The Spine and the Web: Reality-Mapping for Embedded Agents
The Spine and the Web
Reality-mapping for embedded agents, and where it stands to the work that came before
Position note · the architecture of the program, its precedents named, and the one seam left open
Version 1 · June 2026
1. What the architecture is for
The object is a map of reality that an embedded agent can correct. A map is any structure that represents, predicts, or guides action; the territory is whatever the map answers to; the enemy is divergence between them. The full treatment is in The Fidelity Program and is not repeated here. What this note adds is the architecture's shape, and the shape is fixed by three ways the task fails.
The first failure is floating abstraction: conceptual structure built without anchoring to physics, biology, cognition, or evidence. It can be elaborate and internally consistent and still touch nothing. The second is shallow empiricism: facts accumulated without an architecture that says what they mean, how they relate, or how they should constrain action. The third is the one this note is mainly about, because it is the subtlest and the easiest to commit while avoiding the other two: rigid bottom-up reductionism, the picture in which explanation runs only upward — physics fixes chemistry fixes biology fixes cognition fixes society — so that the lowest level is always the real one and the higher levels are bookkeeping.
The architecture answers each. Against floating abstraction, a hard physical spine. Against shallow empiricism, an explicit structure that grades and connects what it holds. Against rigid reductionism, a web rather than a ladder: levels that constrain one another in more than one direction, under a discipline that keeps the whole correctable. The next section states the spine, the web, and the discipline that holds them.
2. The architecture
The spine
The spine carries only the commitments hard enough to bear the weight of everything built on them. They are few. Reality constrains the map: a map nothing can correct is not tracking anything. Maps are physically realized: every cognition, model, sentence, or theory is a physical process, with no exemption for meaning, value, or mind. Agents are bounded and embedded: an agent acts from inside the world it models, through a boundary it maintains, never from nowhere. Information processing has a physical cost: storing, computing, and erasing are not free (Landauer). Persisting systems are multilevel: boundaries sit within boundaries across scales, and the levels interact. Moral weight sits with conscious experience: not with patterns, persistence, or complexity as such, but with what it is like to be a system that can suffer or flourish. And correction requires contact with constraint: a belief sealed from the territory cannot be corrected by it.
These are the load-bearing beams. Most of them are not this program's; they are marked to their sources in §3. The spine's job is to be solid, not original.
From ladder to web
Above the spine the world is not a ladder. The lower level is the substrate of the higher, but it is not always its best explanation, and it is not the only direction causation runs. Denis Noble's principle of biological relativity states the point at the level of biology: a priori, there is no privileged level of causation. Genes constrain organisms; organismal and environmental states — stress, behaviour, context — constrain gene expression in turn. Noble represents the downward direction concretely: higher-level organization sets the initial and boundary conditions on the equations that govern the lower level. He prefers the word scale to level, on the ground that “higher” and “lower” are metaphors — a discipline this program shares and returns to below.
Asserting downward causation incurs a debt at once, and it must be paid here rather than deferred, because the move is load-bearing. The debt is Jaegwon Kim's causal exclusion argument. If every physical effect already has a sufficient physical cause at the lowest level, then a higher-level cause appears to have no work left to do: either the effect is overdetermined by two sufficient causes, or physical closure is violated. The argument is not confined to mind and body; it applies to any hierarchy of levels, including the hierarchy of the sciences, so it bears on this picture directly.
The program's reply is Noble's, and its limits are stated with it. Downward causation enters as boundary and initial conditions, not as a second efficient cause competing for the same work at the same level. A boundary condition selects among the solutions the lower-level dynamics already admit; it constrains which lower-level history is realized, rather than adding a push alongside the lower-level forces. There is a real case that constraint of this kind sidesteps exclusion, precisely because a constraint is not a rival cause. There is also a real reply from Kim's side: constraints are themselves physically realized, so the exclusion worry relocates to the physical realizers of the constraints rather than dissolving. That dispute is open. The program therefore does not claim downward causation is established. It adopts the boundary-condition reading, names exclusion as the standing objection, and marks the question unsettled — which is the method of the corpus applied to the corpus's own move: find the wall, state the reply, claim no more than the reply secures.
The dependency classes
The grading that keeps the structure correctable is a set of dependency classes. The spine is not one undifferentiated mass; its commitments earn different degrees of trust, and what sits above them is replaceable in proportion to how speculative it is.
| Class | Contents | How it is held |
|---|---|---|
| Tier 0 — hinge constraints | Reality constrains the map; contradiction at one index fails; a map can be more or less accurate; correction requires contact. | Presupposed by inquiry; not up for ordinary revision. |
| Tier 1 — high-trust science | Physicalism as the maximally inductively justified standing position; non-equilibrium statistical mechanics; information as physical; evolution; neuroscience as constraint. | Secure; revisable only under extraordinary evidence. |
| Tier 2 — mature cross-level models | Biological relativity; multiscale causation; embodied cognition; organism–environment coupling. | Well-supported; expected to be refined, not discarded. |
| Tier 3 — active frontier models | Free-energy and active-inference interpretations; specific cosmological theories; AI-alignment formalisms; consciousness models. | Held provisionally; may fail without taking the spine with them. |
| Tier 4 — the synthesis layer | The program's own constructions: fidelity as the standard, map–territory divergence, the indexed ought, agency as bounded persistence. | Where the distinctive bets live; carried at the confidence of their cruxes, marked in the apex. |
The point of the classes is the negative one. When a Tier 3 model fails — and frontier models fail routinely — the failure stays local. The spine does not move; the program is not refuted; one module is demoted and, where possible, replaced. The whole is built so that being wrong in a part is survivable, and the structure that makes it survivable is the one Lakatos described for research programmes, named in §3.
What the levels demand of the words
A web of levels imposes a discipline on language, because the same word rarely means the same thing across scales. “Information” in statistical mechanics, in genetics, in cognition, and in speech is not one operational quantity with one referent; “agent,” “selection,” “stress,” “boundary,” and “cost” all shift with the level they are used at. Left unindexed, such a word slides between levels and carries an argument with it that was never made — the failure the corpus elsewhere calls a stripped relatum.
Five distinctions keep the levels from blurring. Substrate dependence: everything is physically realized, without exception. Explanatory level: the scale at which a structure is best captured, which is not always the lowest. Causal coupling: the way levels constrain one another, in both directions, as boundary conditions and as aggregated dynamics. Semantic indexing: the rule that a term must disclose the level and the relation it is operating at, or be treated as unanchored. Correction protocol: the procedure by which a failed module is demoted without damage to the spine. The first two separate “made of” from “best explained by,” and refuse to collapse the second into the first. The third is where Kim and Noble live. The fourth is name-the-relata applied to scale. The fifth is what makes the structure anti-fragile rather than merely robust.
Testing by claim-type
The discipline that anchors the program against floating abstraction is sometimes summarized as “attach a falsifier to every claim.” Taken literally that is wrong, and wrong in a way worth stating, because it would dissolve logic, mathematics, and definitions along with the genuine targets. Not every claim is the kind of thing an experiment can refute. The repair is to test each claim by its type.
A synthetic, model-dependent claim owes falsifiers: predictions, measurements, cases where it says “this is outside my scope.” An analytic or definitional claim is not refuted by observation; it is tested by clarity, by whether its terms refer, by whether it compresses without confusing — a definition can be useless, circular, or too broad, but it is not false the way a prediction is. A foundational commitment is tested differently again: by indispensability, by whether it can be denied without being used in the denial, by whether every rival quietly depends on it, by whether it survives recursive self-audit. This is not a loophole that exempts the program's favourite claims from scrutiny. It is the recognition that a tautology and a measurement fail in different ways, and that demanding empirical falsification of a tautology is itself a category error. The lineage of this correction — the analytic/synthetic distinction, and the fact that the original demarcation criterion was always about empirical science specifically — is named in §3.
Anti-fragility, stated precisely
It is tempting to call the architecture anti-fragile and leave the word to do the work. The word is only earned under a condition. A structure is anti-fragile when it gains from correction, not merely when it survives it, and it gains from correction only if a failed claim is not an identity threat to the one who holds it. That is the hard part, and it is psychological before it is logical: the structure has to be built so that demoting a module, retiring a label, or conceding a point is the system working, not the system breaking. A program that defends its claims because they are its claims has already failed, whatever its content.
A frontier model can fail and the spine remains. A metaphor can overreach and the semantic discipline corrects it. A label can mislead and be retired — as “Thermodynamic Realism” was, and as the Consistency Tax was. A better model can arrive and be absorbed. What is protected is not the furniture. It is the machine that lets the furniture be replaced.
3. Where it stands to the work that came before
Almost every load-bearing move above has a named precedent. Setting them out is not a concession wrung from the program; it is the program's own first rule — locate the prior art before building on the ground. The map below is the result of doing that, and several of the entries are tight enough to count as re-derivations rather than echoes.
| Move in the architecture | Precedent | Scope of the match |
|---|---|---|
| Spine plus replaceable modules; failures stay local. | Lakatos's hard core and protective belt, with the negative heuristic protecting the core and the positive heuristic developing the belt (1970). | Near one-to-one. The dependency classes are the core/belt structure operationalized. |
| The physical spine; never let a higher layer override fundamental physics. | Ladyman & Ross's Primacy of Physics Constraint, and their anti-domestication naturalism (2007). | Tight on the primacy of physics and the refusal of a-priori metaphysics. See the qualification below. |
| Tier-0 hinge constraints. | Wittgenstein's hinges — propositions “exempt from doubt, as it were like hinges on which those turn” — and hinge epistemology (Coliva, Moyal-Sharrock, Pritchard). | The term and the program are the same. “Cannot be denied without being used” is the older retortion argument, from Aristotle on non-contradiction to Apel. |
| Multilevel causation; downward causation as boundary conditions. | Noble's biological relativity: no privileged level of causation (2012). | Cited as source, not claimed. The exclusion debt is Kim (1998). |
| Semantic indexing by level; terms must disclose their relata. | Conceptual engineering (Cappelen) and the symbol-grounding problem (Harnad). | The motivation — unanchored terms float — is the standard one for those fields. |
| Testing by claim-type. | The analytic/synthetic distinction (Carnap; Quine's critique), and Popper's demarcation, which was always about empirical science specifically. | The “refinement” is the correction to naive falsificationism the originals did not need. |
| No view from nowhere. | Nagel (1986); Putnam's “God's-eye view.” | Title and concept. |
One entry needs a qualification, because the match is easy to overstate and overstating it would be the same failure the corpus exists to refuse. Ladyman & Ross is the nearest twin for the naturalist stance and the primacy of physics, and on those it is very close. But their Principle of Naturalistic Closure — that a new metaphysical claim earns its place only by showing how two or more specific scientific hypotheses, one from fundamental physics, jointly explain more than separately — is a specific unification-demarcation criterion, and it is not this program's criterion, which is fidelity and correctability. Their project's aim differs too: theirs is metaphysics of science, an account of what the world is made of given physics; this is closer to an epistemics of mapping with an ethics attached. The honest statement is therefore narrow: near-twin for the stance, different in admissibility criterion and in aim. And the narrowness opens no room for novelty, because the place this program differs from Ladyman & Ross — fidelity under correction as the standard — is itself old, the fallibilist correspondence tradition of Peirce and Popper, on which Ladyman & Ross themselves draw.
Several of these precedents I reached before I had read them. That fact is worth recording, and worth reading correctly. Independent arrival at a structure others already occupy is evidence — of soundness, not of priority. A search process that lands on the same attractor a different process found, without contact between them, is being pulled by something real in the conceptual space rather than wandering at random. That is corroboration: two instruments reading the same value. It is not a claim to have got there first, because convergence is exactly what many independent searchers would produce on real ground. The corpus judges a contribution on its content and not its source, and that rule cuts both ways: the ideas lose no value for having been found elsewhere, and independent rediscovery earns no priority it would otherwise lack. What independent arrival earns is the modest, real thing — evidence that the ground is solid.
4. The one open seam
After the precedents are subtracted, one thing in the architecture is genuinely under-occupied, and it is not the spine-and-web structure, which is largely Lakatos plus naturalized metaphysics plus multilevel causation. It is a question at the seam between the epistemology and the physics: whether semantic and epistemic fidelity can be given a physically grounded cost — whether the accuracy of a map bottoms out, in the end, in thermodynamic constraint on which maps a physical system can build and hold.
The seam splits into a part that is settled and a part that is not. The settled part is real and is not this program's: for a system driven by a stochastic environment, the nonpredictive information its state retains lower-bounds the work it must dissipate — model inefficiency is thermodynamic inefficiency (Still, Sivak, Bell, and Crooks, 2012). That is a specific result for a specific setup, and it refines Landauer's bound. The unsettled part is the generalization: from that bound to a selection principle over maps in general; from “a stochastically driven system pays for nonpredictive information” to “fidelity in general is under thermodynamic selection.” The apex states the same thing in its own register and fences it to the frontier: selection toward fidelity is a tendency, not a law.
What blocks the generalization is worth naming exactly, because it is not the obstacle it looks like. The ground is open — the coarse-graining problem at the heart of the causal-emergence literature is stated by its own authors as open, and the predictive-information thread is under-walked. The obstacle is not that the territory is occupied. It is certification: turning the extrapolation from the stochastic-driving result into a checkable general bound is hard stochastic thermodynamics, a competence I do not yet have and cannot fake. So the seam is a target with a debt attached, not a result in hand. The nearest published neighbour on the metaethical side of the bridge — the Replicator-Optimization Mechanism — shares the instrumental, scale-relative structure and proves a coarse-graining theorem the program lacks, and the one place the program's version goes where it does not is exactly this predictive-information thread. That is the residual, stated as precisely as I can state it: not a discovery, a well-posed problem with its certification debt marked, set down for whoever holds the toolkit.
5. The corpus as demonstration
If the architecture is not a novel result, what is the corpus for? It is the worked exemplar of the method — and a worked exemplar is a recognized kind of contribution, not a consolation. Lakatos's Proofs and Refutations proves no new theorem; it reconstructs the history of one conjecture as a dialogue, and its lasting value is the method made concrete on a hard case, including the method turning on its own provisional concepts under counterexample. The corpus is the analogue. The method — physicalism as the floor, name-the-relata as the diagnostic, fidelity as the standard, correction as the law — is not described in the abstract but run: on the is–ought problem, on personal identity under copying and death, on the boundary of what consciousness-talk can claim, on the cosmological endgame.
And, once, on itself. On 18 June 2026 the method was turned on the corpus, and a literature audit found that six of its load-bearing pieces substantially reinvented existing work; the response was to retire the ambition to priority, demote the overclaims, and record the result. That self-audit is the load-bearing demonstration, more than any single first-order claim, because it is the rare thing the method asserts is possible actually happening: a program demoting its own vocabulary in public, with a timestamp, because the territory said so. A structure that does that is exhibiting the discipline it argues for, which is a different and harder thing than arguing for it.
A note on ambition, since it is the thing most easily mistaken here. A theory of everything is not the goal, and treating it as one corrupts the work. If a wide unification emerges, it emerges as a byproduct of mapping the territory honestly and applying the method well — not as a target aimed at. The full map is in any case too large for one agent to finish; it is immense, multilevel, and growing faster than any individual can track. So the goal shifts, from finishing the map to building the architecture the map can grow inside without losing its spine, its semantics, or its capacity for correction. That is the deeper and the more achievable target, and it is what this note has described.
Compressed: reality exists; agents are inside it; they act through maps; maps diverge from the territory; the divergence has consequences, sometimes including avoidable suffering; so the task is to build map-making systems that reality can correct. The architecture is the standing form of that task. It is not a theory. It is a discipline for becoming less wrong, built so that being wrong in its parts is how it improves.
Standing of this document. A position note, and its parts carry different weight. The architecture (§§1–2) is a synthesis of established tools, placed against its precedents at each joint (§3) and claimed as assembly and demonstration, not as discovery. The treatment of levels engages its standing objection (Kim) and marks the downward-causation question open rather than settled. The one seam (§4) is genuinely under-occupied, and is offered as a well-posed problem with its certification debt named, not as a result. Independent convergence with the prior art is recorded as corroboration of soundness, never as priority. The corpus is the worked exemplar, with the 18 June self-audit as the demonstration that the method runs on itself. On any conflict with the apex synthesis or the modules, they govern. Corrections and counterexamples are welcome and change the document.
References
- Lakatos, I. (1970). Falsification and the Methodology of Scientific Research Programmes. In I. Lakatos & A. Musgrave (eds.), Criticism and the Growth of Knowledge. Cambridge University Press. The hard core, the protective belt, and the negative and positive heuristics; the dependency classes of §2 operationalize this structure. Standard
- Lakatos, I. (1976). Proofs and Refutations: The Logic of Mathematical Discovery. Cambridge University Press. The worked methodology whose contribution is the method enacted on a hard case, not a new theorem; the model for §5. Standard
- Ladyman, J., & Ross, D. (2007). Every Thing Must Go: Metaphysics Naturalized. Oxford University Press. The Primacy of Physics Constraint and the Principle of Naturalistic Closure; the near-twin for the naturalist stance, with the qualification stated in §3. Standard
- Wittgenstein, L. (1969). On Certainty. Blackwell. §341, hinge propositions “exempt from doubt”; the Tier-0 constraints of §2. Standard
- Coliva, A., & Moyal-Sharrock, D. (eds.) (2016). Hinge Epistemology. Brill; and Pritchard, D. (2015). Epistemic Angst. Princeton University Press. The development of hinge epistemology drawn on in §§2–3. Standard
- Noble, D. (2012). A Theory of Biological Relativity: No Privileged Level of Causation. Interface Focus 2(1), 55–64. No privileged level; downward causation represented as initial and boundary conditions; §2. Standard
- Kim, J. (1998). Mind in a Physical World. MIT Press. The causal exclusion argument; the standing objection engaged in §2. Standard
- Carnap, R. (1950). Empiricism, Semantics, and Ontology. Revue Internationale de Philosophie 4, 20–40; and Quine, W. V. O. (1951). Two Dogmas of Empiricism. The Philosophical Review 60, 20–43. Internal/external questions and the analytic/synthetic distinction; testing by claim-type, §2. Standard
- Popper, K. (1959). The Logic of Scientific Discovery. Hutchinson. Demarcation as a criterion for empirical science specifically, not a demand that every claim be empirically falsifiable; §2. Standard
- Nagel, T. (1986). The View from Nowhere. Oxford University Press. No view from nowhere; a spine commitment in §2. Standard
- Harnad, S. (1990). The Symbol Grounding Problem. Physica D 42, 335–346; and Cappelen, H. (2018). Fixing Language. Oxford University Press. Grounding and conceptual engineering; the semantic-indexing discipline of §2. Standard
- Landauer, R. (1961). Irreversibility and Heat Generation in the Computing Process. IBM Journal of Research and Development 5, 183–191. Information processing has a minimum physical cost; a spine commitment. Standard
- Still, S., Sivak, D. A., Bell, A. J., & Crooks, G. E. (2012). Thermodynamics of Prediction. Physical Review Letters 109, 120604. Nonpredictive retained information lower-bounds dissipated work; the settled half of the seam in §4. Standard
- Farzulla, M. (2026). The Replicator-Optimization Mechanism. arXiv:2601.06363. The near neighbour on the metaethical side of the bridge, with a coarse-graining theorem the program lacks; the predictive-information thread of §4 is where the program's version diverges. Verified · June 2026
- The Fidelity Program (Epistemic Forge). The full map–territory–fidelity treatment that §1 compresses. Internal
- Epistemic Forge: The Apex Synthesis. The tiered state of the corpus, and the frontier fencing of the fidelity-selection reading relied on in §4. Internal
- Name the Relata (Epistemic Forge). The diagnostic that semantic indexing applies to scale. Internal
- I Spent Months Reinventing the Wheel. Today I Found Out. (Epistemic Forge). The 18 June literature audit and self-correction recorded in §5. Internal
Comments
Post a Comment