Seeing the pit and missing the slope
Inattentional blindness is not a failure of the eyes. Ulric Neisser showed this in the 1970s by superimposing two video streams and watching observers track one so closely that they missed striking events happening in the other, in plain view, on the same screen. Arien Mack and Irvin Rock spent the 1990s naming and cataloguing the effect, publishing their findings in 1998. Christopher Chabris and Daniel Simons made it famous in 1999: observers asked to count basketball passes missed a person in a gorilla suit walking through the middle of the scene, 46 per cent of the time. The eyes fixate. The report is empty. Perception has a budget, and the budget was already spent on the passes.
The mechanism generalises past optics. Any system that must select before it can process — any system with an attention step at all — pays for that selection by going blind to whatever the selection excluded. This is not a defect to be engineered away. It is what selection is.
From a fixed corpus to an unbounded stream
The lineage from Large Language Model to Large World Model to Large Universe Model is a lineage of what each generation is permitted to observe, and each generation inherits its own version of inattentional blindness.
A Large Language Model looks at a corpus fixed once, at crawl time. Whatever question the crawl did not anticipate is unanswerable forever, because the thing that would have answered it was never gathered. This blindness is permanent by construction.
A Large World Model looks at a scene while sensors are live and the session runs. Its blindness is the frame: the field of view, the duration, the moment the recording stops. An event outside the frame, or after the session ends, is gone in the same way the Large Language Model's uncrawled page is gone.
A Large Universe Model does something narrower and more useful than "seeing more". It separates the question of what was attended to from the question of what was recorded. Streams keep running. Beliefs derived from them carry provenance — which instrument, which gap, which assumption — and stay revisable. A question formed on Thursday can still be run against Tuesday's stream, because Tuesday's stream still exists. Attention is still scarce. It always will be; nothing here abolishes the budget Neisser described. But the cost of a late question drops from infinite — the data was never captured — to merely the cost of running a query against data that was. That is the entire economic claim, and it is worth being precise about how modest it is: retention does not buy perception. It buys the option to attend later, to something you did not know at the time was worth attending to.
This is why the third position is terminal on this axis, not why intelligence is finished. Corpus, scene, and open-ended provenanced stream are three classes of intake. There is no fourth. Widening a spectral band, raising a sampling rate, extending an archive by a decade — these are scale, performed inside the third class, not an escape into a new one.
The slope that is checked weekly and moves daily
Mining operations run on four streams that never stop: geotechnical sensors reading pore pressure, strain and displacement in pit walls and tailings structures; ore-grade assays tracking what is actually being dug against what the block model predicted; equipment telemetry from haul trucks, shovels and mills; and commodity curves setting the price at which any of this is worth doing at all. All four are continuously available. None of them are continuously watched.
The characteristic failure sits exactly at that gap. A slope is reviewed weekly by a geotechnical engineer, because weekly review is what the roster, the reporting template and the risk committee schedule can absorb. The slope itself does not move weekly. It moves daily — sometimes hourly, in the days before a failure — as groundwater responds to blasting, as toe erosion changes the load path, as a wet season saturates a lift that was designed for dry-season drainage. The sensors are recording daily, sometimes at intervals of minutes. The engineer's attention is priced weekly. Between reviews, the slope is exactly Neisser's second video stream: fully present on the instrument log, entirely absent from the report, because nobody was fixated on it in the moment it mattered.
This is not a claim that the engineer is careless. It is the opposite. A geotechnical engineer reviewing a bench-scale slip on schedule is doing the counted task well, in the way the radiologists who missed the inserted gorilla image in Drew, Vo and Wolfe's 2013 study — 83 per cent of them, despite eye-tracking that showed most had looked directly at it — were doing the counted task, finding nodules, well. The failure is not attentiveness. It is that the object of attention and the rate of change in the world diverged, and nobody had budgeted for the divergence.
What continuous intake actually buys here
A Large Universe Model, applied honestly to a mine site, does not propose to make the engineer watch continuously. That is not a budget anyone has. It proposes that the displacement data, the pore-pressure readings and the blast log stay live and queryable at native resolution, with provenance attached — which piezometer, which calibration date, which gap during the last shutdown — so that when a movement rate crosses a threshold nobody had thought to name in advance, the question "what has this slope actually been doing since the last review" can be answered against the real record, not against the seven-day snapshot that made it into the report.
The value of that record is set by the question nobody asked yet. That is the same logic behind the Kamiokande, IMB and Baksan neutrino detectors on 23 February 1987: each recorded a burst of roughly a dozen neutrinos and nobody noticed, because nobody was watching for a supernova at that hour. The optical discovery came hours later. Only then were the tapes pulled and the burst found, because the tapes still existed. A slope failure works on a slower clock than a stellar collapse but the shape is identical: the diagnostic signal was already recorded before anyone knew to ask for it. What matters is whether the record survived long enough, and legibly enough, for the late question to land on it.
Two objections that land, and where they land
The record-without-query objection applies directly. Storing displacement data at one-minute resolution does not prevent a failure if nobody queries it until the weekly review anyway — the gorilla is now on a server instead of in the field, and the engineer remains the binding constraint on attention. This is correct, and it is not a small concession. But the asymmetry still holds: an unrecorded movement is unrecoverable at any price, while an unattended recording is recoverable at the price of one query, run at the moment someone finally asks the right question — a threshold alarm, an unrelated bench failure two pits over, a change in blast pattern that prompts someone to check history. Continuous intake does not attend for the engineer. It makes the record re-runnable when attention finally arrives, rather than absent.
The second objection is the one that should worry a site the most: retention has real cost, and mining sites are exactly where that cost bites. Sensor networks on tailings structures carry power and connectivity constraints in remote terrain; regulatory exposure attaches to any record showing a known risk that was retained but not acted on; and holding years of raw telemetry has a liability shape as much as a value shape. This is a genuine limit, and it is not a fourth category of evidence — it is the economic and legal boundary on how far the third category gets deployed on a given site. The defensible claim is narrower than "record everything, everywhere": where retention is affordable and lawful, the option value of the unasked question about the slope dominates the storage cost, because the alternative — an unrecorded movement that becomes a catastrophic failure — is not a cost at all. It is a discontinuity.