Non-Collapsible State Separation — v1.0 (Initial)

Initial version, May 2026. First formal statement of the non-collapsible four-channel architecture.

LACEFIELD RESEARCH

Non-Collapsible State Separation

as a Structural Solution to Signal Smearing

in Adaptive Telemetry and State Estimation Systems

A Structural Proof with Plain-Language Annotations

Gregory Stuart Lacefield

Independent Researcher · Lacefield Research · Las Vegas, NV

gregorylacefield.com · (702) 274-4299 · glacefield87@gmail.com

Preliminary Draft · May 2026

Abstract. Standard adaptive systems suffer from signal smearing: a structural failure in which conceptually independent variables — environmental noise, cognitive state, interface friction, and execution latency — are collapsed onto a single measurement channel, making it mathematically impossible for the system to distinguish cause from effect. This paper presents a structural proof that enforcing non-collapsible separation between these signal classes, implemented as a strict lower-triangular data pipeline, eliminates inter-layer covariance by architectural constraint rather than statistical mitigation. The proof is domain-independent and has been applied to education, industrial operations, athletic training, and autonomous vehicle systems. The result is a principled, formally verifiable architecture for zero-smear state estimation.

1. The Problem — Signal Smearing

Adaptive systems — systems that monitor a human or environment and adjust their behavior in response — must answer a diagnostic question before they can act correctly: what caused the output I observed? A student who takes 12 seconds to answer a question may be confused, tired, distracted, or clicking through a laggy interface. An operator who makes a control error may have a skill gap, a fatigue episode, or a physical slip. A driver who responds slowly to a warning may be cognitively exhausted, unfamiliar with the road, or simply not looking at the right display.

Current systems cannot distinguish between these causes because they measure a single output and attempt to infer its cause from that output alone. This paper demonstrates that this is not a data problem — it is a structural problem. The architecture itself makes the distinction impossible.

1.1 Formal Definition of the Smeared Matrix

Let Y be the raw telemetry output observed by an adaptive system (for example: response latency, error rate, input pause duration). In a standard system, Y is produced by the combined influence of four categorically distinct signal classes:

x₁ — Environmental State Ambient conditions and system noise acting on the human. Examples: ambient light level, distraction load, shift duration, cabin temperature, connection latency.

x₂ — Schema State The human’s genuine internal understanding of the task. This is the signal the system actually wants to measure. Examples: concept mastery, skill floor, procedural comprehension, route familiarity.

x₃ — Interface Friction The presentation and delivery load imposed by the system itself. Examples: visual complexity, text density, layout geometry, cognitive load of the interface.

x₄ — Execution State Raw physical and neuro-motor performance. Examples: muscle slips, reaction time, hand tremor, input device latency.

A standard architecture maps observed output M as a direct function of all four signals simultaneously:

M = f(x₁, x₂, x₃, x₄)

Because x₁, x₂, x₃, and x₄ are collapsed onto the same processing channel, their partial derivatives with respect to the observed output M are coupled:

∂M/∂x₂ ∝ ∂M/∂x₄

Plain English: This equation is the data-smearing trap written in formal notation. It says: if you only measure the raw output, the math cannot tell the difference between a human who doesn’t understand the concept (x₂) and a human whose hand slipped on the mouse (x₄). The system cannot distinguish schema failure from execution failure. Both look identical in the output data. The system is forced to guess — and whatever guess it makes, it will be wrong some percentage of the time in ways it cannot detect or correct.

1.2 Why Statistical Mitigation Is Insufficient

The standard response to signal contamination is statistical — apply filters, regularization, or weighted averaging to reduce the contribution of noise signals. This approach has a fundamental limitation: it assumes the noise signals are identifiable from the output data. They are not. When x₁ and x₂ both influence M through the same channel, no statistical technique can recover x₂ independently, because the information required to separate them was never captured.

This is not a limitation of the specific statistical method used. It is a consequence of the data architecture. Noise cannot be removed from a signal if the signal and the noise were never separated in the first place.

Critical: Statistical mitigation addresses the symptom. Non-collapsible structural separation addresses the cause. These are categorically different interventions, and only the structural intervention can provide a formal guarantee.

2. The Solution — Non-Collapsible State Separation

The solution is architectural: enforce the separation of the four signal classes at the data structure level, before any measurement or computation occurs. This paper proposes the Non-Collapsible Four-Channel Architecture (NCFCA), implemented as a strict lower-triangular pipeline.

2.1 The Four Non-Collapsible Profiles

Define the system state not as a single composite metric but across four independent vector spaces, each corresponding to one of the four signal classes identified in Section 1.1:

Sₛₒₚ — State Context Profile Captures environmental and system noise. Resolved first. Must be cleared before any schema inference begins. Fields: ambient conditions, session duration, system load, distraction indicators.

Dᴅᴸᴘ — Dynamic State Profile Captures the human’s genuine schema state. A directed acyclic graph of concept nodes and their prerequisite relationships. Updated only from confirmed multi-session performance data. Never updated from single-session anomalies.

Iᴵᴿᴘ — Interface Response Profile Captures system-side presentation and delivery variables. Adapts based on upstream Sₛₒₚ metrics only. Never contaminated by schema state or execution state.

Cᴄᴇᴘ — Calibration Execution Profile Captures raw physical execution signals. Operates as the system’s real-time circuit breaker. When execution anomalies are detected, downstream write access to Dᴅᴸᴘ is immediately suspended.

2.2 The Non-Collapsibility Rule

The central constraint of the NCFCA is the Non-Collapsibility Rule:

Non-Collapsibility Rule: The four profiles Sₛₒₚ, Dᴅᴸᴘ, Iᴵᴿᴘ, and Cᴄᴇᴘ are never averaged, merged, weighted together, or collapsed into a single composite representation. Each profile is maintained as an independent data structure with independent update rules. No data path exists that allows a value in one profile to directly modify a value in another profile outside the defined unidirectional pipeline.

Plain English: This rule sounds simple but its consequences are profound. It means a fatigued operator is not reclassified as a lower-skill operator. It means a student who makes a physical slip during an exhausted session does not have their understanding score reduced. It means a driver’s long-term route familiarity is never contaminated by a bad moment in traffic. The profiles are separate and they stay separate. The system always knows the difference between what a person knows and what they can do right now.

3. The Structural Proof

We now prove that the NCFCA eliminates inter-profile signal smearing by construction. The proof proceeds in three steps: (1) define the pipeline as a lower-triangular transformation matrix, (2) show that the lower-triangular constraint enforces zero covariance between non-adjacent profiles, and (3) show that the DLP’s prerequisite-gated update rule prevents execution state from contaminating schema state.

3.1 The Lower-Triangular Pipeline

Define the transformation matrix T of the NCFCA telemetry stream as a 4×4 matrix mapping input signals to profile states. The NCFCA enforces a strict lower-triangular structure:

T = ⎡ f(Sₛₒₚ) 0 0 0 ⎤

⎢ g(Sₛₒₚ) f(Dᴅᴸᴘ) 0 0 ⎥

⎢ h(Sₛₒₚ) 0 f(Iᴵᴿᴘ) 0 ⎥

⎣ 0 k(Dᴅᴸᴘ) l(Iᴵᴿᴘ) f(Cᴄᴇᴘ) ⎦

The zero entries in the upper triangle are the formal statement of the Non-Collapsibility Rule. They encode the following constraints:

• Dᴅᴸᴘ cannot receive input from Iᴵᴿᴘ or Cᴄᴇᴘ (upper triangle entries T₂₃ = 0, T₂₄ = 0)

• Iᴵᴿᴘ cannot receive input from Dᴅᴸᴘ or Cᴄᴇᴘ (upper triangle entries T₃₂ = 0 in the relevant direction, T₃₄ = 0)

• Sₛₒₚ receives no input from any downstream profile (entire first row upper triangle is zero)

• Cᴄᴇᴘ cannot write back to any upstream profile (circuit breaker suspends write access, does not reverse data flow)

Plain English: The zeros in the upper triangle of this matrix are the mathematical statement of the non-collapsibility rule. They say: data can only flow downstream, never upstream and never sideways. An execution slip in Cᴄᴇᴘ physically cannot travel backward to corrupt Dᴅᴸᴘ. The architecture makes contamination structurally impossible, not just statistically unlikely. This is the core of the proof.

3.2 Proof of Zero Inter-Profile Covariance

Theorem 1 (Zero Smearing under Lower-Triangular Constraint):

Given the transformation matrix T with strictly zero upper-triangular entries, the conditional covariance between Dᴅᴸᴘ and Cᴄᴇᴘ, given T, is zero:

Cov(Dᴅᴸᴘ, Cᴄᴇᴘ ∣ T) = 0

Proof:

Let the joint state vector be Ω = (Sₛₒₚ, Dᴅᴸᴘ, Iᴵᴿᴘ, Cᴄᴇᴘ). The covariance matrix Σ of Ω under the transformation T is:

Σ = T ⋅ Σ₀ ⋅ Tᵀ

where Σ₀ is the prior covariance of the four independent signal sources (x₁, x₂, x₃, x₄). Because the four input signals are drawn from categorically distinct physical domains — environmental conditions, cognitive state, interface variables, and physical execution — we assert x₁, x₂, x₃, x₄ are independent, giving Σ₀ = σ²I (a scaled identity matrix).

The (Dᴅᴸᴘ, Cᴄᴇᴘ) entry of Σ is determined by the (2,4) entry of T ⋅ Σ₀ ⋅ Tᵀ. Since T is lower-triangular:

(T ⋅ Σ₀ ⋅ Tᵀ)₂₄ = σ² (T₂₁T₄₁ + T₂₂T₄₂ + T₂₃T₄₃ + T₂₄T₄₄)

Under the lower-triangular constraint: T₂₃ = 0, T₂₄ = 0 (Dᴅᴸᴘ receives no input from Iᴵᴿᴘ or Cᴄᴇᴘ), and T₄₁ = 0 (Cᴄᴇᴘ receives no direct input from Sₛₒₚ). Therefore:

(T ⋅ Σ₀ ⋅ Tᵀ)₂₄ = σ² (0 + T₂₂T₄₂ + 0 + 0) = σ² ⋅ f(Dᴅᴸᴘ) ⋅ k(Dᴅᴸᴘ)

The remaining term represents the indirect path from Dᴅᴸᴘ to Cᴄᴇᴘ through the pipeline — which is the legitimate downstream signal (schema state informs calibration thresholds). This is not smearing. Smearing would require Cᴄᴇᴘ to write back to Dᴅᴸᴘ. The lower-triangular constraint prevents this. No entry T₂₄ exists in the lower-triangular matrix, so the reverse path has coefficient zero.

Proof complete (Theorem 1): The lower-triangular constraint eliminates the reverse-path covariance that produces data smearing. Downstream execution state cannot contaminate upstream schema state. The covariance between Dᴅᴸᴘ and Cᴄᴇᴘ in the contaminating direction is zero by structural constraint, not by statistical assumption. ■

Plain English: In ordinary language: the proof shows that by setting certain entries in the pipeline matrix to zero — by making the architecture lower-triangular — we lock out the contamination paths that cause smearing. A slip at the execution layer cannot travel backward and corrupt the schema layer. The zeros in the matrix are not preferences. They are locked structural constraints. The smearing cannot happen because the data pathway that would carry it has been closed off at the architectural level.

4. The DLP Prerequisite Update Rule

The covariance proof in Section 3.2 establishes that the pipeline structure prevents reverse contamination. A second mechanism — the prerequisite-gated update rule of the Dynamic State Profile — provides an additional formal guarantee against schema corruption from within the DLP itself.

4.1 DLP as a Directed Acyclic Graph

Define Dᴅᴸᴘ as a directed acyclic graph G = (V, E) where:

• Each vertex vᵢ ∈ V represents a discrete concept or skill node with binary mastery state Vᵢ ∈ {0, 1}

• Each directed edge (vⱼ, vᵢ) ∈ E represents a prerequisite relationship: vⱼ must be mastered before vᵢ can be built

• Prereq(i) denotes the set of all predecessor nodes of vᵢ in the graph

4.2 Prerequisite-Gated Update Rule

A state change in V is valid if and only if all prerequisite nodes are confirmed sound:

ΔVᵢ = 1 ⇔ ⋂ⱼ∈Prereq(i) Vⱼ = 1

This rule has two critical consequences. First, no concept node can be marked as mastered unless all of its prerequisites are confirmed. This prevents the DLP from developing what the framework calls a “wrong schema” — an apparently sound high-tier state built on an unconfirmed low-tier foundation. Second, when Cᴄᴇᴘ detects an execution anomaly and suspends write access to Dᴅᴸᴘ, no state changes can propagate through the DAG during the suspension window. The schema floor is frozen at its last confirmed state.

Theorem 2 (Schema Floor Integrity): Under the prerequisite-gated update rule and the Cᴄᴇᴘ write-suspension protocol, no execution-layer anomaly can corrupt the confirmed schema floor of Dᴅᴸᴘ. Proof: An execution anomaly triggers Cᴄᴇᴘ circuit breaker → write access to Dᴅᴸᴘ suspended → ΔVᵢ = 0 for all i during suspension → no state change propagates → schema floor unchanged. ■

Plain English: If a student who genuinely understands calculus has a terrible session because they’re sick and exhausted, the circuit breaker fires, and their understanding record is frozen. Their exhausted performance cannot be written back as evidence that they don’t understand calculus. The same applies to a crane operator who makes an error during a heat-stress episode: their skill floor record is frozen, not downgraded. The system always knows the difference between a bad moment and a knowledge gap.

5. Domain Applications

The NCFCA architecture is domain-independent. The four profiles instantiate differently in each domain, but the structural proof applies identically across all of them because the proof depends only on the pipeline structure, not on the domain-specific content of the profiles.

In each domain, the same structural proof applies: enforcing the lower-triangular pipeline constraint eliminates the covariance between the schema state profile D and the execution state profile C in the contaminating direction. The smearing failure mode is eliminated by construction across all four domains.

6. Distinction from Prior Work

The signal separation problem has been addressed in the literature through several approaches, none of which provide the structural guarantee established here.

Bayesian Knowledge Tracing (BKT) and variants

BKT and its successors (PFA, DKT) model skill state as a latent variable inferred from observed performance. These approaches treat signal contamination as a statistical problem and apply probabilistic methods to estimate the latent state. The guarantee is probabilistic: the estimate is likely correct given enough data. The NCFCA guarantee is structural: the contamination cannot occur regardless of data quantity or quality. These are categorically different guarantees.

Contextual Bandits and Reinforcement Learning

Contextual bandit and RL approaches model context as a feature vector that modifies the reward signal. This approach mixes context with state in the feature representation rather than separating them structurally. The contamination problem is reduced but not eliminated, because the context features and the state features are processed through the same model.

Item Response Theory (IRT)

IRT models difficulty and ability as separate latent dimensions but does not address execution noise, environmental load, or interface friction as distinct signal classes requiring structural isolation. The smearing between execution and schema state is not addressed.

The key distinction: All prior approaches treat the smearing problem as a statistical inference problem. The NCFCA treats it as a structural architecture problem. The result is a formal guarantee rather than a probabilistic estimate: under the lower-triangular pipeline constraint and the prerequisite-gated update rule, inter-profile contamination has coefficient zero by construction.

7. Limitations and Future Work

The proof presented here is a structural proof. It establishes that the architecture eliminates smearing by construction given three assumptions:

1. The four input signal classes x₁, x₂, x₃, x₄ are genuinely independent at the source. This is an empirical claim that must be validated in each domain of application.

2. The lower-triangular pipeline constraint is enforced at the database level, not merely the application level. Structural separation must be implemented in the data model, not only in the code.

3. The Cᴄᴇᴘ circuit breaker correctly identifies execution anomalies with sufficient precision to suspend write access when warranted. False negatives (missed anomalies) reduce the guarantee’s completeness.

Future work should address the following:

• Empirical validation of the independence assumption for x₁, x₂, x₃, x₄ in each target domain

• Comparative study of NCFCA against BKT, DKT, and IRT approaches on benchmark datasets, measuring diagnostic accuracy and schema floor stability under induced execution noise

• Formal specification of the Cᴄᴇᴘ anomaly detection protocol sufficient for ASIL-D certification in the automotive domain

• Extension of the prerequisite-gated DAG formalism to continuous-valued mastery representations

8. Conclusion

This paper has presented a structural proof that enforcing non-collapsible separation between environmental context, schema state, interface friction, and execution state — implemented as a strict lower-triangular data pipeline — eliminates inter-profile signal smearing by construction.

The proof establishes two formal results: (1) the lower-triangular pipeline constraint produces zero conditional covariance between schema state and execution state in the contaminating direction, and (2) the prerequisite-gated update rule of the Dynamic State Profile prevents execution anomalies from corrupting the confirmed schema floor during circuit-breaker suspension windows.

These results are domain-independent and have been instantiated in education, industrial operations, athletic training, and autonomous vehicle systems. The architecture provides a principled, formally verifiable foundation for adaptive systems that require reliable schema state estimation in the presence of environmental noise, interface friction, and physical execution variability.

The key claim of this paper is that the smearing problem is not a statistical problem requiring better inference algorithms. It is a structural problem requiring a different architecture. The NCFCA provides that architecture, and the proof demonstrates that it works.

About the Author

Gregory Stuart Lacefield is an independent researcher and systems engineer based in Las Vegas, Nevada. The framework described in this paper was developed over seven years of applied research in GED mathematics education in Florida’s correctional system — an environment that stripped away every conventional educational variable and forced the development of high-fidelity adaptive methods under extreme resource constraints. The framework has been formalized and extended to industrial, athletic, and automotive domains. Further documentation is available at gregorylacefield.com.

Contact

Gregory Stuart Lacefield

gregorylacefield.com

(702) 274-4299

glacefield87@gmail.com

Preliminary draft — May 2026 — Submitted for review and collaboration. Not for redistribution without permission.