Mathematics, Apparatus, and the Limits of Description. Toys.

Nicole Flynn, Symfield PBC, September 2026.

The toy is a bounded structure with declared rules. It redirects the flow without owning what emerges. Through exchange, we can expand what we distinguish. Agreement cannot establish what we have understood.

Abstract

A mathematical description can be exact while leaving its source unresolved. This is more than a general warning that models are incomplete. Different underlying states can yield identical observations, and further calculation need not recover the distinctions that observation has removed. This essay examines that boundary through a simple observation map, the data-processing inequality, quantum reduced states, and Markov traces. A small stochastic model then shows how an intervention can separate mechanisms that selected observational statistics leave indistinguishable. The question is what changes when an account becomes more elaborate, its internal consequences, its access to the system, or both. Mathematics makes these differences precise. Its value includes showing where a description succeeds, where identification remains open, and what another apparatus would have to make available.

The brook is not the pond

Consider a brook entering a pond. Its course depends on gravity, terrain, sediment, vegetation, and weather. The pond may also be engineered. Banks, channels, retaining structures, and outlets stabilize some flows and redirect others. The construction participates in the system it organizes.

A tree beside the pond participates through roots, moisture, light, chemistry, season, damage, and growth. A person walking beside it couples through movement, balance, sight, sound, smell, memory, expectation, and attention. An artificial intelligence encounters the scene through language, encoded representations, learned regularities, context limits, and tools.

These are different forms of participation. The tree draws water; the walker feels the ground give beneath a foot; a scientific measurement instrument records a change in moisture. A mathematical model makes selected relationships available for calculation. Each permits a kind of access. Each also establishes conditions under which something can become distinguishable. Here, apparatus names those conditions of access and representation. A nervous system, a sensor, and a formalism do not operate in the same way. The comparison concerns what they preserve, transform, or leave unresolved. Perception belongs within this account; it is not an unconditioned spectator receiving an instrument’s results. A telescope does not invent a galaxy by converting light into an image. Neither does it deliver the galaxy whole. Aperture, wavelength, orientation, calibration, and processing affect the resulting observation. An image can be accurate without containing every distinction relevant to its source.

Mathematical modeling raises a related question. Its precision is indispensable. But what, precisely, has become exact?

An exact answer with an unresolved source

Suppose an apparatus receives two real numbers and reports their sum:

$$C(x,y)=x+y.$$

It reports five. The input could have been $(2,3)$, $(1,4)$, or any point on the line $x+y=5$.

The observer can square the answer, compare it with a threshold, or build a long sequence of conditional calculations around it. If the sum exceeds four, take one branch; otherwise, take another. Within the first branch, introduce further conditions. Nothing prevents this construction from becoming extensive, useful, or mathematically sophisticated. Yet every input on that line still enters the calculation as five. For any subsequent function $f$ of the reported value alone,

$$C(x,y)=C(u,v) \quad\Longrightarrow\quad f(C(x,y))=f(C(u,v)).$$

The calculation may continue indefinitely without separating those inputs. Its limitation is not a lack of possible operations. It is the absence of a distinction those operations could act upon. Mathematically, the apparatus groups inputs into equivalence classes: two inputs belong to the same class when it reports the same result. Its output identifies a class, not necessarily a unique member. An elaborate account can develop within that grouping while leaving the grouping itself unchanged. Another measurement changes the situation. If a second apparatus reports $x-y$, the sum and difference together recover both numbers. Neither measurement is individually sufficient; together they are sufficient for this particular reconstruction. There is no contradiction between extensive possibility and restricted access. A description can support infinitely many operations while remaining unable to answer one specific question about its source.

What processing can and cannot add

Information theory gives this distinction a probabilistic form. Suppose $X$ represents a source, $Y$ an observation, and $Z$ a further processing of that observation. If the processing receives no information about $X$ beyond $Y$, these variables form the Markov chain $X\to Y\to Z$. The data-processing inequality states

$$I(X;Z)\le I(X;Y),$$

where $I$ denotes mutual information. Processing alone cannot increase the mutual information the observation carries about the source. Beaudry and Renner, 2012

This does not make calculation unproductive. A computation can expose a consequence that was inaccessible to a reader, transform an unwieldy record into a useful statistic, or preserve precisely the information needed for a decision. What becomes usable can change greatly even when the information available from the source does not increase. The distinction is between developing the consequences of an observation and acquiring a new observation. New measurements, interventions, or independent information can alter the conditions of the problem. A longer calculation using the same input does not necessarily do so.

This also clarifies the contribution of artificial intelligence. An AI may introduce relevant prior knowledge, identify a neglected result, or propose an experiment. Those contributions can enlarge the inquiry. But repeated reformulation of the same evidence does not by itself make observationally identical sources distinguishable. Fluency and variety are not additional measurements.

A quantum example

Quantum mechanics provides a case in which the restriction is exact. Consider two common joint states of a pair of qubits:

$$|\Phi^+\rangle=\frac{|00\rangle+|11\rangle}{\sqrt2}, \qquad |\Phi^-\rangle=\frac{|00\rangle-|11\rangle}{\sqrt2}.$$

The joint states differ. But an observer restricted to subsystem $A$ obtains the same reduced density operator in either case:

$$\rho_A^+=\rho_A^-=\frac{\mathbb I_A}{2}.$$

For these preparations, every measurement confined to $A$ has identical outcome probabilities. Perfect local equipment and exact local predictions do not identify which joint state was prepared. The difference becomes accessible in correlations: measurements of both qubits in the $X$ basis yield matching outcomes for $|\Phi^+\rangle$ and opposite outcomes for $|\Phi^-\rangle$. Comparing the records resolves what either local record alone leaves open. These are standard consequences of the reduced-state formalism. Preskill, Chapter 2

The local description has not failed. It correctly specifies the statistics available within its scope. The unresolved distinction belongs to a question that exceeds that scope. This result does not establish that dimensions are illusory or that perception creates physical reality. It establishes something narrower and sufficient for the present argument: predictive completeness for a restricted set of measurements need not identify the larger state.

What remains consequential outside a window

Restricted observation does not imply that everything unresolved is dynamically irrelevant. In a Markov trace, an observer registers visits to a subset $A$ of a larger state space. Between successive registered visits, the process may travel through the inaccessible complement $B$. The effective transition law on $A$ includes those excursions. For a finite chain written in block form, when the hidden-state block satisfies the convergence conditions for the inverse, the trace kernel is

$$P_A=P_{AA}+P_{AB}(I-P_{BB})^{-1}P_{BA}.$$

The second term sums contributions from hidden excursions before the next return. With almost-sure return from the relevant states, this gives a stochastic transition kernel on the visible window. The trace indexes visible visits; it need not retain how much original time elapsed between them. The observer therefore encounters dynamics already shaped by what the window does not directly display. Yet distinct larger chains can induce the same trace:

$$P^{(1)}\ne P^{(2)}, \qquad P_A^{(1)}=P_A^{(2)}.$$

Successful prediction of the trace need not uniquely recover its generator. This restricted-access problem is central to the discussion of Hoffman’s program in What Lies Beneath Spacetime?. The mathematical question can be examined without adopting a particular ontology of consciousness.

There is a neighboring result in statistical mechanics. The Mori–Zwanzig formalism expresses reduced dynamics using retained variables, a memory term, and a contribution driven by unresolved degrees of freedom. Eliminating variables from the explicit description need not eliminate their influence. Practical reduced models must then approximate or otherwise handle these terms. Li and Stinis, 2019

These constructions are not interchangeable. A Markov trace records returns to a subset; Mori–Zwanzig projects a dynamical description onto selected observables. Both make the choice of what is retained mathematically consequential.

Before the theorem

Formal papers usually begin after the uncertain part has been cleaned away. Definitions appear first. Lemmas follow. The finished object can look inevitable. It rarely was.

Work may begin with two cases that an existing model treats as equivalent but that appear structurally different; an absence that seems to deform what surrounds it, or a transition that changes not only the next state but what remains possible afterward. At that stage, there is no theorem. There may not yet be adequate language.

Proceeding requires a disciplined form of faith, enough trust in a perceived relation to construct before verification is complete. This is provisional commitment, not a substitute for evidence. A conjecture must be given enough structure to become vulnerable. The resulting equation is neither a finished object simply retrieved from the world nor an unconstrained invention.

As an example, stories give us a way to follow change: something resists, an expectation fails, someone responds. Remove that resistance and the conclusion may remain intelligible, but we lose why reaching it mattered. The interesting complication is that the antagonist need not be wrong. An earlier formulation may work beautifully until a question requires a distinction it cannot preserve. Its limit becomes productive. What defines is what survives translation, counterexample, computation, and intervention. A rejected formulation can be part of that achievement as it reveals which distinction the earlier language could not preserve.

When a changed question separates the mechanisms

The model in Debris as Predictive Compression begins with a distinction: a transition that was available but not realized need not be equivalent to one that was never available.

The toy contains four states and twelve directed edges. Admissible edges leaving the current state receive independent uniform random activations. Activations above a candidate threshold define the prospective alternatives; a fallback selects the strongest edge if none clears it. The strongest candidate is realized. Other candidates contribute activation to edge-specific residue. When residue exceeds a deletion threshold, its edge is permanently removed from the admissible set.

The reported parameters are a candidate threshold $\theta=1/3$, residue retention $\lambda=1/2$, and deletion threshold $\tau=7/10$, starting with zero residue and all edges admissible. Conditional on the initial transition $1\to2$ and a subsequent return to state $1$, the derived probability of repeating that transition on the first return is

$$p_{\mathrm{repeat}}=\frac{2297}{6000}\approx0.3828.$$

The symmetric stationary Markov null gives $1/3$. The difference is $99/2000$, or 4.95 percentage points. This is a conditional repeat probability, not the probability of returning to state $1$.

The result gives the model a detectable signature. It does not uniquely identify the mechanism. The published companion tests compare three alternative memory models that do not retain unrealized candidates. Each is calibrated to reproduce the first-return bias. Later-return statistics separate some of these alternatives more clearly than others; they do not justify saying that all three mechanisms have been identified from the first statistic.

An intervention asks a different question. Fix the initially realized edge’s activation at $0.9$, and another competitor at $0.1$. Change the remaining, unrealized competitor’s activation from $0.5$ to $0.8$. The realized transition remains $1\to2$ in both cases. In one case the competitor survives; in the other it exceeds the deletion threshold. On the first return, its selection probability changes from $1/3$ to zero.

The same change would follow if the competitor remained formally admissible but received zero selection weight; the intervention establishes that the later law depends on the unrealized activation, not which object carries that dependence. The realized initial transition is the same. What was available alongside it changes the later transition law.

Within the toy, this intervention distinguishes dependence on unrealized activation from the tested alternatives whose updates do not depend on it. It does not identify every possible hidden mechanism, and it supplies no evidence by itself for a physical debris field.

The restoration tests turn on timing relative to the first readout. Under the specified update order, residue decays each step and restoration is checked on arrival at the source before selection. At the tested restoration threshold of $0.6$, deleted competitors are restored before selection on the earliest possible return. The first-return bias therefore vanishes by construction; a near-zero simulation estimate is not an independent finding about reversibility. Lower thresholds allow some competitors to remain excluded at that readout and retain a reduced bias. These tests compare restoration timing, not degrees of reversibility. The exact $99/2000$ belongs to the permanent-deletion toy. (See the September 18 $K_n$ analysis, §§3.1 and 4.) What changed was not the observer’s confidence in the same statistic. The intervention made a previously unresolved causal distinction available.

The toy and its continuation

A mathematical toy declares enough structure for consequences to become inspectable. It specifies states, transitions, operations, and assumptions. Those boundaries permit another reader to reproduce a calculation or find a counterexample. A theorem establishes what follows within that specification. A physical application requires a further correspondence between the formal quantities and observations. Neither task is diminished by distinguishing them.

The debris construction could be extended with other memory rules, reversible admissibility, interacting residue, or changing observation maps. Such extensions may be worth pursuing. But their existence as equations would not inherit the evidential status of the tested four-state result. Each would introduce new questions about definition, prediction, and identification. Stopping a construction therefore need not mean that no further mathematics is possible. It can mark the point where the next useful step is an experiment, a sharper observable, or a comparison the present model does not yet support.

The same restraint matters in larger theories. A formalism may have infinitely many states and still leave distinct sources observationally equivalent. It may generate an extensive family of correct predictions while its ontology remains underdetermined. The size of the construction does not settle whether its successful distinctions exhaust those of the system.

What another apparatus can change

The sum-and-difference example shows one way access can expand. A second observation separates inputs that the first groups together. The quantum example shows why the relationship between records may carry a distinction absent from either record in isolation. The debris intervention shows how action can separate accounts that selected passive observations leave compatible.

These are different mechanisms of enlargement. More observers do not automatically produce one. Several instruments may share a calibration error, several models may inherit the same omission, and several people may repeat an interpretation derived from a single source. Agreement is more informative when the ways of arriving at it are sufficiently independent for the question being tested.

Nor must observers always converge. A disagreement can locate a difference in resolution, measurement context, retained history, or experimental access. The task is to determine which difference matters and whether a translation or intervention can make it testable. An artificial intelligence, a human investigator, and a laboratory instrument can contribute to this work without becoming interchangeable. An AI can help expose a missing assumption or generate a counterexample. A person can recognize that the formal question has ceased to address the original distinction. An instrument can return a result neither expected. Each can also obscure something another preserves.

An AI can also make a living argument look as though it arrived already resolved. It may preserve every correction as content while removing the friction that explains why the corrections mattered. An objection becomes a qualification; a failed attempt becomes a sentence about limitations. The conclusions remain, but the reader can no longer see what forced them to change.

This is a loss of argumentative history and emphasis, not automatically an instance of the data-processing inequality. A document may retain the facts while changing their role in the reasoning. Human writers and institutions can make the same compression. The question is whether the resulting clarity still allows a reader to reconstruct what was at stake. As an example, a story can carry precisely that missing history, provided it doesn’t turn every discarded idea into a villain and every surviving one into a hero. Formalization is valuable here because it makes a disagreement portable. It gives others an object they can examine without reproducing the entire encounter from which it arose. The resulting proof or model carries selected relations, not the whole history of discovery. Its clarity should not be mistaken for evidence that everything omitted was irrelevant.

What remains open

Every model must make distinctions before it can derive consequences. The scientific question is how those distinctions answer to what is encountered: which predictions survive, which interventions discriminate, and which observations require a revised account. Sometimes a better calculation is enough. Sometimes the calculation is already exact and the unresolved question requires another kind of access. The difference deserves our attention because both situations can look, from within an inquiry, like a demand for more mathematics. No claim to a final account of reality is needed to contribute something fundamental. It may be enough to make a distinction available that an earlier description erased, or to establish that a favored observation cannot separate the explanations placed upon it. Such results tell the next investigator where to look and what would have to change.

The mathematical toy is useful partly because its boundaries can be stated. The field is not thereby required to share them. Omitting a relation does not establish its absence; including it does not establish its existence. Between those two errors lies the work of construction, measurement, and revision.

The toy is not the field. The brook is not the pond. But a well-placed structure can redirect the flow and change what becomes possible downstream.

References

Beaudry, Normand J., and Renato Renner. 2012. “An intuitive proof of the data processing inequality.” Quantum Information and Computation 12(5–6): 432–441. Open manuscript.

Preskill, John. Lecture Notes for Physics 219: Quantum Computation. Chapter 2, “Foundations I: States and Ensembles.” California Institute of Technology. Chapter PDF.

Li, Jing, and Panos Stinis. 2019. “Mori–Zwanzig reduced models for uncertainty quantification.” Journal of Computational Dynamics 6(1): 39–68. doi:10.3934/jcd.2019002. 2018 preprint.

Flynn, Nicole. 2026. “What Lies Beneath Spacetime? Donald Hoffman, Recursive Trace Logic, and the Mathematics of the Observer.” Symfield PBC, September. Essay.

Flynn, Nicole. 2026. “Debris as Predictive Compression: An Exact Return Bias and an Open Equation Spine for Observer-Relative Coupling.” Technical Research Note v0.3, with companion computational notes I and II. Symfield PBC, September 17 to 18. Research note.

Flynn, Nicole. 2026. “Debris on $K_n$: Exact First-Return Law and the Identifiability of the Unrealized-Transition Channel.” Companion computational note II, in Debris as Predictive Compression, v0.3. Symfield PBC, September 18. Note.

Hoffman, Donald D., and Chetan Prakash. 2014. “Objects of Consciousness.” Frontiers in Psychology 5: 577. doi:10.3389/fpsyg.2014.00577.

Meyer, Carl D. 1989. “Stochastic Complementation, Uncoupling Markov Chains, and the Theory of Nearly Reducible Systems.” SIAM Review 31(2): 240–272. Article.

Zwanzig, Robert. 1961. “Memory Effects in Irreversible Thermodynamics.” Physical Review 124(4): 983–992. doi:10.1103/PhysRev.124.983.

Mori, Hazime. 1965. “Transport, Collective Motion, and Brownian Motion.” Progress of Theoretical Physics 33(3): 423–455. doi:10.1143/PTP.33.423.