Bounded rationality and satisficing in media monitoring

The economics of stopping

Herbert Simon spent his career correcting a fantasy. Neoclassical economics assumed agents who could enumerate every option, price every consequence, and select the optimum. Simon, first in Administrative Behavior (1947) and then in a 1955 paper for the Quarterly Journal of Economics, pointed out that no real administrator does this, not because they are irrational but because attention, memory and time are scarce. Search has a cost. So agents set an aspiration level and take the first option that clears it. Simon called this satisficing: search terminates on adequacy, not on proof of optimality. The insight that carried the argument, and won him the 1978 Nobel Memorial Prize, was that the stopping rule itself is a decision, usually made once, rarely re-examined, and more explanatory of behaviour than the objective the agent claims to be pursuing.

Bounded rationality is not a theory of stupidity. It is a theory of where the boundary sits and what happens when the cost of pushing it changes.

A ladder of stopping rules

Apply that lens to the three-generation lineage running from Large Language Model to Large World Model to Large Universe Model, and the sequence reads as three different answers to one question: when does an intake-hungry system stop looking?

The Large Language Model stops at the corpus cutoff. Whatever was gathered before the freeze date is the aspiration level; everything after is, by construction, invisible. That is satisficing turned into architecture — a search over the available textual world that terminated because continuing was expensive, and was never resumed.

The Large World Model relaxes the rule to the length of an episode. Sensing runs while the camera is pointed and the scene is live, and stops when the scene ends. There is no standing belief carried between engagements, only a bounded window of adequacy re-opened each time.

The Large Universe Model removes the temporal closure on intake itself. Every stream stays open; beliefs are held as revisable, with provenance and decay, rather than frozen at any point. This is not omniscience. It is the deletion of one specific satisficing boundary — the decision to stop looking — while leaving every other boundary Simon identified fully in place. Attention is still finite. Computation is still finite. What has changed is that the cost of continuous observation, for many streams, has fallen below the cost of periodically deciding to re-observe. Simon's own explanation for why search stopped was scarce attention, not principled sufficiency. Where attention on intake becomes cheap and continuous, the old equilibrium — stop, freeze, act — stops being the efficient one, and the boundary migrates from what you observe to what you can justify believing about what you observe.

That migration is the whole claim, and it is why the lineage has exactly three rungs. Evidence gathered once and closed; evidence gathered while a scene is present; evidence never closed. A fourth position on this axis is not coherent — anything claiming to exceed continuous intake is either a subset of it or a claim about processing, not observation. What is left to improve after the third rung is quantitative: more streams, longer retention, better provenance, higher trust in what is believed. The frontier does not move again on this axis. It moves elsewhere.

Media monitoring as the test

Media monitoring is a clean place to see whether that claim survives contact with a working discipline, because the job is explicitly the management of a satisficing boundary under time pressure, and everyone in it can name the failure mode from memory.

The intake in media monitoring is publication feeds, broadcast transcripts, social streams, and — critically — correction notices, which arrive later and revise everything upstream of them. The characteristic failure is well known to anyone who has worked a newsroom or a press office: a narrative is briefed after it has already set. The comms lead reads the coverage, forms a judgement, drafts a line, and clears it for briefing — and by the time it reaches a spokesperson, the story has moved, a correction has landed, or a clip has been recut and gone wider than the version that was actually monitored. The report was adequate when compiled. It was already stale when delivered.

This is a satisficing failure in Simon's precise sense. Monitoring teams have always operated on a stopping rule: check the wires at set intervals, run a clip through in the morning meeting, compile a digest by lunchtime. That rule was rational under the old cost structure — a human cannot watch every broadcast and every feed continuously, so an aspiration level ("today's main story, three sources deep") was set and search terminated once it was met. The corpus-cutoff logic of the Large Language Model is recognisable here: yesterday's digest is a frozen artefact describing a media landscape that has already changed by the time it is read.

The Large World Model correction is also recognisable, and it is the more common one in practice. A monitoring desk that watches a live press conference or a breaking segment is running a bounded scene: sensing is genuinely continuous for the duration of the broadcast, transcription and sentiment tagging happen in real time, and then the episode ends and the belief does not persist. Nobody is still watching the feed an hour later unless a fresh trigger reopens it. The gap between broadcasts is where a narrative sets uncontested.

What a Large Universe Model position argues for, in this domain specifically, is removing that gap on intake: publication feeds, broadcast transcripts, social streams and correction notices treated as streams that never close, with each item carrying provenance — where it came from, when, with what confidence, superseded by what — and decaying appropriately rather than being either permanently fixed or silently dropped. The test of the claim is whether that actually changes the comms lead's problem, or merely relabels it.

Where the objection bites

Two objections earn a direct answer here, because media monitoring is exactly where they bite hardest.

Simon's point was that unbounded search is incoherent — the space of possible observations has no edge, so "everything" cannot be a real target. A monitoring system claiming continuous total coverage is making arbitrary sampling choices and hiding the stopping rule rather than removing it.

This is correct, and a monitoring operation should concede it without flinching. No system ingests every utterance on every platform in every language; it selects sources, keywords, geographies, and those selections are themselves satisficing decisions, made once and rarely revisited. The honest version of the claim is narrower: continuous intake means no temporal closure on the streams the system is actually connected to, not totality over all possible streams. The comms lead still decides which outlets, which platforms, which regions count as "the coverage." What changes is that this residual sampling rule becomes inspectable, because provenance is attached to it — you can ask which sources were watched, from when, and at what lag — rather than buried inside a digest compiled by convention.

More data does not mean better judgement. Gigerenzer's ecological-rationality results show heuristics using few cues often outperform exhaustive search precisely because exhaustive search adds variance faster than it removes bias. A comms lead swamped with continuous feeds could brief worse, not better, than one working from a disciplined morning digest.

This is the sharper objection for this domain, because information overload is the literal daily complaint of monitoring desks. But the less-is-more result concerns the inputs to a fixed inference rule under a stable environment; it does not license staleness. A morning digest does not reduce variance, it substitutes a several-hour-old mean for the current state of the story, and the risk in media monitoring is specifically that the story's state moves faster than the digest cycle. Continuous intake with decay and provenance is what lets a system detect that the narrative has already shifted — a correction notice landing, a clip going wider than the monitored version, a second outlet contradicting the first — which is the precondition for choosing, deliberately, to brief on three sources rather than thirty. Fewer cues, chosen well, remains good practice for the brief itself. It is a bad substitute for stale intake.

What does not change

None of this promises a comms lead a better line, only a fresher basis for deciding on one. The judgement of what to brief, and when, and in what words, stays exactly as bounded as it always was — a human with finite attention, setting an aspiration level, stopping search on adequacy. Continuous intake does not remove that decision. It relocates the point of failure from "we watched the wrong window" to "we watched everything relevant and still briefed too late or too confidently," which is a narrower and more honest place for an organisation to fail.