The Mortal Scientist

Part Part I built the apparatus: a category of theories, a monad that charges for a step, a logic generated from the presentation. What it did not build is anybody to use it. The parts that follow ask what a physics is and why a scientist inside one would be misled; both take the scientist for granted. This chapter builds one.

The constraint that does the work is stated in a single sentence: a computation can only find out about another computation by interacting with it, and interacting costs. Everything else follows from taking that seriously and refusing to add anything not already forced. There is no dataset, because a dataset is somebody else’s observations supplied for free. There is no scorekeeper, because a scorekeeper would be an authority standing outside the world telling the learner how it is doing, and no such authority is available across the cut. What is left is a term in parallel with other terms, holding a finite supply of tokens, spending them to look, and dying when it runs out.

Three things then have to be supplied and only one of them turns out to be a choice. The environment is fixed by ontological isolation. The hypothesis language is not chosen but generated — fed the presentation of the calculus, the OSLF construction emits a logic whose formulae denote exactly the equivalence classes the learner could distinguish by experiment, which is the Pledge’s “anything Bob can do Alice can do” turned into syntax. The motive is the one genuinely free decision, and we make it metabolic: the learner that predicts badly cannot afford to look, and the learner that cannot look starves.

What comes out is stranger than what went in. Destructive measurement forces induction. Foraging bounds concealment, so that perfect privacy is available only to something already dying. Reproduction turns out to be reflection — a genome is a quotation of code, inert because quotation does not reduce — and sex is a defense against a predator’s hypothesis language rather than a mechanism of variation. And the ecology divides, without being told to, into the two kinds of organism Eric Smith finds in the biosphere: those that eat only external energy, and those that eat each other.

21.1 Introduction

21.1.1 The scientist as a learner

Most computational accounts of learning begin with a dataset: a supply of observations that exists independently of the learner, whose acquisition is free, and whose acquisition does not disturb the source. Given such a supply the problem is one of induction — find the hypothesis that best accounts for what is already in hand — and the interesting questions are statistical.

A scientist has none of this. There is no dataset waiting; there is a world, and the scientist must decide what to do to it in order to make it answer. The characteristic activity of the scientific method is therefore not induction but the design of experiments: choosing an intervention whose outcome would discriminate between the hypotheses currently in play, performing it, and letting the result decide. Induction is what happens afterwards, and it is the easier half.

Three features of that activity are ordinarily abstracted away and are exactly what we want to keep.

Observation is action.

To learn something about a system one must do something to it. There is no passive channel. Every measurement is an intervention, and the system afterwards is not the system before.

Inquiry is expensive.

Experiments consume resources that could have been spent otherwise, and a scientist with a finite endowment can run only finitely many. Which questions are worth asking is therefore not a philosophical question but a budgetary one.

The scientist is in the world it studies.

It is not a disembodied observer looking on from outside. It is made of the same stuff as its subject matter, subject to the same constraints, and visible to whatever else is looking.

A model that keeps all three will look less like statistics and more like ecology, and that is the model we build.

21.1.2 What such a model must supply

Before any formalism, here is what has to be on the table.

(R1) An environment, and a ceiling on what can be known of it.

One must say what the learner is learning about, and — crucially — what the best possible outcome of inquiry would be. Without a ceiling there is no way to say whether a learner is doing well.

(R2) A language for hypotheses.

Hypotheses must be stated in something. That something should be able to express every distinction the learner could in principle observe, and no more. A language too coarse leaves the learner unable to say what it can see; a language too fine sends it chasing differences that no experiment could ever settle.

(R3) A mechanism from hypothesis to verdict.

There must be an account of what an experiment is, how a hypothesis determines one, and how the outcome comes back as confirmation or refutation — including an honest account of the case where nothing comes back at all.

(R4) A motive.

Something must make the learner prefer better hypotheses. If this is an external scorekeeper handing out rewards, the model has outsourced the very thing it set out to explain, and one may reasonably ask who scores the scorekeeper.

21.1.3 Four answers

Our answers are as follows, and the remainder of the introduction says why each is less a choice than it appears.

(R1) The environment is other computations.

We take as premise that computations are ontologically isolated: a computation has no access to anything but through interaction. It follows immediately that the environment of a computation is other computations, and that what one can come to know of another is bounded above by the equivalence induced by the contexts one can build against it — context-labeled bisimulation [25]. That is the ceiling R1 demands, and it is derived rather than posited. We fix the model of computation as the \(\rhoc\) calculus [40].

(R2) The hypothesis language is generated, not designed.

Given the ceiling, R2 has a canonical answer: the language must be adequate for context-labeled bisimulation, which is exactly what a logic generated from the calculus’ own presentation gives. Feeding the \(\rhoc\) calculus’ presentation to the OSLF family of algorithms [41] emits namespace logic [39], whose formulae denote bisimulation equivalence classes. We admit the whole of it, including the structural predicates — and admitting those is what gives the learner senses (§21.3).

(R3) An assay is a guarded receipt.

We work in the variant of the \(\rhoc\) calculus whose for-comprehension carries a where clause [24, \S9], so a formula may sit in guard position on a receipt. A hypothesis installed as a guard is an experiment, and a confirmation is a rendezvous. The checker is the reduction rule; nothing is interpreted, and the test is metered by the same apparatus that meters everything else.

(R4) The motive is metabolic.

Cost accounting as a monad [30] makes the learner mortal: every step is paid for, and a computation that cannot pay stops. We then gamify. Tokens are conserved and never minted; the learner holds a bounded internal supply; further stacks lie distributed through the environment, including inside other computations; and breaking a computation open to harvest its stack is the analogue of eating. This is the answer to the scorekeeper problem. A hypothesis says where the tokens are, and the reward for a good one is the tokens themselves. Predictive accuracy and metabolic success become the same quantity, measured in the same units, with nobody adjudicating.

That is the whole of what we put in. The rest of the note is largely the discovery that we had no choices left: §21.14 audits which features are forced by this list and which remain free, and the list of free parameters is short.

21.1.4 What we are claiming, and what we are not

We are claiming that the commitments just listed determine almost all of the apparatus one would otherwise have to invent. The interesting sections below are the ones in which we find we had no choices left.

We do treat learning, and §21.6 is where. What we give there is the geometry of the hypothesis space and the constraints any revision policy must satisfy in it — that the space is an ultrametric tree and so admits no gradient, that incremental revision provably cannot leave the ball it begins in, that the cost of a move is monotone in its distance, and that the branching a searcher can survey is bounded by its wealth — together with a catalog of strategies compatible with those constraints. We are not claiming to have fixed a policy, proved a convergence theorem, or said anything about statistical efficiency. Which policy is best is an ecological question and depends on the distribution of §21.13; the policy itself remains one of only three genuinely free parameters we find. We do treat reproduction (§21.12), but we do not treat inheritance of acquired characteristics: whether a lineage’s transmitted material can carry what an individual learned is left open, since the empirical case against a strict germ–soma barrier [51] is strong enough that we would rather not build on either answer.

We are also not claiming a novel logic. The logic is the one obtained for free from the encoding, in exactly the sense of [23]: given the presentation \((\Sigma, E, R)\) of the \(\rhoc\) calculus, the OSLF algorithms emit a logic, and it is the same logic whether one is classifying knot diagrams or hunting for tokens. What is new here is the reading, the game, and three consequences of the game that we did not anticipate.

21.1.5 Why adequacy is the right demand

R2 asked for a language neither too coarse nor too fine, and it is worth saying why that pair of failures is the right pair to worry about, since adequacy is a technical condition that can look like a formality.

A learner separated from its environment by an interaction cut has exactly one channel of access: what interaction reveals. If its hypothesis language is too coarse, there are distinctions it can observe but cannot state, and it is blind in a way its own faculties do not require. If the language is too fine, it will form hypotheses that differ on environments no experiment could separate, and will spend tokens — which by R4 are its life — discriminating between what are, for it, the same world. Adequacy rules out both at once: the formulae separate precisely what interaction separates.

This is why the logic must be generated from the same presentation that generates the calculus rather than designed. A designed logic has no reason to be adequate. A generated one is adequate whenever its target specification carries the connectives adequacy needs [23, \S1.3], and when it does not, the failure is legible and located.

21.1.6 Ideal logic, checkable fragment, mortal scientist

We inherit the methodological commitment of [23]: the ideal logic, adequate for the idealized calculus, is what defines the invariants; fragments, restricted so that model checking terminates, are what decide them. In the present setting this distinction stops being methodology and becomes the subject matter. A mortal scientist cannot afford the ideal logic. What fragment it can afford is constrained by its budget, and that constraint is a large part of its inductive bias. We return to this in §21.5, and it is the reason this chapter’s central objects are approximants rather than limits.

Two levels are one too few, and the missing one is not the budget.

21.2 Isolation, the cut, and what a computation can know

21.2.1 The premise

We take this as a premise rather than argue for it. Its immediate formal content is that a world factors across an interaction cut. In an interactive graph-structured lambda theory [24], an interaction cut \(K\) separates a program from its environment: application for the lambda calculus, parallel composition for CCS, the pi calculus, and the \(\rhoc\) calculus. We write \[W \;\equiv\; S \mid E\] for a world factored into a scientist \(S\) and its environment \(E\). Nothing here privileges \(S\): the factorization is a choice of which side one is standing on, and in §21.10 we will take seriously that \(E\) contains further scientists for whom \(S\) is environment.

21.2.2 Three grades of access, and only three

The premise has a sharp consequence that the rest of the note leans on. \(S\) does not hold \(E\)’s term. It cannot read \(E\)’s structure by inspection, because inspection is not an operation the calculus provides across a cut. What it has instead is exactly three things, and it is worth being explicit that the calculus offers no fourth.

Perception.

\(S\) may evaluate structural predicates at the interaction surface — the outermost interaction structure of \(E\), the part exposed at the cut. This is what senses are: one sees the skin of a rock, not its interior.

Interaction.

\(S\) may build a probe \(C[-]\), plug \(E\) into it, and observe the steps that \(C[E]\) can then take. This is the only route to behavior. It is worth being careful here, because two different contexts are in play and the note has more than one use for each. The probe \(C\) is what \(S\) constructs and is a context around \(E\). The label on the generated modality \(\langle K_j \rangle\) of §21.3 is a context around the redex — everything in \(C[E]\) except the interacting pair — so it includes \(S\)’s own remaining code and whatever of \(E\) is not participating. The two coincide exactly when the redex sits at the cut, which is the case \(S\) can engineer and the reason probing is informative at all; they come apart as soon as \(S\) attends to a step internal to \(E\). We write \(\langle K_j \rangle\) throughout for the generated modality and reserve \(C[-]\) for probes.

Reflection.

If \(S\) holds a name, it may look inside it, because in a reflective calculus a name is a quoted process and the name predicate \(@\varphi\) sees into that quotation without communicating at all.

The third is the interesting one and it is available only for names \(S\) has come to hold. This gives the distinction the note turns on: observation versus disclosure. Behavior is what one observes across a cut; structure is what one is given, or takes. Perception reads the surface of things one does not hold; reflection reads the interior of things one does. The operation converting the second into the first is the subject of §21.7, and it is fatal to the thing converted.

\(\fitwidth{

}\)

21.3 The logic, and senses

21.3.1 What the generator supplies

We recall only what we use, following [23, \S2]. The \(\rhoc\) calculus has term formers \(0\), \(\ast n\), \(n!(Q)\), \(\texttt{for}(x \leftarrow n)P\), parallel composition, and \(@P\); parallel composition carries an associative–commutative equational theory and there is a single communication rewrite. Feeding this presentation to the generator yields:

Structural connectives.

One connective per term former. We write \(\mathrm{out}(n)\varphi\) for the predicate holding of \(n!(Q)\) with \(Q \models \varphi\), and \(\mathrm{in}(n)\varphi\) for the analogue on receipts. Because \(\mid\) is associative–commutative, the generated connective for parallel composition is a separating conjunction: \[P \models \varphi \mid \psi \iff P \equiv Q \mid R \text{ with } Q \models \varphi,\; R \models \psi .\]

Behavioral modalities.

One primitive modality \(\langle K_j \rangle\) per rewrite rule and choice of redex position, labeled by the one-hole context in which the step fires. The index is not decoration and we keep it. The rho calculus has a single interaction rule, so all the information in a label is in the position: for a world \(W \equiv K_j[\,\texttt{for}(x \leftarrow n)P \mid n!(Q)\,]\) the label \(K_j\) is everything else. Consequently \(\exists j.\,\langle K_j\rangle\varphi\) and \(\forall j.\,[K_j]\varphi\) quantify over the redex positions of \(W\), which is what gives the modal layer its force: an action-labeled logic sees one label where this one sees as many as there are enabled interactions. Where a formula below is written with a bare \(\langle K\rangle\) it abbreviates the existential, and where the quantifier is doing work we write it out.

Name predicates.

A name is a quoted process, so \[n \models @\varphi \iff n = @P \text{ with } P \models \varphi ,\] which is namespace logic [39]: one describes a namespace by a property rather than enumerating it. §21.8 shows this is not a convenience but a requirement.

Restriction, fixed point, grading.

The hidden-name quantifier \(P \models \mathsf{H}x.\varphi\) iff \(P \equiv \texttt{new } x \texttt{ in } Q\) with \(Q \models \varphi\) and \(x\) fresh; the greatest fixed point \(\nu X.\varphi\) with \(X\) positive; and graded separation \(\varphi \mid_k \psi\), which splits a term under \(\mathsf{H}\) with exactly \(k\) shared restricted names [23, \S2.3].

21.3.2 The structural predicates are admitted, and they are senses

In [23] the structural predicates are wielded by an analyst standing outside the term, holding it in hand, free to count its gates and test its decompositions. That is not the scientist’s situation, and admitting the structural predicates without saying where they may be applied would hand the scientist an oracle and destroy the isolation premise of §21.2.

The discipline that saves them is that a sense is a structural predicate evaluated at the surface.

Two comments. First, the depth grading is what keeps perception from being an oracle: seeing deeper costs more, so the interior of \(E\) is not free, it is merely expensive, and past a point unaffordable. Second, \(d(@\varphi) = d(\varphi)\) rather than \(1 + d(\varphi)\) is deliberate and records the third grade of access: opening a name one holds is not a passage through a surface, because one already holds it. Reflection is cheap for exactly the reason predation is not.

21.4 Hypotheses, assays, and the where clause

21.4.1 The embedding, in operational position

An embedding of the logic into the calculus is what lets \(\rhoc\) computations compute over hypotheses. The obvious construction — reify formulae as data, write an interpreter — works but is unsatisfying, because the scientist must then carry the interpreter and the interpreter is not part of the theory. The variant of \(\rhoc\) whose for-comprehension carries a where clause [24, \S9] does better. Its general form is \[\texttt{for}\big( p_{11} \leftarrow c_{11} \,\&\, \cdots \,;\, \cdots \,;\, p_{1n} \leftarrow c_{1n} \,\&\, \cdots \ \texttt{where}\ \vartheta \big) P\] with \(\&\) joining within a group, \(;\) separating groups, and \(\vartheta\) a condition drawn from the booleans over ground formulae, spatial connectives, and the context-labeled modality, in which the variables bound by the preceding clauses may be mentioned. A guard is a formula in operational position. Hence:

The checker is the reduction rule. Nothing is carried, nothing is interpreted, and the cost of testing a hypothesis is metered by the same apparatus that meters everything else — which is essential here, since an unpriced hypothesis test would be an oracle and would collapse the entire motivational story of §21.7.

21.4.2 The assay

   K[ ... ]                                        // the probe: perturb the environment
 | for (x <- p where phi(x)) { confirm!(*x) }       // prediction holds
 | for (x <- p where ~phi(x)) { refute!(*x) }       // prediction fails

The negation in the second guard is what makes this more than an expression of hope. In a concurrent setting a guard that simply fails to fire is indistinguishable from a guard still waiting, so a naive encoding of falsification cannot tell refutation from patience. The complementary pair repairs this.

We stress that \(\bot\) is not a defect. It is the honest residue: the environment did not answer, and no amount of logical apparatus can distinguish that from an environment that would have answered later. What it does is separate that genuine ignorance from falsity, which the naive encoding conflates.

21.4.3 Refutation is budget-relative, and that fixes the fragment

The consequence is the note’s first structural fact about mortality. A scientist with a finite budget can only run assays whose outcome is decided within that budget, so the fragment of the ideal logic it can actually test is the one bounded in modal depth and fixed-point unfolding. That fragment is not adequate — a logic whose model checking terminates never is, for a Turing-complete host [23, \S1.3]. Therefore:

This is not a limitation we regret. It is the formal content of the observation that finite creatures have finite theories, and it locates the inductive bias somewhere unusually concrete: not in a prior, not in a regulariser, but in a token stack.

21.5 The hypothesis space is the relaxation lattice

We do not need to invent a space of hypotheses. [23, \S3.3] constructs one and calls it something else.

Recall the construction. For a process \(P\), a characteristic formula \(\chi_P\) is satisfied by exactly the processes equivalent to \(P\); it turns an equivalence question into a model-checking question. Beneath \(\chi_P\) sits a lattice of progressively weaker formulae obtained by forgetting parts of it — the recursion, the continuations, a private name — each forgetting yielding a coarser predicate that more processes satisfy. In [23] this lattice is where a classifier’s dials live. Here it is the space the scientist moves in.

The stratification is not chosen for convenience: it is how bisimulation is already approximated, it grades with modal depth, and by Proposition 21.2 modal depth is what the budget bounds. Distance, depth, and cost are therefore one graded structure rather than three, which is the main reason to prefer this metric to a measure-theoretic one on denotations.

So \(\chi_E\) — the complete theory of the environment — sits at the top of the lattice as a limit the scientist approaches and never reaches. Science converges along the lattice.

21.6 Searching the lattice

Section 21.5 supplies a space of hypotheses and a metric on it, and says that a scientist’s trajectory is a walk. It does not say how the walk is chosen. That is the revision policy, which we left open as one of the note’s three free parameters (§21.14), and we leave it open still. What we do here is narrower and, we think, more useful: the metric is an ultrametric, and ultrametric spaces are geometrically unlike the spaces search algorithms are usually designed for. Much of what one would reflexively try is unavailable, and one thing one might not have thought necessary turns out to be forced.

21.6.1 The geometry is a tree

Three facts about ultrametric spaces do the work. Balls are nested or disjoint, never partially overlapping. Every point of a ball is a center of it, so no point in a neighborhood is distinguished. And there is no betweenness: for \(\varphi \ne \psi\) there is no formula standing partway from one to the other in the sense a convex combination would.

This has a consequence which we regard as the central fact governing revision policies.

In a Euclidean space, sufficiently many small steps reach anywhere; here they reach nothing outside the ball they started in, however many are taken and however long the scientist lives. A policy of incremental revision is not slow at escaping a bad basin — it cannot escape one at all. Whatever ball a scientist’s early commitments place it in is the ball it will die in, unless it makes a move that is not small.

We take this up again in §21.12, because it turns a biological mechanism into a computational necessity.

21.6.2 The two parameters

So the two parameters the policy must choose are exactly the two the metric grades.

Distance is backtracking depth.

A move of distance \(2^{-n}\) agrees with the current hypothesis on every approximant below stage \(n\) and differs at \(n\). Small distance means a deep backtrack and a fine adjustment; large distance means a shallow backtrack and the abandonment of an early commitment.

Direction is a choice of sibling.

At a node of the tree the available directions are the one-step refinements the generated logic admits at that locus. The branching factor is therefore fixed by which connectives OSLF emitted, and is not ours to choose.

The price schedule follows without further assumption.

A radical revision is expensive in precisely the currency the scientist has been spending: it discards the assays that bought the commitments it abandons. Cheap moves are confined (Proposition 21.4) and unconfined moves are dear, which is the trade-off any policy is negotiating.

21.6.3 Strategies

We catalog rather than recommend, since which policy is best is an ecological question and depends on the distribution of §21.13.

Conservative.

Always move the least distance that the last refutation forces. Cheap by Proposition 21.5, low variance, and by Proposition 21.4 permanently confined to its initial ball.

Radical.

On refutation, abandon a commitment high in the tree. Unconfined, and expensive in exactly the confirmations it throws away.

Adaptive, with refutation rate as temperature.

Jump far when refutations are frequent — a high refutation rate is evidence that a commitment high in the tree is wrong — and refine when they are rare. This is annealing-shaped, but the temperature is not a schedule imposed from outside: it is the ratio of refute to confirm events, which the assays of §21.4 already produce as messages. We regard this as the default policy a designer would reach for, precisely because its control parameter is already an observable of the system.

Direction selection is the second half of the policy, and the candidates differ in what they optimize.

Blind.

Choose a sibling uniformly. Requires no evaluation, and is what a lineage does when the choice function of Definition 21.15 is unbiased.

By formula cost.

Prefer siblings whose formulae are cheap to check. This is the minimum description length bias of §21.16 appearing as a search heuristic rather than as an accounting identity, and it has the merit that the scientist prefers hypotheses it can afford to test.

By discriminating power.

Prefer a sibling \(\psi\) for which some assay separates \(\psi\) from the current \(\varphi\) — that is, design the crucial experiment first and let it select the direction. This is the distinguishing-formula problem for bisimulation inequivalence [42], and it is the most classically scientific of the options: it maximizes the information a single assay returns.

By expected yield.

Prefer siblings which, if true, locate more tokens.

21.6.4 The search is itself metered

One further constraint is easy to overlook and is not a detail. Choosing a direction requires evaluating candidate siblings, and evaluation is model checking, and model checking is priced. A scientist therefore cannot survey the branching it faces.

We note without developing it that this closes a loop with §21.11: a source-feeder whose inquiry converges accumulates the surplus that would buy a wide survey, and has nothing left to search for; a prey-feeder that needs the survey is the one paying an unbounded epistemic rent.

21.6.5 What escapes

Proposition 21.4 leaves an individual scientist with exactly one route out of a bad basin: a move that is not small, at a cost given by Proposition 21.5, chosen from a branching it cannot afford to survey. That is a poor prospect, and it is the reason §21.12 is not an appendix to this note but a part of its argument. A crossover at a locus high in the term tree is, transported along the characteristic-formula construction, a shallow re-branch in the formula tree — a displacement of exactly the kind an incremental walk provably cannot make, and one no individual pays for, because the cost is borne by the offspring that does not yet exist. The escape is not unconditional: by Remark 21.19 how far a swap can displace is fixed by the recombination discipline, and the disciplines that displace furthest are the ones that kill most offspring.

21.7 The game

21.7.1 Conservation

Tokens are conserved and never minted. Write \(\Theta\) for the total supply of the world. The cost monad’s ledger law and the history monad’s record together give, globally, \[\sigma + \kappa \;=\; \sigma_0 \;=\; \Theta ,\] where \(\sigma\) is the total still held in live stacks and \(\kappa\) the total spent. Spent tokens do not vanish; they migrate into the record. So the free supply is monotonically non-increasing, the accumulation of \(\kappa\) is the arrow of time, and there is no primary producer. The whole game is a finite race.

21.7.2 Working in the image

Because \(\rhoc\) is Turing complete, cost-accounted \(\rhoc\) translates back into \(\rhoc\): there is an internalizing encoding \(\enc{-} : \Crho \to \rhoc\) that represents the located token stacks in the undecorated calculus [24, \S6]. Without loss of generality we may therefore avail ourselves of the cost-accounted syntax where that is convenient and of the channels at which token stacks are located where that is convenient.

So eating is a receive, death is starvation, and defense is restriction. All three are already in the calculus.

21.7.3 Predation is the failure of internalization to be faithful

Here is the observation that we take to be the note’s chief structural contribution.

In \(\Crho\) a located stack is inviolable. That is not an accident of presentation; it is precisely what the located-stack discipline and the ambient-authority fix were built to guarantee. No rewrite takes your stack, because the stack is not the sort of thing a rewrite addresses.

In the image \(\enc{\Crho} \subseteq \rhoc\) the stack is a message at a name, and any process holding that name may receive it.

These are not the same situation, and the encoding is not faithful against arbitrary \(\rhoc\) contexts. It is faithful against images of \(\Crho\) contexts — which is exactly the refinement of [25]: an encoding may only be probed by the encodings of source contexts, since arbitrary target contexts see strictly more and would never permit bisimulation preservation. Therefore:

This gives the game a rulebook rather than a stipulation.

The two endpoints are degenerate: \(\mathcal{A} = \enc{\Crho\text{-}\mathrm{Ctx}}\) is decorated cost-accounted \(\rhoc\) with no game at all, and \(\mathcal{A} = \rhoc\text{-}\mathrm{Ctx}\) is total war in which any context may take anything it can address. The interesting point is the middle one:

Since names are unforgeable, derivation is real work, and it is the work namespace logic does: describing a namespace by a property rather than enumerating it. Under capability gating, foraging is name discovery and a hypothesis is a namespace description that either does or does not locate real tokens. Everything below is stated for the capability-gated game.

21.7.4 The foraging inequality

Conservation has an immediate consequence which we will lean on repeatedly: harvest is not automatically profitable, because finding and opening a stack costs tokens drawn from the same conserved supply.

This is optimal foraging theory, and we did not import it; it is what Proposition 21.7 says once concealment is priced by Definition 21.2. It is also the inequality whose sign change draws the phase boundaries of §21.13.

21.8 Destructive assay forces a logic of namespaces

Harvest ends the prey’s computational life. That single decision has a consequence for the hypothesis language which we did not expect and which we now regard as the best available answer to “why namespace logic.”

That is namespace logic’s definition. So the logic is not adopted for elegance or for uniformity with a companion note; it is what survives the game. Induction — the move from the specimen to the kind — is not an epistemic virtue the scientist is asked to display but the only strategy that pays when each specimen affords exactly one measurement.

21.9 Exposure, and the price of being known

21.9.1 The exposure lemma

The prey’s optimal defense in a world of predators looks, at first, unbeatable. Restrict everything. A fully closed computation is a live \(\tau\)-network with no free names; no context can communicate with it; and by the pessimism of [23, \S1] its observable behavior carries almost no information. It cannot be perceived, so its metabolic channel cannot be derived, so it cannot be harvested.

It also cannot eat.

The predator’s counter-move is now legible, and it is a reading of [23, Rem. 2] turned around. Restriction makes a name unobservable but not undescribable; \(\mathsf{H}x.\varphi\) reaches under the binder and speaks about what is hidden. A predator therefore attacks what it cannot observe by describing it, which is why the spatial layer and not the modal one is the hunting instrument, and why senses had to be admitted in §21.3 for the game to be non-trivial at all.

21.9.2 The invasiveness ladder

Table 21.1 ordered the three grades of access by what they reveal. They are ordered in the opposite direction by what they destroy.

This is where the note’s arc closes. A scientist that is very good at forming hypotheses locates metabolic channels efficiently; locating them efficiently is eating efficiently; eating efficiently exhausts the environment that was confirming the hypotheses. The better the theory, the faster it destroys the conditions of its own confirmation. Success is self-terminating, and no additional assumption was needed to make it so.

21.10 Mixed environments

The environment contains a mixture: bare token stacks, non-scientist mortal computations, and other scientists. Each stratum behaves differently and the third is the interesting one.

Bare stacks

are the primary resource: inanimate, undefended, cheap to open, and finite. They are what makes the shallow-and-external regime of §21.13 a regime in which science does not pay — when food lies in the open, senses suffice and a theory is pure overhead.

Non-scientist mortal computations

burn tokens and may have structure, but form no model. They defend passively, by restriction, and are subject to Lemma 21.1 like everyone else.

Scientists

model back, and three consequences follow.

The most epistemically active agents are the most edible. We do not think this is an artefact.

Theory theft.

Reflection makes another scientist’s hypothesis readable. A hypothesis in guard position is a term; a term may be quoted; and \(@\varphi\) inspects the quotation without communicating. So a scientist may acquire a model by reading rather than by experimenting — cheaper, and not fatal to anyone. Whether honest science survives this depends on whether holding a model requires exposing it, and in this calculus it does, because a guard is surface. We record the equilibrium question as an open problem (§21.18) rather than settle it.

Imagination.

Turing completeness plus reflection means a scientist may carry \(\enc{-}\) itself as data and run a model of a prey’s metabolism inside its own budget, paying tokens to avoid paying tokens.

This is the point at which the scientist stops being a reactive forager and becomes a planner, and it is where the stacking of the cost and history monads acquires an interpretation that is not merely formal: the simulated prey is itself a mortal computation, metered inside the predator’s budget.

21.11 Trophic structure

21.11.1 Conservation is the right scope, not a limitation

Smith’s organization of the biosphere [52] distinguishes organisms that consume other organisms from those that consume only external energy, and it is tempting to read the second class as evidence that a model of life needs an open system — a source outside the conserved pool. We do not think that reading survives inspection, and we record why, because the temptation is to apologize for conservation at exactly the point where conservation is doing the work.

The sun is a reservoir of finite extent, situated in a physical system in which energy is conserved. Its apparent externality is an artefact of where the boundary was drawn: put the boundary at the biosphere and the sun is outside it; put the boundary at the solar system and there is one conserved budget being spent down. Openness, here as generally, is a bookkeeping convenience.

So conservation is not a disanalogy with Smith’s picture. It is the picture at the correct scope, and the whole trophic stratification is recoverable without minting anything.

21.11.2 The distinction is an access profile

If the sun is a token stack, then the autotroph and the heterotroph are not distinguished by what they eat. They are distinguished by how the stack they eat may be accessed.

The calculus carries this distinction natively, and no new sort is required to express it.

Same sort, same rewrite rule, different multiplicity. Whether a locus is a source or a prey is therefore a property expressible in the generated logic, and a mortal computation’s diet is fixed by which of these profiles its senses are tuned to derive.

21.11.3 The split is predicted, not stipulated

The payoff is that we do not have to posit the trophic division. Two results already in hand imply it.

Remark 21.6 observes that a recursion following the structure rather than the reduction is exact at a finite unfolding. A rate-limited persistent emitter is precisely such a recursion: a replicated send, structurally regular, non-adaptive. Proposition 21.3 observes that an environment whose recursion follows the reduction admits no characteristic formula exact at any finite unfolding. An adaptive prey, being itself a mortal computation, is precisely such an environment.

Plants do not need brains because the sun is the environment whose complete theory is affordable. We offer this as a consequence of Proposition 21.3 and Remark 21.6 rather than as an observation imported from biology; that it is also true of the Earth we take as corroboration of the modeling rather than as its source.

21.11.4 Why the overlap is thin

The space of mortal computations is plainly large enough to contain those that eat both — nothing in Definition 21.11 forbids tuning one’s senses to both profiles. On the Earth the corresponding region is sparsely occupied, and it is worth asking what the formalism says about why. Four pressures are available, none of them requiring new apparatus.

1. Opposed geometry. Yield from a rate-limited dilute source scales with collecting surface; yield from a concentrated one-shot source scales with reach and speed. Surface against velocity, derived from Table 21.3 alone rather than borrowed from the shapes of leaves and predators.

2. Opposite sides of Proposition 21.3. By Proposition 21.12 the two diets demand different points on the adequacy/checkability trade-off. The source-feeder can live in a terminating fragment; the prey-feeder cannot, since its subject matter is Turing complete and adaptive. One budget cannot be allocated to both regimes at once.

3. Double rent for single-mode yield. A foraging strategy is a hypothesis together with a sensory schedule, and both are held at a cost. A dual-mode computation pays rent on two apparatus while eating with one at a time — the argument of §21.16 applied to strategies rather than to formulae.

4. Hysteresis. Revision is a walk on the relaxation lattice and walking costs. Where the transition cost between strategies exceeds the gain from switching, the mixed strategy is not infeasible but dynamically unstable: it is not a fixed point of the revision dynamics, and specialization is an attractor.

The sharpest statement combines the first two with §21.9.

The region is thin because it is dominated, not because it is unreachable.

21.11.5 The overlap is populated by composites

There is a further and we think better answer, which is that the question contains a false premise.

On the Earth, dual capability is realized almost entirely by symbiosis: corals with their zooxanthellae, lichens, kleptoplasty in Elysia, and — decisively — the mitochondrion and the chloroplast themselves. Very few single genomes do both at scale; pairs of genomes sharing a boundary do it constantly.

In this setting that is \[\texttt{new } m \texttt{ in } (\,P \mid Q\,) ,\] two specialists sharing a metabolic channel under a restriction. Parallel composition makes composites first-class and graded separation \(\mid_k\) measures how much interface they share, so the construction needs nothing added.

Read with Remark 21.15 this is not a special construction but the general situation: the organism boundary is drawn by the restriction, not by the strategy, and what looks like one dual-feeding thing at one granularity is two specialists at another.

21.11.6 Succession

One consequence of taking conservation as the correct scope, rather than as a limitation to be repaired, deserves recording. Because a source is finite — rate-limited, but not inexhaustible — a world does not settle into stable coexisting trophic layers. It runs through them.

The stratification is temporal as much as structural. On a stellar timescale this is a distant prospect and not a modeling artefact; it is the correct statement of the situation at the scope at which energy is conserved.

21.11.7 An ensemble test

Both claims of this section — that the mixed strategy is dominated, and that the composite is not — are checkable inside the framework rather than merely analogical, and the machinery is that of §21.13. Sample ecologies across the \((\iota, \delta)\) plane and ask, first, whether the mixed access profile is ever a viability optimum for an atomic computation, and second, whether the composite of Proposition 21.14 dominates it wherever it is not. We regard this as a sharper first computation than calibrating the phase boundaries, because it can come out either way and one would want to know which.

21.12 Reproduction

Harvest ends a computational life, so an ecology with no births is a monotone depletion of agents and nothing in it persists long enough to be a subject matter. This section supplies reproduction. The mechanism is reflection, and the interest of it is that the genetic apparatus is not modeled but found: a genome is a quotation, a locus is a context, and crossover is the exchange of parands between two quoted terms.

21.12.1 The naive route, and why sex is worth its price

The cheap route is already in the calculus. Replication, guarded recursion, a persistent receipt, a contract: each reinstates a term exactly. This is asexual reproduction, and it is strictly cheaper than the alternative — no partner, no alignment, no disclosure, no recombination to compute. Sexual reproduction produces variants rather than copies, so the two are not different apparatus but the extremes of a single dial: exact reinstatement at one end, recombination at the other. In the sexual regime there is no exact recursion anywhere in a lineage.

Why pay for the dial to be off zero? The usual answers are available, but this setting supplies a sharper one, and it comes from the first half of the note.

This is a Red Queen argument with a mechanism rather than an analogy, and it is available because the hypothesis language was chosen to be a logic of namespaces in the first place.

21.12.2 The germ is a quotation

Everything one wants from the distinction follows from properties the calculus already has.

• \(@P\) does not reduce. A quotation is a name, and names do not step. So a genome does not metabolize, does not age, and burns no tokens while held.

• \(@P\) is transmissible. Names are what sends carry, so a genome may be handed to a partner or to an offspring on an ordinary channel.

• \(@P\) is inspectable. By §21.3 a name predicate sees into a quotation, and by Remark 21.3 it sees all the way down. A genome may therefore be screened without being run — at reflection prices, which by Table 21.1 are the cheapest available.

• \(\ast @P\) costs. Running is what is metered.

21.12.3 The normal form is a tree, and a locus is a context

Up to structural congruence a process is a parallel composition of primes, and each prime is a receive, a send, or a drop: \[P \;\equiv\; \prod_{i \in I} \texttt{for}(p_i \leftarrow c_i)P_i \;\Big|\; \prod_{j \in J} c_j!(Q_j) \;\Big|\; \prod_{k \in K} \ast x_k .\] This is a gross decomposition only. Each continuation \(P_i\) and each transmitted body \(Q_j\) is itself a process with a normal form of the same shape, so the decomposition recurses and a genome is a finitely branching tree whose nodes are parallel compositions and whose edges are prefixes. The parands are the individual primes, at every level.

The vocabulary is not new: a locus is a one-hole context in the same term algebra the modalities of §21.3 are labeled by, so genetics requires no notion of address the note did not already have. It requires a slightly larger one, and the difference is worth stating rather than eliding. A modal label \(K_j\) is a one-hole context whose hole sits at a redex; deleting an arbitrary parand — an unmatched send, a receipt with nothing to receive — yields a one-hole context that is not the label of any modality. The loci therefore properly contain the modal labels, and a crossover at a locus outside that class moves the genome to a place no formula of the generated logic can name.

We flag it here and repair it in §21.12.5, where the repair turns out to be a discipline we were going to want anyway.

Homology is therefore a partial isomorphism of the two trees rather than a matching of two multisets, and the binding clause is a reading-frame condition: an allele’s body may reference only variables its new context binds.

21.12.4 Crossover

We take crossover proper rather than assortment of whole primes. Assortment — exchanging a prefix together with its body — is the safe special case in which the binding condition is automatic; it is Mendel’s independent assortment, and it cannot recombine within a gene. Splitting guard from body is where recombination earns its name.

The gradation is uniform, and it is your own observation about recursion that makes it so: were the genome to contain recursive definitions, a swap inside a recursive body would be a global edit whose effect size was large regardless of depth. Since sexual reproduction yields variants and never exact reinstatement, no such loci occur, and the analogue of the distinction between large-effect regulatory and small-effect structural variation falls out of the term structure rather than being posited as a rate.

21.12.5 Communicating pairs as units of inheritance

Definition 21.15 lets a choice function select any allele at any homologous locus, and §21.12.8 then observes that most of what it produces is inviable. There is a discipline which removes much of that waste, and it changes what the unit of heredity is.

That is worth pausing on, because it obtains something one would otherwise have to impose. Material that must work together is inherited together — linkage — not because a chromosome places it adjacently but because the pairing is the working-together.

It also obtains something we owed. Proposition 21.17 observed that loci properly contain the modal labels, so an undisciplined crossover can move a genome to an address the logic cannot name. The discipline closes the gap exactly.

We regard this as the better argument for the discipline, and a different one from waste reduction. Restricting to communicating pairs does lower the inviable fraction, which is how it was motivated above. What it also does — and what nothing else in this note would have supplied — is make a crossover and a revision the same kind of object: a change of one-hole context, on one side of the characteristic-formula map or the other. Without the discipline the two live in different address spaces and the correspondence is an analogy. With it, it is a bijection.

The discipline does not by itself preserve viability, and the reason is instructive. Suppose the mother contains a race, \[c!(P) \;\mid\; \texttt{for}(x \leftarrow c)Q \;\mid\; \texttt{for}(y \leftarrow c)R ,\] and the swap removes the pair \((\,c!(P),\ \texttt{for}(x \leftarrow c)Q\,)\) in favor of a father’s pair on some other channel. The donated interaction is sound; the residual receipt on \(c\) is orphaned, which is precisely the deadlock half of the balance condition. Removing a pair cannot break the material it donates but can break the material it leaves behind.

So preserving races is not a second benefit alongside preserving balance; it is the side condition under which the balance guarantee holds at all.

21.12.6 Three kinds of channel, and what a swap can touch

The discipline of §21.12.5 operates on some of a genome and not on all of it, and saying which requires a classification that is useful well beyond recombination.

The classification is relative to the cut, as trophic type was (Remark 21.15), and for the same reason: what counts as inside is a question about where the boundary was drawn.

Dialogue

A computation with no dialogue of depth greater than one is a reflex: whatever it emits is not conditioned on anything but the message immediately received. Depth two and above is conversation — the environment’s reply is answered in the light of what the first exchange produced. Two consequences follow, and the first is what the user of this framework will care about.

That repairs something the note has so far taken on trust. The grazer of Appendix 21.19 is not a non-scientist because we declined to give it a hypothesis; it is a non-scientist because its interaction graph contains no depth-two loop through a both-type channel, and the difference is checkable.

It also repairs Proposition 21.10, whose original statement had to normalize by burn rate. The correct statement is structural: deeper dialogue requires more both-type channels, and both-type channels are surface.

What each kind of swap touches

This grading is independent of the depth grading of Proposition 21.19. A recombination policy has two dials: how much structure a swap moves, and how observable the moved structure is.

This sharpens Proposition 21.16. Only the observable fraction of recombination attacks a predator’s hypothesis, because a predator’s hypothesis is a description of the external and both-type sub-namespace and nothing else. But internal variation is the reservoir from which observable variation is later drawn, so the defense is two-staged: accumulate where it cannot be seen, and expose when the exposure is worth its cost.

21.12.7 Names, freshness, and the bound on growth

Because new is sugar (Remark 21.3), names have size, and size is priced. One might therefore fear that recombination inflates them: if a child’s fresh names were manufactured from its enclosing term, and its enclosing term combines two parents, names would double with each generation and a lineage would price itself out of existence within a few dozen.

The calculus forbids it.

The feared scheme is not merely expensive but syntactically impossible, and the bound holds uniformly across every decentralized scheme — which is why we may remain agnostic about which one is taken.

What growth remains is the right kind. At a homologous locus the recombinant takes one allele, not both, so aligned material contributes a pointwise maximum rather than a sum. Growth enters only at non-homologous loci — material present in one parent alone — where inheriting the union is additive per generation. That is gene duplication, it is linear, and by Definition 21.2 it is priced, which is the regime in which selection can act on it.

Under a central name server, or under a calculus of atomic names with global freshness, this route is closed. It is the same fact recorded in Remark 21.14: what denies a locus guaranteed closure is what allows a lineage novelty.

21.12.8 Viability, and why there is selection rather than design

A parallel composition of primes is always a term, so syntactic well-formedness is free. Viability is not, and it is a conjunction of conditions the generated logic can state: closure, the reading-frame condition of Definition 21.14, balance — receives without matching sends deadlock, sends without receipts leak — and liveness in the marked-graph sense. Screening a recombinant is therefore model checking, and the classifier apparatus of [23] is the embryology of this note.

But model checking is metered, and the space is large.

Neither fine granularity nor cost accounting gives this alone; it needs both.

21.12.9 The energy tension: a three-way allocation

Reproduction takes energy, and under conservation the energy comes from the parents and nowhere else. An offspring is endowed out of a parent’s stack, so every child is a portion of the parent’s remaining life.

We take the endowment to be chosen by the parent rather than fixed by the offspring’s code size, which makes the trade-off between many poorly provisioned offspring and few well provisioned ones a strategy rather than a constant.

The tension the reader will have anticipated is not two-way but three-way. One conserved stack funds soma maintenance, germ production, and epistemic outlay: the same tokens buy eating, breeding, and thinking.

This supersedes the reading of surplus offered in §21.16: surplus does not simply buy curiosity, it buys curiosity or children, and which it buys is the life-history problem.

21.12.10 What reproduction does for the science

Two consequences bear directly on the first half of the note.

The lineage is a second learner.

A scientist’s hypotheses are guards; guards are parands; parands are what crossover exchanges. So the genome contains the hypotheses, and they are revised by two mechanisms on two timescales — within a life by walking the relaxation lattice of §21.5, across lives by recombination. The same object is revised by both. A lineage is not merely a means of persistence; it is doing science by other means, and the proximate/ultimate distinction of §21.16 extends to cover it.

The correspondence is tighter than an analogy. A revision move is a locus in the formula tree together with an operation there (Definition 21.6); a crossover is a locus in the term tree together with a choice of allele (Definition 21.15). The characteristic-formula construction of §21.5 carries term structure to formula structure, so a swap at a term-context \(K_j\) induces a revision at the formula-locus \(K_j\) corresponds to — and by Proposition 21.20 the swap is at a \(K_j\) in the first place only because the recombination discipline restricts it to one, without which the induced formula-locus need not exist. The two mechanisms are one move, transported along \(\chi\), and applied at two timescales. What distinguishes them is the accessible distance: by Proposition 21.4 an individual’s incremental walk cannot leave its ball, whereas a swap high in the term tree is a shallow re-branch by construction. Recombination is how a lineage makes the displacements its members cannot.

Reproduction is what makes induction pay.

Corollary 21.3 forced hypotheses to be about namespaces, since destructive assay leaves an individual-level hypothesis worthless after its own first test. But a namespace with no live members is worthless too. Reproduction repopulates the namespace, so a class-level hypothesis retains value across the specimens it costs to test.

21.13 Distributions over ecologies

21.13.1 A world is a partition of a conserved quantity

Because \(\Theta\) is fixed and never minted, a world is a partition of \(\Theta\) across loci together with a structural assignment saying which loci are scientists, which are non-scientist mortal computations, and which are bare stacks. The configuration space is graded by \(\Theta\), and a distribution over it is a microcanonical ensemble in the strict sense — fixed total, distribution over microstates. The analogy is exact rather than suggestive, and it is exact because conservation is exact.

21.13.2 Two order parameters, and a phase

Two quantities govern the qualitative behavior and they are independent:

• \(\iota \in [0,1]\), the fraction of \(\Theta\) already internal to scientists rather than loose in the environment;

• \(\delta \ge 0\), the mean concealment depth of external stacks — how far under restriction and structure they lie, which by Definition 21.2 sets the sensing cost of finding them.

\(\fitwidth{

}\)

We regard this as the most substantive empirical claim the construction makes, and it is a claim the construction makes rather than one we impose: the boundaries of Figure 21.2 are where the foraging inequality of Proposition 21.7 changes sign. The low-\(\delta\) boundary is where \(\kappa_{\mathrm{sense}}\) falls below the cost of holding a theory at all; the upper boundary is where Proposition 21.7 fails for every available \(P\).

21.13.3 Sampling

The machinery for turning a configuration into a distribution and sampling trajectories through it is not new work. Chapter 13 takes a continued interactive GSLT, an adequate spatial–behavioral logic supplying the keys, and a weight map, to a weighted coalgebra (Definition 13.9) which the Gillespie construction samples (Theorem 13.1); assembling the weightings of a fixed theory and taking the Grothendieck construction gives the fibred form. That is precisely what an ecology is: the weight map turns a configuration into a distribution, and the Gillespie construction samples it. §21.15 takes up what that fibred structure means here.

21.14 The audit: what the assumptions already fix

It is easy to build a machine that fits a story. We therefore separate, explicitly, what follows from the commitments and what we chose.

21.14.1 Forced

21.14.2 Chosen

Everything metabolic and everything epistemic follows from cost-accounted \(\rhoc\) together with the logic it generates. What remains to be chosen is the weighting, the initial distribution, and how the scientist thinks.

We used to regard the smallness of that list as the principal evidence that the modeling choices were not tuned to the conclusions, and the list has since grown twice. The last two rows are additions of the present pass, and neither is a detail: the observation signature fixes which distinctions are available to the learner in principle rather than in budget, and the cost-to-rate map is what an earlier version of §21.15.2 claimed the theory supplied for free. A reader auditing this chapter should assume the list is still short of complete, and should treat the “forced” column as a claim about what follows given the choices above rather than as a claim that the choices above are inconsequential.

21.15 Is the assignment functorial?

It is natural to ask whether ecologies are a functor of the theory: is there a functor from the subcategory of cost-accounted graph-structured lambda theories to distributions over computational ecologies of the kind described above? The answer is instructive, and it is no in a way that tells us what the right question was.

21.15.1 Not from the base

Write \(\CGSLT\) for the subcategory of cost-accounted GSLTs. The assignment \(\mathcal{C}(G) \mapsto\) “its ecologies” is not determined by \(\mathcal{C}(G)\). Two different weight maps — two different pricing schedules on the same cost-accounted theory — give different sensing costs, hence different foraging inequalities, hence different phase boundaries in Figure 21.2 and different distributions over trajectories. The object does not determine the image.

This is Chapter 13’s construction with a different label on the target: it already goes from a Grothendieck construction of weightings into weighted coalgebras, and an ecology-with-a- distribution is a weighted coalgebra.

21.15.2 There is no canonical section

There is nevertheless a rescue. Its shape is the interesting part, and it is smaller than an earlier version of this section claimed.

What survives is weaker and still useful. Cost accounting canonically supplies the resource ledger; an ecology additionally requires a pricing or stochastic weighting. Each monotone map from costs to rates defines a section of \(\pi\), and the modeller’s choice among them is a real degree of freedom rather than a departure from the theory’s own answer, because the theory does not have one. That makes it a free parameter, and the audit of §21.14.2 has been amended to list it as one.

21.15.3 The morphisms must be logic-preserving, and that costs nothing

Functoriality on morphisms is where a real correction is needed, and we think it is the most useful thing this section contains.

The morphisms of \(\mathbf{GSLT}\) are bisimulation-preserving maps, where the bisimulation is the context-labeled one and is a congruence relative to a class of admissible contexts (Remark 16.3). But an ecology depends on spatial structure: concealment depth, which computations sit inside which, how a network separates when a locus is removed. And the spatial layer is strictly more discriminating than bisimulation in the base theory — this is the thesis of [23], that separation says things no modal logic can say. So a bisimulation-preserving map may scramble the spatial layer, changing sensing costs and foraging inequalities while preserving behavior exactly.

The obvious repair is to narrow the morphism class by hand to maps preserving the generated logic in both its layers, and note that the narrowing is strict. That is what an earlier version of this section did, and it is the weakest paragraph it contained: a class of morphisms defined by fiat to make an assignment functorial is not a category one has any right to expect adjoints in.

It turns out not to be a narrowing at all.

This is worth more than the repair it replaces. The requirement that a morphism respect what the senses see is not an extra demand imposed on \(\mathbf{GSLT}\) from the ecology’s side; it is the ordinary requirement, made in a theory where the senses are instruments the observer holds. What looked like a narrowing was a change of the theory the morphisms live over. And the practical consequence is that the free and forgetful adjunctions of the machinery part are not obviously lost, since one is still in a category of GSLTs and their bisimulation-preserving maps — a different one.

We state this and do not chase it. The three-line summary for the reader in a hurry is: a fibration, a section for each choice of monotone cost-to-rate map and no canonical one among them, and a morphism class that looks like a restriction of \(\mathbf{GSLT}\) and is in fact \(\mathbf{GSLT}\) over an enlarged theory.

21.16 Proximate and ultimate scores

The original motivation for this construction measured the delta between hypothesis revisions and punished degradation. The game supersedes that mechanism as the ultimate criterion — survival is the score, and it needs no adjudicator, no referee holding a mint, and no exogenous reward signal. Two things fall out that we would otherwise have had to impose.

• Vacuity is punished by hunger, not by rule. The hypothesis \(\top\) predicts nothing, locates no metabolic channel, yields nothing, and starves. No informativeness term is needed in any score function, because there is no score function.

• Occam’s razor is metabolic. A hypothesis is worth holding exactly when expected yield exceeds \(\kappa_{\mathrm{sense}} + \kappa_{\mathrm{assay}} + \kappa_{\mathrm{break}}\) plus the cost of holding the formula itself. Description length and food are denominated in the same currency, so minimum description length is derived rather than stipulated.

The revision delta nevertheless survives, demoted and better placed. A scientist cannot wait until it dies to learn which way to walk the relaxation lattice, so it needs a local signal, and \(d(\varphi_n, \varphi_{n+1})\) together with the confirm/refute counts of §21.4 is exactly one. This is the proximate/ultimate distinction of behavioral ecology: viability is the ultimate criterion, the revision delta is the proximate mechanism that tracks it, and neither is redundant.

21.17 Conclusion: an embodied architecture, and the effectiveness of reason

We close with a speculation that is not a design. Suppose a perceptual front end — a transformer network, wired to the world, doing what such networks are unreasonably good at — were used to encode sensory data as a population of computations and token stacks of the kind this note describes. We are not proposing an interface, and nothing below depends on how the encoding is done. We are asking what would follow if one existed, because we think the consequences bear on a question older than any of the machinery here.

21.17.1 A division of labor

The scientist of this note lives among other computations. The world does not present itself that way, or at least does not present itself that way to us, so something must do the type conversion. That is the front end’s whole job in the proposed arrangement: not to learn, but to render the environment as an environment of the right kind — loci with surfaces, stacks, and channels an interaction can address.

The division is worth stating sharply, because it inverts the usual allocation of labor:

Nothing in §§21.3–21.12 is a learning rule in the gradient sense, and nothing in a perceptual encoder is a hypothesis. Keeping them apart is what makes the arrangement legible.

21.17.2 Why continual learning falls to the ecology

The features that make an architecture of this shape suited to continual learning are not additions to it. They are the results already proved.

Revision is local.

A revision move is an operation at a locus in the formula tree (Definition 21.6), not a global adjustment of every parameter at once. Whatever is not at the locus is untouched, so the failure mode in which new learning overwrites old has no mechanism here. The corresponding cost is that a locus must be found, which is what §21.6 is about.

Exploration and exploitation are a budget, not a hyperparameter.

The allocation between foraging, inquiry, and reproduction (§21.12.9) is a division of one conserved stack, and the trade-off is settled by what the ecology pays for rather than by a coefficient chosen in advance.

A population is required, not preferred.

This is the sharpest of the three. Proposition 21.4 says an individual’s incremental revision cannot leave the ball it starts in — not slowly, but not at all. By Remark 21.19 the escape is recombination, which requires more than one learner and a lineage relating them. So an architecture built on this analysis does not have a population because populations are convenient; it has one because a single learner is provably confined.

Retirement has a criterion.

Mortality is not a scheduling policy imposed from outside. A model that cannot fund its next step stops, and the ledger says when.

21.17.3 The unreasonable effectiveness of reason

Wigner’s question is why mathematics, developed for reasons internal to itself, turns out to describe a world that did not consult it [47]. Hamming’s answer is largely a selection effect: we notice the mathematics that fits and forget the rest [48]. The evolutionary answer is that our faculties were shaped by the world they model. Neither is comfortable, the first because the successes are too systematic and too predictive, the second because the mathematics that works best — for spectra, for spacetime, for objects in no ancestral environment — is exactly the mathematics no ancestor was selected on.

The architecture posited here suggests the question has been asked across a gap that is not there.

Wigner’s framing has mathematics on one side as a free creation, the world on the other as a brute given, and an unexplained correspondence between them. In this setting there is no such separation to explain. The hypothesis language is not chosen and then found to fit: it is generated by OSLF from the presentation of the computational substrate itself, and adequacy is the theorem that it separates exactly what interaction separates — no coarser, so nothing observable is inexpressible, and no finer, so nothing expressible is unobservable. The mathematics available to such a learner is the mathematics of what it can distinguish, because it is the language the distinguishing relation generates.

That sentence has one load-bearing word in it, and until recently this chapter was not entitled to it. “Adequacy is the theorem” presupposes that the generated logic and the observational equivalence are calibrated against each other, and §21.15.3 used to say in so many words that they are not: that the structural layer strictly refines context-labeled bisimulation. An argument cannot rest on an equation a later section denies. Chapter 20 removes the denial rather than the argument — the two claims were about two different relations, and naming the index reconciles them (Remark 21.1) — so what follows now stands on something. It stands on less than it appears to, and the next remark says how much less.

Two things follow which we think are more than restatement.

First, the transfer problem dissolves in the right direction. Adequacy is not adequacy for a particular environment; it is adequacy for interaction as such, since context-labeled bisimulation is the finest equivalence any interacting learner could resolve, whatever it is interacting with. That is why the mathematics carries to domains no ancestor met. It was never fitted to the ancestral environment in the first place — it was fitted to the form of access, and the form of access did not change when the subject matter did.

Second, there is a boundary, and its location is predicted rather than assumed. Effectiveness should degrade precisely where the relation between learner and phenomenon is not one of interaction across a cut — where there is no context to build, no probe to install, no reply to condition on. That the domains which have historically resisted mathematization have tended to be of just this kind is not proof, but it is the right shape of evidence, and it is at least a claim that could fail.

Finally, the categories themselves are accounted for rather than presupposed. Corollary 21.3 forces hypotheses to be about namespaces, because destructive assay leaves an individual-level hypothesis worthless after its own first test, and Proposition 21.29 sustains those namespaces through reproduction. So the traffic in kinds, classes and universals that mathematics and science run on is not read off the world and not stipulated by the knower. It is what a mortal, interacting learner can afford to hold. That is a Kantian conclusion with a ledger attached, and the ledger is what makes it a claim rather than a posture. Chapter 68 takes the remark seriously and reads the whole of Kant’s transcendental project this way.

21.17.4 What is dictated, and what is not

If the mind is a computational apparatus, we think an architecture of roughly this shape is close to forced. Take the premises separately. If a mind is computational, it is ontologically isolated, so its environment is other computations and its access is interaction. If its access is interaction, its hypothesis language must be adequate for context-labeled bisimulation, on pain of being blind or wasteful. If hypotheses are to be tested rather than merely entertained, they must occupy guard position, so that a confirmation is a rendezvous. If the apparatus is finite, its budget bounds the fragment it can hold, and mortality supplies the motive without an external scorekeeper — of which none is available anyway, since a scorekeeper would be an oracle across the cut. And if revision is search on the resulting lattice, Proposition 21.4 requires a population.

Each of those is a derivation in this note rather than a design decision, which is what we mean by saying the shape is all but dictated. It should be equally clear what has not been shown. We have not shown that the mind is computational; that premise is doing all the work and we have assumed it throughout. We have not shown the world is computational either — only that a computational knower’s mathematics will fit whatever it can interact with, which is a claim about the knower and not about what is known. And we have not built the encoder, which is exactly the part where the difficulty of any actual system would be concentrated — a sentence that reads as a modest caveat and is not one. The encoder is where an ontology and a grounded hypothesis language, both of which this note has been spending freely, would have to be paid for. Chapter 56 audits that account.

What the note does claim is narrower and, we hope, more durable: that once learning is modeled as a metered dialogue between isolated computations, the apparatus of the scientific method stops looking like something a mind does and starts looking like something a mind, so situated, could hardly avoid.

21.18 Open problems

1. Inheritance of acquired characteristics. Remark 21.16 declines to read the inertness of a quotation as a germ–soma barrier, and Noble’s case against a strict barrier [51] is a reason for caution rather than a settled verdict. The formal question is sharp: a scientist’s hypotheses are held as data and are therefore quotable and transmissible, whereas its metabolic and sensory machinery is code. Under what pricing does a lineage transmit the first, and what does the resulting two-channel inheritance do to Proposition 21.16?

2. The plagiarism equilibrium. Reading a rival’s model is cheaper than experimenting and kills nobody, but holding a model requires exposing it as a guard. Under what pricing does honest science survive theory theft, and is there a mixed equilibrium?

3. Co-learning and non-stationarity. When the environment is mostly scientists, the target of every hypothesis is itself revising. Convergence becomes a fixed-point question rather than a learning-theoretic one, and this is where we expect the geometry/matter corecursion to be readable as learning.

4. Laxity. Conjecture 21.1: is the induced map on distributions lax, and is stochastic dominance the right order on the target?

5. The revision policy. We left the agent design open on purpose. Is there a policy on the relaxation lattice that is optimal in the ecological sense — maximizing expected residual life rather than predictive accuracy — and does it look anything like the policies learning theory recommends?

6. Is the mixed diet dominated? Propositions 21.13 and 21.14 are the two halves of a claim about trophic sparsity. Testing them in the ensemble of §21.13 — is the mixed access profile ever optimal for an atom, and does the composite dominate it where it is not — is the first computation we would run.

7. Sharpening the phase boundary. Figure 21.2 is drawn from the sign of a cost inequality. Making its boundaries quantitative for a specific pricing schedule, via the sampling construction of Chapter 13, is the first computation to do.

21.19 Appendix: A worked ecosystem

We give a small world containing two large external token stacks, three grazing non-scientist mortal computations, two reproductively isolated species of sexually reproducing scientist, and one asexual lineage. It is written in core rholang extended with the where clause of §21.4; stacks are integers and arithmetic is taken as given. The point of the listing is not that it is efficient but that every construction of the note appears in it as ordinary code, with no apparatus added.

21.19.1 Sources, metabolism, death

// A source (Def. "Source and prey"): rate-limited by being a single contract, non-exclusive
// since anyone may call it, non-adaptive, and FINITE -- conservation is exact.
def Source(s, q, reserve) = {
  for (ret <- s) {
    match reserve >= q {
      true  => { ret!(q) | Source(s, q, reserve - q) }
      false => { ret!(0) }                  // exhausted; cf. Prop. "Trophic succession"
    }
  }
}

// Reclaim - debit - reissue (Def. "Metabolic channel").  Between the receipt and the send the
// stack is a message in flight: metabolism is a race condition (Rem. "Metabolism is a race condition"), and
// the metabolic channel m is exactly the vulnerability of Thm. "Predators are non-image contexts".
def Live(m, burn, step) = {
  for (sigma <- m) {
    match *sigma > burn {
      true  => { m!(*sigma - burn) | *step | Live(m, burn, step) }
      false => { Nil }                      // death is failure to fund a step
    }
  }
}

21.19.2 A non-scientist

// A grazer holds no hypothesis and senses nothing beyond the one channel it
// was born knowing.  By Prop. "The two sciences" its environment is a source, so its science
// was over before it began -- and it pays no epistemic rent at all.
def Grazer(m, src) = {
  Live(m, 1, @{
    new probe in {
        src!(*probe)
      | for (g <- probe) { for (sigma <- m) { m!(*sigma + *g) } }
    }
  })
}

21.19.3 Foraging as an assay

// The complementary guard pair of Def. "Assay", verbatim.  The hypothesis is not
// consulted and then acted on; it sits in guard position, so confirmation IS
// the rendezvous.  If the source is exhausted neither guard fires: that is
// the third outcome of Prop. "Trichotomy", and it is not the same as refutation.
def Forage(m, sns, hyp) = {
  new probe in {
      sns!(*probe)                                 // the probe perturbs
    | for (g <- probe where g |= *hyp) {           // predicted
        for (sigma <- m) { m!(*sigma + *g) }
      }
    | for (g <- probe where ~(g |= *hyp)) {        // refuted
        Revise(hyp)                                // walk the lattice (Sec. 5)
      }
  }
}

21.19.4 Birth: a quotation, then a drop

// The offspring's code is CONSTRUCTED as a name and sits inert on `code'
// -- it does not reduce, does not age, burns nothing (Def. "Genome, phenotype").  Dropping it
// with * is what makes it a phenotype, and what starts the meter.
//
// The guard enforces Prop. "Minimum viable endowment": a parent may not endow below the outlay its
// child needs for a first successful harvest.
def Beget(m, alleles, prv, mate) = {
  for (sigma <- m where *sigma > *prv + floor) {
    m!(*sigma - *prv)                             // the endowment is subtracted
  | match *alleles {
      [s, h, p] => {
        new childM, code in {
            code!( @{ Body(childM, s, h, p, mate) } )      // the germ: inert
          | for (g <- code) { *g | childM!(*prv) }         // the soma: metered
        }
      }
    }
  }
}

21.19.5 Crossover

// Three homologous loci are exposed as parameters: the sensing allele, the
// hypothesis allele, and the provisioning allele.  This is assortment at
// three loci; crossover proper (Sec. "Crossover") splits a guard from its body
// inside an allele and carries the reading-frame condition of Def. "Homology".
def Pick(a, b, bit, ret) = {
  match *bit { 0 => { ret!(*a) } ; 1 => { ret!(*b) } }
}

def Crossover(mother, father, chi, ret) = {
  match (*mother, *father, *chi) {
    ([sm,hm,pm], [sf,hf,pf], [b1,b2,b3]) => {
      new r1, r2, r3 in {
          Pick(sm,sf,b1,r1) | Pick(hm,hf,b2,r2) | Pick(pm,pf,b3,r3)
        | for (s <- r1 & h <- r2 & p <- r3) { ret!([*s, *h, *p]) }
      }
    }
  }
}

21.19.6 A sexual scientist, and an asexual one

// Mating requires DISCLOSURE: the genome goes out on a public channel, where
// anything able to read it can.  Conspecific(s) is a name predicate (Sec. 3.1)
// holding of genomes whose sensing allele names the same source namespace as s.
// Nothing declares a species; reproductive isolation is namespace disjointness
// and nothing more (Prop. "Species is a namespace formula").
def Mate(m, alleles, sns, prv, mate) = {
  mate!(*alleles)
| for (partner <- mate where partner |= Conspecific(*sns)) {
    new chi, ret in {
        chi!( Coin3 )                        // the choice function of Def. "Crossover"
      | Crossover(alleles, partner, chi, ret)
      | for (kid <- ret) { Beget(m, kid, prv, mate) }
    }
  }
}

// The shared developmental program.  Alleles vary; this does not.
def Body(m, sns, hyp, prv, mate) = {
  Live(m, 2, @{                              // scientists burn faster than grazers
      Forage(m, sns, hyp)
    | Mate(m, @[*sns, *hyp, *prv], sns, prv, mate)
  })
}

// The asexual lineage: no partner, no alignment, NO DISCLOSURE, and no
// recombination to compute.  Strictly cheaper -- and by Prop. "Sex attacks the predator's hypothesis language" it presents
// a fixed namespace for anything that cares to learn it.
def Clone(m, sns, hyp, prv) = {
  Live(m, 2, @{
      Forage(m, sns, hyp)
    | Beget(m, @[*sns, *hyp, *prv], prv, @Nil)
  })
}

21.19.7 The world

new sun1, sun2, mate in {

  // Two large external stacks.  sun1 is shallow and gives small quanta;
  // sun2 lies deeper and gives larger ones -- the two order parameters
  // of Sec. 13.2 instantiated as two loci rather than as a distribution.
    Source(sun1, 1,  60000)
  | Source(sun2, 8, 240000)

  // Three non-scientists.
  | new g1, g2, g3 in {
        g1!(50) | Grazer(g1, sun1)
      | g2!(50) | Grazer(g2, sun1)
      | g3!(80) | Grazer(g3, sun2)
    }

  // Species A: hunts sun1, cheap hypothesis, many small offspring.
  | new a1, a2 in {
        a1!(400) | Body(a1, @sun1, @{ q => q > 0 }, @20, mate)
      | a2!(400) | Body(a2, @sun1, @{ q => q > 0 }, @20, mate)
    }

  // Species B: hunts sun2, a discriminating hypothesis, few large offspring.
  // A and B share the mating channel and never interbreed.
  | new b1, b2 in {
        b1!(900) | Body(b1, @sun2, @{ q => q > 4 }, @150, mate)
      | b2!(900) | Body(b2, @sun2, @{ q => q > 4 }, @150, mate)
    }

  // The asexual lineage, competing with species A for sun1.
  | new c1 in { c1!(500) | Clone(c1, @sun1, @{ q => q > 0 }, @60) }
}

21.19.8 What to watch

• Conservation is checkable by inspection. \(\Theta\) is the sum of the two reserves and the eight initial stacks. Nothing in the listing mints; Source only ever moves reserve into a consumer, and Beget only ever moves a parent’s stack into a child’s. Every debit taken by Live is a token leaving \(\sigma\) for \(\kappa\), so \(\sigma + \kappa = \Theta\) throughout.

• The three outcomes are structural, not bookkeeping. In Forage, refutation is an event because \(\lnot\varphi\) is a guard; the case where the source is exhausted produces neither event. Proposition 21.1 is visible as three code paths, one of which is empty.

• Nothing screens. No embryo is model-checked anywhere in the listing. A parent could in principle check its recombinant against the viability formula before dropping it, and does not, because the check is priced against the same stack — Proposition 21.27 as a programming decision rather than an argument.

• Speciation costs one guard. \(A\) and \(B\) share mate and are nevertheless isolated. Delete the where clause from Mate and they hybridize; the offspring inherits a sensing allele naming one source and a hypothesis allele tuned to the other, and starves without ever being refuted. That is the reading-frame condition failing at the level of the ecology.

• The race is real. Live and Forage both take from m and put back. They conserve the stack and cannot deadlock, but they interleave — which is Remark 21.11’s point that living means racing others to your own stack, with the others here being your own body.

• This world is at the autotrophic end. Every organism feeds on a source, so by Proposition 21.12 every lineage’s inquiry terminates and none pays an unbounded epistemic rent. Note how close the other end is: species \(B\)’s sensing allele already names a deeper namespace, and a single reallocation of it toward a grazer’s m converts the lineage to heterotrophy, with Corollary 21.5’s standing cost arriving in the same step.

• And it ends. When both reserves reach zero, Source returns \(0\) forever, no guard on a positive quantum fires again, and the remaining tokens sit in metabolic channels. From there the only credit available is another organism’s stack. Proposition 21.15 is not an asymptotic remark about this listing; it is the last thing that happens in it.