What arrives
An agronomist covering a few thousand hectares of mixed arable land is, at any hour, on the receiving end of four distinct streams. Soil moisture and conductivity probes report every fifteen minutes from fixed points in the field, each one accurate for a radius of perhaps twenty metres and silent about everywhere else. Satellite NDVI passes overhead every two to five days depending on cloud and constellation, giving a canopy-vigour picture that is spatially complete but temporally stale the moment it downlinks. A weather model — often several, blended or not — issues a rolling forecast of rainfall, wind and temperature that revises itself every six hours and is frequently wrong about the next 48. Commodity price feeds update by the minute and say nothing about the crop at all, only about what a decision now will be worth later.
None of these streams stops. None of them was ever going to be reconciled once and filed. That is the condition a Large Universe Model is built for, and it is worth being precise about what "held" means here, because the naive answer — keep a single up-to-date picture of the field — is not available.
What is held
The store is not a map of the field. It is a set of provenance-tagged claims, each with a timestamp, a source, and a confidence, and crucially, no requirement that they agree. A patch of the north block might carry, simultaneously: soil probe P4 saying moisture at 18% and falling (measured nine minutes ago), NDVI saying canopy stress is minimal (measured three days ago, before the last dry spell), and the weather model saying 12mm of rain is expected within 30 hours (issued four hours ago, second revision today). None of these is discarded to make room for the others. Each is a testimony, held as told-true against its own source, and the system's state for that patch is not one number but this bundle of testimonies plus the fact that they do not currently cohere.
This is Belnap's four-valued move at field scale. The proposition "the north block needs irrigation in the next 24 hours" is not simply true or false in the store. It can be told-true by the probe, told-false by the forecast, and the system's job is to carry both flags rather than collapse them into an average moisture-and-rain-adjusted figure that no single source actually asserted.
What triggers revision
Revision is not scheduled; it is provoked. A new soil reading that contradicts the previous fifteen-minute trend triggers a re-check of that patch specifically — not the whole field, not the whole model. A forecast revision that drops the expected rainfall from 12mm to 2mm triggers a re-check of every irrigation decision currently deferred on the strength of that forecast, and only those. A price move on the futures market triggers a re-check of the marginal-value calculation behind any intervention still open for decision, because the cost of delaying now buys or loses a different amount of money than it did an hour ago.
The mechanism that makes this affordable is containment. Each trigger operates inside a bounded region of the belief graph — the patch, the block, the set of decisions dependent on this forecast run — and does not propagate outward by default. A contradiction between the probe and the forecast for the north block is a local condition. It does not, and under a paraconsistent design must not, cast doubt on the unrelated reading of soil pH in the south block three kilometres away. Under classical closure that containment is not available: once P and not-P are both asserted anywhere in the belief set, anything is derivable everywhere, including conclusions about parts of the farm the contradiction never touched. That is explosion, and a working farm cannot run a decision system that is one disagreeing sensor away from useless.
What the operator sees
The agronomist does not see a merged number. She sees the bundle, ranked and dated, with the disagreement made explicit rather than smoothed. The display for the north block does not say "moisture: 14%, adjusted." It says: probe P4, 18% and falling, nine minutes old; forecast, 12mm expected, four hours old, second revision; these two claims are in tension; the system's provisional recommendation, given the tension, is to irrigate now against the pessimistic reading and revisit if the rain confirms.
That last clause is the discipline doing its work. A well-built paraconsistent store does not treat the contradiction as license to shrug. It still wants resolution, still ranks the probe's direct measurement above the forecast's probabilistic claim for an imminent decision, and still flags the unresolved tension as a defect worth investigating rather than an acceptable steady state. The agronomist's judgement is recorded alongside the machine's flag — she is the adjudicator, not a downstream consumer of an already-laundered answer. If she overrides the recommendation, that override becomes another provenance-tagged claim in the bundle, available to whoever checks this decision later.
Just average the two. A 12mm forecast against an 18%-and-dropping reading nets out to something — irrigate at half rate, say, and move on. Precision is false comfort here; farming has always run on rough numbers.
The trouble is that half-rate irrigation is a decision nobody's source actually recommended. The probe, taken alone, says irrigate fully now. The forecast, taken alone, says wait. The average is a third position invented by the system, unattributable to either instrument, and if the rain fails to arrive — which second-revision forecasts do, routinely — there is no source to interrogate about why. Averaging spends the provenance to buy a number, and the number is wrong exactly as often as forecasts are wrong, with none of the audit trail that would tell you which stream to trust more next time.
What it costs
The characteristic failure in this domain is not disagreement itself; it is latency in resolving disagreement while an intervention window is finite. A fungal risk model flags conditions favourable to blight based on humidity and temperature trends. The window for an effective fungicide application is measured in one to three days before infection establishes. If the assessment of that flag — reconciling it against the latest NDVI pass, the latest forecast revision, and the agronomist's own walk of the field — is scheduled as a discrete task rather than run continuously against live streams, the window can close while the ticket sits in a queue waiting for someone to look at conflicting inputs and decide.
This is exactly where the classical escape routes fail in practice. Stopping intake until the disagreement resolves means waiting for the next satellite pass, which may be four days away — the window is gone. Discarding one stream — say, trusting NDVI and ignoring the ground probe because satellite data feels more authoritative — throws away the more current, more local signal for the sake of tidiness, and does so silently, so nobody later can ask why the probe reading was ignored. Averaging manufactures a spray-or-don't-spray threshold that neither the humidity model nor the visual canopy data actually supports.
The paraconsistent alternative costs something too, and it should be named plainly. It costs the agronomist attention: she is shown the tension, not a clean verdict, and adjudicating it is work that a false average would have hidden from her. It costs the system disjunctive certainty — it cannot always tell her flatly "spray" or "don't," because sometimes the honest state of the evidence is "spray now against the pessimistic reading, and here is what would change that." What it buys is a decision made inside the window, attributable after the fact, and auditable when the harvest comes in and someone asks why this block was treated differently from the one next to it.
The objections that matter here
The natural challenge is that a Bayesian farm-management model handles all of this by assigning credences — 70% chance the humidity trend is real, 30% it is a sensor artefact — without needing any talk of contradiction. That works cleanly when both readings share a model of what they are measuring. It works less cleanly when the ground probe and the satellite pass are not expressing uncertainty about the same event but reporting on different event spaces entirely: one measures soil water at a point, the other measures reflected canopy light averaged over a pixel that may contain three crop varieties and a drainage ditch. Blending them into a single credence erases which instrument said what, and an agronomist defending a spray decision to an insurer or an auditor needs exactly that provenance back.
The second challenge is more practical: agricultural data pipelines already merge conflicting feeds constantly, using simple precedence rules — trust the most recent reading, trust the higher-resolution sensor. This is real, and it is often the right call when the merge is cheap and the stakes are low, like reconciling two slightly different rainfall totals for a weekly report. It stops being the right call when the merge is a substantive agronomic judgement — spray or wait, irrigate now or risk waterlogging tomorrow — where the correct resolution depends on information the pipeline does not yet have. A precedence rule chosen in advance cannot know that. A paraconsistent store that defers the merge, flags it, and lets the agronomist adjudicate with full sight of both testimonies is doing the part precedence rules cannot: holding the judgement open until it is actually safe to close.