantage is not proof that every block improves.
The depletion forecast was compared with the farm-only competitor. “Ahead through six weeks on the independent farm, about a season on the estate.”
Transfers with zero labels; a supporting signal, not decision-grade alone.
The research and its qualifiers are available in the investor data room. Independent-farm durability still needs more labels. Research date: September 3 to 4, 2026.
Your record is signed when your adviser approves the order. Signing establishes the record from attestation forward, not the accuracy of the underlying measurement.
Experiment, not a visit schedule
In the cadence experiment, a candidate raised more false alarms.
At the same false-alarm rate, this candidate caught fewer stressed readings than the last chamber reading. The candidate failed that gate. We withheld the cadence candidate from promotion.
This does not establish a reduction in field labor. Shadow results and adviser review must come before a change to the chamber schedule.
Research source: September 5 data room, Model Transfer Tests. Registered versions and metrics come from the production model registry.
See how a field visit becomes a decision