The strongest case against
Start with the objection that should win. An agronomist does not need a century of soil records to know that a field is running down. Yield trend lines, organic matter assays, a species–area-style relationship between hedgerow loss and pollinator decline: these are theory-driven predictions, written down from a handful of seasons, sometimes from none at all. The Rothamsted long-term plots have been running since 1843 and most of what they taught agronomy was extracted in the first fifty years, not the last fifty. If a model of soil degradation can be specified from structural knowledge — cation exchange capacity, tillage history, a known rate of organic carbon loss under continuous maize — then the binding constraint on foreseeing agricultural extinction debt is the quality of the model, not the duration of the watch. Better science, not longer intake. That is the claim to beat, and it should not be waved off.
Where the objection holds
It holds for the shape of the problem. Soil scientists can write the functional form of nutrient depletion under a given rotation without waiting to observe it happen on any particular farm. Extension services routinely warn a grower, from three years of NDVI decline and a known crop coefficient, that a field is heading toward a yield cliff before the cliff arrives. Theory does real, load-bearing work here, and it is dishonest to pretend agronomy is data-starved in the way population ecology sometimes is.
But theory supplies the form, not the coefficients. A model of soil organic matter decline under intensified rotation predicts that carbon falls toward a new equilibrium; it does not tell you, for a specific field with a specific texture, history and microbial community, whether that fall takes four seasons or forty. The equivalent of Ferraz's fiftyfold range in fragment half-life shows up here as the huge spread in reported "time to degradation" across soil types: sandy loams under continuous cropping can lose half their exploitable organic matter in under a decade, while a clay-rich soil with the same management can hold apparent stability for two or three times as long before the same collapse in structure and water-holding capacity arrives. The parameter that actually matters for an agronomist's decision — how long do I have — is not in the textbook. It is only in the record of that field.
The intervention window, and why it closes unseen
The characteristic failure in agriculture is not ignorance of the risk. It is timing. An agronomist commissions a soil health assessment, a satellite-derived NDVI trend review, or a nutrient budget reconciliation, and the request goes into a queue: a contracted lab, a scheduled flyover, a quarterly review cycle. Meanwhile the field is inside a narrow biological window — the point in a drought stress cycle where irrigation still rescues root mass, the point in a nitrogen deficiency where a top-dressing still corrects canopy development, the point in a soil compaction event where a single pass of subsoiling still restores infiltration before the next rain compacts it further. These windows are frequently measured in days. Assessment cycles are measured in weeks. The debt that accrues in a fragmented forest over decades accrues in a stressed field over a single growing season, and the mismatch between the hazard's clock and the institution's clock is structurally the same mismatch Diamond found in land-bridge islands, just compressed.
The commodity price stream adds a second, crueler version of the same failure. A soil moisture deficit that will not show up in NDVI for another ten days is already committing the crop to a yield shortfall; the futures market has not yet moved because the market is also waiting on a satellite pass and a crop progress report. By the time price catches up to the biology, and biology catches up to the visible canopy stress, and the visible stress prompts a scheduling request, and the scheduling request produces an assessment, the window in which any of it could have been acted on has closed several times over.
What a frozen or bounded view cannot supply
A model trained on an agricultural corpus frozen at some cutoff — extension bulletins, historical yield studies, soil survey archives — reports the standing condition of farmland as it was written up, which is nearly always a healthier condition than the one that has since accrued. It can recite the theory of nutrient mining under continuous cash-cropping fluently, and still be blind to the specific field that mined itself out three seasons after the last survey it was trained on.
A bounded scene view — one seasonal pass of satellite imagery, one soil sample set, one weather model run — does better than the frozen corpus, because it is current. But current is not the same as sufficient. A single NDVI snapshot of a field mid-drought-recovery looks identical to a field that recovered fully and a field that is superficially green but has already lost root depth it will not regain this cycle. The scene shows canopy; it does not show trajectory. Relaxation is invisible at the timescale of a look.
The theory already tells us the field will decline under this rotation. What is watched continuously merely confirms the arithmetic. Confirmation is not a new category of knowledge.
The answer is that the arithmetic has no coefficients until something is watched long enough to fit them, and the fitting has to survive changes in instrument. Soil labs change extraction methods; NDVI sensors are replaced with different spectral response curves; a new agronomist reads the same tissue test against different regional benchmarks than her predecessor used. A raw multi-year yield series stitched together across three different assessment vendors is not one measurement, it is three, and without provenance attached to each reading, nobody downstream can tell whether a trend is soil or instrument. This is the same drift problem that dogs relaxation-fauna studies, and it does not go away by observing less. It is addressed only by keeping the record of how each reading was taken attached to the reading itself, so that when a lab recalibrates its phosphorus test in 2019, every prior value can be re-weighted rather than silently contaminating the trend line.
The reserving argument, and its agricultural failure mode
There is a plausible institutional answer that does not require continuous streams at all: insure and reserve. Crop insurance already prices drought and flood risk from historical loss tables without watching every field. Set aside a management reserve for input costs against soil decline, the way an actuary reserves against a known hazard distribution, and the timing problem becomes a pricing problem, solved once, upstream.
This works exactly as well as the hazard distribution is stationary, and agricultural hazard distributions are not. Actuarial tables built from twenty years of drought-loss data are being outrun by shifting rainfall variability faster than most insurers can revise premiums; several regional crop insurance schemes have required repeated upward revision of loss reserves through the 2010s as extreme-event frequency moved outside the range the tables were built on, echoing the asbestos reinsurers who set 1970s reserves against a latency distribution that then widened under them for two more decades. Reserving is a bet on a known distribution. Extinction debt in soil and water systems is precisely the case where the distribution is not known until it has been lived through, because the relevant hazard — cumulative structural loss under a changing rotation and a changing climate — has no closed historical analogue to reserve against.
The narrower claim
None of this makes the agronomist's theoretical toolkit useless, and none of it demands infinite data. Species–area logic, soil carbon kinetics, crop water stress models: these give the functional form, and a short record can often confirm or falsify that form quickly, which is genuine, cheap knowledge. What a bounded or frozen intake cannot do is supply the coefficient that matters for the decision at hand — how long until this specific field's debt comes due — because that coefficient is temporal, field-specific, and only recoverable from a record that spans the interval it is being asked to explain. The Large Universe Model's claim on this domain is narrow: keep the soil sensor, the NDVI pass, the weather model and the price stream open continuously, tag each reading with its provenance so instrument drift can be corrected rather than mistaken for trend, and the intervention window becomes visible before it closes rather than after. That is not a claim that agronomy needs no theory. It is a claim that theory alone cannot tell an agronomist which Tuesday is the last day the subsoiler will still work.