1 Classification is a theory, not a convenience

A classification is a claim about the world, and a claim can be false. This is the whole premise, and it is not rhetorical: to sort knowledge into kinds is to assert that those kinds carve at real joints rather than convenient ones, and that assertion has truth-conditions the way "gold has atomic number 79" has truth-conditions. Foucault's point in tracing the shifting épistémè of successive eras¹ was not that classifications vary but that each one silently encodes the conditions under which its age could take anything to be knowledge at all. An order of knowledge is a compressed epistemology. It is answerable, and it can be wrong.

The ordinary picture — knowledge as an accumulating pile, its arrangement a clerical afterthought performed for retrieval — inverts the actual dependency. The arrangement is prior. It decides what will count as one field rather than two, what will count as knowing rather than merely believing, where the burden of proof sits. This atlas does not arrange knowledge it has first collected; it advances a theory of knowledge and lets the entries fall out of it. If the theory is wrong the entries are misfiled, and the misfiling is detectable.

2 Atlas against library

The worth of an ordering is exactly its exposure to error. Melvil Dewey's Decimal Classification² files religion at 200 and philosophy at 100 and the natural sciences at 500, and its genius — the reason it spread — is that no user is ever asked to believe the adjacency means anything. It is unfalsifiable and therefore empty: a pure convenience, cosmology with the cosmology deleted. A scheme that cannot be wrong tells you nothing, on the same principle by which an unfalsifiable hypothesis is not a scientific one.

An atlas pays the price a library refuses. To map is to commit to a claim about how the territory coheres, and commitment is what makes the map capable of being corrected. The list is spared refutation and earns nothing by the exemption; it can only grow longer, never truer. This atlas chooses the falsifiable form on purpose, because the alternative purchases safety at the cost of saying anything — and every serious classification that has hidden its metaphysics in its shelving has thereby smuggled a philosophy past the reader unexamined. The single honest move is to put the philosophy on the frontispiece and sign it.

3 The unit is the warrant, not the discipline

The atom of the atlas is neither the fact nor the discipline. Facts are inert and unbounded and sort nothing. Disciplines are administrative: the word scientist was coined only in 1833, and the partition of knowledge into the departments we now treat as natural is a nineteenth-century professionalization, an artefact of chairs, journals, and funding bodies, not of the structure of knowing.³ Departments fuse and split for institutional reasons and would carve differently under different institutions; the joints they track are sociological.

The atom is the warrant — the kind of ground a claim offers for counting as knowledge. And warrants are already plural at the root: Ryle showed that knowing how (to ride, to prove, to diagnose) does not reduce to knowing that, that competence is not a stock of propositions,⁴ which means the theory of knowledge cannot be monolithic even before the domains are drawn. Sort acts by the warrant each offers and the disciplines reappear as consequences — chemistry is the acts warranted by controlled intervention on matter, not the contents of a building. Change the building and the warrant is untouched.

4 Two blades — object and mode

There are exactly two blades with which knowledge can be cut, and every taxonomy chooses one usually without noticing. One cuts by the object — what the knowledge is about. The other cuts by the mode — the operation by which it is reached and warranted. The distinction is not invented here: Windelband drew it in 1894 as the split between the nomothetic sciences, which seek laws, and the idiographic, which grasp the particular in its singularity⁵ — a division not by subject but by cognitive aim, the mode-blade wielded openly.

The choice of blade is consequential and non-neutral. Cut by object and you produce a chart of the regions of being; cut by mode and you produce a chart of the operations of mind, a different territory sharing the word "knowledge." Almost every received map cuts by object and forgets it has chosen, which is why the mode-distinctions keep having to be rediscovered — by Windelband, by Dilthey, by Ryle — as if new. This atlas uses the object-blade where the object suffices and the mode-blade where the object fails, and marks the seam where they meet rather than concealing it. That seam is the load-bearing line.

5 Object first, where object fixes method

The object leads wherever it determines the method, and for most of knowledge it does. You cannot run a controlled trial on a theorem, interview an electron, or replicate a poem's meaning across laboratories without destroying what you meant to know; the object dictates the admissible warrant and hands you the mode unasked. Where that determination holds, cutting by object is the deeper cut, because it captures the mode for free and needs no separate axis.

So four domains take their identity from what they know: the abstract, the given, the social, the made. Each object licenses a characteristic warrant — derivation, controlled observation, participant access, construction — and the warrant follows from the nature of the object with something close to necessity. The map would need only the object-blade if the determination were exceptionless. It is not, and the single exception is not a marginal case but the hinge on which the whole structure turns.

6 The exception that forces a mode-domain

There is a class of object whose nature does not fix a single warrant, and there the object-blade fails as a matter of demonstration, not preference. Take one scripture. It sustains a fully natural knowledge — the chemistry of the ink, the neurophysiology of the reader, the demography of the sect — and it sustains an entirely distinct knowledge of what it means to those who live inside it, and no quantity of the first ever converts into the second. Dilthey fixed the distinction as erklären against verstehen, explaining against understanding, and made it the constitutive line of the human sciences.⁶ The decisive move is Gadamer's: understanding is not a softer or preliminary knowledge but a truth-apt achievement with its own structure, the hermeneutic circle in which part and whole determine each other, irreducible to method in the natural-scientific sense.⁷

When one object underdetermines the warrant like this, the object-blade cannot separate the two knowledges, and only the mode-blade can. This is why the fourth domain, alone, is named for how it knows rather than what: not as an aesthetic parallel to the others but because the alternative loses a real distinction between two truth-apt achievements. The asymmetry is not a blemish to be planed away for symmetry's sake. It is the map reporting a genuine fault in the terrain — the fault between fact and meaning — and a taxonomy that smooths it has bought symmetry with a false claim.

7 Aim is the is/ought axis, not a place

A third cut tempts and must be refused as a domain: the cut by aim, pure against applied, knowing against doing. Aim is not an object; it is the is/ought axis laid across every object. Hume fixed the gap that makes this so — no ought follows from any set of is by logic alone⁸ — and "applied" is precisely the introduction of an ought (this end is to be achieved) into a field whose pure form states only what is. Because an ought can be attached to any descriptive field whatever, aim cuts across all of them; made a domain, it collides with each object-domain simultaneously, since anything can be aimed.

So aim is demoted to a tag — a property a text carries, noted and passed. It earns a domain-name in exactly one case: where the aiming has hardened into an object of its own, where what is known is no longer a pure field pointed somewhere but the intervention itself, and where knowing-how has become the operative warrant rather than knowing-that.⁴ That case is the fifth domain. The tag marks a direction; the domain marks the object a direction, pursued long enough, congealed into.

8 I · Formal — certain and undecidable

The object of the first domain is structure considered apart from any instance — number, set, proposition, group, proof — and its warrant is derivation: consistency with axioms, not correspondence with a world. A theorem is certain in a way no empirical claim can be, because it makes no claim on the world that the world could refute; its truth is exhausted by the derivation. This is the domain's singular strength and it is bought at a precise, proven price. Gödel showed in 1931 that any consistent formal system strong enough to express arithmetic contains truths it cannot prove and cannot establish its own consistency from within⁹ — the domain of certainty is, at its foundation, provably incomplete. Formal knowledge is the most certain and the least self-grounding at once.

Two further facts keep the honest account from tidiness. Whether its objects are discovered or constructed remains open and is itself argued by formal, never empirical, means — Platonism against Hilbert's formalism against Brouwer's intuitionism¹⁰ — so the domain cannot say what its own objects are. And its very separateness is contested: Quine's attack on the analytic–synthetic distinction denies any sharp line between the formal-by-convention and the empirical,¹¹ which is why the boundary with Natural is drawn in full knowledge that a major tradition denies it exists. What is not in doubt is the loan: the structures Formal studies turn out to fit the physical world with an effectiveness Wigner called unreasonable and no one has explained.¹²

9 II · Natural — explanation as subsumption

The object of the second domain is the given world, and its verb, erklären, has a precise reconstruction: to explain is to subsume the particular under a general law such that the event could have been predicted from the law and the conditions. This is the Hempel–Oppenheim deductive-nomological model,¹³ and its claim is strong — that explanation and prediction share one logical form, that to explain the apple's fall just is to derive it from the law of gravitation plus initial conditions. The warrant is not verification but survival of attempted refutation: Popper's demarcation makes a claim natural-scientific only insofar as it forbids something observable and stakes itself on the observation not occurring.¹⁴

The domain's regulative ideal is the elimination of the knower — the view from nowhere, a description true regardless of who occupies which standpoint.¹⁵ That ideal is exact where it applies and marks the domain's outer limit precisely: it holds while the object cannot know back, and begins to fail the moment the object forms beliefs about the inquiry into it. Natural borrows its deductive spine from Formal and lends its methods to Applied the instant the aim turns from describing the world to altering it; its purity is the purity of an object indifferent to being described.

10 III · Social — the object that loops

The third domain has an object no other domain possesses: one that changes in response to being known. Hacking calls this the looping effect of human kinds — a classification of people ("the unemployed," "the traumatized," "the gifted") alters the self-understanding and hence the behaviour of those classified, which alters the kind, which forces the classification to move, in a feedback the natural kinds never exhibit.¹⁶ Merton had isolated the mechanism as the self-fulfilling prophecy: the belief that a bank will fail, held widely enough, empties the bank.¹⁷ An electron does not read the paper written about it. A population does, and moves.

This reflexivity is not a nuisance degrading an otherwise natural science; it is the definition of the social object and the reason the domain cannot resolve into either neighbour. It cannot be fully Natural, because there are no laws stable under agents who revise their behaviour upon learning the laws. It cannot be purely Interpretive, because the social also exhibits genuine statistical regularity — suicide rates, price responses, electoral swings — that yields to explanation. Weber built the discipline on exactly this doubleness, defining sociology as the science that interprets social action in order to explain its course¹⁸ — verstehen and erklären yoked in one method. The domain's permanent instability between its neighbours is structural, and the social science that denies either half is the one that has stopped being honest.

11 IV · Interpretive — understanding is truth-apt

The fourth domain knows by verstehen, and its whole standing rests on one contention that must be defended, not assumed: that understanding is a truth-apt achievement irreducible to explanation, not a soft prelude to it. The case is made. Explaining a myth gives its cause — the anxiety it manages, the function it serves; understanding it recovers its sense, and a complete causal account can leave the sense wholly untouched, which is the mark of a distinct cognitive object rather than an unfinished one. Geertz built the method of interpretive anthropology on thick description — a term he takes from Ryle — where the same physical motion (a contracting eyelid) is a twitch or a wink or a parody of a wink depending on a structure of meaning no measurement of the eyelid contains.¹⁹ The winks are not reducible to the twitches, and the reduction's failure is demonstrable, not stipulated.

This is the domain of symbol, rite, image, scripture, and the esoteric traditions — everything known by entering its sense. Its omission from the standard fourfold is therefore not an oversight but a category error: a taxonomy of knowledge with no place for the mode that recovers meaning either loses the interpretive disciplines entirely or misfiles them under Social, treating a myth as a survey with poor sampling and a rite as behaviour awaiting a covering law. Interpretive shares its object with Social — both bend over human culture — and shares its warrant with neither Social nor Natural. That a whole mode of knowing has no drawer in the received map is the received map's deepest and most consequential failure.

12 V · Applied — the privilege of the made

The object of the fifth domain is the made — the artefact, the intervention, the designed system — and its logic is normative by the argument already given: it introduces an ought (the thing shall perform) that no description of the given world entails. Simon chartered these as the sciences of the artificial, the study of the contingent and the designed as against the natural and the necessary.²⁰ But the domain's deepest ground is older and more radical. Vico's verum-factum principle holds that the true and the made are convertible — we can know with certainty only what we have ourselves made²¹ — which inverts the usual hierarchy: the made is not a degraded object of knowledge but a privileged one, knowable from the inside as the given never is. The engineer knows the bridge in a way the physicist cannot know the electron, because the bridge was built to a specification and the electron was not.

The warrant is knowing-how, irreducible to knowing-that (Ryle), and it runs on tacit knowledge that resists full articulation — Polanyi's "we know more than we can tell," the surgeon's hand and the metallurgist's eye carrying competence no proposition captures.²² A domain-V field bears two marks jointly: it is synthetic, fusing many pure parents, and autonomous, contained by none of them. Medicine is not biology aimed sideways — it fuses biology, chemistry, physics, statistics, and psychology toward an end none of them owns and none could reach alone. This is exactly why Applied is a domain while applied-ness is only a tag: the tag is an ought pointed at a descriptive field; the domain is the made object that pursuit of the ought produced, knowable by construction and by hand.

13 Cross-listing is forced by combined warrants

A single text belongs to two domains whenever it deploys two warrants at once, and this is entailed, not tolerated. Mathematical physics derives (Formal) and tests against the world (Natural); game theory proves theorems (Formal) and models strategic agents (Social); actuarial science is probability (Formal) turned to the pricing of contingent loss (Applied). The object may be single while the warrants are plural, and where the warrants are plural the domains are plural, because the atlas indexes warrants.

Forcing each text into one cell would therefore falsify its epistemic structure to preserve the neatness of the partition — asserting one warrant where two operate. Cross-listing is the map declining that falsification: one text, written once, shelved under each warrant it genuinely uses. The seam is not damage to the map; it is the map refusing to lie about how the knowledge is actually secured.

14 The contested borders are real branchings

Some borders are disputed because the terrain itself branches, and history is the exemplary case. Hempel argued in 1942 that historical explanation is covering-law explanation like any other, merely with the general laws left tacit²³ — history as Natural in method. Dilthey and the hermeneutic tradition held it a science of verstehen, the recovery of the meaning an act had for its agent — history as Interpretive. Both describe something real in the practice, which is why every serious historian works somewhere along that line and no ruling settles it.

The other borders are equally principled: the human sciences generally straddle explanation and understanding by their constitution, not their immaturity; philosophical logic leans from Formal toward the interpretive; materials science tips between knowing the alloy (Natural) and forging the blade (Applied) with the aim. The atlas draws these in pencil and says so. A contested border is not a defect in the cartography; it is a real bifurcation in the object, and the honest response is to mark the fork rather than paint a confident line through it.

15 The residue is diagnostic

Every partition leaves a remainder, and the remainder carries information. Some fields are too young to place, hybrids formed at a collision of warrants and not yet stabilized. Some are survivals of dead classifications, kept by inertia. And some are not organized by a warrant at all but by a constituency or a grievance — a proper name attached to a grant line — which is why certain "studies" resist placement: they are administrative unions, not epistemic kinds, and the map's inability to file them is the map correctly reporting that there is no single warrant there to file by.

The honest map keeps an edge where the residue collects and reads it as a signal rather than sweeping it away. Where the existing domains cannot hold a growing body of practice, the pressure marks the site of a possible new domain — as the space between pure science and mere craft, long a residue, hardened into the made once the artefact became an object of systematic knowledge in its own right. The remainder is the map's growing edge, legible in the negative.

16 The provisional names

Two names are unsettled, and leaving them unsettled is a substantive act, not indecision. Interpretive names the warrant precisely but reads as a technique; Humanities names the recognized tradition but concedes in the naming that this is not a science; Human Sciences is Dilthey's own Geisteswissenschaften, the most exact and the least spoken. Applied names the aim but courts the exact confusion the map corrects — that it is a direction, not an object; Praxis honours the made and the doing but carries the odour of the seminar.

A name is a claim about what a thing most essentially is, and where the essence is still under argument the name should stay open, because a premature name would assert a settlement the evidence has not reached. An atlas that concealed its open questions would misrepresent the state of knowledge — the one falsification it exists to prevent. The unfixed name marks, precisely, a place where the theory has not yet finished deciding what it holds.

17 Each age cuts with the blade it trusts

The history of the division of knowledge is the history of which blade an age trusted. Aristotle cut three ways — theoretical, practical, productive — and the third, poiētikē, the knowledge that makes, is our Applied arriving in the fourth century BC, complete with the recognition that making is a distinct intellectual virtue (technē) beside pure knowledge (epistēmē).²⁴ Bacon cut by the mind's faculties — memory, reason, imagination — yielding history, philosophy, and poesy, a mode-cut two centuries before Windelband.²⁵ D'Alembert redrew Bacon's tree for the Encyclopédie and threaded it with renvois, cross-references that let knowledge be traversed as a network rather than a hierarchy.²⁶

Comte then imposed the ladder — the sciences ranked by decreasing generality and increasing dependence, mathematics at the base, his newly coined sociologie at the summit, each resting on those beneath.²⁷ Vico had already supplied the counter-principle in the made; Dilthey supplied the cut at understanding; Simon supplied the artificial's charter. This atlas is the heir that keeps the object-blade and the mode-blade both, refuses Comte's ranking, and takes from the lineage exactly the two moves history kept proposing and dropping: a domain for the made, and a domain for the understood.

18 Two exact failures of the standard fourfold

The map most curricula still teach — formal, natural, social, applied — fails in two locatable places, and the atlas is the minimal repair, not a fresh invention. First failure: no domain for the mode of meaning. The interpretive disciplines are homeless in the fourfold, or misfiled under Social, which is the category error of §11 given structural form — a whole warrant, verstehen, left without a house. Second failure: it lists "applied" as a fourth kind of science beside natural and formal, mistaking the is/ought axis of §7 for an object, when aim cuts across every object and cannot be one of them.

Two edits correct both and nothing else is touched. Add the mode-domain, Interpretive, to house the warrant of understanding. Demote applied-ness to a tag and promote the-made to the genuine fifth object, Applied, on the strength of §12. The result is not a more elaborate map but a more nearly complete and more nearly consistent one — the standard fourfold with its two structural faults removed and everything sound in it retained.

19 Flat by entailment, not by taste

That the five domains form a map and not a ladder is not a preference; it follows from an established result. The ladder is the reductionist thesis — Oppenheim and Putnam's programme of microreduction, each science reducing to the one below until all rest on physics.²⁸ That programme is refuted at its joint by multiple realizability: Fodor showed that the kinds of the special sciences (an economic recession, a mental state, a species) are realized by indefinitely many distinct physical configurations, so their laws cannot be laws of physics and do not reduce, on pain of losing the generalizations that make them sciences at all.²⁹ Anderson made the same point from inside physics — more is different, each level of organization obeying regularities not derivable from the level beneath.³⁰ Irreducibility is a result, not an attitude.

So the flatness of the map is entailed by anti-reductionism, not chosen for evenhandedness. No domain is more fundamental because none reduces to another; Formal is not above Natural for abstraction, nor Natural above Interpretive for exactness, because exactness is one virtue among several and confers no rank. And the deeper reason to refuse the ladder is that a ranking of knowledge has always doubled as a ranking of knowers — positivism's hierarchy licensed a politics of who counts as rational — so a flat map is not merely more accurate but disarms a smuggled hierarchy of persons. The epistemology and the ethics converge on the same refusal.

20 The unity is a question, not a subject

If the domains do not reduce to one another, what binds them into one atlas rather than five? Not a subject — they share none. A question, asked once and answered five ways: what would count as knowing this? Not "what is true of it," which is each entry's own labour, but the prior question of what warrant would even be the right kind here — what it would take for a claim in this region to graduate from belief to knowledge.

The five domains are five answers to that single question, and the answers are warrants, not topics. This is why the unity survives the irreducibility: domains that cannot be ranked or reduced can still be five responses to one interrogation, held together not by a common object but by a common question about the conditions of knowledge. The question is the spine; the domains are the vertebrae articulated on it; remove the question and the vertebrae are bone in a heap.

21 Five warrants, no sixth

The warrants, stated as verbs: derive — to know is to follow without gap from what is granted (Formal). Explain — to subsume the case under a law and thereby have predicted it (Natural, on the covering-law model of §9). Grasp from within — to know a system one composes and can move by knowing (Social, per the looping of §10). Understand — to recover a meaning until it is one's own (Interpretive, the verstehen of §11). Make — to know by constructing the thing and testing whether it holds (Applied, the verum-factum and knowing-how of §12).

The claim is that these five exhaust the warrants, that there is no sixth which is not one of them disguised — and the claim has teeth because it is drawn from independent results, not a tidy list: Ryle already forces make apart from the other four by separating knowing-how from knowing-that, Dilthey forces understand apart from explain, and the reflexivity result forces grasp apart from both. Five distinct warrants, each secured by a distinct argument, and the wager of the atlas is that human knowing has exactly these grammars and no more. It is a strong claim, offered to be tested, which is the only kind worth making.

22 The order is an epistemology, declared

The deepest thing the atlas asserts, it asserts before the first entry: the order is not laid over the knowledge like alphabetization over a dictionary but is a theory of knowledge, an epistemology drawn as a map. To accept the five domains is already to have accepted a first-order claim — that knowing comes in kinds, that the kinds are these five warrants, that none is sovereign. This returns to §1 with the loop closed: Foucault's charge that every classification encodes the conditions of knowledge of its age¹ is here not suffered but met head-on.

Most classifications keep that epistemology hidden in the shelving, where a metaphysics passes as mere arrangement and is never put to the reader for assent. This one drags the epistemology onto the frontispiece and signs it, precisely so that it can be refused. The map knows one thing before any entry is read — the shape of knowing — and it states that thing openly so the statement can be argued with, which is the difference between a theory and a filing convention.

22b The sixth domain, and why it is not a sixth warrant

Added in revision 2.2. The essay above argued that the five warrants exhaust the grammars of knowing and that no sixth exists which is not one of them disguised. The atlas has since acquired a sixth domain. This section exists because that looks like a refutation and is not — and because §24 promised that where a better cut appears the atlas is obliged to move, publicly.³⁷

The occasion was external. Mapped against a rival classification built on a closely related principle — the ground of answerability rather than the warrant — eight of its branches turned out to have no home in this atlas at all, and they clustered in two places: first-person experience, and the study of knowledge itself.³⁸ The first gap was a missing discipline and was filled as one. The second was structural: this essay argues at length that a classification is a theory exposed to error, yet the atlas had nowhere to file a theory of classification, and therefore nowhere to file itself. A map with no place for map-making has no place for the map.

So Domain VI was added. But notice what it is not. It is not a sixth answer to how do you know? An epistemologist arguing about justification is deriving; a laboratory ethnographer is grasping a practice from within; a historiographer recovering why a narrative was emplotted as tragedy is understanding; a methodologist testing whether a result replicates is explaining; a classifier building a scheme is making. The reflexive disciplines borrow all five warrants and add none — which is exactly what §21 predicted of any candidate sixth.

What distinguishes them is not their warrant but their object. Domains I to V are cut by the kind of backing a claim can have; Domain VI is cut by what the claim is about, namely knowing. The atlas therefore divides on two axes rather than one, and this is a real asymmetry that ought to be declared rather than smoothed over: five domains of first-order knowledge, and one domain that takes those five as its subject. The alternative — distributing reflexive work back among the first five, as this atlas originally did — is defensible and was tried. It failed for the reason given above: it left the atlas unable to locate itself, and a scheme that exempts itself is incomplete by its own standard.³⁹

The claim of §21 therefore stands, narrowed and strengthened: five warrants, no sixth — and one further domain, which is not a further warrant but the turn of all five back upon the enterprise that uses them. That the correction was forced by a rival scheme rather than found from inside is itself the finding. A classification cannot audit itself for the categories it lacks, because what is missing leaves no trace in the map that omits it. Only a different cut can show you the hole.

23 Self-similar to the leaf

The structure recurs at every scale. Above the whole stands this super-text; above each domain, a super-text that is to its members what this is to the five; above each sub-domain, another; and so to the leaf. Formal's super-text states what binds logic to number to computation and why they divide as they do — the same argument-form as here, rescoped and recolored, the one-and-many restated in a smaller key.

This self-similarity is not ornament; it follows from the atom being the warrant. Because each node is defined by the warrant it collects, each node is simultaneously a member of the map above and a map of the members below, and the same relation — this warrant, these instances — holds at every depth. To read the atlas correctly is to descend a recursion in which the frontispiece's sentence is spoken again, more locally, at each level. The whole is the largest utterance of a claim the structure repeats all the way down.

24 The shape closes; the contents cannot

The atlas cannot be finished, and this is structural, not a shortfall. New objects of knowledge appear without end — the made itself was once no domain and became a fifth of the whole — so the entries are permanently provisional, the leaves always multiplying. A completed atlas would map a world that had stopped producing knowledge, which is to say a dead one.

What can close is the shape: the five warrants, the two blades, the single question. New objects arrive constantly; a new warrant — a sixth irreducible mode of knowing — is a far stronger and rarer thing, and the atlas stakes itself on there being none. So the shape is fixed while the contents churn, and an atlas rightly built is a form that survives the loss of any of its contents: burn every entry and rewrite them, and if the five warrants are the true ones, each would fall back into the same house. Permanence of shape, not completeness of contents, is what it offers — and the only kind either could have.

25 The map is a refutable theory

The map now turns to face the reader as what it has claimed throughout to be: a theory, and therefore refutable. It is one cut, argued at every joint — the strongest the evidence presently forces — and not the only conceivable one. Borges's Chinese encyclopedia, dividing the animals into those belonging to the Emperor, those embalmed, those that from afar resemble flies,³² is the permanent warning: every taxonomy harbours, somewhere, a line where the ordering mind gave out, and the appearance of naturalness is exactly what a good classification must earn against that suspicion rather than assume. Kuhn's lesson completes it — categories are theory-laden and revisable, and the history of science is in part the history of recarving them.³³

So the reader who locates a truer joint is not the atlas's opponent but its only real collaborator, because a theory is confirmed by surviving the attempt to break it and improved by failing to. Take the blade. Cut again, and where the cut is better the atlas is obliged to move. It has staked everything on being the kind of thing that can be wrong — and the last thing it knows is that it does not know last.