The Structural Intelligence Conjecture and the Structural Observatory

Date: 2026-08-03. Status: research program + three working instruments. The conjecture and the extended formalism (§2–§4) are speculative theory, not established results; they are held to a strict mathematical-claim gate and are not promoted to theorems. The three instruments in §5 are exact, deterministic, and tested — they establish dissociations on solvable toy cases, not the general claims.

Companion (formal, four theorems + six instruments): SIC_MATHEMATICAL_FOUNDATIONS.md — the version of this program that turned three of the extensions below into theorems (existence via minimal sufficiency; cross-task stability under a shared Markov screen; discrete learnability via common-sufficient clustering) and added three more instruments. Companion (plain-English): STRUCTURAL_INTELLIGENCE.md — the same object described with the Trace AI lens.

Both companion docs exist in this repo for a reason: this note is the working version — cautious, extension-labeled, non-theorem — and the formal foundations doc is what it grew into under the mathematical-claim gate. Keeping both prevents the honesty of this document from being retroactively rewritten into the confidence of the other.

Editorial note: this document references source works (AGENTS.md mathematical-claim gate, notes/latent_structures_meta_framework.md, the repo’s symbolic_weakness result, experiments/common/causal_use.py, experiments/world_responds, results/) that live in the paper’s source repository — jawauntb/research-derived-experiments — rather than in the Trace AI repo. Their function here is provenance — the paths are preserved so a reader can locate the underlying material at that repository — not a Trace-AI-side path claim. The originating research note lives at notes/structural_intelligence_conjecture.md.

1. The conjecture

Structural Intelligence Conjecture. A finite adaptive system’s central capability is not predicting states or choosing actions but discovering a quotient $q : X \to Z$ such that (1) the task-relevant dynamics descend to $Z$, (2) irrelevant variation is confined to the fibres $q^{-1}(z)$, (3) useful interventions can be specified compactly in $Z$, (4) those specifications can be re-instantiated through a compiler, and (5) their consequences remain stable across substrates and contexts.

In one line: intelligence finds the level at which the world becomes both compressible and controllable. This is the operational form of the Manifest-Invariant Principle: find the coordinate in which the hidden invariant is a coordinate axis.

2. The master object: a stochastic fibration with a compiler

The common object across the corpus is not a latent space $Z$ by itself. It is a stochastic fibration together with an abstraction–realization pair:

$$X \xrightarrow{q} Z, \qquad Z \xrightarrow{K} X, \qquad \operatorname{supp} K(\cdot \mid z) \subseteq q^{-1}(z).$$

• $X$ — the space of concrete realizations;
• $Z$ — the space of structures / functions / observables;
• $q$ — says which concrete differences do not matter (coarse-graining);
• $q^{-1}(z)$ — the fibre of possible embodiments of structure $z$;
• $K(dx \mid z)$ — a compiler: a distribution over realizations inside the fibre.

This generalizes the meta-framework’s $R \dashv C$. There, realize $R$ and coarse-grain $C$ were (near-)deterministic functors. Here $q = C$ and the compiler $K$ is a stochastic section of $R$: it does not pick one realization but a distribution over the fibre. The biology paper (Kiiskinen–Kivinen–Rivas) is exactly this: the genome names $z$, the physics substrate is the compiler $K$ that “computes the samples,” and “selection acts on the statistics of the resulting ensemble.” Its threshold theorem is the statement that below a critical coarse-graining $C^*$ the fibre $q^{-1}(z)$ is too large to address with the specification budget — the fibre is real and irreducible.

The stochastic-fibration formulation is the director’s synthesis; no single source states it in this form. Category theory (source A) supplies the compositional implementation/verification/compilation maps; Wigner (D) supplies invariance under irrelevant conditions plus the warning that theories may be local and non-unique; the structural-realism thread (F) says structures become causally effective only through embodiment and approach truth when preserved relations survive independently-defined contexts.

A note on rigor (the director’s correction). Cross-domain resemblance is not isomorphism. The honest hierarchy of “sameness,” strongest to weakest: isomorphism $\supset$ bisimulation $\supset$ functor $\supset$ natural transformation $\supset$ adjunction / Galois connection $\supset$ Morita-like equivalence $\supset$ simulation-at-a-resolution. Most relations among these works are adjunctions, simulations, and shared diagram shapes, not object-level isomorphisms. Every proposed connection must answer: exactly what kind of sameness is this, what does the map forget, and what would have to be proved to make the analogy a theorem?

3. Concern as a geometry on the fibre

A compiler gives a baseline distribution over embodiments $K(dx \mid z)$. A concern state $c$ need not change which realizations are possible; it changes which are salient, viable, reachable. Model it as a reweighting:

$$K_c(dx \mid z) = \frac{e^{\beta U_c(x, z)} K(dx \mid z)}{\int e^{\beta U_c(x', z)} K(dx' \mid z)}$$

where $U_c$ is the concern field — it makes some regions of the fibre more consequential. This induces an information geometry on concern states, e.g. the Fisher metric $g_{ij}(c) = \mathbb{E}_{K_c}[\partial_i \log K_c \cdot \partial_j \log K_c]$, measuring how distinguishable nearby concern configurations are by their effect on realized possibilities.

Precisely:

• meaning is a deformation of relevance over possibilities, not a representation;
• concern is a field/measure that changes the effective fibre geometry;
• agency is the capacity to alter the kernel $K$ or the field $U_c$ so viable regions become reachable;
• selfhood concerns which changes to the kernel are attributed to the system’s own interventions.

This is the exact continuation of the source repository (jawauntb/research-derived-experiments)’s Geometry of Concern and Gauge-Fixed Transport of Concern: concern transport between contexts becomes transport between kernels over corresponding fibres, and holonomy measures the failure of transported concern to return unchanged around a loop of contexts. It is now fittable from interventions and trajectories, not merely a metaphor.

4. Ten constructs the master object generates

A research program, ranked by fertility, practical use, and beauty. Each is stated as a target, with its status and its relation to the instruments in §5.

1. Concern geometry (§3) — fit $K_c$, its Fisher metric, and concern holonomy from interventions. Fertile; theory + fit.
2. Conditional rate–distortion control limit. Robust control of $Y = f(X)$ from a coarse spec $Z = q(X)$ requires $H(Y \mid Z) \approx 0$, or, under tolerated distortion $D$, extra addressed bits $B_{\text{extra}} \ge R_{Y \mid Z}(D)$. Why micromanagement fails: the controller tries to distinguish outcomes its channel cannot address. This generalizes the biology theorem; conjectural, not yet proved.
3. Abstraction frontier. Replace “find the best $Z$” with the Pareto frontier of representations trading task-sufficiency $I(Y; X \mid q(X))$, dynamical closure $I(Z_{t+1}; X_t \mid Z_t, A_t)$, cost $H_0(Z)$, and control regret. Explains why two representations can both be “right” yet incomparable (an antichain, as in the biology paper’s threshold). Seeded by Instrument 1.
4. Fibre audit (adversarial). Instead of testing held-out prediction, vary the allegedly irrelevant degrees of freedom while holding $q$ fixed and measure the interventional discrepancy $\Delta_q(z) = \sup_{x, x' \in q^{-1}(z)} d(P(Y \mid \operatorname{do} x), P(Y \mid \operatorname{do} x'))$. Large $\Delta_q$ $\Rightarrow$ the abstraction collapsed a causally important distinction. The operational form of Wigner’s relevant/irrelevant split and F’s truth criterion. Seeded (non-interventional core) by Instrument 1; the interventional version is next.
5. Theory atlas (sheaf/stack gluing). Treat theories as local charts $M_i$ on contexts $U_i$ with translations $T_{ij}$; test the cocycle $T_{jk} \circ T_{ij} = T_{ik}$. Where it holds, the charts glue; where it fails, the obstruction is informative (missing latent, scale transition, phase boundary, category error). Wigner’s non-unification worry, made constructive.
6. Compiler tomography. Given many $(s_i, x_i)$ with $x_i \sim K(\cdot \mid s_i)$, infer the shared compiler and compact specs by MDL: $\min_{K, \{s_i\}} \big[L(K) + \sum L(s_i) - \sum \log p_K(x_i \mid s_i)\big]$. Variation across $i$ is specification; shared regularity is the compiler; residual is unresolved state/randomness. Then compiler ecology: $K_{t+1} = U(K_t, \text{outcomes})$ — build a compiler under which good outcomes are cheap, rather than re-specifying good outcomes. Formal language for education, institutions, long-horizon training.
7. Causal semantics. Two symbols are equivalent when they induce naturally equivalent update operators $\Psi_{m, c}$ across independent contexts. Groups sentences by how they reorganize reachable possibility, not textual co-occurrence. An operational meaning layer ordinary embeddings omit; extends Instrument 3.
8. Representation-repair calculus. A library of failure signatures $\to$ minimal structural lifts (scalar loses multiplicity $\to$ lift to operator/PSD; global norm loses location $\to$ localize to a measure/sheaf; quotient hides degeneracy $\to$ restore the fibre/stabilizer; affine mishandles infinity $\to$ projectivize; exact target exceeds capacity $\to$ move to ensemble control; symmetry breaks identifiability $\to$ gauge-fix / moduli). Turns “try another approach” into diagnose the lost invariant, then apply the minimal lift.
9. Alignment as ensemble governance (not trajectory scripting). A finite spec cannot address every fine trajectory of a long-horizon agent; target a viable region $V \subseteq Z$ with $\Pr[q(X_t) \in V \forall t] \ge 1 - \delta$ under a broad family of unresolved compiler/environment states. A fibre audit becomes the alignment evaluation. Conjectural generalization, flagged as such. Extends Instrument 3 and the source repo’s causally-grounded-agents line.
10. Autocatalytic artwork. A work $S_t \xrightarrow{K_t} E_t \xrightarrow{\text{experience}} K_{t+1}$ whose early movements teach the grammar by which later movements become legible — an autocatalytic symbolic structure that produces part of the machinery required for its own fuller instantiation, across harmony / shader / navigation / language / social interaction. The literal reading of F’s closing proposition; extends Instrument 2.

5. The three built instruments

Each is exact, deterministic, tested, and public-safe. They are the first three modules of the Observatory and establish the dissociations the program rests on.

Instrument 1 — experiments/representation_search (Fiber Finder)

Over a Boolean world with a known invariant, it enumerates a lattice of quotients and three selectors. Result: only minimal_sufficient (sufficient, then minimal description length) recovers the ground-truth invariant; mdl_only collapses the obstruction; accuracy_only never compresses. Establishes that sufficiency, description length, and accuracy dissociate — the discrete, non-interventional core of the fibre audit (extension 4) and the seed of the abstraction frontier (extension 3). Connects to the source repository (jawauntb/research-derived-experiments)’s symbolic_weakness result: weakness is a counit/fibre quantity, MDL is a spec-length quantity; they are not the same, and the fibre quantity is the one that governs generalization.

Instrument 2 — experiments/structure_compiler (one invariant, many embodiments)

An abstract automaton with accumulation $\to$ phase-transition $\to$ hysteresis is compiled into music, a visual field, text, and spatial navigation; each medium’s readback recovers the same abstract trajectory (verified structural identity, fidelity 1.0). Establishes cross-substrate structural identity as a checkable property ($q_i \circ F_i = \text{id}$), not mood matching — the ensemble-compiler and Gesamtkunstwerk idea, and the base case for the autocatalytic work (extension 10).

Instrument 3 — experiments/symbolic_causation (agency science)

An exact finite-state world treats a symbolic model $m$ as an operation on the future-trajectory distribution and separates signal ($\Delta_{\text{KL}}$), control (goal_gain), knowledge (predictive_accuracy), and agency (control + calibrated self-attribution + transfer). Seven conditions each realize a distinct metric signature; a false_credit condition improves the outcome while its true do-effect is zero and its self-attribution is miscalibrated; a brittle controller controls but does not transfer. Establishes that no single scalar (“behavioural influence”) identifies agency — the measurement core for concern geometry (extension 1), causal semantics (7), and ensemble-governance alignment (9). Connects to experiments/common/causal_use.py and experiments/world_responds.

Cross-reference to Trace AI. Instrument 3 is the exact-solvable formalization of what TRB Track 3 (self-attribution) measures on trained models. The false_credit condition is the ground truth for Track 3’s balanced-accuracy score; see BENCHMARK_SPEC.md §3 Track 3.

6. Capstone application: conscious / reliable agents (the honest version)

The program bears directly on the two questions above — can we make agents conscious, and can we make them never wrong. The disciplined answer is not in the literal sense of either, but the framework gives a precise, non-inflated construction target. It factors as two nested systems:

$$\text{Concerned Self-Modeling Core} \;\subset\; \text{Proof-Carrying Reliability Shell}.$$

Inner core = an active, self-maintaining $T$-algebra. A persistent agent that maintains a world model and self-model, represents concern-weighted futures (§3), globally broadcasts selected information, remembers commitments, predicts the consequences of its actions, performs false-credit tests on its own causal claims (exactly Instrument 3’s calibration metric), and reports uncertainty and internal conflict. In the framework this is precisely the passive $\to$ active threshold: a $T$-algebra that runs $K$ and $q$ in a closed loop and optimizes its own counit gap online. It operationalizes proposed consciousness indicators (global availability, metacognition, self-modeling, agency) — and that is the ceiling of the claim. Functional selfhood $\not\Rightarrow$ subjective consciousness: satisfying architectural/behavioural indicators does not establish that there is something it is like to be the agent, and there is a real welfare/ethics wrinkle in building systems that strongly satisfy such indicators.

Outer shell = the fibration with a verifier on the counit. A system that converts goals into contracts (compiler tomography / spec compilation), separates observation from inference, attaches provenance to every claim, and commits an action only under machine-checkable evidence: $\operatorname{execute}(a) \iff V(s, a, \varphi) = \operatorname{PASS}$, else abstain / ask / simulate / escalate. Its correctness envelope is a fibre audit (extension 4): vary everything the spec claims is irrelevant and check the property survives. Verified-in-a-bounded-domain $\not\Rightarrow$ universal infallibility: verification certifies the formal statement, not that the statement captures what a human wanted, and reliability also depends on model, harness, tools, environment, and budget.

So the honest breakthrough available now is not conscious, perfect agents. It is agents with experimentally measurable selfhood and formally bounded error — functional selfhood ⇏ consciousness, and verified bounded behavior ⇏ infallibility, kept as two separate, non-inflated claims.

7. The umbrella: a Structural Observatory

One system that, given a problem, theory, agent, organism, or artwork, returns: its candidate quotient maps, realization fibres, automorphisms, abstraction frontier, fibre-audit results, inferred compiler, and possible cross-substrate embodiments. The three instruments above are its first three modules (quotient search, cross-substrate compiler, causal/agency measurement); §4 lists the remaining modules as pre-registered future instruments.

8. Limitations and rejected alternatives

• The instruments are toys by design. They establish dissociations on exactly-solvable cases (finite Boolean worlds, tiny MDPs, lossless encoders). They do not establish the general conjecture, the conditional rate–distortion limit (extension 2), or the alignment claim (extension 9), all of which remain conjectural. Instrument 2’s fidelity is 1.0 by construction: it tests that the compilers are faithful functors, not that faithful compilation is hard.
• The master object is posited. The stochastic fibration $(q, K)$ organizes the corpus; that the sources’ relations are genuine adjunctions/simulations rather than looser analogies is exactly what §2’s sameness question demands be proved case by case. Do not cite the fibration as if it were any source’s theorem — only the biology paper proves a specialized version (its $C^*$ threshold and “no fourth category” accounting).
• Rejected — collapse to one scalar. “Just measure behavioural influence / compression / prediction.” Rejected: Instruments 1 and 3 exhibit exactly the cases where a single scalar misidentifies the target (MDL collapses the obstruction; false-credit influence is uncaused). The dissociations are the content.
• Consciousness caution retained. Nothing here licenses a claim of machine phenomenology; the core is a testable functional model of selfhood, and the gap to subjective experience is left open on purpose.

9. One-line synthesis

The object is a stochastic fibration $X \xrightarrow{q} Z$ with a compiler $K : Z \rightsquigarrow X$ filling the fibres; concern is a reweighting of the fibre, agency is the licensed alteration of the kernel, truth is fibre-invariance under independent tests, and intelligence is the search for the quotient where the world is at once compressible and controllable — of which conscious selfhood is one measurable functional face and formally bounded error is the other, with neither collapsible into infallibility or phenomenology.

v0.1, 2026-08-03. See DECISIONS.md (D29) for provenance. This document is the research-program-status sibling of SIC_MATHEMATICAL_FOUNDATIONS.md; when they disagree, that disagreement is intentional — this one is the working note, the other is the formalized paper.