The soil-moisture advisory's backtests, 2026-10-11 to 2026-10-19#

The record of what the soil-moisture advisory was tried on, on the virtual farm as each day had it, and what each try showed: the tables that once stood in its README, kept here as they were last written. The README holds what the advisory does now and the evidence for it as it is now, from trials.py; this keeps how it got there.

Read each table with its date. The simulator changed under them: on 2026-10-15 its soil began to move water below field capacity; on 2026-10-16 its crop began to use its water as FAO-56's does, from the top of its roots, and its bare soil to dry; on 2026-10-17 its crop began to slow by the water near its roots; on 2026-10-18 its roots began to make up for each other, its crop to run short sooner on a hot day, and the backtests to water probe-led beds at the weather beds' rate, with forecasts as they stood (the audit). And the advice changed with what they showed. Since retired, and so found here only: Bed.adjust_p (FAO-56's lower refill point on a hot day), the root signs' stair-step, learning field capacity by the median of the drain-downs or from soakings alone, and the simulator's soil before 2026-10-15.

The backtest bed, 2013#

§11 asks for it before anything goes live. On McAllister Farm's recorded summer of 2013 (June to August), a 30 cm root zone in a silt loam, the drill's forecasts in the log, the advisory followed to the letter, beside the drip timer most farms run (sim/irrigation_drill.py). The backtests are backtest.py, backtest with probes and weather_backtest by the weather, kept apart from the advisory, which never imports the simulator:

62% less water and 59% less drained, with no hour of real stress. The cost is a drier root zone on average — mean stress 0.12 against 0.06 — which is the trade a refill point makes, and the reason p is the crop's to set. (On the simulator of 2026-10-16, below; on the earlier one the advisory used 156 mm. The drought of 2012 took the timer to 51 hours short; the advisory, on 319 mm, to none.)

Each simulated probe draws its own noise (noise_streams, since 2026-10-14). Before, all of a backtest's probes drew from one stream, so a run with a probe more or less, or with one probe silent for a while, read different noise on the probes it shared with another, and two runs set side by side differed by more than the probe. Now each draws from its own; the first keeps the seed's, so a run with one probe reads exactly as before, to the last bit. The tables here with more than one probe were made with the shared stream; those re-measured on the new streams say so.

What that showed: hours short are noisy. The water a bed was given moved by a few per cent at most; the hours short moved by several times in some rows, from nothing but which draws of noise each probe read (the code before the change gave the old numbers again, exactly, for the two rows tried). Over five draws of it, tomatoes led by probes at 10 and 30 cm were short in 1 or 2 summers of thirty, for 16 to 45 hours; on the one draw the tables were first made with, 3 for 178. So a difference of a few summers or a few dozen hours between two rows can be the noise's, and the comparisons below that rested on one have been said again. The large ones stood: the pan found by the root signs, the crop known against not, a probe among the young plants. All of that was on the simulator's earlier soil, which moved no water below field capacity. The tables below are the simulator of 2026-10-16, three draws each: its soil moves water below field capacity (below), and its crop uses water as FAO-56's does, draws it most from the top, and leaves bare soil to dry in the sun (below). Several results moved a long way with it, and the advice was changed for what they showed.

On Harrow Bend, eight days, six beds: the beds are told apart on the first morning; the far bed node's radio outage turns its beds' advice into "unseen" and no one else's; and a probe lifted out of bed 6 into bed 2 leaves bed 6 with "no probe" and keeps bed 2 advised through its own probe's outage.

pip install -r modules/soilwater/requirements.txt      # cel-python, via modules/rules
cd modules/soilwater && python3 -m unittest discover -p 'test_*.py'

advisor = soilwater.Advisor(log)
beds = [soilwater.Bed(land_id, field_capacity=0.30, wilting_point=0.12),
        soilwater.Bed(yard_id, 0.38, 0.25, layers=((0.10, 0.20, 0.09),),
                      root_depth_mm=150, surface="turf", planting="turf")]
advisor.advise(state, beds, at=morning, forecast=rain_forecast_id)
soilwater.render(event.payload, "es", state)   # "Riegue cama 3 hoy: …"

The bed's own soil#

A soil survey maps the native soil. A landscaped yard, a raised bed or a market garden on an old house lot is often something else:

• topsoil brought in over a subsoil the builders compacted;
• sod with its own soil over the yard's;
• a compost mix over the ground.

Set up with the survey's values, the advisory misjudges such a bed. So each probe learns its own field capacity from its own drain-downs (learned_capacity):

• A wetting is the probe's reading, smoothed by a running median of three, rising 0.04 or more within four hours.
• A drain-down runs from the wetting's peak to the next wetting, or 72 hours. It counts when it is at least 16 hours long with no gap over three hours.
• Each drain-down is fitted as water above field capacity draining away while the crop dries the soil steadily: θ(t) = fc + a·exp(−t/τ) − r·t.
• Only drain-downs from clearly above field capacity count: at least 0.02 above it (a ≥ 0.02). A watering that only refilled the soil part way never drained, and says nothing.
• Only drainage that can be told from the crop's drying counts. A τ shorter than the readings' interval is not tried: on hourly readings, half an hour's "drainage" is one noisy reading. And a drainage slower than a quarter of the drain-down's span must carry 0.06 or more (SLOW_DRAIN_MIN_EXCESS). A crop draws by day and less as the soil dries, so the soil dries along a gentle curve, and a slow, small "drainage" fits that curve. Counted, it put field capacity below the level the bed had been watered to; the next watering stopped there, and the learned value sank round by round (below). Water perched on clay does drain slowly, and with more: 0.12 and 0.13 in the yard's subsoil.
• The probe's field capacity is the upper quartile of the levels the last 60 days' drain-downs settled to, once there are three (Bed.learn_from, "upper"). It replaces the configured value for that probe's slice of the root zone. On a soil that goes on moving water after it has drained, a light watering over drier soil spreads down after its drain-down ends, and the sun dries the top: such drain-downs settle below the soil's field capacity, and their median read it low. On the simulator of 2026-10-16, before the second look and the lawn's ground cover (below), a lawn led by one probe at 7 cm and watered back to its median was short in all 30 summers (14,175 hours); learning from soakings alone, in 29 (8,422); by the upper quartile, in 25 (4,125). With those, by the upper quartile, in 1 or 2. "all" keeps the median, and "soakings" the median of the drain-downs from 0.06 or more over it.

Until then, and for a value within 0.03 of its soil's wilting point, the configured value stands. Bed(learn=False) turns learning off.

What it learns is the level the bed actually holds a day or two after a soaking, the in-place sense of field capacity: where drainage has become negligible (Twarakavi, Sakai and Šimůnek 2009, Water Resources Research 45, doi:10.1029/2009WR007944). A laboratory's value can sit below it.

A layered bed. Two probes whose learned field capacities differ by 0.08 or more (LAYER_GAP) are in two soils. When the bed was given one soil's values, the advisory then writes a second decision beside the morning's, soil-layered:

bed 3's soil is layered: its probe at 5 cm drains back to 23% water after a soaking, the one at 12 cm to 40%. One soil's values cannot describe it, and the advice may wait too long for the layer whose water the roots cannot take. Give the bed each layer's field capacity and wilting point, from a soil test of each layer or the soil survey's horizons.

Bed.layers takes those values: (bottom_m, field_capacity, wilting_point), top down. Each probe's slice of the root zone is then judged at its own soil's limits, and the note stops. tools/soil_survey.py gives the survey's values by depth.

Backtested on the virtual farm. The test bed is a yard (physics.LayeredBed): 10 cm of brought-in sandy loam over a compacted clay subsoil. Its roots are in the top 15 cm, and it ran over McAllister Farm's summers of 2013 and 2012 (June to August), with the advisory followed. It was set up three ways:

• with the survey's silt loam, field capacity 0.30 and wilting point 0.12;
• with the topsoil's own nominal values, 0.20 and 0.09;
• with a soil test of each layer: the topsoil as above, the clay 0.38 and 0.25.

The soils' values are round, typical ones, assumed for the simulator rather than measured. Water applied, drained below the roots, and hours with half the available water gone:

Measured on the simulator of 2026-10-16 (below), three draws of the noise: the water and drainage are their mean, the hours short the least and most of the three.

What the table shows:

• On the survey's values, a probe in the topsoil drowns the yard. The topsoil holds 0.23 to 0.25, which the advisory reads as a sixth to a quarter short of 0.30, so it waters most mornings. Learning cuts the water by three fifths.
• A soil's nominal values can starve it, and here barely did. Sandy loam's 0.20 is where its drainage stops, but in the simulator drainage slows to almost nothing sooner, and a day after a soaking the bed still holds about 0.23: refills to 0.20 left the roots short in the 2012 drought for up to 17 hours, where on the soil of 2026-10-15 it was 193 to 286. Learning, by the upper quartile, takes it to none, for a third more water.
• Learning field capacity does not fix a layered bed on its own. With a probe in each layer, the survey's 0.30 happens to sit near the two layers' depth-weighted mix, and learning did not help: 8 to 244 stressed hours against 7 to 16 in the 2012 drought. The trouble is the wilting point: one value for two soils counts water the clay holds that the roots cannot take, so the advice waits too long. The note is what helps here. Given each layer's values, as the note asks, and with learning on, the yard was short for at most 2 hours in 2012 and none in 2013.
• Learning leaves a correctly set-up bed alone. On the advisory's own backtest bed, a bucket of the soil it is set up with, the 2013 summer took 208 mm learning and without, and 2012 319 mm both ways, to the millimetre.

A uniform sandy fill, 60 cm of the same sandy loam with a probe at 5 cm, behaved like the single-probe yard on the earlier soil: 730 mm on the survey's values against 175 learning in 2013. On its nominal values it had 87 stressed hours, and none learning. (Not measured again: sandy loam passes almost nothing below its field capacity, so the new soil changes it little.)

What is still not right. Over McAllister Farm's thirty summers, 1991 to 2020, the yard set up with each layer's values and learning was short in 19 to 22 (119 to 222 hours, up to 32 in one summer) on 275 mm, where on the soil of 2026-10-15 it was 2 to 7. Two summers do not show that; thirty do. Its root zone holds about 17 mm its roots can use, its roots are densest at the top, where the sun also dries it, and a hot day can take 8 mm: a third gone in the morning can be past half by the afternoon. Watering at a fifth gone instead (Bed.depletion_fraction 0.2) left it short in 4 to 7 summers (14 to 32 hours) on 301 mm, 10% more; it is a setting for such a bed, not yet the default.

Looking ahead to tomorrow. So the advice also waters a bed that, drying as it has been since it was last at field capacity, would pass Bed.short_fraction of its water (half, FAO-56's depletion fraction for most field crops) before the next morning:

Water bed 3 today: 26% of the water its plants can use is gone, and drying at 8.1 mm a day it would be short of water before tomorrow morning. About 7 mm brings it back to field capacity.

(A test's bed: 15 cm of roots, 7 mm gone, drying 8 mm a day.) A deep bed rarely dries that fast, so it changes little; a shallow one is watered a day early. Over the thirty summers:

Measured on the simulators before 2026-10-16: the yard with it on the soil of 2026-10-15, three draws (on the earlier soil, one draw, 202 mm, 8 summers, 141 hours), the rest on the earlier soil. On the simulator of 2026-10-16, with the second look of the day and the other changes since, the yard was short in 19 to 22 summers on 275 mm (above), and the backtest bed in 0 to 3 for at most 4 hours, on 289 mm.

The deep bed never dried fast enough to be watered early, which is the point. On single summers the change is mixed: the table above is with it, and the one-probe yard on the survey's values went from 16 stressed hours to 53 in 2012, while the two-probe yard on the survey's values went from 116 to 31.

Looking ahead to the refill point itself (short_fraction set to the depletion fraction) did more, and cost more: in a first run the yard was short in 2 summers, never more than 12 hours, and the backtest bed never, but the backtest bed took 241 mm a summer, 28% more, as much as the weather alone gives a bed with no probe. The drying rate is timed from the The drying rate is a whole day's loss, each end the mean of the readings within two hours of it. A crop draws by day and hardly at night, so any part of a day misstates it: a fit over the night since the soil left field capacity read a heat wave's 8 mm a day as 1.3, and the advice did not look ahead to it. A day with a wetting in it is skipped for the day before, and a day that began above field capacity begins there, so drainage is not counted as the crop's.

A learned value that sank, and why it no longer does. A bed watered by its probe is watered back to the field capacity it learned, so after most waterings nothing drains. The crop then draws by day and less as the soil dries, and the old fit read that gentle curve as a slow drainage, with field capacity below the level the bed had been watered to. The next watering stopped there, and round by round the learned value sank: from 0.248 to 0.206 over the 2012 drought, on a probe whose true level was 0.25. Now such a slow, small "drainage" does not count (above). On 200 simulated three-day curves with no drainage at all, the old fit took 34 for drain-downs, down to 0.23 on a bed watered to 0.27; the new one 10, none below 0.253. The advisory's backtest bed with its probe, over the thirty summers:

On the simulator of 2026-10-16 the probe 0.05 low did not learn its level at all in some summers, and was watered as an unlearned probe is: hence its water.

The 25 mm that ran off the yard in 2013 was rain on topsoil already holding water over the clay, the same with or without learning. A watering can do the same when it is run faster than the clay takes water. In the simulator, the same 50 mm given as four quarter-hour runs an hour apart sheds less than two hours in one go (sim/test_physics.py), which is cycle-and-soak.

A probe that reads off. Learning forgives a probe that reads low: one 0.05 low called for 1,045 mm unlearned and 422 learned, against a true probe's 208, on 2013. It cannot forgive one that reads high. That probe says the bed is fuller than it is, so the bed is watered late and little, 127 mm, and the probe sees too few soakings to learn from: 293 hours of stress. A second probe in the bed is what catches it (soil sensors).

These results come from the simulator. A real yard's probes are what would check them (UNTESTED).

A tension sensor beside a probe measures field capacity instead of learning it. Each hour the two at one depth give a point on the soil's own retention curve: this much water, held this hard. Over the tensions a Watermark is calibrated for, 10 to 75 kPa (sdk/tension.py), water falls close to linearly with the logarithm of tension, so the points are fitted and field capacity read at Bed.capacity_kpa, 33 kPa by default (retention.py). It needs 24 hours of points spanning a factor of two in tension and reaching to within half again of 33, and no soaking. On a silt loam's curve (Carsel and Parrish's mean, through van Genuchten's equation) it reads the 0.239 at 33 kPa to within 0.005, and within 0.01 with a probe's noise. A measured value stands before a learned one for that probe's slice; the decision says which ("how": "measured"). A tension sensor pairs with a probe installed within 2 cm of its depth; its raw series reports tension and is calibrated with a = -1, as fieldmesh records water potential, never above zero.

A crop that grows#

A seedling draws little water from a few centimetres of soil; a full canopy draws more than the reference grass, from deep. A bed told its crop and the day it was sown (Bed.crop, Bed.planted) follows FAO-56's curve for that crop through its season (crop_day), with the 2025 revision's values (FAO-56 Rev.1, data/fao56rev-crops.csv) and the 1998 edition's stage days (data/fao56-crops.csv):

• its crop coefficient by stage (the revision's Tables 6.1 and 6.2): the initial value through the initial stage, rising in a line to the mid-season value over the development stage, holding, and falling to the end value over the late season;
• its roots from 15 cm at sowing (the 1998 Annex 8: "0.15 to 0.20 m" for many annual crops) to the shallow end of the revision's range of maximum depth as mid-season begins, growing with the crop's development in a line, as the revision's Example 8.3 still grows them;
• its stages in the revision's degree-days (Table 6.11), for the 10 of these crops it has field-observed values for, where the farm's station gives the temperatures (below); otherwise the 1998 edition's days (its Table 11).

Where the revision gives a crop two seasons, the plain name is the first and the other is named after a colon: tomato is fresh-market tomatoes and tomato:processing the processing crop; lettuce, potato, broccoli, corn (sweet corn) and muskmelon are the short season and …:long the long; pea is fresh peas for market and pea:industry the processing crop. Against the 1998 edition, tomatoes' roots now reach 0.6 m, not 0.7, and their end-of-season coefficient is 1.00 for the fresh market, not 0.80; lettuce's mid and end values are 1.05, not 1.00 and 0.95.

Lawns: the 1998 values, and the revision's as turf_lawn. The 2025 revision gives a mown lawn (cut under 10 cm) 0.50, 0.70 and 0.50 through its season (Table 6.4), against the 1998 edition's 0.95 for cool-season turf and 0.85 for warm; turf_lawn holds its 0.70 all year, as the 1998 lawns hold theirs. Which is right for a given lawn matters only where no probe measures it: by the weather, a lawn told 0.70 is given about a third less water, and if it truly uses 0.95 it goes short. On the thirty summers, the lawn drill's tall fescue with 15 cm of roots, the advice by the weather and the virtual lawn using one value or the other:

Measured on the simulator of 2026-10-16 (below), three draws of the forecasts' noise. Told 0.70, a lawn is given a third less water either way, and if it truly uses 0.95 the yard's goes short in nearly every summer. On the soil of 2026-10-15 the same was short in 3 summers for 27 hours, and the case for turf_cool as the default looked weak; on the earlier soil, which moved no water below field capacity, in 7 (87 hours). Now a lawn's crop uses its water whole until half is gone, as FAO-56's does, and turf_cool stays the default: told too little, a lawn by the weather is starved; told too much, it is given water that drains. Even told rightly, the yard's 15 cm of roots in sandy topsoil over clay went short in 6 or 7 summers by the weather, its margin (below) included. By a probe the value makes no difference at all, to the millimetre: the probe measures what the lawn used. A lawn with a probed bed nearby as its reference can learn that its rate is faster still (Bed.reference, below). A lawn led by one probe at 7 cm, short in 13 summers of 30 on the earlier soil and 1 to 8 on the soil of 2026-10-15, was short in 1 or 2, for 7 to 31 hours, on 181 mm: half the water by the weather.

The day's root depth sets the root zone the probes are weighed over and the water it holds, and the day's coefficient drives the weather balance; an established lawn (turf_cool, turf_warm) holds its mid-season values all year. A probe below a seedling's roots counts for nothing until the roots reach it: ten days after sowing, tomatoes' 15 cm of roots do not reach a probe at 20 cm. Nor does a probe at the root zone's very bottom count (weights): it reads the soil the roots are only reaching, half of it below them. On the simulator's earlier soil, which moved no water below field capacity, tomatoes' roots at their deepest, 60 cm, with a probe at 60 cm counted as a quarter of the root zone, were short in 13 summers of 30 for 1,303 hours; left out until the roots are past it, in 0 or 1 over five draws of the probes' noise. On the simulator of 2026-10-16 the probe at 60 cm makes no difference to how often they are short (below).

The revision's depletion fractions are in the file; the advice keeps a third, lower than its tables as it says suits frequent watering, and with Bed.adjust_p lowers it on a hot day as the revision's Eq. 8.5 does (below).

Backtested on McAllister Farm's thirty summers, sown on 1 May, the virtual crop growing by the same curve in the 1998 days, in a silt loam 1.5 m deep, probes at 10 and 30 cm. "Not known" is the bed as it was before: 30 cm of roots and a coefficient of 1 all season.

Measured on the simulator of 2026-10-16 (below), three draws of its noise each: the water is their mean, the summers and hours short the least and most of the three. On the soil of 2026-10-15 no row was short in any draw, on half the water or less (tomatoes known, 300 mm by the weather and 247 by probes).

• Not knowing the crop now costs it. By the weather, a crop not known is charged a coefficient of 1 all season: a seedling's wet soil evaporates more than that early on, and tomatoes at mid-season use 1.10. It was given less water than knowing the crop, and went short in a third of the summers. By probes, a crop not known is advised for 30 cm of roots, and its probe at 30 cm, at that root zone's bottom, does not count; tomatoes, whose roots draw down to 60 cm, went short in two summers of three.
• Known, by the weather, neither crop was short in any summer. The balance's margin (below) is part of that, and so is the water.
• Known, by probes, both were short in some summers, all in a few days at the end of the initial stage. The young crop is watered at a fifth of its water gone (young_fraction) through its initial stage, 30 days for tomatoes and 20 for lettuce; on the day it ends the advice goes back to a third, while the roots are still the 15 cm they were sown with and the crop begins to use more. In the three summers traced, among the worst, every short hour fell within three days of that: tomatoes on days 30 to 33, lettuce on 20 to 23. A third probe at 60 cm, in the deeper half of tomatoes' roots, changed nothing in those days.

The virtual crop grows by the very curve the advice is told, so this shows what knowing the curve is worth when the curve is right, not how right FAO-56's curve is for a given field: a real crop's stages run early or late with the season, which is what its clock (below) and its probes are for.

By the season's warmth#

The 1998 edition's stages are days, for a typical season; a warm year runs a crop through them sooner and a cool one later. The 2025 revision gives them in growing degree-days instead (its section 6.8.1). Given the farm's station, the advisory does the same (thermal.py, Advisor.clocked):

• A day's degree-days are its mean temperature less the crop's base, nothing below the base, and no more than its upper threshold less the base (the revision's Eq. 6.23), each crop's thresholds from its Table 6.10: tomatoes 7 and 28 °C, lettuce 4 and 28, potatoes 2 and 30, sweet corn 10 and 32.
• For the 10 of these crops the revision has field-observed degree-days for (its Table 6.11: tomatoes, lettuce, carrots, onions, potatoes, broccoli, peppers, melons, peas and sweet corn; its beans are dry beans, not green), each stage ends when the crop has gathered its degree-days since sowing. Fresh-market tomatoes reach full cover and full rooting at 985 and end at 2,065: on McAllister's summers, sown 1 May, that is 59 to 74 days (median 66) and 113 to 132 (median 122), against the 1998 edition's 70 and 145.
• For the other ten, the 1998 days are taken as right for the farm's normal year: the mean of each day of the year over a record of past years (climate in the config: a daily file as tools/weather_csv.py writes from NASA POWER). On any day the crop is as far through its curve as a normal year would have taken it to gather the degree-days this one has given it since sowing, which is how the revision's own Table 6.12 turns those days into degree-days, at a representative place rather than this farm.

A day the station missed counts as a normal one; without the farm's record, the mean of the last week's, and a crop is clocked only when the station read nine in ten of its days since sowing. Each morning's decision says how far through its curve the crop is (crop_gdd, or crop_days) beside the calendar (calendar_days), and whether by warmth (by_warmth). Without a station it is the calendar.

Backtested on the thirty summers, sown on 1 May, by the weather with no probe. The virtual crop goes through its curve by the season's warmth; the advice either keeps the calendar or keeps the same clock:

Measured on the simulator of 2026-10-16, three draws of the forecasts' noise; the single summers are the furthest of the three.

On average it changes little; in a single year it moves the water by up to an eighth, toward what the virtual crop used. The calendar's advice was never short either: the balance waters to field capacity with margin, so a crop running early or late on this farm costs drained water rather than stress. Where it would matter more is a crop whose stages are near their water's limit, and a farm far from its normal year.

Plantings at different stages#

A bed often holds plants at different stages side by side: a succession sown a few weeks apart, a replant in a gap, lettuce under tomatoes. One curve does not fit them all. A bed can name each (Bed.plantings; in the config "plantings": [{"crop": "tomato", "planted": "2026-05-01"}, {"crop": "lettuce", "planted": "2026-06-15"}]), each on its own clock, and the advice works to the range of depths their roots span (grown):

• When to water is set by the youngest roots. The bed is due when the soil as deep as its shallowest planting's roots has used its share of its water, a third, as the whole root zone's is.
• How much is set by the oldest. The refill brings the soil back to field capacity down to the deepest roots, so watering on the young plants' account never shorts the old ones.
• The crop coefficient for the weather's balance is the plantings' mean.

A planting not yet sown counts for nothing. Each decision gives the depth the watering was timed by (timing_depth_cm) beside the root zone's (root_depth_cm), and the plantings' stages (inputs.plantings). When it is the youngest roots that make the bed due, it says so:

Water tunnel bed 3 today: its shallowest roots, in the top 15 cm, have used 39% of the water they can reach. About 20 mm brings the bed back to field capacity, down to its deepest roots at 50 cm.

By probes, the young roots' soil is read by the probes in it, or, where none is that shallow, by the shallowest. A probe among the young plants, as shallow as their roots, is what reads them truly. A planting can name the probes set among it ((crop, planted, (sensor id, …)); in the config, "probes": [...]): the youngest planting's own probes then time the watering, the oldest's, with any no planting names, measure the refill, and the refill is at least what the young planting's soil has lost.

Backtested on the thirty summers: tomatoes sown on 1 May, lettuce sown among them on 15 June, each in its own patch of the same silt loam, watered whenever the bed is:

Measured on the simulator of 2026-10-16 (below), three draws of the noise: the water is their mean, the summers and hours short the least and most of the three. The tomatoes' hours short are the ones at the end of their initial stage (above). Before 2026-10-15 the backtest grew the lettuce's patch as a seedling from the start of the season and counted its hours short from then, so lettuce sown on 15 June was "short" in May, and every table here before then overstated its shortfall (backtest's second).

• By the weather, the lettuce must be named. Told the tomatoes alone, the advice left it short in 22 summers of 30: its 15 cm of roots, in a patch now drying at its surface, use their water long before the tomatoes' deep root zone is due. On the soil of 2026-10-15 it was never short, water rising from below carrying it. Told both, never, on a seventh more water.
• By probes among the tomatoes, told the tomatoes alone, the lettuce went short in 22 to 24 summers; told both, in 2 to 5.
• A probe in the lettuce's patch. Until the lettuce is sown its patch is bare. On the simulator before 2026-10-16 a bare patch lost no water at all, and a probe there, unnamed, read as one of the tomatoes', held their watering back: they went short in 6 to 8 summers, 477 to 1,266 hours. Now the bare patch dries at its surface, as a real one does, and an unnamed probe there misleads far less: moved there, it leaves the tomatoes short somewhat more (5 to 12 summers, 95 to 233 hours) on a fifth more water; added as a fourth, the plantings fare about as well as with each probe named for its planting ((crop, planted, (sensor id, …))). Naming still keeps a probe from counting before its crop is sown.
• Two plantings this far apart in age are still better on two valves.

Roots that stop#

The crop's curve is for a soil that lets roots go. Something can stop them sooner:

• a hardpan or plough pan;
• the compacted fill under a new garden;
• rock;
• the floor of a raised bed.

Advice that follows the curve then counts water the roots cannot reach. It waits too long, and refills less than the roots used. How deep roots run through a season, and what shows it, is in the root-depth note. There are two ways to give the advice the right depth:

• Bed.rooting_limit_mm, set by hand once the farmer has dug and found the pan. The crop's curve stops at it. A probe at that depth or below it sits in the pan, reading soil the roots do not enter, so it is not counted.

• Root signs from the probes (root_signs, rooted), over the last two weeks. A probe's dry days are those that started no wetter than its field capacity, with nothing wetting it; its field capacity is the lower of the bed's and its own. A probe whose soil fell less than 0.002 a day over three or more of them is level: nothing draws on it. One that fell faster has roots, or is draining, and two tests tell which:

  • by field capacity, the default: a probe that went 0.02 under field capacity has roots. Just past field capacity a soil still drains about that fast, but only roots, or the sun near the surface, take it well below;
  • by the stair-step, with Advisor(stair_step=True) and the farm's timezone (its station's, or Advisor(timezone=…)): roots draw by day, so a probe that does 70% or more of its falling per hour between 08:00 and 18:00, rather than between 20:00 and 04:00, has roots, and one that falls at least as fast by night has none; drainage, and water creeping up to roots above, run day and night alike. Between the two it says nothing.

  Take a probe without roots that the crop's curve says the roots passed at least 15 cm ago, under one that has them. The roots likely stop between the two, and a note says so. A stop seen on three of the last seven mornings stands on mornings that cannot see it, when waterings leave too few dry days to judge by.

  Tunnel bed 3's crop should have roots to about 60 cm by now, but its probe at 30 cm has stayed level through its dry days while the one at 10 cm has dried: something may be stopping the roots, a hardpan or compacted soil. The advice now waters when the top 20 cm has used its share, and still refills to 60 cm. If you know how deep the soil lets roots go, set the bed's rooting limit.

  A stop only changes when to water, never how much. The watering is timed by the soil above half way to the level probe, as for a young planting (above); the refill is still each probe's own deficit down to the curve's depth. Under a real stop the probes read wet and add nothing to it; if the roots are there after all, they still get their water. Counting the roots only to half way, as a first version did, also cut each refill to the top 20 cm: on the thirty summers with no pan at all, the stops it saw wrongly then left the crop short in 3 summers, for 714 hours (with the 1998 curve).

Backtested on the thirty summers, tomatoes sown on 1 May with probes at 10, 30 and 60 cm, the virtual roots stopped at 30 cm whatever the curve says, the 2025 revision's curve taking them to 60 cm. Two pans: one that stops the roots alone, the soil below it the same silt loam; and a dense one, as a plough pan is, 5 cm of compacted silt loam passing 0.5 mm of water an hour (growing_bed(pan_mm=…)). Three draws of the noise each; the water is their mean.

Measured on the simulator of 2026-10-16 (below). The hours short shared by most rows fall at the end of the initial stage (above), before any pan matters.

• Over a pan that only stops roots, the curve alone does not starve the crop. The soil below the pan dries as water rises out of it to the roots above, so the bed's average falls and the advice waters. On the simulator's earlier soil, which moved no water below field capacity, the soil under the pan stayed at field capacity, held the average up, and the crop was short in 29 summers of 30 for eleven days on average (7,563 hours). That was the case the root signs were built for.
• Over a dense pan it does go short, in 10 to 18 summers: water rises through the pan too slowly. The root signs cut the hours short by up to half, for a quarter more water; the rooting limit set by hand does best, on less.
• But the root signs no longer tell a pan from none. With or without one, they see a stop on 74 to 87 of a season's 145 mornings, the stair-step on 59 to 88. With roots densest at the top, as now, the probe at 30 cm is drawn on lightly and reads near level in a soil with no pan at all. On the soil of 2026-10-15, roots drawing evenly, the test by field capacity saw a stop that was not there on 9 to 12 mornings, the stair-step on 24 to 29. Where they help, over the dense pan, it is by timing the watering by the top 15 cm, which a young crop's or a shallow bed's earlier refill point would do as well (above).

So on the simulator as it now is, the root signs do not do what they were built for: their stop is right over a pan and wrong without one about as often. Whether a real crop's roots read so, and whether the level line should be judged against what the roots' own density predicts, is not known (UNTESTED). The test by field capacity stays the default, and the stair-step an option, until that is decided; it is one of the maintainer's open choices. A first version counted the probe at 30 cm, set into the pan; leaving it out, and now any probe at the roots' front, fixed that.

Hot days, and beds under cover#

On a hot day. FAO-56 adjusts the depletion fraction for how fast a crop is drying (the 2025 revision's Eq. 8.5): p + 0.04 (5 − ETc), ETc in mm a day. With Bed.adjust_p, the advice lowers its third that way on a day drying faster than 5 mm, from the probes' drying rate or, by the weather, the day's Kc · ETo, so the crop is watered with more water left. FAO-56 also raises p on a mild day; that is not done here. On the thirty summers:

Measured on the simulator of 2026-10-16, three draws of the noise. On the earlier soil the backtest bed was short in 3 summers (72 hours), and 2 (33) with adjust_p; Eq. 8.5 both ways, raising p on mild days too as FAO-56 does, saved water there and left it short in 7, and that option was taken out.

On the backtest bed it took the few hours short to about none for 3% more water. On the shallow yard it changed nothing at all: its 15 cm of roots never dry 5 mm a day, though they are short often; what that bed needs is an earlier refill point (above).

Whether to make it the default was tried across the backtests above, each run again with it on (thirty summers each; water, summers short, hours short):

It only acts where the crop draws well over 5 mm a day, which on this farm is tomatoes at mid-season; a lawn seldom does. On the simulator of 2026-10-16 it ended none of the crops' shortages, which fall at the end of their initial stage, before mid-season, and added a half more water, most of it drained; for the lettuce among tomatoes it cut the lettuce's hours short to a third, on the tomatoes' extra water. On the simulator's earlier soil it ended most of the shortages there were, at a fifth to a third more water. So it stays off by default: a bed whose probes show it running short in hot spells can be given it. (Measured on the simulator of 2026-10-16, three draws of the noise each.)

A bed under cover. Rain forecast for the farm does not reach a bed under a tunnel or a roof, and "rain is coming: wait" is the wrong advice for it. Bed.rain_share says how much of the rain reaches the bed, 0 under cover; the forecast that holds a watering, and the weather balance's rain, are scaled by it (rain_reaching_mm). Left unset, it is all of the rain, unless the bed's probes show it covered: through the last three storms of 15 mm or more in the farm's rain gauge (Advisor(rain_gauge=…), or the station's) that fell while its valve stayed shut, none moved a probe in its root zone by 0.03, as a mean of three readings, within 12 hours of the rain's end (covered_by_its_probes). Only a bed whose valve reports its waterings is judged: without it a watering could not be told from the rain. A share between none and all is not learned.

On the thirty summers, the backtest bed under a tunnel, its forecasts the open sky's, watered by a valve that reports:

Measured on the simulator of 2026-10-16, three draws of the noise, the mornings over the thirty summers. Told, it was never held for rain that could not reach it. Learning took the first part of each summer to see three storms, and held it on a quarter to a third fewer mornings. In the open the learning never took the bed for covered: the water and the hours short are the open bed's without it.

Water below field capacity#

The simulator's soil moves it, since 2026-10-15. Before, its layered soil (physics.LayeredBed) drained a layer wetter than its field capacity into the one below and moved no water at all between layers drier than that. A real soil does: slowly, but not slowly beside a lawn's days. Now each horizon holds its water along van Genuchten's curve, fitted through its own saturation, its field capacity at 33 kPa and its wilting point at 1500 kPa, and passes it as Mualem's conductivity from its saturated conductivity lets it. Water moves between neighbouring layers by Darcy's law, suction and gravity, at the geometric mean of the two layers' conductivities (physics.Horizon). Every table here with a crop or a lawn was run again on it, three draws of the noise each.

What found it. A lawn led by one probe at 7 cm (the lawn drill's tall fescue, 15 cm of roots, in silt loam) was short in 13 summers of 30, 255 hours, where the same lawn by the weather never was. Morning by morning in 2008, each watering, sized to what the probe read, wet the top 10 cm a little past field capacity, and the grass used that before it drained. The bottom 5 cm of the roots stayed near 0.19 while the probe read 0.28, and the probe never saw them. Carsel and Parrish's (1988) mean silt loam, through van Genuchten and Mualem, passes 2.5 to 6 mm a day between such layers 2.5 cm apart; the old soil passed none.

What moved then.

Led by probes, beds were short far less often, and most took less water. Water held near capacity at the top reached the bottom of the roots on its own, and water below the roots rose to them. On the simulator of 2026-10-16 several of those results went back the other way (below): the soil was moving water rightly, but the crop was using it too gently.

What it rests on. The conductivity between two layers is their geometric mean, a choice: the arithmetic mean would move water faster, the harmonic slower. Since 2026-10-16 each soil's residual water and Mualem's L are Rosetta3's for its texture and bulk density (physics.SANDY_LOAM, COMPACTED_CLAY, the growing bed's silt loam), where they were none and 0.5. Checked against HYDRUS-1D on the same soils, the simulator's water content was within 0.0002 to 0.002 while water spread and rose, and 0.002 to 0.012 while a watering soaked in (soil models).

Learning reads a soil that moves water low. On this soil a light watering over drier soil goes on spreading down after it has drained, so the probe settles lower than the soil's field capacity, and the median of many such drain-downs read low: in 2008 a lawn's probe learned 0.27 where the soil's is 0.30, and watered back only to that, the lawn sat near half its water gone. The upper quartile of the drain-downs is now learned instead (above); learning from the drain-downs after soakings alone was tried, and read lower still.

A probe in another soil needs that soil's values. In the builder's yard, sandy loam over compacted clay at 10 cm, one probe at 7 cm kept the lawn watered in every summer, on 493 mm. A second at 12 cm, in the clay, with the bed given one soil's values, read wet all summer: 25 to 26 summers short (1,865 to 2,946 hours). Given each layer's values (Bed.layers, above), 16 to 20 (45 to 140 hours) on 270 mm. (The simulator of 2026-10-16.)

The crop's water use and the soil's surface#

The simulator's crop, since 2026-10-16. Checked against the models agriculture uses (soil models), its soil moved water well but its crop was too gentle, and its bare soil never dried. It now has, after FAO-56 (Allen et al. 1998) and its revision:

• the crop's whole use to its p, then less in a line to the wilting point (FAO-56's Ks, Eq. 84), each crop's own p; before, it used less as soon as the soil was below field capacity;
• its roots densest at the top, drawing from each layer in a line falling to none at the roots' depth (Prasad 1988); before, from each layer by the water it could give;
• the soil's surface drying, FAO-56's dual coefficient: the crop's basal coefficient drawn by the roots, and the top 10 cm evaporating apart from it as the canopy leaves it open (Eqs. 69 to 79); bare soil dries in two stages. A mown lawn covers 0.95 of its ground, where FAO-56's estimate from its coefficient (Eq. 76) would leave a third bare;
• a dense pan as an option, a horizon that passes water slowly as well as stopping roots (above).

At USCRN's Crossville station, under grass, the simulated soil followed the probes hour by hour (correlation 0.85 to 0.96), and dried on rain-free days at 0.74 of the rate the probes saw at 5 cm, 0.85 at 10 and 1.26 at 20, where the crop before dried them at 0.56, 0.71 and 1.12 (soil models).

What it showed, and what the advice took from it. A crop that uses its water whole until its refill point, from the top first, empties a shallow zone faster than the advice had been built for, and a young crop's wet soil loses more than its coefficient said. So:

• learning by the upper quartile of the drain-downs, not their median (above);
• a second look at 13:00, at the top of this page;
• a young crop watered at a fifth of its water gone through its initial stage (Bed.young_fraction), its few roots in the soil the sun dries: tomatoes' first month went short in 25 summers of 30 without it, 9 with it, for 4% more water. Among older plantings, it is the refill point of the young one's roots alone;
• by the weather, the dual coefficient for a bed whose crop is named, so the balance counts a young crop's wet soil, and the crop's own p for when it slows; and a margin, the crop's use counted a fifth above FAO-56's estimate where nothing measures it (below);
• a silent probe's bed carried from the mean of its last three hours of good readings, not from one (below).

What moved, the lawn drill's tall fescue and the backtest crops, thirty summers, three draws (water a season; summers short; hours short):

Every crop takes about twice the water it did: on the simulator before, a crop used less as soon as its soil was below field capacity, and a bare or young bed lost nothing at its surface. Most of the new shortages fall in a few days at the end of a crop's initial stage (above), and on beds with little water to spare. By the weather, with its margin, no bed whose crop is named went short.

By the weather, with no probe#

A bed with no probe used to get "no probe" and nothing else. Given the farm's own station (eto.Station: its site and the series it keeps), the advisory keeps FAO-56's daily root-zone balance instead (weather_balance).

• Each day's reference ET is FAO-56 Penman–Monteith from the station's readings that the checks passed (sdk/eto.py, which reproduces FAO-56's Example 18: 3.88 mm). Missing sunshine, humidity or wind are replaced by FAO-56's own stand-ins, and each decision lists which it used.
• The balance: each day the depletion falls by the gauge's rain and the watering, and rises by Ks · Kc · ETo, Kc the bed's crop_coefficient and Ks slowing the crop once it has used its own p (eq. 84). For a bed whose crop is named, (Ks Kcb + Ke) ETo, FAO-56's dual coefficient: the crop's basal coefficient and its wet surface's evaporation, which runs well above a young crop's single coefficient when the soil is wetted every day or two.
• A margin: with nothing measured, the crop's use is counted a fifth above FAO-56's estimate (WEATHER_MARGIN). The estimate is not the bed: the station's reference ET and the crop's coefficient can each be a tenth or more out, and a balance that counts low waters late and short, and its shortfall adds up until rain, since each watering refills only what it counted. Counted high, it waters a little early and a little over, and the excess drains.
• Watering is the valve's recorded minutes times the bed's application_mm_per_h. Without that record, the advisory's own advice to water is taken as followed, and the advice says so. A balance that never sees the watering it asked for asks again every morning: on the backtest bed, on the simulator before 2026-10-15, that called for 1,027 mm in 2013.
• Two of the last three days without a full day of weather, and it says it cannot tell, rather than guess.

Water bed 3 today, by the farm's weather: about 50% of the water its plants can use is gone, with no probe to confirm it. About 27 mm brings it back to field capacity.

(The test's bed, five dry days after a soaking.)

Backtested with no probe on the same bed and summers, with an hourly station:

• air temperature, humidity and rain as the virtual farm's recorded weather gives them;
• each day's recorded sunshine spread over its hours;
• no wind sensor;
• the watering unrecorded (weather_backtest).

Why the margin. The virtual bed loses water by Hargreaves' formula and a crop coefficient of 1.05; the balance counts Penman–Monteith and 1. Over the thirty summers the bed so used about a fifth more than the balance counted, as a farm's real bed may, either way. With no margin it was short in 21 summers of 30 (1,945 hours) on 249 mm; a tenth, 6 (101 hours) on 287; a fifth, none, on 329. Watering a fifth more each time with the balance unchanged left 8 (118 hours) on 298 mm: it still watered late. Where the balance was not low, the margin cost water and nothing else: a lawn 27% more, tomatoes 21%, lettuce 14%. (The simulator of 2026-10-16; on the one before, whose crop used less as its soil dried, the bed by the weather alone was never short, the balance's own excess its margin.)

When a bed's probes go silent or untrusted, the advisory used to say "unseen" or "untrusted" and give no advice until they came back. With the station, it now carries the bed on the weather from the probes' last good readings:

• that reading's depletion, at each probe's soil and learned field capacity, is the balance's starting point, the reading the mean of each probe's good ones in the last three hours before it fell silent (ANCHOR_HOURS): on the 2012 drought one reading 0.027 over the soil, taken alone, started the balance at half the depletion the bed had, and it went short 15 hours in the outage's first two days;
• the days after it are stepped by the weather;
• every advice says to check the probes.

bed 3's probes have given no reading to trust for 81 hours. By the farm's weather since their last good one, about 55% of the water its plants can use is gone: water it today, about 30 mm. And check the probes.

(The test's bed, its probe silent from 11 July.)

It runs as long as the balance does, sixty days, because a bed anchored on its last reading is never worse off than one that never had a probe. On the 2012 drought, with the backtest bed's probe silent for the two weeks from 1 July:

(The simulator of 2026-10-16, on 446 mm; before, 13 mornings without advice and 142 hours.)

A probe nearby corrects the balance#

The objection to soil probes for lawns and fields is that a probe is a point: one square metre does not speak for a zone. The balance covers the zone but cannot see the soil. So a probe lends the balance only what carries from bed to bed: its rate, not its level.

A bed with no probe can name a bed nearby, of the same crop on the same station, as its reference. Each morning reference_factor reads that bed's probes over the last 30 days. It counts the days with no rain on the gauge, no watering (its valve shut, its probes not rising) and less than the readily available water gone. On each, it measures how much the water missing below field capacity grew, midnight to midnight, against the balance's Kc · ETo for the same day. Water held above field capacity after rain is not depletion, which the probe sees and the balance does not.

The ratio corrects the bed's crop coefficient: all the way when the probe sees the bed drying faster than the balance says, half way when slower, because more water is the safe error; and never below the margin: a probe's rate can raise the balance past the fifth it counts over FAO-56, but a slower one is recorded and not used (margin_held). It is not used:

• with fewer than five such days, or under 10 mm of the balance's ET over them;
• until every probe in the reference has learned its own field capacity (above). A probe reading high sees the first of each spell's depletion as water above field capacity, and never learns;
• when the ratio is outside 0.5 to 1.5: then the probe or the station is wrong.

The same rate, read from a bed's own probes before they fell silent, carries the fallback above, and every decision records the factor, its days and its standard error.

bed 3 is due water by the farm's weather, with its water use set at 85% of the weather's estimate from a probe in a bed nearby: about 19 mm. Its watering is not recorded, so this takes the last advice to water as followed; if it was not, it needs more.

(The test's bed, 19 days into a summer with rain every fifth day, its reference's probe drying at 70% of the balance's rate.)

Backtested on McAllister Farm's thirty summers, 1991 to 2020 (NASA POWER). The backtest bed has no probe and its watering unrecorded; a bed of the same soil and crop beside it has a probe at 20 cm and is watered on its own advice (weather_backtest(reference=True)):

The simulator of 2026-10-16, three draws. The probe-led bed beside it was given 398 mm a summer, more than the bed it corrects, and seldom read a rate above the margin. On the 2012 drought, watered often on its own advice, its probe read its bed drying at 85% to 111% of FAO-56's estimate; taken half way to that, the margin dropped, it left the bed short 33 hours; held at the margin, never. On the simulator before, without the margin, the correction saved 2% (237 mm against 241), never short.

Why only half way, and never below the margin. The rate a probe saw over a month does not hold through a hot spell, and a probe whose bed is watered often reads its rate low: the soil above its learned level after a watering is not counted. Corrected all the way, on the simulator before, the bed went short in one summer in six. A probe reading high, used without the learning check, took the bed to 397 stressed hours on the 2012 drought. With it, the probe was never used.

The balance's ET is Penman–Monteith, while the virtual bed loses water by Hargreaves' formula. So it runs on a model of the bed that is not the bed, as it would on a farm. In the dry summer it watered a tenth more than the probe did, and still well under the 552 mm a drip timer uses. A probe reads the bed itself; this reads only the sky. A pyfao56-style reset of the balance from a probe in the same zone is not built.

A soak test: field capacity on the day#

Learning a probe's field capacity takes three soakings it drains from; the soak test makes one on purpose, and checks the probes on the way (regimens/soak_test.star). A person starts it for a bed:

1. It asks for enough water to wet the root zone 0.05 past its field capacity, over a square metre centred on the probes, where a millimetre is a litre:

   Soak test for bed 3: pour 39 litres of water over a square metre centred on its probes, a can at a time, letting each soak in before the next. The probes are checked in 3 hours, and the result comes two days later.

2. Three hours on, the probes must have risen by half what that water should add, and read no more water than a mineral soil can hold (0.55). Either failing is said: a probe out of contact, water run off, or a probe reading high.

3. Two days on, the level the root zone drained back to, fitted as any drain-down is, is reported: the bed's field capacity, measured. Its drain-down also counts toward the probes' learning.

(The test's bed at 0.22, field capacity 0.30, 30 cm of roots: 39 litres; drained back to 28%.) Read-only like everything here: it asks a person, and checks what the probes saw.

Did the water arrive?#

The first regimen (Q10; modules/rules): regimens/watering_check.star, a few lines of Starlark. When the morning's advice is "water today", it waits for the watering to happen and soak in, then checks that the root zone rose by at least half of what that watering should add. If not, it says so, and says both things a probe cannot tell apart:

bed 3's probes show almost no new water since this morning: the root zone rose 0.4 points, where 18 mm of watering should add about 6.0. If it was watered, check the drip line and its emitters; if it was not, it still needs water.

It was backtested on the recorded summers with some waterings blocked, as through a blocked drip line: the person waters, and the bed gets nothing.

On the simulator of 2026-10-16, with the afternoon's look: a watering blocked in the morning is advised again that afternoon, and blocked again, so a blocked day is often two checks. On the earlier simulator the closest calls were 0.74 and 0.27 in 2013, and a watered bed at 0.55 in 2012.

With the valve reporting, it says which. A bed can name the series of the valve that waters it (Bed.valve, the valve_open term). An OpenSprinkler controller reports each station opening and shutting over MQTT, and the bridge puts that in the log (sdk/mqtt_bridge.py). The check then tells the two cases apart:

bed 3's valve ran 60 minutes this morning, but its probes show almost no new water: the root zone rose 0.4 points, where 18 mm should add about 6.0. Check the drip line and its emitters for a blockage, and the line for a leak.

bed 3 was due 18 mm of water this morning, and its valve did not open. It still needs water.

On the 2013 summer, with three waterings blocked and two forgotten, it named each one rightly, all five, and said nothing on the ten that arrived. A shutting reported for a valve never seen opening is a lost message, so the check falls back to saying both rather than guessing.

The 2012 false alarm is the honest caveat, foreseen when it was a near miss on the earlier simulator. A bed that soaks its water up slowly, a probe set deep, or a crop that draws hard in the hours after a watering can fall below half on a day the water did arrive. The share is a parameter, and a season of a real bed's waterings is what would set it.

A probe out of the soil#

A probe lifted by a mower or a frost, or pushed in beside an air gap, reads the air: almost no water. Every health check passes it, since the reading is in range and still moves, and the advice would read the driest bed on the farm and ask for a full refill every morning.

Nothing in soil reads that dry. Below the surface only roots dry the soil, and they stop at its wilting point. The sun dries the top further, and FAO-56 takes that layer, the top 0.10 to 0.15 m, to dry "halfway between oven dry (no water left) and wilting point" (Rev.1 Eq. 7.5, the 1998 edition's Eq. 73). So a reading more than 0.03 below half the wilting point within the top 15 cm, or below the wilting point under them (soil_floor), 0.03 being a factory calibration's usual error, is not read. With no other probe the advice says why, and with the farm's station the bed is carried on the weather from the last reading that was in the soil:

bed 3's probe at 20 cm reads 2.0% water, drier than its soil can be (9.0% at least): it may be out of the soil, or in an air gap beside it. There is no advice for bed 3 today. Push the probe back in and firm the soil around it.

With another probe still in the soil, the advice comes from that one, and a note beside it says the same about the first.

So it can tell a probe in the air from dry soil only where the soil holds some water at its wilting point. In sand it cannot, and says nothing. That is the case of the openSenseMap LSE01 reading 2.1 to 2.4% (notes/test-data-sources): out of the soil in a loam, possible in a dry sand, and its keeper has not said which. Not built: the other sign, a probe that rises after neither the rain the gauge recorded nor the watering the valve did, which would catch it in sand too.

The trials of 2026-10-17#

Before the audit of 2026-10-18: the probe-led beds' refills went on within the hour and ran off, the advice read forecasts issued after it, and the crop did not make up for a dry top. Its figures stand as they were run.

trials.py on code 4475640: ten beds, each with the choices that bear on it, through McAllister's and Davis's recorded summers of 1991 to 2025, each summer on two farms drawn for it (backtest.VirtualFarm), the farm's water loss Penman–Monteith with the day's own wind; 5,320 backtests. "Tuned" is McAllister's 1991 to 2020, which the defaults were chosen on; "held out" the rest, nearly all Davis. Made by python3 trials.py report.

Each choice against the default it would replace, on the same site, summer and farm: the difference's mean and its 90% interval over the pairs (water and lost transpiration in mm a summer, hours a summer with under a tenth of the root zone holding air).

Measuring more, 2026-10-18#

Before the audit of 2026-10-18: the probe-led beds' refills went on within the hour and ran off, the advice read forecasts issued after it, and the crop did not make up for a dry top. Its figures stand as they were run.

The same trial on code c1de215, with what a farm might measure besides (trials.py's MEASURES): the station's wind (weather_backtest(wind=True)); the crop's water use, one number for the season measured to a lognormal 0.07 or 0.15 (use_measured); a tension sensor beside each probe, each sensor a tenth out for good and a twentieth each hour (backtest(tension=True)). Each crossed with the choices it bears on, on the same farms and summers, 9,380 backtests more; the 5,320 above were not run again, and three of them run again on c1de215 agreed to the bit. Each against the bed as built, measuring nothing more: the difference's mean and its 90% interval. Made by python3 trials.py report.

Three of the tomato summers by probes, run again to see each probe's field capacity (the soil's, plus the probe's own error, as the probe reads it), learned and measured:

The audit of 2026-10-18#

The simulator, the advice and the trials were each read by someone set to break them, and each finding checked before anything changed. What was found, and what was done:

The simulator#

• The crop did not make up for a dry top. Its use was its demand times its roots' Ks layer by layer, so a dry top layer cut the whole crop's use however wet the soil below was; the docstring said the deeper roots made it up, and they did not. Swapping in FAO-56's Ks on the root zone's mean, a Davis summer's tomatoes by probes lost 1.0 mm for 16.8. Now the roots in wetter soil make up for those in drier as Jarvis (1989) has it, as HYDRUS does (Šimůnek and Hopmans 2009), each farm's crop as far as a compensation drawn for it between much (0.5) and none (1) lets it (physics.LayeredBed.compensation, VirtualFarm.compensation).
• Evaporation took the top 25 mm first, below its residual water. Its suction was then without end, nothing rose into it from below, and the most roots went without. Now it is spread through the top 10 cm, the nearer the surface the more, and stops a little above the residual water.
• The crop's p was the table's all summer. FAO-56 lowers it on a day the crop draws fast, p + 0.04 (5 − ETc): at Davis's 8 or 9 mm a day, tomatoes' 0.4 is 0.2 to 0.28. The virtual crop's p now follows its day's demand (SoilBed.p_follows_demand).
• A rain day lost a sixth of its reference ET. The hours it rained had 15% of their share and the rest was lost, where the day's Penman–Monteith already counts its cloud; at McAllister, a station's balance counted 6% above the farm's water use for no other reason. Now the rain hours' share goes to the dry hours.
• Found and not changed: above field capacity water drains by a cubic rule, not Mualem's conductivity, and a soil at field capacity never drains further; the soil below the roots, at field capacity, gave the crop 60 to 86 mm a season; rain always soaks in, at 2.5 mm an hour; no canopy catches rain; watering is perfectly even; both start full. Each bears on the numbers below and is in UNTESTED.

The advice#

Each of these was a hand-built bed for which the advice was wrong; each is now a test (TestWhatTheAuditOf20261018Found).

• Rain forecast held a bed with 97% of its water gone, 11.5 mm against a 102 mm deficit, as long as the forecast kept sliding. Now rain holds a watering only if it would leave the bed short of half its water by tomorrow.
• A silent probe left the others to speak for the whole root zone: a dry top probe 12 hours stale over a wet one, and the bed had plenty. Now it stands for its slice at its last reading, less what the others have dried since, and the decision says so.
• A wet subsoil cancelled a dry top: two thirds of the top's water gone over water perched at 0.48, and it said too wet. Now each probe's deficit counts on its own.
• A probe in percent was read as a fraction, 15 as 15, and the bed never watered. Now a series not in a volume fraction is not one of the bed's probes.
• Read every five minutes, a probe had no drying rate: the window was a count of readings, 18 hours of them. Now it is hours, the readings thinned to one an hour.
• A reference drying faster than 1.5 times the balance was dropped, and the bed counted at the margin's 1.2. Now it is taken at 1.5.
• The weather balance's margin, and a probe's correction, scaled the crop's transpiration alone: for a young tomato the 20% margin was 3%. Now they scale the whole of its use.
• A day whose rain or watering filled the bed counted its whole use besides: the water came off first and stopped at full. Now the day's water and its use are taken together, FAO-56's Eq. 85, what the root zone cannot hold draining.
• The rain since midnight was in neither the balance nor the forecast. It comes off now.
• By the weather the advice looked no further than its refill point, where by probes it looks a day ahead. It does now, by the balance's last day.
• A cover the probes had seen held for the forecast alone, not for the balance that carries the bed when they fall silent. It holds for both now.
• A bed's layers were left out of the weather balance's available water. They are in.
• A bed configured past sense (no rain skipping a watering, a share of rain over one, a crop that uses nothing) gave advice. It is refused.
• Found and not changed: the soil above the shallowest probe is read as that probe reads (a probe at 3 cm cut one Davis summer's loss from 38 mm to 9, on the simulator before the audit; tried again below); a dead or misreporting rain gauge goes unseen; the refill reaches past the deepest probe to the roots' depth by assumption; the crop's days are counted in UTC. In UNTESTED.

The trials#

• Watering by probes went on within the hour, as much as 40 mm an hour on a soil that takes 8, and what it could not take ran off: 67 to 89 mm a Davis tomato summer, unscored, and none by the weather, watered at 12 mm an hour. Now both water at 12 mm an hour.
• Each morning's advice read forecasts issued after it: the season's issues were all in the log from its start, and the advice read the latest. Now it reads the forecast as it stood (State.forecast(as_of=...)).
• The forecasts all but never missed a day's rain: each hour's miss was drawn alone. Now a rain day in four is missed whole at a day's lead, and a dry day in twelve forecast a spell (backtest.forecast_skill).
• The station read the farm's own weather exactly, so by the weather the advice's only errors were the wind and the crop. Now each farm's station is off by its sensors and its siting (VirtualFarm.station_*).
• The lawn's and the yard's soils, given whole, did not depart. They do.
• Each summer's probe noise and forecasts were drawn from the draw alone, the same every summer. Now from the summer's farm.
• "Held out" was Davis, and has been spent: the fixes above were found looking at it. The defaults were chosen on McAllister and Davis and then tried, once, on two sites kept sealed until then, Bushland, Texas, and Tifton, Georgia.
• The scores: runoff, the soil's water at the end against the start, the crop's demand (for the share of it lost) and hours waterlogged (under a twentieth of air) are now scored. Hours under a tenth of air, a few points above field capacity in these silt loams, counted a watering's first day.
• Fine: the pairing, the code stamps, the forecast cache; resampling whole summers or five-year blocks widened the intervals by up to two-fifths and turned no conclusion.

The trials of 2026-10-18 and 19#

On the code after the audit: ten beds at McAllister and Davis, 1991 to 2025, one farm a summer (trials.py, its BEDS and MEASURES), 5,880 backtests on 51dffc9; then, its reading of the crop's use fixed, the hot-day refill point's probe configurations again (1,680, as committed in 735d41f); then, once, the sealed sites (1,330 on 735d41f). Made by python3 trials.py report.

How the defaults were to be chosen#

Written before the open sites' trial ran, and amended once, before the rerun, as below.

Groups: McAllister 1991–2020, McAllister 2021–2025, Davis. Pairs are the same summer's farm, each against the bed as built (margin 0.2, 2 m/s wind, no hot-day refill point, no bound on learning). "Lost" is lost transpiration, mm a summer.

1. Weather beds, no anemometer. Among {margin 0.1, 0.2} × {2 m/s, the wind's monthly normal} × {hot-day refill point off, on}, the default is the one using least water such that, in every group and for every weather bed, its mean lost is no more than the bed as built's + 0.5 mm, and the 90% interval of the difference reaches no higher than +1.0 mm. If none beats the bed as built, it stays.
2. Weather beds, with an anemometer. The same rule among margins 0.1 and 0.2 with the wind measured, giving the advice for a station that has one.
3. Probe beds. The hot-day refill point becomes the default if, at Davis, it cuts the mean lost of the crop beds (tomatoes, lettuce, dense pan) with the interval below 0, and in neither McAllister group raises any probe bed's mean lost by more than 0.5 mm. Learning within 0.05 becomes the default if, added to whatever (1)–(3) chose, no probe bed's mean lost rises by more than 0.5 mm in any group and it cuts the water or the hours waterlogged of at least half the probe beds.
4. Hardware. A tension sensor or a probe at 5 cm is recommended where it cuts the bed's mean lost at Davis by a third or more of the bed as built's.
5. Sealed. Bushland and Tifton are then run once: the bed as built against the defaults chosen, and the hardware recommended. Reported as they come; a default that does worse there than as built is said so and not quietly kept.

Amendment, 2026-10-19 01:07, before the hot-day refill point was tried again.

The open sites' trial showed rule 3 weighed only lost transpiration: the hot-day refill point, reading the crop's use from the probes' drying, took drainage for use in wet summers and more than doubled six McAllister tomato summers' water, nearly all of it drained. Its use is now read only from a day that began at least 0.02 below field capacity. Tried again on the open sites (its configurations alone), it becomes the default only if, besides rule 3, in neither McAllister group does it (with learning within 0.05, as chosen) add more than a tenth to any probe bed's mean water; else it stays off, and the bound alone is the default.

What the rules chose#

• By the weather, no anemometer: margin 0.2 and FAO-56's 2 m/s, as built. Margin 0.1 added 23 mm a summer to Davis's bucket's loss; the site's monthly normal wind at 0.2, 0.8 mm to McAllister's bucket's (interval to 1.7); the hot-day refill point at 0.2 passed but used more water.
• With an anemometer: margin 0.2, as built: at 0.2 the measured wind added 1.1 mm to McAllister's bucket's loss (interval 0.2 to 2.1).
• The hot-day refill point by probes: as first written it passed rule 3 and was set as the default; then its water was seen, more than twice the as-built in six McAllister tomato summers, nearly all drained, its reading of use having taken drainage for it. Its use read only from a day that began 0.02 below the lower of the learned and configured field capacity, it cost no water (the amendment), and cut tomatoes' lost transpiration at Davis by 1.8 mm (interval −3.5 to −0.3) and the dense pan's by 4.3 (−7.1 to −1.9), and lettuce's not (+0.1, −2.5 to +2.6): off.
• Learning within 0.05: on its own, it raised Davis's lettuce's mean lost by just over 0.5 mm (+0.5, −0.1 to +1.4): off. It was also seen, by the tests, to hold a bed whose survey is badly wrong short of its own soil.
• Hardware: a tension sensor for the lawn, the yard and lettuce, which it cut at Davis by 47, 15 and 4 mm of 55, 28 and 11.5; a probe at 5 cm for none.

The open sites#

Code: 51dffc9, 51dffc9+crop-drying. Each summer one farm.

The sealed sites#

Bushland, Texas, and Tifton, Georgia, run once, the advice as built against the hardware recommended. The tension sensor's help for the lawn held at Bushland (−13.4 of 23.5); for lettuce and the yard it did not, both losing 1 to 2 mm more at Tifton. An anemometer, not recommended for the calm open sites, cut every bed by the weather's loss at windy Bushland.

Code: 735d41f. Each summer one farm.