The Well and the Archive

Every act of knowing has two sides.

There is reality as encountered: particular, contextual, resistant, not exhausted by any one description of it. Call this the well.

And there is what we retain and construct from the encounter: concepts, measurements, categories, models, rules, memories, maps, procedures. Call this the archive.

The archive is indispensable. Without it, knowledge could not accumulate: no calculation, no bridges, no teaching, no proof. Every technical civilization depends on turning encounter into stable representation.

But the archive carries an intrinsic danger. A representation can become so useful that we forget it is a representation. Korzybski warned that the map is not the territory; the warning is old, the failure perennial. The model becomes the world. The record becomes the event. The description of the valve being closed becomes indistinguishable from the valve being closed.

I did not arrive at this distinction by philosophizing about it. It emerged from formal work on reasoning errors and kept reappearing where I had not put it: in engineering failure, in AI inference, in rhetoric, in bureaucracy, in moral psychology. At some point one has to ask whether a framework that keeps finding itself everywhere has become so broad it can explain anything.

But there is another possibility: these fields converge because they share a root. All of them concern the relationship between reality and the representations cognition constructs of it, and that relationship is fundamental enough that its characteristic failures should recur wherever anyone, human or artificial, tries to know, reason, or act.

Which possibility holds is not a matter of taste. It can be tested. I will come back to the test.

A claim about a claim

The formal work behind this essay (Paper II) isolates one precise form of the danger: a claim about a claim does not establish the claim. That a proposition is widely reported, was once true, is probably true, or resembles the truth are all operations on the representation, and none of them discharges into reality. A hedge quietly dropped, a “was” promoted to “is,” a simulation’s output treated as a measurement: representation-level moves credited with reality-level force.

That is one schema among five. The framework distinguishes five structural ways a boundary between representations can be illegitimately crossed: one kind of truth substituted for another, a rule stretched past its license, the wrong kind of evidence for a kind of claim, an unjustified jump between levels, a false opposition imposed on a domain that does not support it. It reached them the hard way, by cataloguing fallacies and failures until the classification stabilized. Only afterward did the deeper question become answerable: why five? Paper III derives them from a single premise. Given any system capable of treating one thing as different from another, exactly these conditions must hold for that capacity to remain coherent. Not given language, not given human cognition. Given distinction. That is why the constraints cross substrates: sensors, grammars, logics, and software hit the same walls because they share the act of distinguishing, not a subject matter. And it recasts the essay’s central boundary. The line between a representation and what it represents is itself a level distinction; map versus territory is the stratification condition at the largest scale, the boundary between the system and the world. The well and the archive are not a metaphor laid over the framework. They are its deepest condition, seen from outside.

The test: convergence as evidence

A theory that can reinterpret every phenomenon in its own vocabulary has demonstrated very little. Reinterpretation is cheap. The stronger test is whether independently developed disciplines discover structurally equivalent constraints for reasons internal to their own work: because their objects of study forced the distinction on them, not because they borrowed a vocabulary.

By that standard, the clearest demonstrated case is stratification: levels of description must not be collapsed into one another. Tarski, confronting the Liar Paradox, found that a language making unrestricted claims about its own truth predicate collapses, and that the repair is a hierarchy of levels. Centuries earlier, the scholastic logicians built the machinery of suppositio to keep claims about terms distinct from claims made with them. The Buddhist two-truths tradition, working on entirely different problems, held that statements do not travel freely between conventional and ultimate description. Kant drew his own version of the boundary. And the cross-linguistic work in Paper I found the constraint embedded in grammar itself: languages across unrelated families refuse constructions that would collapse levels. Five arrivals at the same wall, from logic, medieval semantics, Buddhist philosophy, critical epistemology, and empirical linguistics; none derived from the others. (Paper IV traces the full convergence.)

That is what convergence-as-evidence looks like: named discoverers, internal motivations, one constraint.

Once seen, the structure is easy to recognize elsewhere, and here the difference between evidence and illustration matters. A bureaucracy treating the record as more real than the person recorded; a control system acting on a stale representation of physical state; a language model producing coherent output that tracks nothing: these are illustrations. They show the failure’s shape in familiar settings. They do not establish the framework, and I do not offer them as though they did.

When the well pushes back

The most literal grounding for the metaphor is what Paper I calls the negative space of language. Across sixteen languages from ten families and three isolates, what languages consistently refuse to say proved as robust a finding as what they permit. The prohibitions are invisible in any corpus, because they are precisely the sentences no one says; they appear only when probed for. Resistance counting as information, in the most literal form I know of: the well pushing back through the archive of grammar itself.

Presentation and re-presentation

Reality is first presented; it is encountered. It is then re-presented: made present again to thought through abstractions we can grasp and manipulate. (Iain McGilchrist has one articulation of this distinction; versions of it appear in phenomenology well before him.) The pathology is not abstraction. Re-presentation is how encounter becomes transmissible. The pathology begins when re-presentation forgets its dependence on presentation: when the representation declares itself complete.

The same failure appears wherever a system moves from given these premises, this conclusion follows to therefore the matter itself is settled. A calculation may be finished without reality being exhausted. A model may be internally resolved while the question it represents remains open.

The remainder

Moral reasoning meets this boundary too. The Trolley Problem’s arithmetic is straightforward: five lives on one track, one on the other. Represented in those terms alone, the calculation seems nearly complete. Five is greater than one.

Yet many people experience a remainder: a sense that the calculation, even if relevant, has not exhausted the situation. The remainder is usually dismissed as emotion interfering with reason. But perhaps the discomfort is information, an early signal that the ontology of the model is thinner than the ontology of the situation, which included agency, causal authorship, the difference between permitting harm and redirecting it, obligation, relationship.

There is a second remainder, so common it has its own dismissive name: fighting the hypothetical. I’d look for a third option. I’d shout a warning. This is treated as evasion. But notice what it detects. The option space has been closed by stipulation (two tracks, one lever, nothing else), and closure of that kind is licensed only where a definition constitutes the boundary, never where the world does. A represented set of options is not thereby an exhaustive set of options. Fighting the hypothetical is boundary sensitivity aimed at the frame rather than the calculation: a detection of imposed closure, not a failure to engage.

Together the two remainders reveal what the trolley problem actually measures. It was never an instrument for discovering what is right; it is an instrument pointed at the person asked. What it extracts is the structure of the respondent’s moral archive: which distinctions their commitments maintain that the arithmetic dropped. Different people’s remainders differ because their distinction structures differ. Which is why surveys of trolley responses yield facts about respondents, not verdicts about the act; sliding from most people would pull the lever to pulling the lever is right validates a normative claim with the wrong kind of support.

Feeling that something is missing is not proof that something is missing; intuition can encode bias as easily as insight. The remainder is a hypothesis, not a verdict. What lets it earn standing is that articulation has a target: the felt discrepancy is vindicated when it can be cashed out as a specific, nameable omission or boundary violation, whether a dimension left out, evidence of the wrong kind admitted, or a level silently crossed. And when the discomfort cannot be articulated as any identifiable omission, that too is information of a different kind: the signal may have been noise.

That is what separates a discipline of boundary sensitivity from an injunction to trust one’s gut. Encounter → felt discrepancy → articulation → structural diagnosis → revised understanding. The middle step aims at a vocabulary of specific failure types, and it can fail, publicly, when nothing in that vocabulary fits.

The loop

Revised understanding feeds back into perception. The novice engineer who once heard only an unfamiliar noise learns to hear cavitation. The reader who merely sensed an argument had cheated learns to see the joint at which it cheated. Formal knowledge educates intuition; the archive becomes better at listening to the well. The result is a loop, not a transfer: reality → representation → testing → revision → better encounter.

Call it epistemic homeostasis, a name that only gathers an older lineage: Peirce’s fallibilism, Popper’s conjecture and refutation, the cybernetic principle of correction through feedback. A healthy cognitive system does not merely accumulate representations. It continually exposes them to correction by what they represent.

The most dangerous condition is not ignorance. It is premature closure. Ignorance still knows something is missing. Closure has forgotten that anything could be.

The archive becomes self-validating, and the disciplines give that one failure their different names: model overreach, dogmatism, bureaucracy, hallucination.

The archive keeps a hole open

Reality exceeds every archive constructed from it. Cognition succeeds not by abolishing the gap but by managing it faithfully: categories open to distinction, conclusions proportional to what their premises license, representations forbidden from silently inheriting the properties of their referents, and resistance from the world permitted to count as information.

An essay that says all this owes the reader one demonstration, applied to its own framework. Here is one.

The framework includes a map classifying the classical fallacies into its five types and claiming the decomposition is exhaustive. Row 31 is the non sequitur, and its classification reads, in effect, the inference is invalid. Under stress-testing that entry did not survive as a decomposition at all: every violation type ultimately produces an unlicensed inference, so saying “the inference failed” restates the thing to be explained, and unlike its neighbours the row generates no specific repair. It is a placeholder in a table that claims completeness: a hole in the archive, currently marked open rather than papered over.

That admission is not a footnote. It is the framework operating as designed. A healthy archive is not the one with no holes; it is the one that can still see them.

Beneath the vocabularies of logic, engineering, linguistics, philosophy, and AI lies one problem: how can a finite system of representation remain faithfully related to a reality it can never completely contain?

A healthy archive does not answer that question once. It remains in conversation with the well.

The formal framework discussed here is developed across four papers: Paper I (the cross-linguistic empirical foundation), Paper II (the fallacy decomposition), and Papers III–IV (the conditions for coherent distinction-making, and truth-kinds with the convergence of the philosophical traditions) — forthcoming.