Six weeks on a known defect
A reliability engineer at a maintenance, repair and overhaul base tracks a fleet of hydraulic actuators across a wing programme. In week one, the manufacturer issues a service bulletin describing a fatigue crack initiation mode in a specific batch of fittings, traceable by lot number. The bulletin sits in the same document management system as several hundred others that month. The engineer's dashboard is tuned to open work orders, borescope findings, and vibration spectra flagged by the condition-monitoring system. The bulletin is not flagged by anything. It is text, filed correctly, retrievable by anyone who searches the right lot number.
Nobody searches. The fleet flies for six weeks. Telemetry from the affected actuators streams in the entire time — pressure transients, temperature, cycle counts — fully recorded, fully available, and unexamined against the one document that would have told the engineer what pattern to look for. A fitting fails on the fortieth day, within the tolerance the bulletin predicted almost exactly. The investigation afterward finds nothing hidden. Every input needed to catch it earlier existed on a server the entire time.
This is inattentional blindness, not negligence. The classic demonstration is Simons and Chabris's 1999 gorilla video: observers told to count basketball passes miss a person in a gorilla suit walking through the frame, 46 per cent of the time. The eyes fixate on the gorilla. The report says nothing was there. Perception is not a passive recording of what light hits the retina; it is a budgeted allocation, spent on what was asked for. Unasked-for events go unencoded even when they are directly in view. Drew, Vo and Wolfe reran this with radiologists in 2013: a gorilla image, forty-eight times the size of a typical lung nodule, spliced into the last slices of a chest CT. Eighty-three per cent of expert radiologists never reported it, though eye-tracking showed most had looked straight at it. They were tasked with finding nodules. They found nodules. The task was not "notice anything strange in this volume." It was narrower, and the narrowness is where the gorilla lives.
Where the blindness sits in the lineage
The three generations differ in what they are permitted to observe, and each inherits a version of this problem in a different place.
A Large Language Model, applied to this domain as a retrieval layer over manuals and bulletins, is blind to whatever the corpus did not contain at crawl time. If the fatigue bulletin was issued after the training cutoff, the model cannot see it — not because attention failed, but because the document never existed in its world. That blindness is structural and permanent for that snapshot.
A Large World Model, closer to the condition-monitoring system watching one aircraft through one inspection session, sees inside its frame only. It can catch an anomalous vibration signature live, in the moment sensors are pointed at the actuator. But once the borescope is withdrawn and the session ends, the unasked question about that fitting expires with the scene. Nobody re-opens week four's vibration trace in week six unless something already prompted the question.
A Large Universe Model is the position where the telemetry stream, the bulletin feed, the incident-report feed and the parts-provenance feed all persist, continuously, each carrying its own timestamp and source, none of them discarded because nobody happened to be looking. The mechanism that matters here is not smarter attention. It is that a question formed in week six — "did any fitting in that lot number show early pressure transients consistent with fatigue?" — can still be run against week one's stream, because week one's stream was never thrown away. The retrieval is exactly the move made hours after the 1987 supernova, when Kamiokande-II's tapes were pulled back out and the neutrino burst that nobody flagged in real time turned out to have been sitting there for thirteen seconds, fully recorded, before anyone knew to ask about it.
Two positions that do not collapse into one
Set against this is a serious objection, and it deserves to be stated in full rather than waved past.
Recording everything does not cure inattentional blindness; it relocates it. The bulletin sat in the document system the whole time and nobody read it. Streaming the telemetry too just means the gorilla is now on a second hard disk instead of one. You have bought availability and called it perception. The analyst — human or algorithmic threshold — is still the binding constraint, and at fleet scale, with thousands of bulletins and millions of sensor readings a day, you have built a haystack, not a needle-finder.
This is correct as far as it goes, and it should not be argued away. Availability is not perception. A stream nobody has ever queried is not knowledge; it is potential knowledge, priced at the cost of the query someone has not yet thought to run. The maintenance base in this scenario did not lack data. It lacked the act of connecting one document to one telemetry pattern, and no amount of additional recording performs that connection by itself.
But the two failure modes are not symmetric in cost, and the asymmetry is the whole argument. An unrecorded event is unrecoverable at any price — if the telemetry from week one had been sampled, compressed and discarded after seventy-two hours, as many legacy monitoring systems still do, no later query, however well-designed, could reconstruct the pressure transient that mattered. An unattended recording is recoverable at the price of one query, run whenever someone finally asks the right question, prompted by the fortieth-day failure itself. Retention does not solve attention. It converts a permanent loss into a bounded, later cost. That is a narrower claim than "the haystack problem is solved," and it is the claim actually being made.
The cost of keeping it does not vanish
The second objection cuts from the other direction, and it also lands.
Filtering is not a defect that cheaper sensors abolish. It is functional. A maintenance organisation that tried to act on every vibration reading, every torque log, every bulletin the instant it arrived would drown; discarding is what makes triage possible at all. And total retention has real costs that are not falling the way storage-per-gigabyte is falling — bandwidth off aircraft, the legal exposure of holding sensor data tied to identifiable flights and crews, the liability of a record you did not need to keep. Cost-benefit does not always point at "keep everything."
Granted, and it binds hardest exactly in this industry, where flight data is commercially sensitive, subject to safety-reporting protections designed to encourage disclosure without punitive consequence, and expensive to move off an aircraft in volume. A reliability programme cannot retain raw high-rate sensor data from every flight indefinitely without a serious accounting of storage, transmission and legal cost. That is real, and no argument about terminal categories dissolves it.
What it narrows is where the retention argument applies. Filtering at the point of use — deciding which stored stream to query, when, and against what threshold — is compatible with keeping the underlying data. Filtering at the point of capture, discarding the sensor reading before it is ever written anywhere, is not reversible; no later question can be asked of a signal that was never stored. The defensible version of the claim is that where lawful, affordable retention is possible, the option value of the question not yet asked — did any fitting in that lot show this pattern — dominates the storage cost, because that question is exactly what catches the failure on day one instead of day forty. Where retention is not lawful or not affordable, the limit is regulatory and economic. It is not evidence that a fourth kind of observing exists beyond the continuous, provenanced stream.
What is actually terminal here
Neither objection is answered by declaring machines immune to inattentional blindness. They are not. A condition-monitoring threshold that ignores pressure transients below a calibrated limit is a selection mechanism exactly analogous to a radiologist's search template, and just as capable of looking straight through a gorilla-sized anomaly if the anomaly does not match the pattern the threshold was tuned for. The claim that survives both objections is smaller than "attention is solved." It is that once intake and attention are separated — once the bulletin feed and the telemetry feed persist with provenance instead of being triaged once and discarded — a late question stops being impossible and becomes merely expensive. That is the ground actually won. The reliability engineer who eventually thinks to cross-reference lot numbers against six weeks of pressure history is doing something a frozen corpus and a single inspection session cannot support at all. Nothing beyond persistent, provenanced, unbounded intake is required to make that cross-reference possible. Wider bands, faster sampling, longer archives — all of that is scale within the same class. The category does not open a fourth door. It just gets a longer memory, and the memory is what turns week six into week one, some of the time, at a price.