Skip to main content

The models behind the field work

The field record behind the models.

Trained models estimate crop stress and the hours each block’s pump ran, from the probes, weather and pump records. Your agronomist reviews the evidence and makes the irrigation call.

The signing method is patent-pending.

IllustrationA record from the same orchard blockStem water potential (SWP) readings are matched to the same block's soil probes and weather. Those field records support crop-stress estimates reviewed by an adviser. Probe summaries are signed separately from the irrigation order.THE SAME ORCHARD BLOCKstem water potential(SWP) readingsSoil probesWeatherCrop-stress estimatesYour adviser makes the callSigned probe summaries keep the record
Signed daily probe record. The adviser makes the call; the order itself is not signed.

What runs every week

Estimates, a draft, then the adviser’s call.

Three trained models run every week. The Thursday draft water order comes from rules mined from our agronomist’s past orders for each block; it does not read that week’s weather, sensor data or any model.

01 · SENTRY / WATCHDOG

Which blocks need a closer look?

Two trained models estimate crop stress. Your adviser weighs the estimates against local stem water potential (SWP) readings. The stress models have not beaten the simple rules a grower already has, such as the block’s latest SWP reading.

02 · PULSE

How many hours did each pump run?

Probes, weather and pump records support an estimate of the hours each block’s pump ran. A model output is not a flow-meter reading.

03 · PRESCRIBE

What should the irrigator run?

The Thursday draft comes from rules mined from past orders. Trained versions are being tested.

04 · ADVISER REVIEW

The irrigation call stays with an adviser.

On Platform, your own Certified Crop Adviser or adviser makes the call. On Managed, our Certified Crop Adviser does.

What each model does and where it is used

SENTRY: which blocks need a closer look?

SENTRY is a trained classification model that estimates crop stress, trained on our own stem water potential (SWP) readings matched to soil probes and weather. It is meant to help the agronomist choose where to check between visits. It needs local SWP readings and does not replace your SWP readings.

In the decision path

PULSE: how many hours did each pump run?

PULSE is a trained regression model that estimates the hours each block’s pump ran, from probes, weather and pump records. An estimate stays an estimate; a model output is not a flow-meter reading.

In the decision path. A new version of the pump-hour model has been running since November 11, 2026. Its evidence so far is on past data. A check on fresh data can be read no earlier than December 9, 2026, and we publish no accuracy figure for it until then.

WATCHDOG: is a block without SWP readings under stress?

WATCHDOG is a trained classification model that estimates tree water stress from soil, weather and rootstock data, with no SWP readings among its inputs. We use it on blocks without recent SWP readings. Your adviser’s readings remain the measurement of record. It is not a sensor-fault detector: separate rule-based sensor health checks flag probes that stop reporting or return error values.

In the decision path

PRESCRIBE: what should the irrigator run?

The Thursday draft water order comes from rules mined from our agronomist’s past orders for each block; it does not read that week’s weather, sensor data or any model. On Platform, your own Certified Crop Adviser or adviser makes the call. On Managed, our Certified Crop Adviser does. We test trained models against the current rules before using them to draft orders.

Rules mined from past orders, not a trained model. Trained versions are being tested.

SALT FLUSH: will an irrigation move salt below the roots?

We are testing a trained model for salt flushing. It does not make field recommendations, and we have not published an accuracy figure for it.

In testing, not in the weekly run. No field recommendations.

How the models fit Platform and Managed

Three trained models run every week: two estimate crop stress, trained on our own stem water potential (SWP) readings matched to soil probes and weather, and one estimates the hours each block’s pump ran, from probes, weather and pump records. The stress models have not beaten the simple rules a grower already has, such as the block’s latest SWP reading. The Thursday draft water order comes from rules mined from our agronomist’s past orders for each block; it does not read that week’s weather, sensor data or any model. On Platform, your own Certified Crop Adviser or adviser makes the call. On Managed, our Certified Crop Adviser does.

Trained models estimate crop stress and the hours each block’s pump ran, from the probes, weather and pump records. Your adviser reviews those estimates alongside the farm’s own SWP readings, each judged against its block’s stress line. Separate alerts flag a missed set or a block that is dry and overdue.

Use the sensors and crew you already have with Platform, after we confirm the sensor connection. AgWaterAI adds crop-stress estimates, signed probe reading summaries and, where our agronomist’s past orders for a block are loaded into the draft’s rules, a draft water order. With Managed, we install and own the sensors and run the field program.

New in November 2026

A soil projection, a new water model and a pump switch check.

Three additions since November 11, 2026, each with the test behind it and what it does not show.

Trained model · on farm dashboards

Where will this probe’s soil moisture be in three days?

A trained model projects each probe’s soil moisture three days ahead if no irrigation or rain occurs. On farm dashboards since November 11, 2026. On a second farm never used in fitting, for probes with their own history, its three-day projection in periods when the probe saw no wetting was closer than every plain rule we tested, in the test plan and after it: its miss was about seven tenths of the best plain rule’s. For a brand-new probe with no history (tested on established probes shown none of their history) a plain rule does as well or better.

Trained model · under a fresh-data check

A new water model is running.

A new version of the pump-hour model has been running since November 11, 2026. Its evidence so far is on past data. A check on fresh data can be read no earlier than December 9, 2026, and we publish no accuracy figure for it until then.

A rule, not a trained model

Is this pump switch logging?

When the soil wets on dry days and the pump switch records no run, the dashboard asks whether the switch is logging. It is a plain rule, with no model involved. On the dashboard of our longest-running orchard since November 11, 2026.

How the soil moisture projection was tested

On a second farm never used in fitting, for probes with their own history, its three-day projection in periods when the probe saw no wetting was closer than every plain rule we tested, in the test plan and after it: its miss was about seven tenths of the best plain rule’s. For a brand-new probe with no history (tested on established probes shown none of their history) a plain rule does as well or better.

Only periods when the probe saw no wetting, on one farm and one probe make, and not a fresh blind test: that ground had been scored before. With that farm’s locations weighted equally, the model’s lead over the best plain rule is not clear. With irrigated periods removed the lead held; most remaining periods could not be checked, and on checked dry ones it is unclear. On the checked dry periods the lead is in the top 60 centimeters; deeper down, a plain rule does as well. Its printed range can be too narrow on ground it was not fitted on. The first forward score is November 23, 2026. It is a projection, not a call to irrigate.

It was fitted on the probes of one estate only, with its test plan fixed before training. On ranches held out of training its three-day projection was closer than each plain rule in that plan. On a second farm never used in fitting, in periods when the probe saw no wetting, it was closer than every plain rule we tested when it had the probe’s history. For a brand-new probe with no history (tested on established probes shown none of their history) a plain rule does as well or better. Methods, plans and figures are available in diligence.

What the new water model has and has not shown

A new version of the pump-hour model has been running since November 11, 2026. Its evidence so far is on past data. A check on fresh data can be read no earlier than December 9, 2026, and we publish no accuracy figure for it until then.

The water model figures on this page are those of the last version that passed its check. It estimates pump hours from the block’s own pump record, not a metered volume, and it has not been shown on a new farm.

How the pump switch check was tested

In a check registered before scoring, it caught about four in five silent switch months we planted in that orchard’s record, usually within a week or two, with few unexplained questions. On other loggers we hold it read about the same in a first look that is too small to call proof.

A question about the hardware, never a finding. The silent periods were planted; no real silent switch has yet been confirmed by a person. It needs a run record for each block and a rain record, and it is weaker outside the irrigation season.

In research, not in the product

In research, on past data and provisional, not in the product: a trained model for checking run records. In front of a crew, the plain rule is what runs.

Your adviser’s readings remain the measurement of record. We do not claim water savings from our models. Every figure is labeled measured or estimated.

Research comparison, estate data

The labels are the asset.

Our models draw on years of SWP readings matched to the same blocks’ soil probes.

12,000+

stem water potential (SWP) readings through Oct 6, 2026.

4,000+

of them on blocks with their own soil probe logging that day (through Oct 6, 2026).

2016

Daily soil records since 2016.

Production test results, by test method

We report each model's test method, evaluation data and limits separately. Holding out whole blocks, or whole site groups, is a harder test than scoring later weeks of the same blocks. For the water model the held-out block still supplies its own pump history, so it is not a cold start. The stress figures are visit-day nowcast metrics: same-day inputs scored against that day's SWP reading, never weekly forecasting accuracy. Neither predicts results for a new customer. AUC describes classification; R² describes a regression fit.

Production registry metrics
ModelIn-season or walk-forward testHeld-out test
PULSER² 0.7375 in-seasonMedian R² 0.7324 on held-out blocks on the same estate, each keeping its own pump history; served seed 7, the highest of five seeds; five-seed median 0.7292
SENTRYRaw AUC 0.9226 walk-forwardSite median AUC 0.9287; pooled 0.9153 (visit-day nowcast metrics with same-day inputs, not weekly forecasting)
WATCHDOGAUC 0.8038 walk-forward (median of 11 seeded runs of the served recipe, Oct 10 2026; visit-day nowcast)Held-out site groups mean AUC 0.7873; pooled 0.6993 (unweighted mean of three folds; Sep 22 2026 artifact run)

Metrics come from the production model registry. The water model figures on this page are those of the last version that passed its check. A new version of the pump-hour model has been running since November 11, 2026.

Tested against the trees, October 2026

Can a language model call crop stress? We tested it against the trees.

Three ways to call stress on a block, scored against the adviser's own SWP reading on 1,790 held out block weeks from one orchard.

One orchard, three seasons, scored within each week; pooled across weeks, Claude and Gemini handed the same readings did better than the record. Nothing here claims water savings. The adviser's SWP reading remains the instrument of record, and visits do not change on the strength of a score.

  1. The block's own reading record, a rule0.845
  2. Claude or Gemini, handed the same readings0.81
  3. Claude or Gemini with no field readings0.55
The comparison starts at 0.5. These are research results from one orchard.
Research comparison: design and scope

1.0 is perfect, 0.5 is a coin flip. The two frontier language models, Claude Opus 5.5 and Gemini 4 Argon, ran inside our own cloud project and no farm data left it. The design was registered before the first call.

Without the readings, the frontier models are a coin flip. Handed the same readings, they scored below the top bar within the week and above it pooled across weeks. The top bar is the block's own reading record: a rule on the block's SWP readings, not a trained model.

Method note, October and November 2026

How we tested these results.

On a new farm, your adviser weighs the model estimates against local readings. SWP readings, when available, are judged against the block's stress line.

Research result, not a customer promise.

A soil probe on a farm the model has never seen

The test. A test with its plan fixed before training: the projection was fitted on the probes of one estate and scored on ranches held out of training and on a second farm never used in fitting.

The frame. Only periods when the probe saw no wetting. Each projection starts from a probe’s reading and says where its soil moisture will be three days later if no irrigation or rain occurs; the score is the miss against what the probe then read.

The comparison. The comparisons are plain rules. On the second farm, for probes with their own history, the projection was closer three days ahead than every one of them, in the test plan and after it; its miss was about seven tenths of the best plain rule’s. For a brand-new probe with no history (established probes shown none of their history) a plain rule does as well or better. The test plan’s comparisons were scored a second time with new code, from the same tables, and agreed. Methods, plans and figures are available in diligence.

The date. Scored November 11, 2026, and again against more plain rules on November 12, 2026, on one farm and one probe make. Not a fresh blind test: that ground had been scored before. With that farm’s locations weighted equally, the model’s lead over the best plain rule is not clear. With irrigated periods removed the lead held; most remaining periods could not be checked, and on checked dry ones it is unclear. On the checked dry periods the lead is in the top 60 centimeters; deeper down, a plain rule does as well. Its printed range can be too narrow on ground it was not fitted on. The first forward score is November 23, 2026.

Your adviser’s readings remain the measurement of record.

The Monday list against a fixed rotation and a hand sort

The test. A pre-registered replay of past seasons, week by week. A fixed rotation reads every block in turn; the Monday list puts first the blocks most likely to be under their stress line. A registered re-test of November 12, 2026 added the orders a farm could follow without software: a careful crew sorting last week’s SWP readings by hand, and smarter rotations that ignore the readings. Every order gets the same number of readings each week.

The frame. One citrus estate, our longest-running orchard, the 2023 to 2025 seasons pooled. A stressed block-week is a week where the block was at or below its stress line.

The comparison. Each order gets the same shorter round. A stressed block counts as reached when it is read in the week it is stressed or early the next. On that count the list reached about 11 percent more stressed block-weeks per reading than the fixed rotation, somewhat more than the strongest rotation that ignores the readings, and was level with the careful hand sort. Within seven days no difference was shown, and a plain rotation catches at least as many. Of 102 deep-stress block-weeks in the replay, the shorter round would have left about 20 unread in their week in a fixed rotation and 3 with the list. A careful crew sorting last week’s SWP readings by hand would have left about as many as the list. Smarter rotations that ignore the readings closed most of the gap. The crew read every one of these blocks in its week, nearly all by midweek: the count exists on the shorter round alone, and at the full round every order leaves none. Half of the 102 sit on three blocks. Counts this small do not separate these orders. These figures are the replay’s. The list in service sets the stress line another way, so fewer block-weeks count as stressed and too few as severely stressed to compare orders. Scored its own way, it again reached more stressed blocks in their week than a fixed rotation and was not separated from a careful hand sort or from the strongest rotation that ignores the readings. Scored on the block-weeks the replay counts as stressed, its lead over a fixed rotation was smaller. There it reached fewer stressed blocks than the order as replayed. The replay’s counts were scored a second time with new code, from the same tables, and agreed. Methods, plans and figures are available in diligence.

The date. First run October 19, 2026; rerun November 10 and 11, 2026; scored November 12, 2026 against a careful hand sort and rotations that ignore the readings. A development result, not a field trial. The number of SWP visits stays the same; the list says where to start.

Where it comes from. This result comes from one citrus estate, the one the list was developed on.

A second farm, November 12, 2026. On November 12, 2026 the list was replayed, unchanged, on a second farm’s 2026 SWP readings: one season and 25 reading sets, on a shorter round, with the plan fixed before scoring. Within seven days no difference was shown. The plan’s word for this replay is no verdict, for two reasons fixed before scoring: one season held too few severely stressed block-weeks to test, and the plan’s check on sudden onsets did not pass (the list was not shown to catch more of them than a rotation). It is a development replay, too small to settle. Against a careful hand sort of last week’s readings the list was not separated there either. The count of severely stressed block-weeks has not yet been shown on another farm.

What it does and cannot do. The list works from the stress already seen in each block’s readings. It cannot see a block that turns suddenly in a week it was not read.

Deep-stress block-weeks a shorter round would have left unread in their week: about 20 with a fixed rotation and 3 with the Monday list, of 102, at the same number of SWP readings. A careful crew sorting last week’s SWP readings by hand would have left about as many as the list.

See the Monday list case study

What is signed and what a signature shows

Probe readings are sealed into a signed, tamper-evident chain once each logger day closes, on the sites where daily sealing runs today. Well-meter readings that reach us through telemetry are sealed into a signed, tamper-evident chain as they arrive, checked every 10 minutes. Today that chain is separate from the probe chain, and a meter outside the telemetry feed is not sealed. Each issued water order is kept with the readings behind it; the order itself is not signed. A signature shows a record has not changed since signing; it does not prove a measurement was correct.

The signing method is patent-pending.

The measurement of record

Your adviser’s readings remain the measurement of record.

The signed record

A signature shows a record has not changed since signing.

It does not prove a measurement was correct. The order itself is not signed.

From a field visit to a decision

See the evidence reach the adviser.

Your agronomist reviews the evidence and makes the irrigation call.