The Operator Compendium¶

A thorough treatment of the fourteen discovery operators¶

The instrument manual for the program described in Operator-Driven Discovery of Prime Abstractions. Read that paper for the framing and the yield law; read here for the operator-by-operator detail.

Abstract¶

The discovery program in the companion paper is driven by fourteen operators — deterministic moves that transform an existing prime, or combine several, into a candidate new one. This document treats each operator in turn: what the move actually is (built up from everyday intuition), where it comes from, how you concretely apply it, worked examples, and what running it across the catalog taught us. It is written to be legible to a reader not already immersed in the project, and it doubles as a reference for anyone wanting to reproduce or extend the search.

How to read a section¶

This document is primarily an elucidation of the operators — what each move actually is, and how you would concretely go about applying it to a prime — and only secondarily a report of what happened when we ran it. Every operator is treated under five headings, in that order of emphasis.

What it is. The move, built from the ground up: the plain-language intuition first, then the structural statement, then why it is a genuine and distinct axis of variation. Sub-flavors are explained as different senses of the same move (run as separate passes), each with its own intuition. The goal is that a reader who has never thought about the operator can, after this subsection, recognize the move and see why it is natural.

Where it comes from. The operator's prior-art lineage — the mathematicians, engineers, and theorists who named or used the move long before we did. None of the fourteen is original to this project; placing each one in its intellectual pedigree is part of understanding what it is.

How to apply it. The concrete procedure. Given a prime, what question do you ask? What do you hold fixed and what do you vary? And — the load-bearing part — how do you tell a genuine transformation from a degenerate one that only looks like the move? Each operator has a "did it take?" test, and it is different for each; getting that test right is most of the craft.

Worked examples. Two or three applications, with the reasoning walked from input prime to candidate, not merely the input→output pair.

What sweeping the corpus taught us. A compact coda — the operator's characteristic collapse (how the move fails, retained as audit signal), its yield and where it sits in the yield law, and the method lesson running it produced. This is the empirical, "what we learned by actually doing it" material, kept subordinate to the elucidation above it.

The operators divide into three families: unary (transform one prime along one axis, operators 1–8), n-ary (combine primes, operators 9–12), and catalog-internal (operate on the graph rather than the primes, operators 13–14). A summary matrix closes the document.

A note on how a sweep runs¶

The "What sweeping the corpus taught us" coda in each section uses a few terms from the pipeline; here is the whole of what you need. Each operator is run through the same cheap-to-expensive funnel. A cheap, fast model generates one candidate per prime — this generation pass is the crank. Most candidates are collapses: the move returned the input essentially unchanged (a fixed point), or an abstraction already in the catalog (an alias), or something that is not a genuine pattern. Collapses are kept as signal, not thrown away — a fixed point tells you the prime is already maximally general, an alias confirms the catalog is well-balanced at that point. The thin stream of genuine survivors is then judged by a stronger, more expensive model (the adjudicator), and finally by a human. A slug is just the catalog's internal identifier for a prime (e.g. random_walk). Two recurring machine errors are named in the codas: the cheap model sometimes declares a candidate "already in the catalog" while naming a slug that does not actually exist (a false collapse), and the strong model sometimes rejects a real find because the original prime is its nearest neighbour (the retrieval-proximity fallacy — when a transform is genuinely new, its input is naturally its closest relative, so proximity is not evidence of redundancy). That is enough to read every coda; the full pipeline lives in the companion paper.

A note on reflexivity¶

A striking property of these fourteen operators is that they are themselves prime abstractions. Dualization is duality; generalization is abstraction; self-application is reflexivity; composition and analogy are primes by those very names. The discovery machine is built out of the same structural vocabulary it was built to catalog — which is exactly what one would expect if the vocabulary is genuinely closed. That correspondence is not merely elegant: when we checked it deliberately, exactly one operator — interpolation — had no home prime, and following that single thread revealed a real gap in the catalog (interpolation itself was missing, its safe pole hidden behind a failure-mode framing of its dual, extrapolation). The method, turned on itself, found a prime. Where an operator corresponds to a named prime, its section notes it.

That self-check was not itself one of the fourteen operators. It was an instance of a separate discovery mode — exogenous reference-coverage: take an authoritative external list (here, the operator set), align each element to a corpus prime, and mint the ones with no image. It is not a fifteenth operator, because an operator transforms primes already in the corpus whereas this mode maps an outside list onto it; the companion paper's closing section places it in the full trichotomy of ways the catalog grows.

Part I — Unary operators¶

Transform one prime by moving it along a single axis of variation. Each applies to every prime in the corpus. These are the operators that yield early and then become completeness instruments.

1. Dualization — the polarity axis¶

Status: full sweep, all three sub-flavors, 3,973 applications, cheap-model tier. Yield: 6 accept-recommendations + ~25 holds (provisional).

What it is. Dualization is the move you make when you look at a pattern and ask: what lives at the far end of this same dimension? Every structure sits somewhere along some axis, and often the opposite end of that axis is a structure in its own right — one that may or may not already have a name. The essential discipline is that you flip one thing to its extreme opposite while holding everything else constant; you are not looking for something merely unrelated or contrasting, but for the specific pole that co-defines the input by standing against it.

The reason "opposite" needs care is that it means three different things, which is why dualization has three sub-flavors, each run as its own pass. Logical negation (1a) is the presence-versus- absence flip: does the thing happen or not, is the element in or out. Redundancy (many independent backups) flips to the single point of failure (exactly one, no backup); the set operation "intersection" (belongs to all) flips to "union" (belongs to at least one). Opposing force (1b) is not about presence but about two tendencies pulling in opposite directions, each meaningful only against the other: centralization and decentralization, exploration and exploitation. Neither pole is the "absence" of the other; they are a matched pair. Failure-mode (1c) is the subtlest — it is the characteristic way a pattern breaks, treated as its own structure: the dark twin that a pattern's own success produces. Optimization, pushed too hard, becomes overfitting; redundancy, when the backups share a hidden dependency, becomes common-mode failure. The failure is not the opposite of the pattern; it is the shadow the pattern casts.

Polarity is a genuine axis — not a grab-bag of contrasts — because opposed structures come in pairs, and catalogs systematically name one pole and forget the other. We talk about redundancy far more than about common-mode failure, about action far more than inaction; the more salient pole gets the name, and its twin sits unnamed. So a swept polarity axis is really a hunt for asymmetries in what has been named: wherever one pole of a real pair is present and the other absent, there is a candidate.

Where it comes from. Dualization has the deepest formal pedigree of the fourteen. In category theory, every construction has a dual obtained by reversing all its arrows, and the "opposite category" hands you a dual of every theorem for free — duality is a first-class structural principle, not an afterthought. Its familiar finite face is De Morgan's laws (the negation of an AND is an OR of the negations), which is exactly the intersection↔union flip. The opposing-force sense (1b) is the ordinary dialectical intuition that many concepts are defined by what they stand against; the failure-mode sense (1c) is the engineering discipline of failure-mode-and-effects analysis, which treats each way a system can break as a structure worth naming and anticipating. And dualization is itself a prime — the catalog holds duality, along with its relatives complementarity, symmetry, and inversion — so the operator is one of the catalog's own entries put to work.

How to apply it. Given a prime P, the procedure is: (1) find the single dimension along which P sits at an extreme — redundancy sits at the "many" end of a multiplicity dimension, intersection at the "all" end of a quantifier dimension; (2) choose which sense of opposite you are testing (the three sub-flavors are separate passes, because the same prime can have a logical-negation dual, an opposing-force dual, and a failure-mode dual, all different); (3) describe the pole at the other end of that one dimension as a positive structure — state what lives there, not what is missing.

The "did it take?" test for dualization is the crucial craft: a genuine dual has its own anatomy, describable without the word "not." Union is not "not-intersection"; it is "belongs to at least one," a rule with its own algebra. Surjectivity is not "a map that fails to be injective"; it is "every target is hit," a property you can state and check on its own terms. If the best you can say about your candidate is "P doesn't happen" or "the absence of P," you have produced a degenerate negation, not a dual, and it should collapse. The test, in one question: can you name the dual's mechanism without referring to P? If yes, it is a real pole; if no, it is just P switched off.

Worked examples. intersection, flavor 1a: intersection is AND over membership (an element qualifies only if it is in every set); flip the connective to OR and you get "in at least one set," which is union — a co-equal operation with its own De Morgan algebra, and one genuinely missing from the catalog when the operator proposed it. optimization, flavor 1c: optimization succeeds by driving a system toward the best value of an objective; push that success past the point of generalization and it fits the sample's noise as if it were signal — that failure structure is overfitting, a real abstraction with its own anatomy, not "optimization turned off." action_bias, flavor 1b: the systematic tendency to prefer doing something has, as its co-defining opposite, the systematic tendency to prefer doing nothing — omission bias — the two being deviations of equal standing in opposite directions.

What sweeping the corpus taught us. Dualization collapses in two informative ways: a fixed point, where the prime is its own dual (symmetry negated is still symmetry), logged as a structural fact about the catalog; and an alias collapse, where the dual is real but already named (decentralization is present), which confirms the catalog is well-balanced at that point. The costly error — the one that took several rounds of prompt redesign to suppress — is the false collapse, where the cheap crank waves a genuine dual away as "already in the catalog" while naming a slug that does not actually exist. As an axis-step operator, dualization yielded a concentrated handful of real finds (eager_evaluation, risk_seeking, confabulation, pareidolia, hold_up_problem, consequentialism) against a long tail of collapses that fed the corpus audit and the training-negative set. The method lesson: dualization built the safeguards every later operator inherited — a rule that verifies an "already exists" claim against the actual catalog before trusting it, the finding that embedding-similarity alone cannot separate true aliases from genuine new candidates, and above all the protection rule: a candidate that several sub-flavors independently propose, or that an earlier pass already flagged, may never be killed by the cheap crank on its own say-so. It is the operator that taught the program not to trust the cheap verdict.

2. Self-application — the reflexive axis¶

Status: full sweep, 1,325 applications, capable-model tier. Yield: 15 holds (provisional); the highest fixed-point census of any operator (41%).

What it is. Self-application is turning a pattern on itself — taking a structure that normally operates on ordinary things and pointing it back at its own kind. The name sounds abstract, but the everyday instances are completely familiar once you see them: learning to learn is the pattern "learning" applied to learning itself; a rule about rules is the rule-concept applied to rules; wanting to want something different is a preference operating on your own preferences; knowing what you know is knowledge turned on knowledge. In each case a first-order structure — one that acts on the world — is lifted to a second-order structure that acts on the first. That lift "one level up" is the move.

There are two senses in which you can point a pattern at itself, and both are run. The reflexive sense is P about P: the pattern takes its own outputs as its subject. A preference over your preferences (do I endorse wanting what I want?); a belief about your beliefs; an uncertainty about how uncertain you are. The machinery sense is P applied to the process that produces or sustains P: instead of turning the pattern on its own outputs, you turn it on the mechanism that generates it. Natural selection acts on organisms; apply the selection idea to the very capacity to evolve and you get selection for evolvability. A norm governs behavior; apply the norm-concept to whether norms get enforced and you get a metanorm — a norm requiring the punishment of those who break norms.

Why "one level up" is a real axis, and not a trick of wording: the second-order structure genuinely behaves differently from the first, in ways you cannot reduce away. A preference over outcomes and a preference over your own preferences are not the same shape — the second introduces conflict between levels (the smoker who prefers to smoke and simultaneously prefers not to have that preference), and that inter-level conflict is the entire structure of self-command, willpower, addiction, and value change. It has anatomy the first order simply lacks. So wherever a structure can take its own type as input, there is a potential second-order prime hiding — and the whole point of sweeping this axis is to find the ones whose second-order form is load-bearing enough to deserve its own name.

The characteristic subtlety, which you must understand to apply the operator at all: sometimes turning P on itself changes nothing. Symmetry-under-symmetry is still just symmetry; recursion-of-recursion is recursion; closure applied to closure is closure. These are fixed points, and they are not failures of the move — they are a discovery in their own right. A prime that is its own self-application is a maximal generalizer, a kind of "top" element of its region: there is no higher level for it to go to. Roughly two in five primes turn out to be fixed points under this operator, and that census is a direct map of the catalog's most self-referential structures.

Where it comes from. The formal home of self-application is the corner of mathematics that studies structures which take themselves as arguments: fixed-point theory (a fixed point is exactly a place where applying an operation returns its input) and, in computer science and category theory, recursion and monads (a monad is, loosely, a device for wrapping a computation inside more of its own kind). The informal home is Douglas Hofstadter's "strange loops" — the self-referential tangles (a hand drawing the hand that draws it, a sentence about itself) he argued underlie consciousness. Self-application is itself a prime, and pointedly so: the catalog holds reflexivity_self_reference, recursion, and fixed_point, so this operator is quite literally the discovery method applying one of its own entries to the rest of them.

How to apply it. Given a prime P: (1) ask what P takes as input, and whether that input slot can be filled by P itself (the reflexive route) or by the process that produces P (the machinery route); (2) construct "P of P" and describe it as a concrete structure — the norm that enforces norms, the preference over preferences; (3) apply the "did it take?" test, which for this operator is sharper than for any other: does the second-order version have new anatomy that the first order did not? Second-order preference introduces cross-level conflict — new anatomy — so it took. Symmetry of symmetry introduces nothing — it is a fixed point. (4) If there is new anatomy, you have a candidate; if not, log it as a fixed point (a maximal-generalizer finding), not a failure.

The failure mode to guard against here is specific and seductive: the vacuous meta-slogan. "Meta- learning," "meta-cognition," "meta-strategy" are easy to generate and sound like second-order primes, but many have no distinct mechanism — they are the word "meta" stapled to a prime with nothing new underneath. The discriminator is always the same question: can you name a mechanism the first order did not have? If you cannot, the "meta-" is decoration, and the candidate should collapse.

Worked examples. preference, reflexive route: a preference over one's own preferences is second-order preference. The new anatomy is inter-level conflict — you can want X while wanting not to want X — which is the structure of self-command and addiction, absent from a plain preference over outcomes. It took, and it was independently reached from five different inputs, which is what marked it as a genuine gap rather than a slogan. norm, machinery route: apply the norm-concept to the enforcement of norms and you get a metanorm — a norm requiring punishment of norm-breakers (and of those who fail to punish). The new anatomy is that a metanorm is what makes a first-order norm self-sustaining; without it, norms decay. symmetry, either route: symmetry applied to symmetry is still symmetry — no new anatomy — so it is a fixed point, logged as evidence that symmetry is a maximal generalizer rather than as a discovery.

What sweeping the corpus taught us. The fixed point dominates — 41% of primes return themselves, the largest and cleanest fixed-point census in the program, because recursion, closure, reflexivity, and symmetry are maximal generalizers. The genuine finds are all high-convergence (second_order_ preference from five inputs, metanorm from four, meta_learning and structural_stability and evolvability from two each); the individual "meta-P" coinages are trivial, and only the convergence separates signal from noise. The method lesson: self-application forced a safeguard against a specific self-inflicted error. Because "P about P" always references P, the adjudicating model is tempted to wave a genuine second-order prime away as "mere iteration" — the reflexive-axis version of the retrieval-proximity fallacy. The fix is a rule about when a dismissal can be trusted: if the model actually wrote out a candidate structure and then called it trivial, the candidate is sent on to be judged properly anyway; only a dismissal that comes with no candidate at all is taken at face value. Without it, the operator's strongest find — second_order_preference — would have been discarded in silence.

3. Temporalization — the time axis¶

Status: full three-flavor sweep, 3,971 applications, cheap-model tier. Yield: 5 holds (provisional); the second-largest fixed-point census in the program (59% — the catalog is deeply dynamical).

What it is. Temporalization is the move you make when you look at a pattern that describes a state of affairs and ask: what does this look like while it is happening? Most abstractions name a thing that sits still — a balance, a configuration, a stock, a resting point. Temporalization sets that thing in motion and asks whether the motion has a structure of its own worth naming. The everyday instances are familiar the moment you see them. A bank balance is a number frozen at an instant; the throughput moving through the account — deposits and withdrawals per unit time — is a different, dynamic thing. A photograph of a pendulum tells you where it is; its trajectory tells you where it is going and how fast. Standing still is one structure; the shape of the motion is another, and the second is not contained in the first.

"Add a time axis" means three genuinely different things, which is why temporalization has three sub-flavors, each run as its own pass. Level→flow (3a) turns a quantity held into the rate at which it changes — a stock becomes a flow, an amount becomes a derivative. Inventory (how much is on the shelf) becomes throughput (how fast it moves); a reservoir's level becomes the flow filling or draining it. The new structure is a rate, with its own units and sign. State→trajectory (3b) is the richest of the three, and the operator's discovery engine: a static configuration or resting state becomes the path by which the system reaches it. An equilibrium is a place where forces balance and change stops; its trajectory is the approach — the transient by which a disturbed system settles back, complete with a settling time, a possible overshoot, and a rate of return, none of which the resting point itself contains. Present→anticipation (3c) turns a present-focused structure into its forward-looking version: a reactive control becomes a predictive one, a present choice becomes planning over future choices, a preference becomes a preference over time. The new structure is an expectation or forecast reaching into the future that the present-tense original does not have.

Time is a genuine axis, and not just "the same thing, observed for a while," because motion has anatomy that stillness lacks. A resting equilibrium and the relaxation dynamics of returning to it are not the same shape: the second has a transient, a rate, an overshoot, a relaxation time — a whole vocabulary of dynamical structure the endpoint simply does not possess. Wherever a static prime has a dynamic counterpart with structure of its own, there is a candidate. And wherever the prime is already dynamic — where adding a time axis returns the input unchanged — that too is a finding, of which more below.

Where it comes from. Temporalization's lineage is thinner than dualization's or self- application's, but it has a clean home in the engineering literature: it is one of the classic TRIZ trends of technical-system evolution — the observed tendency of engineered systems to become more dynamic over time (rigid parts giving way to hinged, then flexible, then fluid ones). Here that "increasing dynamism" trend is applied to abstractions rather than machines. Temporalization corresponds directly to a catalog prime — it is time put to work, and its state-to-trajectory sense draws on state_and_state_transition — so, like the other unary operators, it is one of the catalog's own structural entries turned on the rest.

How to apply it. Given a prime P: (1) decide which sense of the move you are running — is there a rate hiding in P (3a), a resting state whose approach you can trace (3b), or a present-tense structure you can point at the future (3c)? (2) Construct the dynamic counterpart and describe it as a concrete structure — the relaxation transient, the throughput, the planning horizon — not as "P, but over time." (3) Apply the "did it take?" test, which for temporalization is the dynamics rule: name one temporal structure the dynamic version has that the static input does not — a rate, a lag, a transient, a trajectory shape, a path-dependence, a convergence, or an anticipation. If you can name it, the move took; if the only thing you can say is "P, but happening over time," it did not.

Two failures to guard against. The first is the fixed point: many primes are already dynamic — a half-life is already a rate, a Markov process is already an evolution, feedback and diffusion and oscillation are already motion — so adding a time axis returns the input. This is not a failure; it is a census entry recording that the prime is inherently temporal. The second, seductive one is the "rate-of-P" coinage: the cheap crank mechanically prefixing "the rate of" to any static prime, producing something grammatical but empty. The dynamics rule is exactly what separates the two — a genuine rate is a structure you can describe; a "rate-of-P" coinage has nothing to say about the rate itself.

Worked examples. equilibrium, flavor 3b: an equilibrium is the state where opposing influences balance and change stops. Trace the approach to it — the path a disturbed system takes on its way back — and you get a transient with a settling time, a damping, and a possible overshoot: transient response (relaxation dynamics). The endpoint has none of that anatomy; the return does. It took, and it was reached from a dozen distinct inputs, which is what marked it as a real gap rather than a one-off. criticality, flavor 3b again: as a system nears a tipping point, the very rate of its return to equilibrium slows — a dynamical early-warning signature that is distinct from the tipping point itself. That is critical slowing down, and it came in alongside transient response. half_life, any flavor: a half-life is already the constant time for a quantity to fall by half — a rate over time — so temporalizing it returns it. Logged as a fixed point, evidence that the prime is inherently temporal, not as a discovery.

What sweeping the corpus taught us. Temporalization's dominant outcome is the fixed point, and that is the point: 788 of the 1,325 primes — 59% — turn out to be already inherently temporal (feedback, diffusion, oscillation, random walk, and their kin all return themselves), so the sweep's primary product is a census of the catalog's dynamical core rather than a pile of new primes. A further 357 applications collapsed as aliases — the dynamic counterpart is real but already named, which confirms the catalog's dynamic layer is well-placed. Against that audit-heavy background the genuine discovery is modest: one strong, high-convergence find (transient_response) and four moderate dynamic primes. This is the signature of an axis-step operator whose axis was already well-covered — it yields a little and audits a lot. The method lesson: temporalization proved that model tier is cheaply predictable. Before committing to the sweep, a fifteen-prime probe checked whether the cheap model agreed with the expensive one on the hard judgments (is this prime already temporal? is this a real dynamic prime or an empty "rate-of-P"?); it agreed 12 out of 15, so the entire sweep ran on the cheap model, making temporalization the least expensive full operator in the program. That probe became standard practice — and, one operator later, its worth was proven from the other direction.

4. Spatialization / scaling — the extent axis¶

Status: full three-flavor sweep, 3,975 applications, capable-model tier. Yield: 6 holds (provisional); fixed-point census 37%. The most expensive unary operator (~4.5M tokens).

What it is. Spatialization is the move you make when you take a pattern defined at a point — at one place, for one item, at one scale — and ask: what happens when it is spread out, or when there are enough of them to form a mass? Many abstractions are implicitly local: a value, a single interaction, a discrete count. Spatialization gives them extent and asks whether the extended version has structure the point version lacks. The everyday instances are immediate. The temperature at a spot is a number; the temperature across a room is a field, with warm and cold regions, gradients between them, and a boundary at the walls — structure that a single reading cannot have. One person's choice to buy is an individual act; the demand of a whole market is an aggregate that behaves according to laws no single buyer follows. Spreading a thing out, or piling enough of it together, produces genuinely new structure.

"Add extent or scale" means three different things, run as three passes. Point→field (4a) spreads a localized quantity over a whole space, so that it takes a value at every point — and with that come gradients (how fast it changes from place to place), boundaries and edge effects, and spatial correlation (nearby points resembling each other). A force at a point becomes a force field; a value becomes a value everywhere. Discrete↔continuous (4b) flips granularity. Push a countable structure to its limit and the sum becomes an integral, the count becomes a density, the jagged becomes smooth — a continuum with derivatives and integrals the finite version could not support; or run it the other way and a smooth quantity becomes quantized. Micro→macro (4c) aggregates an individual-level structure into its collective counterpart: many interacting units become a mean field in which each one feels only the average of all the others, and the aggregate obeys laws unlike any individual's. Local rules become global patterns; per-item quantities become per-unit-extent densities.

Extent is a real axis because a distributed structure carries machinery a point cannot. A single value has no gradient, no boundary, no correlation length; a field has all three. A finite sum has no derivative; its continuum limit does. An individual has no notion of "the average of everyone else"; the mean-field aggregate is built from exactly that. Wherever a local prime has a genuinely extended counterpart — one with a field's structure, a continuum's operations, or an aggregate's emergent laws — there is a candidate. And wherever a prime is already spatial or scale-free, adding extent returns it, which is a census entry rather than a discovery.

Where it comes from. Spatialization has two distinct lineages. The engineering one is again TRIZ, whose "transition to the micro-level" trend describes systems evolving toward finer, more-distributed structure. The cognitive-science one is more pointed: Peter Gärdenfors' conceptual- spaces program treats concepts themselves as regions of a geometric space, so that reasoning is literally movement through an extent — spatialization made into a theory of concepts. Spatialization corresponds to the catalog primes scale_invariance and dimension; it is those entries applied to the rest.

How to apply it. Given a prime P: (1) choose the sense — is P a point quantity you can spread into a field (4a), a discrete or continuous structure whose granularity you can flip (4b), or a micro-level structure you can aggregate (4c)? (2) Construct the extended counterpart and describe it as a concrete structure — the field with its gradients, the continuum limit with its integrals, the mean-field aggregate — not as "P, but bigger." (3) Apply the "did it take?" test, the extent rule: name one spatial or scale structure the extended version has that the input lacks — a gradient, a boundary or edge effect, a correlation length, a per-unit-extent density, an aggregation statistic, a continuum limit with new operations, or a scale-dependence. If you can name it, it took.

The two failures mirror temporalization's. The fixed point: many primes are already spatial or scale-free — a fractal is already scale-invariant, a network is already a distributed relational structure, diffusion is already a field process — so spatializing them returns them, and that is a census entry. The seductive one is the "field-of-P" coinage — the spatial twin of "rate-of-P," a cheap model gluing "a field of" onto any prime with nothing to say about the field. There were roughly a thousand of these, and the extent rule is what demotes them: a real field has gradients and boundaries you can describe; a coinage has only the phrase.

Worked examples. a pointwise value, flavor 4a: spread a single quantity across a space so that it takes a value at every point, and you get gradients, boundaries, and spatial correlation — the structure that "a field of P" instantiates for countless P. Hundreds of inputs converged on this same missing genus, field, which the catalog conspicuously lacked: it held gradients and diffusion acting on fields, and a receptive field as an instance, but no field itself. non_zero_sum_game and convolution and hysteresis, flavor 4c: replace each unit's detailed interactions with the average effect of all the others, and the collective is described by one self-consistent aggregate — mean_field, reached from those three unlike inputs. fractal_geometry, any flavor: a fractal is already self-similar across scales, so adding scale returns it — a fixed point, logged as evidence that the prime is inherently multi-scale rather than as a find.

What sweeping the corpus taught us. Like temporalization, spatialization is mostly audit: 495 primes (37%) are already spatial and return themselves, and 200 more collapse as aliases of an existing spatial prime. The catalog's spatial layer is well-populated — though, tellingly, less saturated than its temporal layer (37% vs 59%), which is why spatialization found a genuine hole where temporalization mostly confirmed. The discovery is small but real: three strong finds (field, mean_field, continuous_relaxation) plus a few moderate ones, isolated from ~1,000 "field-of-P" coinages by the same discriminator every axis-step operator relies on — a canonical name plus convergence from many unlike inputs. The method lesson: spatialization is the case that proved the tier probe earns its keep in both directions. Its fifteen-prime cheap-vs-capable probe agreed only 5 times out of 15 — because deciding "is this prime already spatial?" and "does a genuine field emerge here?" needs judgment the cheap model does not have — so the probe correctly routed the whole sweep to the expensive model. That made spatialization the most expensive unary operator in the program, but the cost was one the probe let us anticipate rather than discover after paying for a failed cheap run.

5. Generalization / lift — the abstraction axis¶

Status: full sweep across three strata, 1,326 applications, capable-model tier. Yield: ~47 holds (provisional) — the highest-yielding operator, and the strategically most important.

What it is. Generalization is the move you make when you look at a specific pattern and ask: what is this a special case of? Almost every abstraction is a species of some broader abstraction — a narrower thing that inherits its shape from a wider one. Generalization walks up that relationship: it strips away exactly one constraint, or the particular domain the pattern lives in, to reach the genus — the nearest parent of which the input is an instance. The everyday version is completely ordinary. A robin is a bird; "bird" is the genus you reach by forgetting the robin's particulars. A random walk — a path built by adding independent random steps — is one kind of stochastic process, the genus that also contains Markov chains and Brownian motion; you reach it by forgetting the specific step rule and keeping only "a family of random variables indexed by time." The move is: hold the input, let go of one thing, and name what remains.

The single most important distinction for this operator is the difference between a genuine lift and its two impostors. A real lift climbs to a strictly more abstract structure that the input and other, unlike things all instance. The first impostor is the restatement: a paraphrase of the input at the same level — "a random walk is a wandering path" — which climbs nowhere. The second is the overgeneral umbrella: a jump so far up that the genus is vacuous — lifting "random walk" to "process," or "dialectic" to "interaction," reaches a word that would parent half the catalog and therefore says nothing. A valid genus threads between these: more abstract than the input, but still narrow enough to exclude most things. The test that pins this down is the sibling test — a genus is only real if you can name two or more other instances of it that are structurally unlike the input. "Stochastic process" passes because random walks, branching processes, and queues are its siblings and they are genuinely different from one another. A restatement fails because its only "sibling" is the input reworded.

Now the deep structure that makes generalization the program's most important operator: the corpus is stratified, and the lift behaves differently on each stratum. Some primes are structural — their content is pure abstract form, and they already sit close to their genera, so lifting them yields little (they are already near the top of their branch). But many primes are framed — their content is bound to a particular human practice or domain (a concept from law, from critical theory, from management), and underneath the frame there is almost always a genuine structural parent that has not been named. Lifting a framed prime strips the domain frame and exposes that parent. This is what makes generalization strategic rather than merely productive: framed primes live in an open, ever-growing layer (there is no end to human practices to name concepts from), while structural primes form a closed core. Every successful lift of a framed prime moves mass from the open layer into the closed core — it converts an endlessly-multipliable domain concept into a bounded structural one. We called this the staircase: generalization is not a treadmill running in place but a staircase carrying the catalog upward toward structural closure. That is why it yields most, and why its yield matters most.

Where it comes from. Generalization has the most venerable pedigree of any operator in the program. It is the first of George Pólya's three master heuristics of mathematical discovery (generalization, specialization, analogy), the move he placed at the head of How to Solve It — "it is often useful to consider a more general problem." Long before that, it is the genus–differentia definition of Aristotle (to define a thing, name its genus and the difference that distinguishes it within that genus) — the lift is that definition read backwards, dropping the differentia to recover the genus. And it is the everyday engine of mathematics itself, where naming the general structure behind a family of special cases (the group behind the symmetries, the topology behind the metric) is most of the subject. Generalization corresponds to the catalog prime abstraction — it is, quite exactly, the catalog's entry for the abstraction move applied to its own entries.

How to apply it. Given a prime P: (1) identify one constraint or the domain frame you can strip — for a framed prime, the frame itself is usually the thing to drop. (2) State the genus that remains: one sentence, domain-general, describing the parent structure. (3) Apply the sibling test, which is this operator's "did it take?": name two or more instances of the proposed genus that are structurally unlike the input. If you can — the genus is real; if you cannot, you have produced a restatement and it collapses. (4) Check the genus is not a vacuous umbrella (does it still exclude most of the catalog?) and not already an ancestor of P in the hierarchy (in which case the lift is a trivial rediscovery, logged as a consistency confirmation).

The failure to guard against is the restatement disguised as a lift — and it is disguised well, because a paraphrase and a genuine genus can read almost identically until you demand the siblings. This is exactly why the sibling test is load-bearing rather than decorative: it is the only thing that reliably separates "a random walk is a wandering path" (no siblings — restatement) from "a random walk is a stochastic process" (three unlike siblings — lift).

Worked examples. random_walk → strip the specific increment rule and keep "a family of random variables indexed by time governed by probabilistic law," and you have stochastic process — a genus whose siblings (Markov chains, branching processes, queues) are genuinely unlike a random walk. This is the exemplar the operator was calibrated on, and the same move done by hand years earlier had been the catalog's single most productive source of new primes. triangulation → strip the investigative frame and the genus is "credibility earned when methodologically independent sources converge," whose siblings include sensor fusion, corroboration, and replication — an established idea (Whewell's, then Wilson's) that the catalog simply lacked: consilience. directed_acyclic_graph → strip the graph representation and a DAG is revealed as a species of partial_order — the foundational order-theoretic genus that the catalog, remarkably, had never named on its own. Each of these is a genus whose species were present but whose parent was missing — which is precisely the shape of a real gap.

What sweeping the corpus taught us. Generalization was run in three strata — a 53-prime pilot, then the 236 framed primes, then the ~1,037 structural and unlabeled ones — and it is the highest-yielding operator in the program, with about 47 holds. The yield law reads straight off the strata: the framed sweep produced 25% new candidates, the structural sweep only 15%, exactly as the staircase predicts — framed primes hide un-named structural parents, structural primes already sit near theirs. The framed finds (exploration_exploitation, consilience, referent_drift, and others) are the staircase in action; the structural finds skew toward one-off coinages with a lower keep rate. A large byproduct came for free: 259 candidate hierarchy edges across the sweeps — cases where the named genus already exists but is not yet drawn as an ancestor of the input, a densification harvest that rivals the prime yield.

The method lesson — and generalization produced two of the biggest in the whole program. First, the capability floor. The cheap model, run on the pilot, produced zero genuine genera in 53 applications: it paraphrased every input instead of abstracting, because a real lift requires holding the input and its prospective siblings in mind at once, which is beyond the cheap model's reach. The fix was two-fold — move the generation to the expensive model, and add the sibling-evidence rule (no genus counts unless the model can name two structurally-unlike instances of it). With that, the same pilot went from 0 to 8 genuine lifts. Some operators can run cheap; this one cannot, and the pilot is what revealed the floor before a full cheap sweep was wasted on it. Second, the retrieval-proximity fallacy, in its purest form. On the framed sweep, the expensive adjudicator rejected every one of the 35 survivors on its first pass — and about 40% of those rejections reasoned "the input is the top search hit for this candidate, so the candidate must be a restatement of the input." That is exactly backwards. When a genus is genuinely absent from the catalog, the input is necessarily its own nearest neighbor — a species has no closer relative than itself when its parent is missing — so retrieval proximity is not evidence of redundancy; it is the expected signature of a real find. The sibling test, not search rank, is the discriminator. Once that correction was written into the adjudicator's instructions, all eight framed finds were recovered, and it became a standing amendment every downstream operator inherited.

Status: full three-flavor sweep, 3,975 applications, capable-model tier. Yield: 4 holds (provisional); fixed-point census 22% — the least self-saturated axis, yet the highest alias absorption.

What it is. Modalization is the move you make when you take a pattern that describes what is and ask instead about what could be, what is likely, or what is required. Ordinary abstractions often name a settled fact — an outcome that happened, a rule that holds, a value that is fixed. Modalization shifts that fact onto one of the axes philosophers call modality — possibility, probability, or obligation — and asks whether the shifted version is a distinct structure. The everyday instances are familiar once named. The decision you made is an actuality; the option you held before deciding — the right, but not the obligation, to act, worth more the more uncertain the future — is a modal structure with its own value. A fixed price is a number; the expected value of an uncertain payoff is a different number computed by integrating over everything that might happen. A law that must be obeyed is one thing; a default that holds unless overridden is another, with a built-in mechanism for its own exceptions.

"Shift the modal status" means three different shifts, each run as a pass. Actual→possible (6a) moves from what is realized to the space of what could be: an outcome becomes the set of possible outcomes it was drawn from, a fact becomes its counterfactual alternatives, a commitment becomes an option. The new structure is a space of alternatives, each with an accessibility or a likelihood. Deterministic→stochastic (6b) moves from certain to probabilistic: a fixed relation becomes a distribution over outcomes, a fixed value becomes an expectation integrated over that distribution, a clean rule becomes a noisy one. The new structure is a probability space with an expectation and a variance. Mandatory→default (6c) moves normative force from binding to defeasible: a hard constraint becomes a soft presumption that holds unless sufficient reason overrides it. The new structure is a defeasibility — an explicit account of when the rule bends.

Modality is a real axis because a possible, probable, or defeasible version of a structure carries machinery the flat actual version cannot. A committed outcome has no optionality; an option has an asymmetric payoff (upside kept, downside capped) over a space of futures. A fixed value has no room for an expectation; a random one is defined by integrating over its distribution. An inviolable rule has no override conditions; a defeasible one is built out of them. Wherever an actual prime has a genuine modal counterpart with that kind of new machinery, there is a candidate — and wherever a prime is already modal (a probability is already stochastic, an optionality is already about possibilities), shifting modality returns it, which is a census entry.

Where it comes from. Modalization has a rich and threefold pedigree, one strand per sub-flavor. The actual→possible sense is modal logic — specifically the possible-worlds semantics of Saul Kripke, in which "possible" and "necessary" are made precise as quantifiers over a space of accessible worlds. The deterministic→stochastic sense is probability theory itself, the mathematics of reasoning under uncertainty. And the mandatory→default sense is default and non-monotonic logic — Raymond Reiter's default logic and its relatives, which formalize reasoning that can be withdrawn when new information arrives. Modalization corresponds to the catalog primes counterfactuals and the modality family; it is those entries applied to the rest.

How to apply it. Given a prime P: (1) choose the sense — can you open P into a space of possibilities (6a), make it probabilistic (6b), or make it defeasible (6c)? (2) Construct the modal counterpart and describe it as a concrete structure — the option with its asymmetric payoff, the expectation over a distribution, the presumption with its override conditions — not as "P, but maybe." (3) Apply the "did it take?" test, the modal rule: name one modal structure the shifted version has that the input lacks — a space of alternatives with likelihoods, a distribution with an expectation, an option that is a right-not-obligation, a counterfactual dependency, or a defeasibility. If you can name it, it took.

The failures are the now-familiar pair. The fixed point: probability, randomness, optionality, counterfactual reasoning are already modal and return themselves. The seductive one is the "modal word attached" coinage — "P, but it might happen," "P, but random," "P, but optional" — the modal twin of "rate-of-P" and "field-of-P," grammatical and empty. The modal rule demotes it: a genuine modalization names a distribution or an option or an override condition; a coinage names only the adverb.

Worked examples. a deterministic value, flavor 6b: replace a fixed quantity with its probability-weighted average over everything it might be, and you get expected value — the single number a random quantity "amounts to" in the long run, defined by an integral over a probability space the fixed value had no room for. (A subtlety the sweep flagged: expected_utility is already in the catalog, so expected_value is either its genus or a fold — the kind of judgment call left for curation.) a hard rule, flavor 6c: shift a rule from inviolable to overridable and you get defeasibility — a presumption that holds unless defeated, the structure underneath default logic, legal presumptions, and non-monotonic reasoning. It was reached from six unlike inputs, and it generalizes the catalog's existing mandatory_vs_default_norms (which is just the deontic-norms special case) to the general defeasible-reasoning prime. probability, any flavor: a probability is already a measure over a space of outcomes, so modalizing it returns it — a fixed point.

What sweeping the corpus taught us. Modalization is the operator that forced a distinction the program needed: there are two ways for an axis to be saturated. Temporalization was saturated by returning the input — a high fixed-point census. Modalization is the opposite: its fixed-point census is the lowest of any axis (only 22% of primes are inherently modal), yet its alias absorption is the highest — over 400 modalizations collapsed as aliases of an existing prime. The reading is subtle and specific: the modal layer is covered not by many inherently-modal primes but by a small set of unusually powerful ones (probability, optionality, counterfactual reasoning, expected utility, mixed strategy) that soak up nearly every modalization thrown at them. A few primes doing an enormous amount of work. That is why modalization's genuine discovery yield is the smallest of the operators — four holds, led by the high-convergence genus possibility (alethic modality itself, absent as a standalone prime) and defeasibility. The method lesson: modalization sharpened the vocabulary of saturation itself. "Saturated" had been used loosely; modalization showed it means two structurally different things — the input is already the answer (a fixed point) versus a handful of powerful existing primes already name the answer (an alias). Both are completeness results, but they say different things about how the catalog is organized, and the fixed-point census and the alias count are precisely what tell them apart.

7. Perspective inversion — the viewpoint axis¶

Status: pilot only (53 applications), retired as saturated, capable-model tier. Yield: 1 framed find (interpellation, provisional).

What it is. Perspective inversion is the move you make when you notice that a pattern is being told from one side of a relationship, and you ask what the same relationship looks like from the other side. Almost every relation has two seats, and we habitually name only one of them — usually the more active, more visible one. The seller advertises; there is also a buyer who must decide whether to believe the advertisement. A speaker sends a message; there is also a listener who must infer what was meant. A predator hunts; there is also prey that must detect and evade. In each case the second seat is not the absence of the first — it is a full structure in its own right, facing its own problem, with its own logic. Perspective inversion swaps to that second seat and asks: is what the other party is doing a distinct pattern that deserves its own name?

The everyday version is completely familiar. When someone signals their quality to you — a job candidate flashing a prestigious degree — you, the observer, are not merely "the person being signaled at." You are running a genuinely different operation: deciding which signals to trust, designing tests that separate real quality from fakery, screening. The signaler's problem (how do I credibly reveal what I know?) and the observer's problem (how do I extract what they're hiding?) are mirror images, but they are not the same shape — one emits, the other infers. That asymmetry is the whole point. Where a relation is genuinely two-sided, each seat is a candidate prime.

Unlike most operators, perspective inversion has no sub-flavors — the move is always the same single gesture: identify the relation a prime sits in, identify which seat it is read from, and swap to the complementary seat. What varies is only the type of relation (sender/receiver, principal/agent, predator/prey, teacher/learner, attacker/defender), and those are illustrations of one move, not different moves.

Why is this a genuine axis and not just a rewording? Because relations that name only one seat leave a real, describable structure unnamed — and the test of whether the second seat is genuine is whether it faces a different problem than the first. If swapping seats hands you back the same structure, the relation was symmetric (a mutual exchange, a handshake) and there was never a second seat to find. If swapping seats hands you a structure the catalog already names under another word, the pair was simply named in full. And if the prime was never relational to begin with — a pure property like symmetry, which has no "other side" — there is no seat to swap to. So a swept viewpoint axis is a hunt for asymmetric relations whose passive seat was forgotten.

Where it comes from. The intellectual home of perspective inversion is the study of relations where the two parties face different problems: game theory and mechanism design, where every principal has an agent and the designer's problem is distinct from the player's; and information theory's sender/receiver duality, where encoding and decoding are separate operations on the same channel. Perspective-taking is itself deeply catalogued — the corpus holds perspective, frame_of_reference, figure_ground, duality, complementarity, and inversion as named primes, which is exactly why one expects the relational seats to be well-covered too. As a discovery move, perspective inversion corresponds most closely to the primes frame_of_reference and perspective: it is one of the catalog's own viewpoint entries turned into a search instrument. It also borders dualization's opposing-force sense (operator 1b) — an attacker and a defender are both a pair of seats and a pair of opposed forces — and that overlap is part of why the axis saturates.

How to apply it. Given a prime P: (1) ask whether P is relational at all — does it describe one party's role in an interaction between two parties? If not, it is a non-relational fixed point and there is no move. (2) Identify the relation and the seat P occupies (P is read from the sender's seat, the principal's seat, the observer's seat). (3) Swap to the complementary seat and describe the problem that seat faces as a positive structure — not "the other side of P," but the distinct operation the other party performs.

The "did it take?" test — the load-bearing craft — is the seat rule: a genuine inverted prime must name a complementary seat with its own problem-structure, not already in the catalog. State what the other party is actually doing, and check that it is different from what the first party does. The receiver's inference problem is genuinely unlike the sender's signaling problem — that took, and indeed both seats got named (signaling and screening are both in the catalog). If you swap seats and get the same structure back, that is a symmetric fixed point, logged as a fact about the relation, not a discovery. The seductive failure to guard against is mistaking a symmetric relation for an asymmetric one: cooperation, exchange, and reciprocity all feel like they have two parties, but both parties are doing the same thing, so there is no second seat to find.

Worked examples — of the collapse, because the axis is saturated. signaling, the canonical case: read from the sender's seat, it is "emit a costly action that credibly reveals hidden quality." Swap to the receiver's seat and you get "elicit or filter for hidden quality you cannot observe directly" — a genuinely distinct problem. But the catalog already names it: screening. So the move took (the seat is real and distinct) and yet returned nothing new — it landed on an existing prime. This is the operator's dominant experience: the canonical relational pairs — signaling/screening, principal/agent, moral hazard, adverse selection, compellence/deterrence (both folded into coercion) — are already both-seats-named, because relational primes are almost always introduced as pairs in the first place. reciprocity: swap the seat and the structure is identical — returning like for like is symmetric, each party does what the other does — so it is a fixed point, logged to the audit as evidence that this relation has no hidden second seat. symmetry: not a relation at all, no seat to occupy, a trivial collapse. Across the 53-application pilot the tally was 7 symmetric fixed points, 8 alias confirmations (both seats already named), 37 rejects, and a single survivor.

What sweeping the corpus taught us. Perspective inversion is a saturated end of the unary family, and it was retired on the pilot without a full sweep. The dominant collapse is the alias: the complementary seat is real but already named, because relational primes come in pairs and the catalog's relational core was built pair-complete. The one survivor, interpellation — the seat of the one addressed by a grand narrative, constituted as a subject by being hailed (Althusser's "hey, you there!") — is framed, bound to critical theory rather than domain-general, and it is provisional. As an axis-step operator that reaches the ends of its axis, perspective inversion belongs with threshold/phase (operator 8): both yield essentially nothing because both target regions the catalog already fills. The method lesson it sharpened is that saturation is often predictable before you spend a single token: a five-minute check of the catalog's relational pairs, done at build time, showed both seats already present for nearly every canonical relation, and the near-zero pilot yield simply confirmed the forecast. A cheap pre-pilot grep is a reliable saturation oracle — run it before committing to a sweep.

8. Threshold / phase — the linearity axis¶

Status: pilot only (106 applications, two sub-flavors), retired as saturated, capable-model tier. Yield: 2 gaps captured (catastrophe, order_parameter, provisional).

What it is. Threshold/phase is the move you make when you look at something that changes smoothly and ask what happens when the smooth change suddenly snaps. Most of the world varies by degrees — water gets hotter and hotter — but at particular critical points a gradual change triggers an abrupt, qualitatively different regime: at 100°C the water doesn't just get hotter, it becomes steam. That discontinuity, where a small further push across a critical value flips the whole system into a new state, is a structure in its own right, and it is genuinely different from the smooth ramp that led up to it. The everyday instances are everywhere: a rumor spreads person by person until it suddenly "goes viral"; a bridge bears more and more load until it collapses; a crowd's mood shifts gradually until, at some tipping point, a riot erupts. The move is: take a gradual, continuous prime and find its critical-point version.

There are two senses of "cross a threshold," run as two passes. Gradual→tipping (8a) is the snap: a smoothly-varying quantity crosses a critical value and the system jumps to a new qualitative regime, often irreversibly (you cannot un-boil an egg). The new anatomy the tipping version has, and the ramp lacks, is a critical value plus a qualitative regime change across it — and often hysteresis, where the system doesn't snap back when you retreat below the threshold. Continuity→discreteness (8b) is quantization: a smooth continuum becomes a set of distinct, separated levels, like a dimmer replaced by a light switch, or energy that comes only in discrete packets. This second flavor overlaps heavily with spatialization's discrete↔continuous move (operator 4b), which was already run at full corpus — a fact that turns out to matter enormously.

Why is this a genuine axis? Because the critical-point version has structure the smooth version simply does not: a threshold, a regime change, and frequently an order parameter — a quantity that is zero on one side of the transition and non-zero on the other, measuring the qualitative event. A pattern and its threshold version are different shapes, not the same shape described twice. So wherever a smooth prime exists, there is potentially a critical-point prime hiding — and the sweep hunts for the ones whose threshold form is load-bearing enough to name.

Where it comes from. The formal home of this operator is the physics and mathematics of sudden qualitative change: the theory of phase transitions and critical phenomena (Landau's order parameters, universality, critical exponents), catastrophe theory (Thom and Zeeman's fold and cusp geometries, which classify the ways a smooth control surface can produce a discontinuous jump), and percolation theory (the critical density at which local connections suddenly span the whole system). This is one of the richest and most-worked regions of applied mathematics, which is a strong prior that the corpus will already have it covered. As a discovery move, threshold/phase corresponds to the prime tipping_points_or_phase_transitions — the operator is, quite literally, that catalog entry turned into a search over the rest.

How to apply it. Given a prime P: (1) ask whether P is already critical — if P is tipping_points, percolation_threshold, symmetry_breaking, or nucleation, then criticalizing it returns itself, a fixed point. (2) If P is smooth, find the critical value across which it would flip to a new regime, and describe that regime change. (3) Apply the "did it take?" test — the critical rule: a genuine threshold prime must exhibit a critical structure the input lacks — a named critical value with a qualitative regime change across it, an order parameter, a bifurcation of the solution set, or quantized levels. State it concretely. The seductive failure is the "P-but-past-a-threshold" coinage: attaching a threshold word to a smooth prime without any new critical structure — and worse, the resulting structure is almost always already named by tipping_points or threshold, so it collapses as an alias rather than a discovery.

Worked examples — of the collapse, because the axis is saturated. tipping_points_or_phase_ transitions itself: apply the operator and there is nothing to do — it is already a phase transition, so criticalizing it returns it unchanged. A clean fixed point, one of nineteen the pilot logged, each a confirmation that the catalog's critical-point layer is already dense. a gradual accumulation of pressure: the natural threshold version — "the point at which the accumulation snaps" — is exactly what tipping_points already names, so it collapses as an alias. This is the operator's characteristic experience: the region is the most saturated in the entire catalog. Tipping points, threshold, percolation threshold, critical mass, critical juncture, nucleation, self-organized criticality, hysteresis, metastability, symmetry breaking, regime change, punctuated equilibrium, cascade, and supersaturation are all already present. any continuous quantity under 8b: its discretized version lands on discreteness or discrete_vs_continuous_quantization — primes that spatialization's operator 4b had already surfaced — so flavor 8b was pure redundancy, returning 0 new over 53 applications. The two survivors that were captured are refinements, not new territory: catastrophe (the control-surface geometry that refines tipping points) and order_parameter (the observable that measures a transition, complementing symmetry breaking).

What sweeping the corpus taught us. Threshold/phase is the clearest saturated end of the unary family, and it was retired on the pilot — 8a returned two candidates that both dissolved on inspection, 8b returned nothing but op-4 duplicates — saving an estimated 4M tokens a full sweep would have cost. Its dominant collapses are the fixed point (the prime is already critical) and the alias (the threshold version is already named), and both are completeness results: the catalog's critical-transition layer is essentially finished. Like perspective inversion, this operator sits at the end of its axis, where the corpus is already saturated, which is why an axis-step operator that usually yields early here yields nothing. The captured gaps (catastrophe, order_parameter) are provisional refinements of existing primes; a possible bifurcation split and a domain-specific critical_exponent were noted for later rather than stubbed. The method lesson: a near-zero pilot is not a failed operator — it is a completeness measurement. Retiring an operator the moment its saturation is demonstrated, rather than grinding through a full sweep to confirm the obvious, is the program's standing rule, and threshold/phase is its textbook case.

Part II — N-ary operators¶

Combine two or more primes into a candidate. These do not brute-force the pair space — 1,325² is 1.76 million pairs — but run on a deterministically pre-filtered candidate set. Three of the four saturate; one is a discovery hero. Together they are the sharpest test of the yield law.

9. Composition — conjunction into an emergent unit¶

Status: pilot (88 pairs), retired as saturated. Ran last of all fourteen operators. Yield: 0 new primes; 23 hierarchy-edge flags.

What it is. Composition is the move you make when two patterns appear together so reliably, and interlock so tightly, that they have stopped being two things and become one. The paradigm case is in the catalog already: variation (offspring differ from parents) and selection (the fitter ones survive) and retention (the survivors pass on their traits) each mean something on their own — but bolt them together and you get natural selection, a single engine with a behavior none of the three parts has alone: it accumulates adaptation over time. That accumulation is emergent — it belongs to the assembled unit, not to any piece. Composition takes two tightly co-occurring primes and asks whether their combination is a genuine emergent unit like this, deserving its own name.

The everyday intuition is the difference between a heap and a machine. Flour and water and yeast, on a counter, are a conjunction — three ingredients that happen to be nearby. Combine them under heat and they become bread, which rises and browns and does things no ingredient does alone. The whole craft of the operator is telling those two situations apart: a mere conjunction, where two primes just happen to be used together (both are generically useful, so of course they co-occur), versus a true composite, where the parts fuse and the combination exhibits behavior that is not in any part. The pre-filter did the mechanical work of finding co-occurring candidates — it selected the 4,166 prime-pairs that share three or more archetypes at complementary embedding similarity (close enough to interact, distinct enough not to be mere aliases of each other) — and a diverse 88-pair sample, capped at four pairs per prime so no hub prime could dominate, tested the space.

Why is this a genuine move and not just "list two primes"? Because emergence is a real, checkable property: a composite has a name and a mechanism that neither part supplies. Natural selection ratchets; a feedback loop stabilizes or explodes; a market discovers prices. Where two primes genuinely fuse, the fusion is a structure in its own right — and the sweep hunts for fusions the catalog has not yet named.

Where it comes from. The intellectual home of composition is Fauconnier and Turner's conceptual blending — the cognitive-science account of how two mental "input spaces" project into a third, blended space that has emergent structure present in neither input (their stock example: a debate in which you argue "against Kant," blending a modern speaker and a dead philosopher into a single impossible conversation). More formally it draws on the mathematical notion of composing two operations into one, and on systems theory's core claim that the whole can be more than the sum. As a discovery move it corresponds exactly to the prime composition — it is one of the catalog's own entries applied to the rest, which is fitting, since composition was the last of the fourteen operators run.

How to apply it. Given a co-occurring pair P1, P2: (1) ask what property the combination has that neither part has alone — name the emergent behavior, or find there is none. (2) If there is an emergent property, check whether the composite is already named (variation∘selection is natural_selection; the move found the unit but the unit exists). (3) Apply the "did it take?" test — the emergence rule: can you name a mechanism or behavior that belongs to the assembled unit and to neither piece? If yes, and it is unnamed, you have a candidate. Two seductive failures guard the gate. The first is the conjunction trap: the pair co-occurs, but only because both primes are broadly useful (feedback and observability show up together in a thousand systems without fusing) — no emergent unit, just proximity. The second is subsumption: one prime turns out to be an aspect of the other, which is not a composition at all but a hierarchy edge — the parts don't combine side-by-side, one contains the other.

Worked examples — of the collapse, because there are no finds. variation + selection: they co-occur in every evolutionary account, and their composite is unquestionably a genuine emergent unit — it ratchets adaptation over time, which neither variation nor selection does alone. The move took perfectly. And it returned nothing new, because that composite is already named natural_selection. This is the operator's dominant experience, and it is instructive: the pairs whose composite would be worth naming have already been named, precisely because a fusion worth naming got noticed and named during the years the catalog was built. feedback + observability: these co-occur constantly, but there is no emergent unit — you can observe a system and have it feed back on itself, but "observed feedback" is not a new engine with its own behavior; it is a conjunction, two useful primes in the same neighbourhood. That was the dominant kill. A third pattern: some pairs collapsed as subsumption, where inspection showed one prime was an aspect of the other — not a composition to name but a hierarchy edge to draw, and 23 such pairs were flagged and fed to the edge harvest. The final tally over the 88-pair pilot: 0 emergent composites, 88 collapses.

What sweeping the corpus taught us. Composition is a recombination operator — it builds a candidate by recombining two primes already in the catalog — and the yield law says exactly why it saturates: anything worth naming that could be assembled from parts already present was already assembled, reactively, during construction. When a builder noticed that variation and selection fuse into an accumulating engine, they named natural selection on the spot; the reactive naming already harvested every emergent composite worth having. So composition belongs with tension-resolution, interpolation, and cluster-without-parent — all recombination operators, all saturated at zero, because a mature catalog already implies their outputs. Its one non-trivial byproduct is the 23 subsumption flags, which are not compositions but genuine missing hierarchy edges, adding to the program's ~295-edge densification harvest. The keystone: composition ran last, and its zero was the final datapoint that let the yield law be stated in full and symmetric. By the time it was run, the other three recombination operators had all saturated and the two new-ground operators (analogy, lattice) had both yielded — so composition's zero completed the pattern: an operator yields only when it reaches a structural position not already implied by the corpus. The corpus is complete with respect to recombination, and composition's zero is what proved it.

10. Tension-resolution — the dialectical synthesis¶

Status: pilot (53 applications), retired as saturated. Yield: 0 new primes.

What it is. Tension-resolution is the move you make when you find two forces genuinely pulling in opposite directions and ask what third thing resolves the pull. It is the ancient shape of dialectic: a thesis, an antithesis that opposes it, and a synthesis that transcends both. The everyday version is the experience of a real dilemma and its escape. You want to explore new options and exploit the good option you already have; you cannot fully do both, and the tension is real — but sometimes there is a specific structure that dissolves it (a phased strategy, a portfolio, a reframing that shows the opposition was less sharp than it seemed). The operator takes an opposed pair and proposes either the resolution that synthesizes them into a new structure, or the dilemma itself named as a load-bearing pattern.

The critical thing to understand — the thing that makes or breaks the operator — is that "resolving a tension" almost always means one of two cheap things that are not primes. The first cheap thing is generic balance: "trade them off," "find the sweet spot," "moderate between the extremes." Every tension in the universe has a balance point; naming the balance point of a particular tension adds nothing, because the pattern "balance the two poles" is the same regardless of which two poles. A genuine resolution must instead be a specific, structurally distinct move — a reframing that makes the opposition dissolve, a higher-order structure that contains both poles, a named third way with its own anatomy. The second cheap thing is that the "resolution" is really just one pole's opposite — which is a dualization, already swept by operator 1. So the whole discipline of the operator is refusing both cheap escapes and demanding a genuine third structure.

Why is this a genuine move at all, given how easily it degenerates? Because real syntheses do exist and do have their own anatomy — a genuine dilemma-prime or resolution-prime names a structure you can describe, reason about, and recognize across domains. The sweep hunts for those; the difficulty is that they are rare and the degenerate look-alikes are everywhere.

Where it comes from. The intellectual home is the oldest of the combinatorial moves: Hegelian dialectic — thesis, antithesis, synthesis, the engine by which a contradiction drives the emergence of a higher concept that preserves and transcends both sides. As a discovery move, tension-resolution corresponds to the primes dialectic and paradox, both of which the catalog holds — so this operator, like the others, is one of the catalog's own entries turned on the rest. It also overlaps heavily with dualization's opposing-force sense (operator 1b): the opposed pair that a tension sits on is itself a dualization, which means much of what tension-resolution could reach was already reachable by a cheaper unary route.

How to apply it. Given a prime C that sits in a tension: (1) name the A↔B opposition C embodies (exploration vs exploitation, centralization vs decentralization). (2) Propose either the resolution that transcends the opposition or the dilemma itself as a named structure. (3) Apply the "did it take?" test — the synthesis rule: is the proposed resolution a specific structure, or is it merely "balance the two poles"? If you cannot state a mechanism beyond moderation, it is not a prime, because every tension admits a balance and naming each one is empty. Two seductive failures guard the gate: the generic-balance trap (the only "resolution" is a point on a spectrum between A and B), and the dualization re-run (the "resolution" is just B, C's opposite, already swept by operator 1). A candidate survives only if it is neither.

Worked examples — of the collapse, because there are no finds. exploration opposed to exploitation: a textbook tension, and it invites a synthesis. But the catalog already names both the dilemma — ambidexterity_exploit_vs_explore, the organizational capacity to do both — and the general shape, dialectics and paradox. The only "resolution" the operator could propose beyond what exists was generic balance: "trade exploration off against exploitation." That is not a prime, because every tension has a balance point, so it collapses. This is the operator's dominant experience: the catalog is already rich in named tensions and their resolutions — risk_return_tradeoff, collingridge_dilemma, social_dilemma, jevons_paradox, braess_paradox, calibrated_rule_vs_moving_world, and dozens more — so a genuine unnamed synthesis is vanishingly rare. Another pattern: for many inputs the proposed "resolution" was simply the opposite pole, a dualization re-run that operator 1 had already swept. Across the 53-application pilot the tally was 0 genuine resolutions, 53 collapses — the collapse reasons falling into exactly the four predicted buckets (generic balance, dualization re-run, already-resolved-by-an-existing-prime, and domain-specific).

What sweeping the corpus taught us. Tension-resolution is a recombination operator — it builds a candidate by combining two opposed primes already present — and it saturates for the same reason composition does: a mature catalog already names the syntheses worth naming, because they were resolved reactively during construction, and it already names the tensions worth naming as dilemmas. It belongs with composition, interpolation, and cluster-without-parent, all zero-yield recombination operators. Its saturation is doubly overdetermined: it overlaps dualization 1b (the opposed pair) and the existing prime dialectics, so most of what it could reach was reachable by a cheaper route already run. The method lesson it sharpened is a discriminator the whole funnel now uses: "balance between two poles" is not a prime. Every opposition admits a trade-off, so a resolution earns primehood only when it is a distinct structure rather than a point on a spectrum — and that test now guards against a whole class of plausible-but-empty synthesis coinages that would otherwise flood the survivor stream.

11. Analogy completion — the cross-domain diagonal¶

Status: full sweep, 1,325 applications, capable-model tier. The one n-ary discovery hero. Yield: 7 genuine finds (provisional) — the only n-ary operator with real yield.

What it is. Analogy completion is the move behind every "A is to B as C is to _____" question you ever saw on an aptitude test. Hand is to glove as foot is to sock. You are given a relation that holds between two things, a third thing that plays the same role as the first, and you are asked to supply the fourth thing that stands to the third exactly as the second stood to the first. The everyday version is a word game; the version that matters here is the same move performed on abstractions. Given that two primes stand in some structural relation, and that a third prime occupies the analogous starting position, you ask: what is the fourth term that completes the pattern? — and, crucially, whether that fourth term is a structure the catalog has never named.

What makes this a genuinely distinct move — and not just one of the unary operators wearing a disguise — is that the relation A→B must be a real two-place mapping, a structure-preserving correspondence between two whole configurations, not a single step along one axis. That distinction is the entire craft of the operator. If the relation A→B is "B is the opposite of A," then completing the analogy for C just hands you C's opposite — which is dualization, already swept. If A→B is "B is the more general version of A," you get C's genus — generalization, already swept. Those completions collapse by design, and correctly so: they confirm the unary sweeps already covered that ground. The move earns its keep only when the relation A→B is a composite, cross-domain mapping — the kind mathematicians call a functor, a correspondence that carries a whole web of relationships intact from one domain to another. A fourth term reached that way lives on the diagonal: it is a position no single axis-step can arrive at, because getting there requires holding two relations in register at once. That is why analogy is the operator that reaches cross-domain composite primes — structures like ergodicity (the precise bridge between a time-average and an ensemble-average) that are literally definable only as the term that completes a two-relation mapping. Wherever the catalog holds three corners of such a square and not the fourth, there is a candidate hiding on the diagonal.

Where it comes from. Analogy is the third of Pólya's three master heuristics for mathematical discovery — after generalization and specialization — and he treated it as a first-class method for guessing a theorem's shape before proving it: find a simpler problem with the same structure, solve that, and carry the solution across. Its rigorous modern form is the functor of category theory, a structure-preserving map between two categories that sends objects to objects and relations to relations while keeping the whole diagram of arrows intact — which is exactly what a genuine four-term analogy is when it holds. This connects the operator directly to the formal structure-mapping account of analogical reasoning (Gentner), on which an analogy is good precisely to the degree that it preserves a system of relations rather than surface features. And analogy is itself a prime: the catalog holds analogy and its more exact relative isomorphism, so this operator is one of the catalog's own entries turned loose on the rest.

How to apply it. Given a prime C, the procedure is: (1) find a well-known structural relation A→B between two other concepts — a relation with real internal machinery, not a one-word contrast; (2) check that C stands in the A-position of an analogous relation, so that the same machinery could run again with C as its input; (3) name the fourth term D that completes it, and describe D as a positive structure with its own anatomy; (4) run the operator's decisive filter.

The "did it take?" test for analogy is the operator-re-run filter, and it is the sharpest gate in the whole n-ary family: could D have been reached by a single unary move on C? If D is C's opposite, C's genus, C-applied-to-itself, C-over-time, C-as-a-field, C-made-modal, or C-at-a-threshold, then the "analogy" is a unary operator in disguise and it collapses — that ground is already swept. D survives only if it requires the composite, cross-relational mapping to reach — if you genuinely cannot get there in one axis-step from C. The seductive failure to guard against is the dressed-up axis-step: a candidate that looks like a rich four-term analogy but, once you write out the A→B relation plainly, turns out to be "B is A's dual" or "B is A generalized" — a single operator move wearing an analogy's clothing. Write the relation out in one flat phrase before trusting it; if the phrase names one of the eight unary moves, it did not take.

Worked examples. ergodicity, from the mapping time_average : ensemble_average :: stationarity : ? — the relation is "the condition under which the first average equals the second." A time-average follows one system along its whole history; an ensemble-average polls many copies at one instant; these agree only under a special condition, and stationarity is a necessary-not-sufficient piece of it. Completed, the fourth term is ergodicity: the exact condition under which a time-average equals an ensemble-average — which is the precise statement of when expected-value reasoning is even valid, a deep complement to expected value itself. It is canonical (Boltzmann, Birkhoff; Peters's ergodicity economics), was verified absent, and is reachable by no single axis-step — it is a genuine diagonal composite of stationarity plus the averaging bridge. allostasis, from thermostat : single-variable regulation :: homeostasis : ? — a thermostat holds one variable at a fixed setpoint; homeostasis is that idea lifted to a body's many variables; complete the analogy toward the anticipatory, setpoint-shifting form of regulation and you reach allostasis (stability through change — shifting setpoints ahead of predicted demand, with allostatic_load as its cost). It is homeostasis's sophisticated sibling, a composite of homeostasis plus anticipation plus variable setpoints, not merely homeostasis-over-time. construal_level, from spatial_distance : perceived_size :: time_preference_discounting_future : ? — just as spatial distance shrinks apparent size, the catalog already names how temporal distance discounts future value; complete the analogy across all distance dimensions at once and you reach construal_level, the genus of which the temporal discounting the catalog already held is only one species: value and abstraction shift with distance along any dimension — time, space, social, hypotheticality (Trope-Liberman).

What sweeping the corpus taught us. Analogy completion collapses in two ways, both instructive. It collapses when the completing term already exists (a plain alias), and — the dominant kill — when the completing term is a single-operator re-run, a dual or genus or field-version already reached by the unary sweeps; 99% of applications (1,305 of 1,325) went one of these two ways. That near-total collapse is not failure but confirmation: it shows the catalog was built largely by cross-domain mining, so most deep analogies are already realized. The thin residue is the whole point. Seven genuine finds survived — ergodicity and allostasis (both strong), plus boundary_object, construal_ level, coherentism, generativity — and they clustered in the framed, social-science families, the same region generalization's staircase marked as the yield engine. In the yield law this is the operator that pays where the other three n-ary operators do not, and it pays for a precise reason: it does not recombine two existing primes into a unit or fill a gap between them, it reaches a new structural position defined by a two-relation mapping — exactly the kind of position a reactive, one-domain-at-a-time construction cannot stumble into. The method lesson: analogy is the empirical core of the yield law, because it proved the tempting slogan "n-ary operators saturate" false as a blanket claim. The recombination n-ary operators saturate; the diagonal one does not. The whole law turns on that distinction between recombining structure already present and reaching a position not yet implied — and analogy is the operator that forced the distinction to be stated exactly.

12. Interpolation — the missing intermediate¶

Status: full pre-filtered run (61 candidate edges), retired as saturated. Yield: 0 new primes; 13 missing-edge flags + 13 mis-parent flags.

What it is. Interpolation is the move you make whenever you look at two things at different levels of a hierarchy and feel that something belongs between them. It is the everyday intuition that there ought to be a rung on the ladder you can't see: an animal is a living thing, but "mammal" sits usefully between them; a number is a mathematical object, but "integer" belongs in the gap. Given a parent concept and a child concept — where the child is a genuine special case of the parent, but the conceptual distance between them is wide — you ask: is there an intermediate concept that is more specific than the parent and more general than the child, a real level that the ladder skipped? On abstractions the move is the same: take a subsumption edge parent→child with a large gap, and propose the missing intermediate prime that should sit on it, so that parent→intermediate→child becomes a smooth three-step chain instead of one long leap.

The reason this is a real, distinct move and not just generalization run twice is that the intermediate is pinned from both sides at once. Generalization strips a constraint off one prime and looks upward for its genus, unbounded. Interpolation is bounded above and below: the intermediate must be a species of this parent and a genus of this child simultaneously, threading a specific gap. That two-sided constraint is what makes it an n-ary, catalog-internal move — it operates on an edge of the hierarchy, a relation between two existing primes, not on a prime in isolation. A candidate either fits snugly into the named slot between them or it does not; there is no room to drift upward into a vacuous umbrella, because the child fences it from below.

Where it comes from. The intuition traces to Gärdenfors's conceptual spaces, on which concepts are convex regions of a geometric space and between any two of them lies a betweenness region that may itself be a nameable concept — interpolation is the literal reading of "what lies between." But interpolation holds a singular and slightly embarrassing place in this program, and honesty requires naming it: of all fourteen operators, interpolation was the one operator with no home prime. Each of the others turned out to be a prime the catalog already held — dualization is duality, self-application is reflexivity_self_reference, analogy is analogy — but when we checked the correspondence deliberately, the slot for interpolation was empty. Following that single missing thread revealed that the catalog itself had a gap: interpolation was a genuine prime, absent, its safe pole hidden behind a failure-mode framing of its dual (extrapolation). The discovery method, turned on itself, found the prime — interpolation was accepted into the catalog on 2026-07-04. The operator that fills gaps between concepts was, fittingly, discovered by finding the gap it left in the operators.

How to apply it. This operator does not sweep the whole corpus; it runs on a deterministically pre-filtered worklist. The procedure is: (1) take the subsumption edges whose parent and child are most dissimilar — the entire large-gap set, 61 edges at embedding cosine below 0.65, not a sample; (2) for each, ask whether a genuine intermediate belongs in the gap, and name it as a positive structure with its own anatomy; (3) run the "did it take?" test.

The "did it take?" test for interpolation is strict two-sided placement: can you name a level that is strictly more general than the child and strictly more specific than the parent, with anatomy of its own? If the best you can produce is a paraphrase of the child dressed up a rung higher, the gap was apparent, not real, and it collapses. And there is a second collapse the test must route around: sometimes the intermediate you name is genuinely correct but already exists in the catalog — in which case the finding is not a new prime at all but a missing edge, the fact that parent→intermediate→child is a chain the hierarchy should draw but hasn't. The seductive failure to guard against is the child-restatement: the wide gap tempts you to invent a middle term that, on inspection, just re-describes the child at the same level of generality. State the child's mechanism and the candidate's mechanism side by side; if they are the same mechanism, nothing new took.

Worked examples — of the collapse, because there are no finds. monitoring → (large gap) → formative_assessment: the intermediate that belongs between watching a process and using observations to reshape teaching is feedback — but feedback already exists in the catalog. So the interpolation produces no new prime; it produces a missing-edge flag (monitoring → feedback → formative_assessment, a chain the hierarchy should draw). Thirteen runs came out exactly this way. adaptation → tolerance: the intermediate is habituation — the gradual dampening of response to a repeated stimulus — but that too is already present (habituation_to_repeated_signal), so again a missing edge, not a new prime. And adaptation → contextual_mode_switching: the proposed intermediate ("detect context, then select a mode") turns out on inspection to simply restate the child — the gap was apparent, the middle term hollow, the verdict collapse. Across the entire large-gap set, all four candidates the crank flagged as "new" resolved into one of these two harmless modes.

What sweeping the corpus taught us. Interpolation collapses in two ways: the intermediate is a child-restatement (the gap was illusory), or the intermediate is real but already named (a missing edge, not a prime). Zero genuine new intermediates emerged from the entire large-gap set — and that zero is itself a finding, because the catalog's hierarchy turned out to be tight: of its roughly 717 subsumption edges, none has a parent-child gap below cosine 0.5 and only twelve fall below 0.6, so there is almost no room for an intermediate to sit. In the yield law, interpolation is a recombination operator — it fills between structure already present — and it saturated for the same reason its siblings did: that infilling was done reactively, by hand, across the years the hierarchy was built, whenever a curator noticed a family needed a middle term. Its real output is densification: 13 missing-edge flags plus 13 mis-parent flags (edges whose wide gap signals a questionable parent rather than a missing child), both feeding the edge-review harvest. The method lesson: interpolation confirmed, from the hierarchy side, exactly what cluster-without-parent confirmed from the clustering side — the recombination operators were already exhausted reactively during the catalog's own construction, and what they offer a mature catalog is edge densification, not new primes. Two independent operators reaching the same verdict is the strongest evidence the program has that its recombination ground is genuinely covered.

Part III — Catalog-internal operators¶

Operate on the graph of primes rather than on individual primes. One recombination operator that saturates, and one empty-cell operator that yields.

13. Cluster-without-parent — the missing ancestor¶

Status: full run (82 coherent clusters), retired as saturated. Yield: 0 new primes; 12 hierarchy-edge flags.

What it is. Cluster-without-parent is the move a librarian makes on noticing that a shelf's worth of books keep getting pulled together yet have no shared section heading — the books clearly belong to one family, but the family has no name. Given the whole catalog laid out by similarity, you look for tight little groups of primes that plainly hang together — that share mechanism, that keep co-occurring — and then you check the hierarchy to see whether they have a common ancestor that unites them. When a coherent cluster has no shared parent drawn above it, that absence is the signal: either the family needs a genus that nobody has named, or it needs edges up to a genus that already exists. Naming the missing ancestor is the move.

This is a catalog-internal operator in the fullest sense: it does not transform any single prime and it has no per-prime crank. It consumes the entire corpus — every prime's similarity signature and the whole subsumption hierarchy — in one pass, and its unit of analysis is the cluster, not the prime. That makes it the mirror image of interpolation: interpolation works downward-and-between on a known edge; cluster-without-parent works upward from an unlabeled family, hunting for the genus that should crown it. It is a distinct move because the thing it discovers — a shared parent — is defined purely by the set of children beneath it, not by any transformation of a single input.

Where it comes from. The formal ancestor is Formal Concept Analysis (Wille, 1982), which takes a table of objects and their attributes and mechanically computes the complete concept lattice — surfacing every implied concept, including ones no human bothered to name, as a formal consequence of which objects share which attributes. Cluster-without-parent is that idea with a fifty-year-old algorithm behind it: find the groupings the data already imply, and read off the concepts that should sit above them. Among the catalog's own primes, the operator corresponds most nearly to classification (and its relative taxonomy) — the move is the act of classification turned back on the catalog to find the classes it has left implicit.

How to apply it. The procedure is a pipeline, not a prompt: (1) embed every prime into a shared similarity space and cluster the whole corpus (here, k-means over fresh signature embeddings); (2) keep only coherent clusters — those whose members are meaningfully tighter than a random grouping would be — which filters the raw clusters down to the ones worth trusting (82 survived); (3) for each coherent cluster, check the subsumption hierarchy for a common ancestor; (4) where none is drawn, name the parent that should sit above the family. The "did it take?" test is a genuine-genus check on that proposed parent: is the named ancestor a real abstraction that all the cluster's members are species of, and is it absent from the catalog? The characteristic — and, it turned out, universal — failure to guard against is the parent-already-exists case: the family's natural genus is real and correctly named, but it is already in the catalog, so the finding is not a new prime but a set of missing edges linking the children up to it.

Worked example — of the collapse. A cluster pulls together fail_safe, fault_tolerance, error-proofing (poka-yoke), and validation — plainly one family, all concerned with making a system withstand faults, and no common ancestor drawn above them in the hierarchy. The operator names the parent that should unite them: robustness. But robustness is already in the catalog. So the cluster reveals not a missing prime but a missing-edge worklist — each of those four should link up to robustness, edges the hierarchy simply hadn't drawn. Every one of the operator's coherent clusters resolved this way: a family whose natural genus (pedagogy, historicism, rule_of_law, fairness, natural_selection, and others) already existed, needing edges rather than a new name.

What sweeping the corpus taught us. Cluster-without-parent has essentially one collapse mode, and it accounts for every case: the natural ancestor already exists, so the finding is a set of missing edges rather than a new prime. Zero new primes emerged — the twelve coherent clusters produced twelve parent-hub edge sets and nothing else. In the yield law this is the archetypal recombination operator: naming the shared genus of things already present is exactly the move the catalog's builders made by hand whenever they noticed a family needed a parent, so the operator had, in effect, already been run reactively for years. The method lesson: cluster-without-parent produced the program's first hard saturation datapoint, and it was informative precisely because it returned zero. A zero here is not a dead end — it confirmed the up-front prediction that recombination operators would find nothing new on a mature catalog, and it turned that confirmation into the start of the roughly-295-edge densification harvest. A zero that is predicted, and that throws off a useful byproduct, is a successful operator, not a failed one.

14. Lattice / symmetry completion — the empty cell¶

Status: single frontier reasoning pass (~15 candidate gaps), run to completion. The second n-ary discovery hero. Yield: 5 genuine finds (provisional).

What it is. Lattice completion is the move Mendeleev made when he stared at the periodic table, saw a hole where a known regularity said an element should be, and predicted its properties before anyone had found it. The power of the move is that a pattern in what you already have can tell you about something you don't have yet. Two everyday shapes it takes. The first is the product space: whenever concepts vary along two independent either/or dimensions, they lay out on a grid — {costly, costless} crossed with {communication types}, say — and a grid with three corners filled and one empty is an announcement that the fourth corner should exist. The second is the small typology family: a set of options that is meant to be complete — the classical responses to a problem, the members of a recognized triad — where the catalog holds the problem and all-but-one of the responses. In both shapes you find the empty cell and name the prime that belongs in it.

What makes this a real and distinct move is that the empty cell is specified entirely by the filled ones. You are not transforming a prime, combining a pair, or lifting a family to its genus; you are reading a symmetry in a near-complete structure and inferring the exact member the symmetry demands. The gold-standard case is the k-1-of-k gap: a genuine, closed, mutually-exclusive family with all but one member present, so the missing member is not a free invention but the precise complement the existing members leave a hole for — its role, and often its name, is fixed before you go looking. That is why lattice completion, like analogy, reaches new ground: an empty cell is by definition a structural position the catalog has not yet occupied, and a symmetry argument can point straight at it without any single prime ever leading there.

Where it comes from. The canonical instance is Mendeleev's predicted elements — eka-aluminium, eka-silicon — inferred purely from a hole in a periodic regularity and later found to exist with nearly the predicted properties: the paradigm case of deriving a missing entity from a symmetry among the known ones. Its formal cousin is again Formal Concept Analysis, whose lattice completion computes every concept the object-attribute structure implies, empty cells included. Among the catalog's own entries the operator corresponds to symmetry (with its relatives duality and complementarity) — the move is symmetry-reasoning applied to the catalog to find the members its symmetries require.

How to apply it. This is not a per-pair sweep but a single, disciplined frontier-model pass: (1) seed a strong model with the catalog's explicit oppositional dimensions — its 23 named "X_vs_Y" primes — plus its recognized typology families (the ethics triad, the inference triad, the epistemic-regress responses, the alethic-modality square, sender/receiver, governance modes); (2) have it build genuine product spaces and families, mark which cell each existing prime fills, and identify the empty cells; (3) for each empty cell, name the missing prime; (4) deterministically existence-check every candidate against the corpus before trusting it. The "did it take?" test is the k-1-of-k discriminator: is this a genuine closed family with all-but-one member present, where the empty cell is the exact complement of the present members and carries a canonical name? A cell passes only if the family is real and the gap is real. The seductive failure to guard against is the fake grid: a 2×2 whose axes are not actually independent (feedback crossed with sign), or a "family" that is not jointly exhaustive, or a cell that looks empty but is filled by a prime under another name. These manufacture empty cells that mean nothing — the discipline is to reject any cell whose family is not a true, orthogonal, closed structure, and to existence-check the rest.

Worked examples. infinitism, from the epistemic-regress family. The problem of the infinite regress of justification has three classical responses: foundationalism (stop the regress at a basic belief that needs no further support), coherentism (let beliefs support each other in a mutually reinforcing web), and infinitism (accept an infinite, non-repeating chain of reasons). The catalog held the problem (infinite_regress) and two of the three responses — so infinitism (Klein's "continue" response) was a textbook k-1-of-k empty cell, and naming it completes the triad. It was independently surfaced by analogy completion as well, which is the strongest signal the program can produce. cheap_talk, from the communication-games family. The catalog held costly signaling — communication made credible by being expensive to fake — but not its opposite pole, cheap_talk: costless communication made credible by interest-alignment rather than by cost (Crawford-Sobel). Two poles of a {costly, costless} axis, one named, its complement a clean empty cell with its own distinct credibility mechanism. common_pool_resource, from the governance-modes family. The catalog held the two failure modes of shared resources — tragedy_of_the_commons (overuse) and anti-commons_tragedy (underuse) — but not the working institution: common_pool_resource, Ostrom's self-governed commons, the positive third mode alongside state and market provision. The empty cell was the institution that succeeds, hiding behind the two named ways of failing.

What sweeping the corpus taught us. The characteristic collapse here is the frontier model's own self-rejection: most candidate gaps it considered turned out to be cells that looked empty but were already filled (analogical_inference — but analogy is present; punctuated_equilibrium — present), or "families" that were not real product spaces at all (the vacuous "true-outcome" corners of a confusion matrix, non-orthogonal 2×2s). The discipline held — the operator proposed only where a genuine symmetric family had a genuine hole — and five genuine finds survived: infinitism and cheap_talk (strong), plus common_pool_resource, positional_good, and type_iii_error. In the yield law, lattice completion is the second n-ary operator to pay, and it pays for the same reason analogy does: an empty lattice cell is a structural position not yet reached, so completing it reaches new ground rather than recombining old. And like analogy, its finds cluster in the framed, social-science families — governance, welfare, epistemology, game theory — confirming that the catalog's structural core is essentially lattice-complete while its framed families still hold a few predictable empty cells, the same yield region generalization's staircase identified. The method lesson: lattice completion validated a cross-operator convergence signal. Several of its empty cells had already been reached by another operator — infinitism was independently proposed by analogy completion; heterarchy, one of its candidate cells, was already a dualization hold reached three separate ways. When an empty-cell prediction and an independent operator's find land on the same prime, the candidate is as strong as the program can make it: the structural argument (the symmetry demands this cell) and the generative argument (an operator independently produced it) agreeing on the same answer.

Summary matrix¶

#	Operator	Family	Move	Status	Provisional yield	Regime
1	Dualization	unary	flip polarity	full sweep	6 acc-rec + ~25 holds	axis-step (yields)
2	Self-application	unary	apply P to its own machinery	full sweep	15 holds; 41% FP	axis-step (yields + audit)
3	Temporalization	unary	add time	full sweep	5 holds; 59% FP	axis-step (audit-heavy)
4	Spatialization	unary	add extent/scale	full sweep	6 holds; 37% FP	axis-step (audit-heavy)
5	Generalization	unary	strip to genus	full sweep	~47 holds	axis-step (highest yield)
6	Modalization	unary	shift modal status	full sweep	4 holds; 22% FP	axis-step (alias-heavy)
7	Perspective inversion	unary	swap the seat	pilot → retired	1 find	axis-step (saturated)
8	Threshold/phase	unary	continuous→critical	pilot → retired	2 gaps	axis-step (saturated)
9	Composition	n-ary	fuse a co-occurring pair	pilot → retired	0	recombination (saturates)
10	Tension-resolution	n-ary	synthesize an opposed pair	pilot → retired	0	recombination (saturates)
11	Analogy completion	n-ary	A:B :: C:? diagonal	full sweep	7 finds	new ground (yields)
12	Interpolation	n-ary	fill a hierarchy gap	full → retired	0; 13 edges	recombination (saturates)
13	Cluster-without-parent	catalog	name a missing ancestor	full → retired	0; 12 edges	recombination (saturates)
14	Lattice completion	catalog	fill an empty symmetry cell	frontier pass	5 finds	new ground (yields)

The pattern the matrix makes visible. The three regimes of the yield law read straight down the last column. Axis-steps (1–8) yield a few high-convergence finds each and then turn into audit instruments, saturating at the ends of their axes (7, 8). Recombination operators (9, 10, 12, 13) saturate at zero, because a mature catalog already implies their outputs. Only the two operators that reach a genuinely new structural position — the cross-domain diagonal (11) and the empty symmetry cell (14) — yield on a mature corpus, and both find their primes in the same framed, social-science region. That is the whole program in one table: the corpus is complete with respect to axis-steps and recombination, and incomplete only along the diagonals and in the empty cells of its symmetry families.

Companion to Operator-Driven Discovery of Prime Abstractions. Provisional counts pending final curation.