Paraconsistent reasoning: why continuous ingestion follows

The rule that breaks everything

Classical logic contains a rule with a violent consequence. It is called explosion, and it says that from a contradiction, anything follows. Accept both P and not-P as premises, and you can derive, by entirely valid steps, that the moon is cheese, that you owe your neighbour money, that this sentence is false and true at once and therefore so is every other sentence. The proof is short and correct. That is the problem. A single inconsistency, anywhere in a body of beliefs, licenses every conclusion, however unrelated. One false note and the whole score is noise.

Paraconsistent logics are the family of formal systems built to deny this. They do not deny that contradictions can arise, and most do not even insist that no contradiction is ever true. What they deny is the inference from a local contradiction to global chaos. Several routes get there. Relevance logics restrict inference so that a conclusion must share content with its premises, blocking the slide from "P and not-P" to "the moon is cheese" on the simple ground that lunar dairy has nothing to do with the contradiction. Newton da Costa's C-systems keep classical logic almost entirely intact but weaken negation, so that a proposition and its negation can both hold without detonating the rest of the theory. Nuel Belnap's four-valued scheme assigns each proposition one of four statuses: told true, told false, told neither, told both. Contradiction becomes a value a proposition can have, sitting in its own corner of the system, rather than a rupture that voids the enterprise.

The shared move, across all these approaches, is to convert contradiction from a global catastrophe into a local condition. A theory can be inconsistent in one corner and perfectly usable everywhere else. That is the entire idea, and it is worth sitting with before going anywhere near a machine.

Origins: committees, inconsistent theories, unreliable informants

The lineage is older than computing and was never aimed at it. Stanisław Jaśkowski published a discussive logic in 1948 to model the beliefs of a committee: several members, each internally consistent, collectively disagreeing, and Jaśkowski wanted a logic that could represent the group's combined position without every disagreement between two members collapsing the whole record. Newton da Costa developed the C-systems through the 1960s in Brazil, motivated by a more abstract worry — that inconsistent mathematical theories, historically common at the frontier of a field before the frontier is tidied up, deserved to be studied rather than dismissed outright. Alan Anderson and Nuel Belnap attacked the same target from the direction of relevance, insisting that valid inference requires a real connection between premise and conclusion, not merely the formal trapdoor that classical logic leaves open once a contradiction is on the books.

Belnap's own contribution sharpened further still. His 1977 paper, "A Useful Four-Valued Logic," was written explicitly to handle computerised databases fed by multiple informants of uncertain reliability — a problem statement written half a century ago that reads, today, like a design brief.

The turn

Move from logic to intake, and a pattern appears that none of these logicians were describing but all of their machinery fits.

A Large Language Model reads a corpus assembled once and frozen at a cutoff. Contradictions inside that corpus — and there are many, since the internet disagrees with itself constantly — are never surfaced as contradictions. Two documents differ; the training process averages across them in weight space; nothing in the resulting model can point at the disagreement or name its sources. The inconsistency is real but invisible, resolved by arithmetic before anyone could inspect it.

A Large World Model takes in a bounded scene: a room, a road, a moment. When its sensors disagree — one camera's depth estimate against another's — the disagreement is usually resolvable, because the sensors share a target and a timestamp. Fusion algorithms exist precisely for this, and they work well, because the contradiction is local in time as well as in content. It can be filtered out as noise without much cost, because there is, in the relevant sense, a single fact of the matter available right now to check against.

A Large Universe Model takes in every stream that is still running, with no cutoff and no single scene. Streams arrive from different instruments, different owners, different latencies, and they do not stop. At some point, inevitably, two of them will say incompatible things about the same fact, and there will be no third fact available yet to adjudicate between them. This is not a design flaw to be engineered away. It is a structural feature of unbounded, continuous intake. A system built this way faces the situation Belnap described in 1977 as its normal operating condition, not its exception.

The claim, stated plainly: continuous intake forces paraconsistency. Under classical logic, the moment of contradiction destroys the belief set. The only classical escapes are to stop reading, to discard one report before adjudication, or to average the two into a synthetic third value nobody actually reported. Stopping abandons the continuity that defines the third position. Discarding erases provenance and makes the choice unauditable — who decided, on what basis, and can it be reviewed. Averaging manufactures a belief that no source holds, which is a specific kind of dishonesty when the sources are named and disagree for reasons. What remains is the paraconsistent option: record the contradiction as a contradiction, tag it with its sources, contain its consequences to the neighbourhood of the dispute, and carry it until evidence resolves it or the question stops mattering.

Air traffic control already runs on this discipline without naming it. When a Mode S transponder reports an altitude 300 feet from the radar-derived figure, controllers do not average the two numbers into a comforting midpoint. The discrepancy is annunciated, both values persist on the display, and separation is maintained against the more pessimistic reading until the aircrew confirms which is right. The contradiction is carried, attributed, and acted on, sometimes for minutes. Hospital medication reconciliation runs on the same discipline: the GP record, the pharmacy dispensing history and the patient's own account disagree in over half of admissions, and the reconciled list a pharmacist produces is not a merge but a document recording each claim, each adjudication, and each item still marked pending. Bank reconciliation is older still — two ledgers of the same cash position, routinely disagreeing because of cheques in transit, held apart in an itemised statement rather than forced into agreement.

Objections worth taking seriously

Bayesian probability already handles this. Assign 0.6 to P and 0.4 to not-P and there is no contradiction, only distributed credence.

True, and this narrows the claim rather than defeating it. Where disagreeing sources share a likelihood model, conditioning is the right tool and paraconsistency adds nothing. But conditioning presupposes a single joint distribution, and disagreeing custodians frequently do not supply one — a transponder and a ground radar are not two noisy estimates of the same variable, they are two instruments with different failure modes and different event spaces. Assigning a credence also erases who said what. Probability handles uncertainty about a fact. Paraconsistency handles conflict between testimonies about a fact. The two compose. Neither replaces the other, and a fair reading has to concede this is often, not always, the paraconsistent case.

Weakening the logic is expensive. You lose disjunctive syllogism and much of the reasoning anyone actually relies on.

This is the sharpest objection and it should not be waved off. The loss is real. But the working answer is hybrid, not wholesale: reason classically inside each internally consistent subset of belief, paraconsistently across the seams where subsets collide, and flag the boundary explicitly. Discipline is kept everywhere it is affordable and suspended only where two streams actually meet in disagreement.

This is a plumbing problem. Vector clocks and conflict-free replicated data types solved conflicting writes decades ago without any non-classical logic.

CRDTs resolve conflicts by construction, which means the resolution rule is fixed before the conflict is understood — fine for a shopping cart, not fine for a patient's medication list or a reactor's temperature reading, where the merge is a judgement that must be deferrable and reviewable. Paraconsistency names the structure that can hold that judgement open without corrupting everything the store already knows. Plumbing and logic, not one instead of the other.

The misreading to disown

The wrong version of this idea says contradictions are fine, truth is plural, anything goes. That is dialetheism at its loosest, and it dissolves the concept rather than using it. Paraconsistency is not permission to be incoherent. It is a containment discipline. A well-built paraconsistent store still wants consistency, still ranks a resolved belief above a conflicted one, and still treats a contradiction that stays unresolved for too long as a defect demanding investigation. The gain is that the defect stays local. It does not license the moon being cheese.

What this does and does not establish

This establishes that a system committed to unbounded, continuous intake needs a logic that can carry a live contradiction without collapsing, and that classical logic, taken straight, cannot do this. It does not establish that any such system currently exists at scale, nor that paraconsistent logic alone is sufficient — provenance tracking, decay functions and adjudication procedures are doing at least as much work. It narrows, rather than proves, the case for the third position on the intake axis: not that more computation makes contradiction disappear, but that a certain kind of intake makes contradiction permanent, and only a certain kind of logic was ever built to survive that.