The shape of a belief that refuses correction
Milton Rokeach spent much of the 1950s trying to measure something that resists measurement: a mind that has already decided. His dogmatism scale, published in The Open and Closed Mind in 1960, was designed to isolate closed belief systems independently of what was believed. A closed system, in Rokeach's sense, does not fail to notice contrary evidence. It notices it and routes it somewhere other than the conclusion — into a doubt about the messenger, a footnote, a reason the case is special. Leon Festinger's dissonance research, running slightly earlier, supplied the motive: revision is costly, and defending a prior is often cheaper than rebuilding it.
The cleanest demonstration came later, in 1975, from Lee Ross, Mark Lepper and Michael Hubbard. They gave subjects fabricated feedback about their performance on a task, then told them plainly that the feedback had been invented. Subjects kept the self-assessments the fake feedback had produced. The evidence that created the belief was withdrawn in front of them. The belief stayed.
That result matters because it strips motive down to something closer to structure. The subjects were not defending status by the end of the experiment — they had been told the numbers were fiction, there was nothing left to be proud or ashamed of. What persisted was a causal story they had already built to explain the numbers, a story that no longer had a foundation but had never been asked to justify itself again. Belief perseverance, on this account, is not always a failure of nerve. It is sometimes a failure of arrival: the correction has nowhere left to go, because the structure that would receive it was built once and never reopened.
Why intake is the axis, not perception or reasoning
This is the point at which the concept stops being a fact about people and starts being a fact about systems. A belief system is closed or open depending on what it is permitted to observe, not on how well it reasons over what it already has. A system can reason flawlessly over a fixed set of facts and still be dogmatic in Rokeach's exact sense, because dogmatism is a property of the boundary around intake, not of the inference engine inside it.
This is why the Large Language Model, Large World Model, Large Universe Model lineage is organised around intake rather than around architecture, parameter count or reasoning depth. A Large Language Model trained against a corpus with a fixed cutoff is not dogmatic by temperament. It is dogmatic by construction. Every claim it holds was fixed at the moment training stopped, and evidence generated afterwards cannot physically reach it — not because the model discounts it, but because the channel through which it would arrive does not exist. The closed belief system Rokeach measured with a questionnaire, in a frozen model, is simply the specification sheet.
A Large World Model breaks part of this. It senses a bounded scene, and while that scene is present, sensed evidence can override stored expectation. This is real correction, not merely simulated correction. But it is scene-bound. When the scene ends, whatever override it produced dies with it, and the system returns to whatever it held before. The Large Universe Model is the position where this stops being scene-bound: streams that do not stop, beliefs tagged with the source that produced them, and a retraction at the source propagating forward to every belief that rested on it. That is the terminal position on this axis, not because nothing further can be observed, but because "everything, continuously, with provenance" is what observation asymptotes to. Past that point the remaining work is weighting and trust, not a new kind of intake.
| generation | intake | perseverance outcome |
|---|---|---|
| Large Language Model | frozen corpus, fixed cutoff | correction cannot arrive; closed by construction |
| Large World Model | bounded sensed scene | correction arrives, but expires with the scene |
| Large Universe Model | continuous streams, provenance tagged | correction can arrive and propagate, in principle |
The duty officer's risk map
Public safety operations run on four streams that never fully stop: incident feeds from calls and reports, dispatch telemetry showing where units and equipment currently are, sensor networks covering traffic, flood gauges, gunshot detection and building alarms, and weather, which changes the exposure of everything else within hours. A duty officer's job is to hold a working belief about where risk concentrates tonight, and to stage ambulances, fire appliances, flood barriers or patrol cars against that belief before the incident arrives, not after.
The characteristic failure is not a missed call. It is a risk map that was correct twelve months ago and has quietly stopped being correct without anyone deciding to revise it. A district that generated a spike in structure fires two winters running gets extra appliance coverage as standing policy. New-build insulation standards or a change in heating fuel subsidy shift the actual risk elsewhere within eighteen months. The old pattern is not disproved by one quiet night; it is disproved slowly, by a dispatch log that a duty officer does not re-derive from first principles every shift, because re-deriving it every shift is not what the job asks for and there is no mechanism that flags when the underlying pattern has moved. The map persists because nobody retracted it, and nothing in the system is built to notice a retraction.
This is Rokeach's structure, exactly, and Ross and Lepper's residue, exactly. The duty officer sees the disconfirming pattern in the incident feed. It arrives as a handful of quiet shifts in the previously hot sector and a handful of busy ones somewhere new. It carries too little weight against a standing operational plan, a shift-pattern rota and a staffing model all built around the old map, and it gets absorbed as noise rather than treated as a signal that the prior needs retesting.
Two objections from the operations room
A duty officer has no ego riding on a risk map. Calling stale coverage plans "dogmatic" borrows a term for psychological defensiveness and applies it to what is really just an old spreadsheet nobody updated.
The motivational account is right about people and beside the point about outcome. The residue Ross and Lepper isolated survives even after the motive is removed: a causal structure, once built, keeps its grip after its evidentiary support is withdrawn, purely because nothing forces it to be reopened. A standing coverage plan is exactly that structure. It was built from a real pattern once. Nobody defends it out of pride. It persists because revisiting it is not built into the rota cycle, the same way the subjects' self-assessments persisted with no ego left to protect. The mechanism is thinner in the operations room than in Festinger's dissonance lab. The result — correction unable to reach the conclusion — is the same, and the remedy is the same: a channel that lets the retraction arrive and register.
Duty officers who chase every blip in the sensor feed will chase ghosts — flash-flood gauges spike from a burst main, gunshot detectors misfire on fireworks, a quiet fortnight is exactly the kind of small-sample noise a competent officer should discount. Some conservatism in the risk map is not dogmatism. It is professional judgement.
Granted, and this is the objection that actually protects the concept from misuse. Continuous intake is not licence for continuous credulity. What a duty officer needs is not more raw stream, but provenance on each claim within it: a gunshot alert corroborated by two independent sensors and a call to the incident line stands differently from one flagged sensor at 1am on New Year's Eve; a heating-fuel shift reported by the district energy office and confirmed across three consecutive quarters stands differently from a single warm winter. The failure in the standing-coverage-plan case was never that the officer weighted contrary evidence too lightly by careful judgement. It is that the operational system gave that evidence no registered route to the conclusion at all, so no weighting, careful or careless, ever took place. Rational conservatism presupposes a channel through which the anomaly can arrive to be weighed. The stale risk map lacked the channel, not the caution.
What provenance would actually change
A Large Universe Model, in this domain, is not a smarter forecast. It is a belief structure in which each element of the current risk map carries its source and its age: this sector's weighting rests on eleven incidents logged across two winters, last revised on this date, from these feeds, at this confidence. When the underlying pattern shifts — new-build stock changes the fire load, a gauge network is recalibrated, a heating subsidy alters exposure — the retraction of the source propagates to the belief that rested on it, automatically and visibly, rather than sitting unnoticed in a log a duty officer has no structural reason to re-read.
This does not relieve the officer of judgement. It relieves the risk map of the one failure mode that judgement cannot fix on its own: a conclusion that has quietly outlived the evidence that produced it, with no route back to the test.