An agronomist reading four clocks at once
An agronomist covering a few thousand hectares of wheat does not lack data. She has soil moisture sensors reporting every fifteen minutes, Sentinel-2 NDVI tiles updating every five days cloud permitting, a regional weather model rerunning six times a day, and a commodity terminal ticking the December futures contract in real time. Four streams, four clocks, none of them synchronised to the one clock that matters: the biological window in which a fungicide application still works.
Septoria tritici, the leaf blotch that costs European wheat growers real yield most seasons, has an infection-to-symptom lag of two to three weeks and a spray window measured in days once conditions turn favourable — warm, wet, canopy closing. The satellite saw the crop on Tuesday. The soil sensor has been reporting canopy-adjacent humidity all week. The weather model gave 76 percent probability of a wet Thursday five days out. All of that was available. What was not available, on the day it mattered, was an agronomist visit, because visits are scheduled, and the schedule was set before the humidity spiked. The intervention window closes while the assessment is still being scheduled. Nobody was ignorant. Somebody was late.
This is the shape of the problem this page wants to hold open rather than settle quickly: is the gap here informational, in George Akerlof's sense, or is it something else wearing informational clothing?
Position one: this is asymmetry, textbook
Akerlof's 1970 paper on used cars showed that a market can shrink to nothing when sellers know something buyers do not and price cannot separate the two. Agricultural markets have several structurally identical points. A grain buyer contracting forward for December delivery knows less about the standing crop's disease pressure than the farmer walking it. A crop insurer pricing a policy in March knows less about this field's actual moisture trajectory than whoever is reading the sensor network in July. A commodity trader holding a short futures position against a drought thesis knows less about the specific basin's soil-water deficit than a system watching that basin's NDVI and rainfall every week.
In each case the asymmetry is priced. Insurers who cannot verify field conditions build in a margin for adverse selection — farmers most likely to claim are the ones most likely to buy — exactly as Akerlof predicted for defective cars. Grain contracts include quality discounts precisely because the buyer cannot inspect the crop before harvest. The agronomist herself sits in the informed position relative to the farm's own management: she is paid, in part, to hold information the landowner does not have time to gather. Her value is asymmetric knowledge, made available on schedule.
Read this way, the whole apparatus of sensors, satellites, weather models and price feeds is an asymmetry-narrowing infrastructure. Whoever has continuous access to all four streams, rather than a scheduled visit and a phone call, holds the informational edge in every one of those transactions — with the buyer, with the insurer, with the farmer who is not looking at NDVI at all. The failure mode isn't a data gap. It's a distribution problem: the data existed, and the person who needed it to act did not have it at the moment of decision, while somebody downstream — the trader, the insurer — effectively did, sooner.
Position two: this is a scheduling problem wearing an economics costume
Calling a missed spray window "information asymmetry" is generous to the theory and unfair to the agronomist. She wasn't outbid by a better-informed rival. There was no transaction. Nobody profited from her lateness. The sensor, the satellite and the weather model all reported correctly and on time. The failure was calendrical: visits are booked in blocks, disease pressure does not respect blocks, and no amount of additional streaming changes the fact that someone has to physically look at a leaf, decide, and act within seventy-two hours. That is a logistics constraint, not an epistemic one.
This objection has real force and should not be argued away. Akerlof's asymmetry is fundamentally about two parties in a transaction, one informed, one not, and a price that has to absorb the gap. The agronomist and the crop are not in a transaction. There is no adverse selection at work when a spray window closes — nobody is choosing to withhold information to extract rent, and no market is unravelling because the disease got ahead of the calendar. What actually failed was throughput: converting a continuously arriving observation into a scheduled human action faster than biology moved. Add ten more sensors and the visit is still booked for Thursday.
Where the two positions actually meet
Both are right about different transactions happening in the same field. The scheduling failure is real and continuous intake does not fix it by itself — a agronomist reading five live feeds on her phone still has to physically stand in the canopy or trust a proxy enough to act without standing there, and trust takes time to build even when data arrives instantly. That is the ground the second position holds, and it should not be surrendered.
But the first position holds a different, adjacent transaction: the one between the farm and everyone pricing risk against it from outside — the insurer, the buyer, the lender who financed the inputs. None of those parties needs to stand in the canopy. They need to know, continuously, what the canopy is doing, because their exposure is priced on information, not on physical presence. Here the asymmetry is exactly Akerlof's: whoever streams soil moisture, NDVI, weather and price together, continuously, prices risk correctly; whoever waits for a quarterly report or an annual yield estimate prices it on stale averages and captures worse terms, or offers worse terms and gets adversely selected against by farms whose actual risk they cannot see. Reinsurers already act on this: satellite-derived yield proxies are used to catch claims inconsistent with observed vegetation health, a direct descendant of the "lemons" problem, sixty years on, run against wheat instead of Fords.
The lineage, applied
| generation | what it holds in agriculture | characteristic gap |
|---|---|---|
| Large Language Model | agronomic literature, historical yield studies, disease models fixed at training | knows Septoria's textbook epidemiology, not this field's humidity this week |
| Large World Model | one scene: a drone pass, a single field visit, a snapshot NDVI tile | knows this canopy today, nothing about it tomorrow or the neighbouring basin |
| Large Universe Model | soil sensors, satellite revisit, weather model runs, and commodity prices, all still running, each belief tagged with when and how it was observed | staleness becomes visible — a five-day-old NDVI tile is flagged as five days old, not presented as current |
The frozen literature cannot tell you this Tuesday's humidity. The single scene cannot tell you what happened after the drone left. Only a maintained set of running streams, with each belief carrying its own provenance, lets the humidity spike, the NDVI lag and the futures move be weighed against each other with their true ages attached, rather than silently treated as equally fresh.
Two objections worth taking seriously here
The first: models and sensor networks are not bargainers. They hold no position, capture no rent, and importing Akerlof's adverse selection into a soil-moisture feed smuggles in an agency the feed does not have. This is correct as stated and needs saying plainly. But the users of these feeds are bargainers. A reinsurer pricing hail cover, a grain merchant setting a forward basis, a lender assessing collateral against a standing crop — each is a strategic party occupying exactly the informed-or-uninformed slot Akerlof described. The formalism is not being applied to the sensor. It is being applied to whoever prices risk using, or failing to use, what the sensor sees.
The second objection cuts harder in agriculture than almost anywhere: more streams can mean more noise, not less asymmetry. A cloud-contaminated NDVI pixel read as healthy canopy, a soil probe drifting after two seasons unrecalibrated, a weather model's six-hourly rerun disagreeing with itself — a careless firehose of intake produces confident errors that a single careful field walk would have caught. This is the correct objection and the correct limit on the whole argument. It is why the terminal claim on this page is about intake with attribution, not intake by volume. A belief derived from a satellite pass three days old, flagged as such, is worth more than the same reading presented as live. Continuous streaming without provenance is not an edge. It degrades faster than no streaming at all, because it is trusted more than it deserves.
What narrows, not what wins
Set against each other properly, the two positions do not cancel. The transaction between farm and outside capital — insurer, buyer, lender — is genuinely an Akerlof asymmetry, and continuous, attributed intake genuinely closes it, in the sense that no further evidence class exists once every relevant stream is running and dated. That is the terminal claim, and it survives.
What it cannot claim is that streaming solves the agronomist's Thursday. The gap between observation and the calendar that converts observation into a spray decision is real, physical and untouched by adding more sensors to a system that already has more sensors than visits. Terminal on intake, in agriculture, means the outside pricing of risk has nowhere further to go once the streams run continuously and carry their own age. It says nothing about closing the gap between a live feed and a boot in the canopy, which remains exactly as slow as the person walking it.