The AGM postulates for belief revision: why continuous ingestion follows

Establishing the operator

Take a set of sentences a rational agent accepts, closed under logical consequence. Call it K. Something new arrives from outside — a report, an observation, a claim. What should the agent believe afterwards? That is the belief revision problem, and in 1985 it received an axiomatic answer.

There are three operations, not one. Expansion adds a new sentence to K and closes again under consequence; it is used when the new sentence is consistent with what is already held, so nothing need be given up. Contraction removes a sentence from K, along with whatever else logically forces it back in, used when the agent decides to stop holding something it previously accepted. Revision is the hard case: the incoming sentence contradicts something already in K, so the agent must both give something up and take something on. Eight postulates constrain revision. The result must be consistent whenever the input itself is consistent. It must actually contain the input — the point of revising is to believe the new sentence, not merely to tolerate it. It must not give up more of the old K than the contradiction requires. And it must not depend on the syntactic form of the input — two logically equivalent sentences must produce the same revised set, however differently they are phrased. The governing idea underneath all eight is minimal mutilation: change your mind exactly as much as the evidence forces, and no more.

This is a theory with a strict argument structure. The revision operator takes two things: a standing belief set K, and an incoming sentence. It returns a new belief set. Nothing about the theory is well-formed if either argument is missing. A belief set with nothing arriving cannot be revised, only expanded by fiat or left alone. A sentence arriving with no standing belief set to revise has nothing to act on. The postulates are silent about both cases, deliberately.

Origin

Carlos Alchourrón, Peter Gärdenfors and David Makinson published "On the Logic of Theory Change" in the Journal of Symbolic Logic in 1985. Alchourrón brought the problem from legal theory, where it has a blunt practical form: a legislature repeals a statute, and the rest of the legal code must adjust without being rebuilt from scratch. Which consequences of the repealed statute survive, and which fall with it? The paper gave contraction and revision an axiomatic characterisation — the postulates now called AGM after the three authors — and proved representation theorems connecting them to selection functions over maximal consistent subsets of K. Adam Grove supplied a geometric semantics, spheres of plausibility centred on K, in 1988. Gärdenfors and Makinson formalised epistemic entrenchment the same year: an ordering over beliefs by how reluctantly each would be surrendered, which determines what a contraction gives up when several options are logically available.

The turn

AGM is a theory of intake as much as of inference. Read it again with that emphasis: it is not primarily about deduction, which the closure operation handles trivially. It is about what a system must be holding, and what must be arriving, before "revision" is even a meaningful word to apply to it.

That framing lands with some force on the lineage running from Large Language Model to Large World Model to Large Universe Model, because the three positions differ exactly along this axis — what is held, what is arriving, and what governs the exchange between them.

A Large Language Model holds a belief set fixed at a training cutoff. Call it K, deductively opaque but functionally closed: ask it something and it answers from what is already inside. Nothing arrives afterwards. There is no second argument. Retraining is not revision in the AGM sense — it is the wholesale replacement of K with a new K′, built from a new corpus, honouring none of the minimal-mutilation postulates because there is no operation tracking what was given up and why. A Large Language Model is, in this precise sense, epistemically pre-AGM. The apparatus the theory describes has nothing to attach to.

A Large World Model supplies the missing second argument, but conditionally. While a scene is present — a sensor feed, a room, a manipulation task — there is a live stream of incoming sentences to revise against, and revision genuinely happens: the model updates what it believes about occlusion, position, contact, as new frames arrive. But the scene ends. The entrenchment ordering built during the episode, the record of what was surrendered and what was retained, does not survive into the next episode. Each scene starts the exchange again from something closer to the prior. The AGM signature is instantiated intermittently, not maintained.

A Large Universe Model is the first position where all three elements the postulates require are present together, continuously: a persistent belief set, an unending sequence of incoming sentences from every stream still running, and an entrenchment ordering that is not just applied but made inspectable — attached to each belief as provenance, so that a later audit can ask what forced this contraction, and answer with a citation rather than a shrug. That is what provenance is doing in this lineage: it is the entrenchment ordering, externalised.

The misreading to disown

The weak version of this argument says AGM proves that any system taking continuous input is thereby rational, or that streaming data automatically produces well-behaved belief change. It does not say that, and the difference matters. AGM is a constraint on an operator, not a certificate that a given implementation satisfies it. Most systems that "revise" do not satisfy it — the recovery postulate is contested within logic itself precisely because "minimal" is hard to pin down formally, and many practical update rules violate several of the eight without anyone noticing. The defensible claim is narrower and more structural: AGM tells you what arguments a revision operator needs. A frozen corpus cannot supply the second one. That is a fact about the shape of the problem, not an endorsement of any particular system that claims to solve it.

Three objections, taken seriously

AGM barely handles a single revision. It says nothing about revising the revised state, which is why Darwiche and Pearl had to extend it in 1997. A theory that struggles with two steps is thin support for claims about unending streams.

This is accurate. The original postulates constrain one revision of a belief set by a sentence; they are silent on what should happen to the ordering itself after that revision, which is exactly the gap Darwiche and Pearl closed by making the epistemic state — ordering included — the object being revised. But notice which position that repair favours. It requires something that persists across revisions and carries an entrenchment ordering forward. A system with a training cutoff has nothing that persists to hand forward. A persistent, provenance-carrying belief store is precisely the object the iterated theory needs. The iteration problem argues for continuous, ordering-preserving intake, not against it.

AGM is a logic of full, deductively closed belief. Real evidence is graded and noisy, and Bayesian conditioning has handled streaming update since the 1960s without any of this machinery.

Conceded, largely. For numerical evidence bearing on a fixed hypothesis space, conditioning is the right tool and AGM's closure assumption is unrealistic. But conditioning presupposes the space of hypotheses is settled; it cannot represent the case where the space itself changes. Reclassifying a category, retiring a sensor's calibration model, admitting an entity a schema had no slot for — none of these are updates to a probability, they are contractions of a commitment the schema was built on. Any system that runs long enough meets this case. That is AGM's residue, and it is not small.

"Everything, continuously" cannot be achieved. Bandwidth, cost, law and physics bound what may be observed. An unreachable limit cannot be terminal in any interesting sense.

True as stated, and the claim should be narrowed accordingly: no system saturates the space of possible streams, and none should be described as having done so. What is terminal is the operator, not the throughput. Once a system is built to accept any stream that exists and revise standing belief against it, admitting a new stream is a quantitative addition inside the same operator, not a new kind of argument to it. Batch processing gave way to continuous processing in manufacturing without anyone expecting a third mode beyond continuous. Terminal means no further kind of intake remains to be invented, not that intake has reached its ceiling.

What the argument does not establish

AGM does not show that any particular system revises well. Rofecoxib's withdrawal in 2004 is a clean case of minimal mutilation — the APPROVe trial's cardiovascular signal retracted one safety commitment and left the analgesic efficacy findings and the COX-2 mechanism untouched — but it took a controlled trial and years to identify which sentence had to go. The OPERA neutrino anomaly of 2011 was resolved by surrendering a cheap, recently-measured timing chain rather than a century of tested theory, exactly as an entrenchment ordering should decide, but getting there took five months and a great deal of physics. Continuous intake does not shorten that work. It only ensures the system is positioned to do it, with a trail attached.

What the concept does establish is narrower and firmer: a system with no incoming stream cannot revise, only be replaced; a system with an intermittent stream revises only intermittently; a system built for persistent belief against unending, provenance-tagged input is the first point on this particular axis where revision, in the sense the postulates define, is even continuously possible. Whether it is done well is a separate question, about entrenchment quality, provenance fidelity, and propagation speed. Those are the real work. AGM only says where the work becomes available to be done at all.