From Relational Observability to Relational Information Accounting
Dynamic Container Encoding, Relational Closure, Synergistic Information, Causal Identifiability, and Intervention-Based Information Flow
John Swygert
Ivory Tower Publishing
September 12, 2026
Research Note
This paper continues the experimental sequence developed after From Dynamic Container Encoding to a Candidate Relational Inverse Operator. The earlier work asked whether the time-resolved geometry of a dynamical system could preserve recoverable information about hidden external influences acting through nested environments.
The initial results established a useful but limited finding: geometric history can contain substantially more information about a hidden outer cause than a final snapshot. However, the first proposed relational observability quantity, based on a singular-value ratio, failed under changes of sampling, delay depth, noise, and representation. Likewise, the proposed relational inverse operator did not consistently outperform conventional delay-based reconstruction methods.
The present paper therefore does not attempt to rescue those failed candidates.
Instead, it develops the direction that emerged from their failure:
Rather than asking whether one reconstruction algorithm succeeds, ask what happened to the information itself as the system transformed.
The resulting investigation progressed through information preservation, subsystem/environment accounting, nested relational closure, multiple hidden causes, fundamental non-identifiability, higher-order synergy, adversarial confounding, causal intervention, and directed information propagation.
Several results are positive, but they must be classified carefully. Much of the successful mathematics belongs to established information theory, causal inference, dynamical systems, latent-variable analysis, and system identification. No new physical law is claimed here.
The contribution of the present work is therefore primarily the construction and stress testing of a unified relational-accounting framework and the explicit identification of its success conditions and failure boundaries.
Abstract
A sequence of computational and mathematical experiments was performed to investigate whether information about hidden causes remains recoverable after transformation within nested dynamical systems. Earlier Dynamic Container Encoding experiments showed that time-resolved geometric histories can preserve information that is absent from final snapshots, but a proposed singular-value-based relational invariant failed under changes of representation, sampling, delay depth, and noise.
The investigation was therefore reframed as an information-accounting problem. Controlled experiments demonstrated that information apparently lost from a subsystem can remain accessible in a larger environment; that the smallest surrounding system required to recover a hidden cause can be treated as a provisional relational closure boundary; that information can be distributed across several layers such that no individual layer or pair carries the target while the joint relational state does; and that different hidden causes become fundamentally indistinguishable when they generate identical observable relational effects.
Adversarial testing then established a decisive limitation. Statistical information about a target can be distinguished from redundancy, irrelevant correlation, shared noise, and some forms of synergy using joint and conditional information measures, but passive observational statistics cannot universally establish causal attribution. Observationally equivalent causal models can contain identical mutual information while differing completely under intervention.
This leads to a two-ledger framework. Statistical relational closure concerns how much information about a target is present within a registered boundary. Causal relational closure concerns how much information is actually transmitted from the target under intervention or otherwise valid causal identification. Controlled intervention experiments further showed that directed propagation through nested systems can be reconstructed, including branching and shortcut paths.
The strongest surviving conclusion is that information accounting across transformation must distinguish at least five questions: what information remains accessible, where it is distributed, whether it is redundant or synergistic, whether it is causally attributable to a specified source, and how it propagates through the system. The resulting framework remains compatible with established mathematics and does not yet constitute a new physical invariant, but it provides a sharper basis for further TSTOEAO investigations of transformation, relational closure, and information flow.
01 Introduction
The Dynamic Container Encoding program began from a simple question:
What information about the forces and history of a system's environment is encoded in the geometry the system produces?
The motivating picture was not a static container surrounding a static object. It was a dynamically changing system embedded within other moving systems, each contributing constraints, boundary conditions, forces, rotations, flows, histories, and transformations.
A system observed at one instant may conceal much of that history.
A system observed dynamically may preserve more.
Initial computational experiments supported this weaker claim. Hidden outer drives could often be reconstructed substantially better from time-resolved histories than from final states.
However, the first attempt to turn that result into a portable relational invariant failed.
The quantity
ρ_R = σ_min / σ_max
was highly sensitive to delay depth, sampling interval, noise level, and representation. Large changes in ρ_R frequently occurred while actual recoverability changed very little.
The proposed relational inverse operator
T_rel = W* ◦ Π_out ◦ D_L ◦ Φ
also failed to establish a uniquely relational advantage. In some systems it performed well, but standard full-delay regression generally matched or exceeded it.
These failures shifted the research question.
The important issue was no longer:
Can this particular operator reconstruct the hidden cause?
It became:
What happened to the information about the cause during transformation?
This change proved productive.
02 From Reconstruction to Balanced Information Accounting
Suppose an initial hidden state H contains a known amount of information.
The system transforms that state through some dynamical process:
H → dynamical evolution → observable state Y.
A failed reconstruction of H from Y does not by itself establish that information was destroyed.
Several alternatives exist.
The information may have:
- remained directly accessible;
- changed representation;
- become distributed among multiple variables;
- moved into an environment not being observed;
- become accessible only through temporal history;
- become encoded synergistically among several components;
- become hidden beneath noise or coarse-graining;
- become statistically recoverable but causally ambiguous;
- or actually become unrecoverable within the registered system.
The accounting problem is therefore broader than reconstruction.
A useful conceptual ledger is:
input information → transformation → accessible information → distributed information → hidden information → inaccessible information → genuinely unrecoverable information.
The distinction between these categories became the central experimental objective.
03 Controlled Information-Preservation Test
A controlled experiment began with eight equally probable hidden states.
Eight equiprobable states contain
H(H) = 3 bits.
The hidden state was transformed through a dynamical system and then reconstructed from different representations of its resulting geometric history.
In the noiseless transformation, all three bits remained recoverable.
This demonstrated a fundamental distinction:
Transformation alone did not imply information destruction.
When moderate observational noise was introduced, the full time-resolved geometry retained substantially more recoverable information than the final endpoint alone.
Reversible transformations of the observation—such as coordinate rotations, relabeling, or reversible reordering—did not materially change the recoverable information.
Lossy operations did.
This suggested a first accounting principle:
A change of representation should not be confused with a loss of information when the transformation is reversible.
04 System Versus Environment
The next experiment deliberately moved information out of the measured subsystem.
The hidden state affected both a central subsystem and an environmental degree of freedom.
When only the subsystem was examined, the original state appeared to have vanished.
When the subsystem and environment were analyzed jointly, the complete information could again be recovered.
The important lesson was straightforward:
Information missing from a subsystem may have been transferred rather than destroyed.
The registered observational boundary therefore matters.
This immediately raised a new question:
How far outward must the boundary be extended before the information account becomes complete?
05 Relational Closure
Consider nested boundaries
B₁ ⊂ B₂ ⊂ B₃ ⊂ ... ⊂ Bₙ.
For a target hidden state H, define the statistical information available at boundary B_k as
R_stat(B_k) = I(H ; B_k).
A normalized form is
C_stat(B_k) = I(H ; B_k) / H(H).
The incremental information obtained by enlarging the boundary is
Δ_stat(k) = I(H ; B_k \ B_{k−1} | B_{k−1}).
This motivates the provisional concept of a relational closure boundary:
the smallest surrounding registered system within which the available information is sufficient to account for the target to a specified tolerance.
If B_full is the largest measured boundary, a practical relative closure measure can be written as
C̃_stat(B_k) = I(H ; B_k) / I(H ; B_full).
Then, for tolerance ε,
k* = min { k : C̃_stat(B_k) ≥ 1 − ε }.
This does not imply that information physically resides at one specific boundary.
It means only that the chosen collection of relationships has become sufficient for the registered information account.
06 The First Russian-Doll Test
The relational-closure concept was tested in explicitly nested systems.
Two different hidden causes were constructed so that their effects were indistinguishable within the inner layers.
Inside those smaller observational boundaries, classification remained near chance.
As progressively larger surrounding layers were included, a boundary was eventually reached at which the two hidden histories left different relational traces.
At that point distinguishability rose sharply.
Thus:
Two causes can be indistinguishable within one container while remaining distinguishable within a larger relational system.
This gives a precise interpretation to an important distinction:
“The information is unavailable here” is not equivalent to “the information no longer exists anywhere in the registered system.”
07 Split-Information Russian-Doll Test
A stronger experiment was then designed.
Instead of placing the missing information in a single outer layer, the target information was distributed across several layers.
The construction was arranged such that:
- no single layer identified the hidden target;
- no pair of layers identified the hidden target;
- the complete collection did.
This is a higher-order synergistic encoding.
A canonical binary example is
X₁ = R
and
X₂ = H XOR R,
where R is an independent random bit.
Individually,
I(H ; X₁) = 0
and
I(H ; X₂) = 0.
Jointly,
I(H ; X₁, X₂) = 1 bit.
The information is therefore not properly described as residing in either component.
It exists in their relationship.
The generalized lesson is:
Information can be absent from every individual component while remaining completely present in the relational state of the whole.
This phenomenon is established information theory rather than a new TSTOEAO law, but it is central to any serious relational-accounting framework.
08 Why Information Cannot Simply Be Added
The split-information result also revealed why an accounting sheet cannot be constructed by summing information measured separately in each layer.
Redundant variables may cause overcounting.
For example, suppose
X₂ = X₁.
If X₁ contains 0.5 bits about H, the pair does not contain 1 bit merely because two observations exist.
Instead,
I(H ; X₁, X₂) = I(H ; X₁).
The second variable adds no new target information:
I(H ; X₂ | X₁) = 0.
Conversely, synergistic variables can individually contain zero information while jointly containing substantial information.
Thus information accounting must distinguish at least:
- redundant information;
- unique information;
- synergistic information.
This is closely related to Partial Information Decomposition.
09 Innovation Points and Hidden Entry Regions
The next experiments returned to dynamical systems.
A hidden external drive was injected at an unknown layer in a nested coupled system.
Rather than revealing the injection location to the reconstruction algorithm, local prediction residuals were examined.
The idea was that a genuinely external contribution should appear as unexplained innovation relative to the internal dynamics.
In controlled systems, the strongest local innovation region frequently corresponded to the boundary at which target information increased most sharply.
Moving the hidden source to different layers moved both features.
This suggested a provisional correspondence:
local relational innovation and global information closure may sometimes be different views of the same hidden input event.
However, later adversarial tests demonstrated that this correspondence is not universal.
Innovation can also be produced by confounders, correlated noise, model error, or feedback.
Therefore:
innovation is evidence of an unexplained contribution, not proof of a specific causal source.
10 Multiple Hidden Causes
The network was then made more difficult.
Multiple hidden causes were introduced with:
- different strengths;
- different entry points;
- correlated source histories;
- overlapping influence regions;
- branching;
- loops;
- nonlinear responses;
- process noise;
- measurement noise.
Simple local detection degraded substantially as these complications increased.
Whole-system latent-factor approaches remained more effective.
In blind computational trials using residual factorization and independent-component-style methods, hidden entry regions could often be detected with high precision under moderate conditions.
Performance degraded when the hidden causes became weak, strongly correlated, or observationally similar.
This created the next critical question:
Was the failure caused by an inadequate algorithm, or had the system itself ceased to contain enough information to separate the causes?
11 Algorithm Failure Versus Fundamental Indistinguishability
Two hidden causes were progressively constructed to produce more similar observable effects.
When their relational footprints were substantially different, they were separable.
As those footprints approached one another, separability degraded.
The key limiting case occurs when two hidden causes generate identical observable distributions.
If
P(Y | H₁) = P(Y | H₂),
then no decoder operating on Y can distinguish H₁ from H₂.
This is not algorithm weakness.
The required information simply is not present in Y.
The distinction can be summarized as:
Algorithmic failure: information exists in the observation, but the chosen reconstruction procedure fails to extract it.
Information-theoretic failure: the observation itself does not contain sufficient information to distinguish the alternatives.
This became one of the strongest conceptual advances of the sequence.
12 Relational Equivalence Classes
The failure above suggests that a transformed system naturally defines equivalence classes of hidden histories.
Two hidden causes are observationally equivalent with respect to a registered observation Y when they produce indistinguishable observable distributions.
Symbolically,
H_a ~_Y H_b
when
P(Y | H_a) = P(Y | H_b).
Within that observational boundary, the causes belong to the same relational equivalence class.
Expanding the observational boundary may break that equivalence.
Thus a pair of causes may satisfy
H_a ~_{B₁} H_b
but not
H_a ~_{B₂} H_b,
where
B₁ ⊂ B₂.
This provides a rigorous interpretation of the Russian-doll result.
The information did not necessarily “appear” at B₂.
Rather, the larger relational boundary contained distinctions absent from B₁.
13 Adversarial Closure Testing
The relational-closure idea was then subjected to an explicitly adversarial test.
Systems were constructed in which apparent closure could result from:
- strong ordinary correlation;
- unrelated common causes;
- redundant copies;
- shared observational noise;
- feedback loops;
- correlated hidden causes;
- synergistic higher-order structure;
- observationally equivalent causal models.
Several naive closure ideas immediately failed.
Correlation among observed layers, for example, is not sufficient.
If an unrelated hidden variable Z drives two measurements,
X₁ = Z XOR N₁
and
X₂ = Z XOR N₂,
then X₁ and X₂ may be strongly correlated even though
I(H ; X₁, X₂) = 0
for the actual target H.
Thus:
inter-layer correlation is not target information.
14 Joint Target Information Survives Several False Closures
When the target H is known for validation, joint mutual information performs considerably better than internal correlation.
Define
J(S) = I(H ; X_S).
The information gained by adding variable X_j is
Δ_j(S) = I(H ; X_j | X_S).
This automatically prevents several forms of double counting.
Redundant copies contribute no additional conditional information.
Target-unrelated correlations contribute no target information.
Repeated information circulating through a feedback loop does not multiply the source entropy.
Joint synergistic information can be detected even when all marginal target-information terms vanish.
Thus the statistical accounting framework survived several adversarial attacks.
But a more serious failure remained.
15 The Causal Confounding Problem
Suppose two hidden variables share a common cause U.
Let
H₁ = U XOR N₁
and
H₂ = U XOR N₂.
If an observation X₂ depends on H₂, then X₂ may contain substantial information about H₁.
Yet H₁ may not cause X₂ at all.
The relationship may be entirely due to their shared ancestor U.
Therefore:
information about H is not the same thing as information transmitted by H.
This distinction proved decisive.
Mutual information measures statistical dependence.
It does not automatically establish causal origin.
16 Observational Equivalence: The Hard Boundary
The strongest falsification was constructed using two causal systems with identical observational distributions.
Model A: Genuine transmission
H → X → Y.
Suppose deterministically
X = H
and
Y = X.
Then observationally,
H = X = Y.
The system contains one bit of information about H.
An intervention on H changes X and Y.
Model B: Common cause
Let U generate all three:
U → H
and
U → X → Y.
Again,
H = X = Y
observationally.
The complete observational joint distribution can be identical to Model A.
Therefore every statistic calculated solely from
P(H, X, Y)
is also identical.
But intervention separates them.
Changing H directly in Model A changes X and Y.
Changing H directly in Model B does not alter U and therefore does not alter X or Y.
Thus the systems are observationally identical but causally different.
This creates a hard impossibility result:
No rule based only on passive observational statistics can universally determine whether information about a target was causally transmitted from that target.
This is not a weakness of a specific TSTOEAO method.
It is a causal-identifiability boundary.
17 Two Separate Ledgers
The experimental program therefore requires two distinct information ledgers.
Statistical Ledger
The statistical ledger asks:
How much information about target H is available within observational boundary B?
A suitable quantity is
R_stat(B) = I(H ; B).
This ledger can describe:
- accessibility;
- redundancy;
- synergy;
- incremental information;
- observational closure.
Causal Ledger
The causal ledger asks:
How much information appearing within B is actually attributable to interventions on H?
Let π(h) be a chosen intervention distribution.
Define the intervention ensemble
P_π(B) = Σ_h π(h) P(B | do(H = h)).
Then the causal information transmitted from H to B can be represented as the mutual information in the intervention ensemble:
R_do(B) = I_π(H ; B).
The corresponding normalized causal closure measure is
C_do(B) = R_do(B) / H_π(H).
These quantities answer fundamentally different questions.
A system may have a perfectly balanced statistical information ledger while possessing a completely different causal ledger.
18 Statistical Relational Closure
The weaker relational-closure concept survives.
For nested boundaries,
B₁ ⊂ B₂ ⊂ ... ⊂ Bₙ,
statistical relational closure concerns the smallest boundary at which essentially all recoverable target information within the registered universe has been captured.
Using the full registered system B_full,
C̃_stat(B_k) = I(H ; B_k) / I(H ; B_full).
Then a tolerance-dependent closure boundary is
k*_stat = min { k : C̃_stat(B_k) ≥ 1 − ε }.
This is a legitimate measure of statistical sufficiency.
It does not establish causal history.
19 Causal Relational Closure
When intervention data or valid causal identification is available, a stronger boundary may be defined.
Let
C̃_do(B_k) = I_π(H ; B_k) / I_π(H ; B_full).
Then
k*_do = min { k : C̃_do(B_k) ≥ 1 − ε }.
This can be interpreted as:
the smallest registered relational boundary containing at least 1 − ε of all causally attributable target information accessible within the registered system.
This definition preserves the Russian-doll concept while respecting the causal-identification limit.
20 Controlled Intervention Test
The next experiment directly addressed the observational-equivalence problem.
Two systems were constructed to produce essentially identical passive relationships.
In the first, the nominated variable genuinely drove the downstream system.
In the second, the nominated variable and downstream state were correlated only because of a common hidden source.
Passive observation could not reliably distinguish them.
A controlled intervention was then applied to the nominated cause independently of the confounding mechanism.
The outcomes diverged.
In the genuine-causal system, downstream states changed in response to the intervention.
In the confounded system, the downstream state did not follow the manipulated variable.
The intervention therefore supplied information unavailable in the passive observational distribution.
This confirmed the necessity of maintaining separate statistical and causal ledgers.
21 Nested Intervention and Information Flow
The intervention framework was then extended to a layered dynamical network.
Rather than perturbing only the presumed hidden source, individual layers were deliberately perturbed one at a time.
The resulting propagation patterns were recorded.
In the controlled network, interventions revealed:
- direct downstream paths;
- branch points;
- shortcut connections;
- delayed propagation;
- directionality.
The recovered structure was therefore more informative than static correlation.
An intervention at one layer generated a measurable response at connected downstream layers, while interventions at downstream layers did not necessarily reproduce the reverse path.
This provided an operational way to distinguish:
where information appears
from
the direction through which causal influence traveled.
22 Information Distribution Versus Information History
The results reveal an important distinction.
A final system state may contain the same information in several places.
But this does not reveal how that information arrived there.
Consider a branching process:
A → B
A → C
B → D
C → D.
Information about A may eventually appear in B, C, and D.
The final statistical distribution does not by itself specify the propagation history.
Interventional perturbations can help reconstruct the directed transaction structure.
This leads to a more complete relational information ledger:
source → transformation → distribution → redundancy/synergy → causal attribution → propagation history.
23 Toward a Relational Transaction Ledger
The earlier concept of balanced accounting can now be sharpened.
A complete relational transaction ledger would attempt to record:
-
Initial information
What distinctions existed in the source state?
-
Transformation
What dynamical process acted on those distinctions?
-
Representation
In what form did the information subsequently appear?
-
Distribution
Across which components, histories, or environmental variables was it spread?
-
Redundancy
Which apparent copies represented the same underlying information?
-
Synergy
Which distinctions existed only jointly among several variables?
-
Accessibility
What portion could be recovered from a particular observational boundary?
-
Relational closure
How large a boundary was required to recover the available target information?
-
Causal attribution
Which information was actually transmitted from the nominated source rather than merely correlated with it?
-
Propagation history
Through which directed paths did the information move?
-
Irrecoverability
Which distinctions were genuinely absent from the registered observational universe?
This is substantially richer than asking whether one inverse model successfully reconstructs a hidden variable.
24 What the Experiments Do Not Show
The experiments do not demonstrate that information is universally conserved in every physically meaningful sense.
They do not prove that all information apparently lost in physical systems can always be recovered by enlarging an observational boundary.
They do not establish a new physical conservation law.
They do not establish a novel universal relational invariant.
They do not show that relational closure can autonomously discover causal truth from passive observations.
They do not defeat causal-identifiability limits.
They do not demonstrate that TSTOEAO presently contains scientific content unavailable to established information theory, system identification, or causal inference.
These boundaries are important and should remain explicit.
25 What Has Been Demonstrated
Within the controlled systems examined, the following statements survived testing.
25.1 Time history can contain information absent from a final snapshot
Dynamic history can substantially increase recoverability of hidden drives.
25.2 Transformation is not equivalent to information destruction
Reversible representation changes can preserve the complete account.
25.3 Information may leave a subsystem without leaving the larger system
Environmental degrees of freedom can carry information apparently lost locally.
25.4 Recoverability depends on the registered relational boundary
Expanding the observed system can restore distinctions absent from smaller boundaries.
25.5 Information can be purely relational
Individual components may contain zero target information while their joint configuration contains the complete target.
25.6 Redundancy must not be counted as new information
Joint and conditional information measures prevent naive double counting.
25.7 Different causes can become fundamentally indistinguishable
If they leave identical observable distributions, no algorithm can uniquely separate them using those observations alone.
25.8 Statistical dependence is not causal transmission
A variable can carry information about a target without being causally downstream of that target.
25.9 Passive observation has an exact causal boundary
Observationally equivalent causal structures cannot always be distinguished without additional assumptions or interventions.
25.10 Intervention can reveal direction of causal propagation
Controlled perturbations can separate correlation from transmission and help reconstruct information-flow paths.
26 Implications for Dynamic Container Encoding
Dynamic Container Encoding originally emphasized the possibility that geometric histories retain information about hidden outer environments.
That claim survives only in a qualified form.
The geometry or trajectory of a system can indeed encode information about hidden influences.
But several distinctions must now be maintained.
The observed information may be:
- directly generated by the target;
- indirectly correlated through another cause;
- redundant;
- synergistic;
- environmentally transferred;
- dynamically propagated;
- or observationally indistinguishable from another hidden explanation.
Therefore geometry alone cannot be treated as a self-authenticating causal record.
The correct question is no longer merely:
Can the hidden cause be reconstructed from geometry?
It becomes:
What target information is accessible in the geometry, what additional information emerges when the relational boundary is enlarged, and under what causal assumptions can that information legitimately be attributed to a particular source?
That is a considerably stronger scientific formulation.
27 Implications for TSTOEAO
The TSTOEAO relational lens continues to be useful as an organizational framework.
Its recurring emphasis on:
- larger relational domains;
- nested boundaries;
- transformations;
- preserved relationships;
- equivalence classes;
- observer-dependent accessibility;
- cross-representation invariants;
- and explicit failure boundaries
helped generate the experimental progression.
However, the tests repeatedly show that the existence of a useful relational interpretation does not establish a new theory.
At present, the successful machinery is largely recognizable as established mathematics.
This is scientifically valuable because it clarifies the boundary between:
a useful lens
and
new excess scientific content.
The next stage must therefore ask whether the framework can predict something nontrivial that is not merely recovered after inserting known information-theoretic or causal machinery.
28 A Refined TSTOEAO Information Principle
The strongest defensible relational principle emerging from this work is:
A transformed system should not be said to have lost information merely because that information is absent from one representation, one component, one instant, or one observational boundary. The accounting must be performed over the registered relational system and must distinguish accessibility, redundancy, synergy, causal attribution, and genuine non-identifiability.
A second principle follows:
The smallest relational boundary sufficient for statistical reconstruction need not be the smallest boundary sufficient for causal attribution.
And a third:
The final distribution of information and the historical path by which information propagated are separate objects requiring separate measurements.
These principles are compatible with established theory but provide a clearer foundation for subsequent TSTOEAO experiments.
29 Proposed Next Test
The natural next experiment is a full relational transaction-history test.
A dynamical network should be constructed in which information:
- enters at more than one point;
- propagates inward and outward;
- splits across several paths;
- loops backward through feedback;
- recombines later;
- is redundantly copied in some locations;
- is synergistically represented in others;
- encounters correlated noise;
- partially enters an unobserved environment.
Controlled interventions would then be performed at multiple locations and times.
The goal would not merely be to infer the final causal graph.
The goal would be to reconstruct an explicit time-resolved ledger of information transactions:
source → path → transformation → split → recombination → final representation.
The critical question would be:
Can the accounting framework reconstruct the actual history of information movement rather than merely identify where information is present at the end?
Success would still need comparison against conventional causal-state reconstruction, directed information, transfer entropy, state-space identification, and related established methods.
30 Conclusion
The work following the Candidate Relational Inverse Operator produced a substantial change in direction.
The first relational observability invariant failed.
The candidate inverse operator did not demonstrate superiority over conventional reconstruction.
Those negative results redirected the research toward a deeper question:
What happens to information when a dynamical system transforms it?
The subsequent experiments showed that apparent information loss can result from incomplete observation, subsystem restriction, temporal truncation, noise, coarse-graining, or distributed encoding.
Nested “Russian-doll” experiments demonstrated that distinctions absent from smaller relational boundaries may become recoverable in larger ones.
Split-information experiments demonstrated that target information can exist entirely in relationships among variables rather than in any component individually.
Multiple-cause tests distinguished algorithmic reconstruction failure from genuine observational non-identifiability.
Adversarial testing then established a hard boundary: passive statistical information cannot universally determine causal origin.
This required the framework to divide into two ledgers:
statistical relational closure and causal relational closure.
Controlled interventions restored distinctions that passive observation could not provide and allowed directed information propagation through nested systems to be reconstructed.
The resulting picture is therefore not a universal information-preservation theorem.
It is a progressively sharpened accounting framework.
The central lesson is:
Information transformation, information distribution, information accessibility, information synergy, causal transmission, and information loss are not the same thing.
A scientifically defensible relational accounting system must keep them separate.
The long-term TSTOEAO objective remains more ambitious: to determine whether some domain-independent relational structure can predict when information will survive, where it will become distributed, when relational closure will occur, and when causal distinctions will become identifiable or fundamentally impossible.
That result has not yet been obtained.
But the experimental path toward testing it is now considerably clearer.
References
Cover, T. M., & Thomas, J. A. (2006). Elements of Information Theory (2nd ed.). Wiley.
Granger, C. W. J. (1969). Investigating causal relations by econometric models and cross-spectral methods. Econometrica, 37(3), 424–438.
Kullback, S., & Leibler, R. A. (1951). On information and sufficiency. The Annals of Mathematical Statistics, 22(1), 79–86.
Massey, J. L. (1990). Causality, feedback and directed information. Proceedings of the International Symposium on Information Theory and Its Applications, 303–305.
Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press.
Schreiber, T. (2000). Measuring information transfer. Physical Review Letters, 85(2), 461–464.
Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27, 379–423, 623–656.
Williams, P. L., & Beer, R. D. (2010). Nonnegative decomposition of multivariate information. arXiv:1004.2515.
Wiener, N. (1956). The theory of prediction. In E. F. Beckenbach (Ed.), Modern Mathematics for Engineers. McGraw-Hill.
