dynamicsystemsarchitecture.org

NCFCA Formal Specification — Version.001

State Separation and Causal Write Ordering in Multi-Source Estimation. Dated June 27, 2026.

PurposeThe original formal specification of the architecture, stated as a control-theoretic estimation problem.
StatusSuperseded by Version.002. Version.001 is never edited; a revision is a new, separate version.
ProvenanceReproduced as received on October 4, 2026. Wording unchanged; only markdown formatting (headings, code blocks, tables) was applied for display.
EvidenceA specification, not a test. No outside party has tested or reviewed it.

NCFCA: Formal Specification

State Separation and Causal Write Ordering in Multi-Source Estimation

Status: Control-theoretic grounding, defensible claims, prior-art aware Author: Gregory Stuart Lacefield Date: June 27, 2026

1. THE REAL PROBLEM (Precise Statement)

Given

The Failure Mode

A naive estimator tries to recover X from a mixture:

ŷ(mixed) = αX + βη_system + γη_interface + δη_execution

Where the estimator treats this as:

X̂ = f(ŷ(mixed))

The problem: ŷ(mixed) is non-identifiable. Multiple distinct X values could have produced the same ŷ(mixed) given different noise compositions.

Result: Downstream estimates of X are biased and contaminated by unrecoverable correlation between components.

Why This Matters

In real systems (education, justice, medicine, hiring):

2. THE STRUCTURAL FIX (Minimal Rule Set)

Design Principle

Separate causal origins BEFORE inference, not after.

Four Required State Variables

S : context/environment state
D : inferred latent capability
I : interface/measurement design parameters
C : execution output trace

Each is maintained separately, with separate update logic.

Rule A: Channel Separation (Write Isolation)

Each channel has exclusive write authority from specific sources:

Update rule for S:
  S(t+1) ← f_S(S(t), external_environment_observations)
  No other channel can write to S

Update rule for D:
  D(t+1) ← f_D(D(t), S(t), d_channel_probe)
  C observations do NOT directly update D
  Only filtered summary through circuit breaker can update D

Update rule for I:
  I(t+1) ← f_I(I(t), S(t), D(t), interface_diagnostic_observations)
  Again: C cannot directly update I

Update rule for C:
  C(t+1) ← g(S(t), D(t), I(t), execution_event)
  C is observable, not a state variable to estimate
  C is logged for diagnostics

The key constraint: C does NOT create a feedback loop into D or I except through the gate function (Rule D).

Rule B: Directional Update Constraint (DAG Property)

Update dependencies must respect a partial order:

Time flows left → right:

S(t) ──→ D(t) ──→ I(t) ──→ C(t)
 ↓        ↓         ↓
 S(t+1)   D(t+1)    I(t+1)    C(t+1)

Forbidden updates:

Allowed updates:

Rule C: Observation Channel Specification

Each channel has designated observation sources:

S-channel observations:  environmental instruments
  (temperature, noise, time-of-day, socioeconomic context)

D-channel observations:  direct capability probes
  (knowledge tests on fixed content, work samples, assessments)
  These must be designed to isolate D from S, I

I-channel observations:  interface diagnostics
  (response time variance, semantic drift, pattern frequency)

C-channel observations:  execution events
  (test scores, decisions, outputs, behavioral traces)

Design requirement: D-channel probes must be insensitive to S and I variations. This is hard engineering (the test design itself must separate channels), but that's the point.

Rule D: Gate Function for D-Channel Updates

Updates to D are gated:

D(t+1) ← D(t) + α · Δd_t,  where Δd_t passes filter φ

where:

φ : (d_probe, history, telemetry) → {pass, hold, revert}

Specifically:

telemetry ∈ {latency(t), semantic_distance(t), pattern_frequency(t)}

If telemetry suggests I-saturation:
  φ returns "hold"
  D stays at schema floor S_floor(D(t-k)) for some k

Else:
  φ returns "pass"
  D updates normally

This is your circuit breaker, but as a formal gate function

3. What This Achieves (Formally)

Observation 1: Causal Identifiability

By enforcing write isolation, we preserve observability of the causal structure:

Observation 2: Information Preservation

The original information that was "lost" in ŷ(mixed) is now visible in:

These residuals are now diagnostic, not corrupting.

Observation 3: Bias-Variance Tradeoff

The tradeoff is favorable because bias was destroying actionable information.

4. What This DOES NOT Claim (Precise Boundaries)

Claim: We achieve "zero cross-channel effect"

False. In reality, S affects D (environment enables learning), D affects performance C. These causal relationships exist and are not denied.

What we claim: We prevent these causal relationships from corrupting the estimate through estimation-stage contamination. The causal effects are present in the observable residuals, where they're useful for diagnosis.

Claim: "Four channels are categorically independent"

False. In any human system, S and D are correlated (environment enables capability development). What we require is that the estimator doesn't treat correlation as evidence of identity.

Claim: "Information lost in collapsed measurement can be fully recovered with separation"

False. If measurement apparatus is fundamentally non-identifiable (all channels observe through identical sensor), no architecture fixes that.

What we claim: If measurement apparatus can be designed to separate channels (which is hard engineering, not mathematics), then separation is recoverable.

5. Comparison to Known Frameworks

Distributed Kalman Filtering

Our structure is similar to:

Difference: We enforce causal ordering constraints that standard Kalman fusion doesn't require.

Causal Inference (Pearl/SCM)

Our DAG constraint (Rule B) is similar to:

Difference: We also enforce write isolation (not just causal order)

Multi-Model Estimation (MME)

Our approach resembles:

Difference: We enforce exclusive write paths (no model blending)

Byzantine Fault Detection

Our circuit breaker (Rule D) is similar to:

Difference: We use channel-specific diagnostics, not generic outlier detection

6. Where the Architecture Actually Breaks (Stress Testing)

Case 1: Feedback Loops in Reality

What if the true system has feedback?

This is a true feedback cycle in human learning.

Our architecture: DAG constraint forbids this.

Reality check: The cycles do exist. Our separation assumes feedforward structure, which is false.

Fix: Allow cycles but with time-delay constraints. D(t+1) can depend on C(t), I(t) (past values), as long as they're delayed. This trades perfect separation for realistic dynamics.

Case 2: Unobservable S-Channel

What if environmental state is not fully observable?

Our S-channel update rule assumes we can measure the environment. In practice (classroom noise, socioeconomic context), some S components are latent.

Fix: Treat S like we treat D — infer it from residuals. This creates a dual-estimation problem.

Case 3: D-Channel Probe Design Failure

What if we cannot design a D-channel probe that's insensitive to S and I?

If every assessment of capability is corrupted by anxiety, noise, or time-of-day effects, then D-channel separation is not achievable in practice.

This is the hardest case. It's not a mathematical problem; it's an engineering problem.

Mitigation: Use multiple independent D-probes and vote. If all probes agree despite different S/I conditions, confidence in D is higher.

7. Formal Definition Summary

State Space

X(t) = (S(t), D(t), I(t), C(t))

where:
  S : ℝⁿ (context state)
  D : ℝᵐ (capability estimate)
  I : ℝᵖ (interface parameters)
  C : ℝᵍ (output)

Update Equations

S(t+1) = f_S(S(t), obs_S(t))

D(t+1) = D(t) + α · φ(d_probe(t), history, telemetry(t)) · ΔD

I(t+1) = f_I(I(t), S(t), D(t), obs_I(t))

C(t+1) = g(S(t), D(t), I(t), action(t))

Constraints

1. Write isolation:
   S written only by f_S
   D written only by gated update (no C → D direct)
   I written only by f_I
   C is observation (no equation for C(t+1) as state update)

2. DAG constraint:
   Let E_dep = {(X_i, X_j) : X_j depends on X_i in update equation}
   Requires: E_dep is acyclic

3. Gate function:
   φ : {0, 1} → {close, open}
   Closes if telemetry exceeds thresholds

4. Observability:
   C(t) = H · (S(t), D(t), I(t)) + η(t)
   Requires: H has full rank (all hidden states observable through C)

8. The Real Insight (Non-Technical Summary)

The breakthrough here is not "new mathematics."

It is:

Organization matters more than algorithm.

Even a naive estimator works well if information is organized correctly at the input level. Conversely, a sophisticated estimator fails if contaminated information is fed to it.

The hard part isn't the math. It's the engineering:

Most real systems fail not because the estimator is bad, but because the input stage mixes channels.

This structure forces the input stage to do the hard work.

9. Next Steps (If Taking This Seriously)

For Theory

For Applications

For Implementation

10. Relationship to Original NCFCA Claim

The original claim was:

"Signal smearing is architectural, not statistical. Separated channels are 6-7x better."

Tightened claim (this document):

"If you can design measurement systems that separate causal origins, estimation accuracy improves significantly. The improvement is bounded by observability, not by the estimator algorithm."

The original claim oversold. This claim is grounded.

References (Grounding in Known Theory)

Causal Inference:

Control & Estimation:

Multi-Agent Estimation:

Byzantine Fault Tolerance:

Appendix: Example (Education Domain)

The Problem

Student takes a math test. Score is 65.

Naive interpretation: Student understands math at 65% level.

Reality: Score = f(understanding, classroom_noise, anxiety, sleep, test_format, ...)

All mixed.

The Corrected Structure

S-Channel: Classroom conditions
  - Noise level: 72 dB (moderate)
  - Time: 9 AM (optimal)
  - Distractions: 2 (low)

D-Channel: Direct probe on isolated content
  - Give student problem-solving task in quiet room
  - No time pressure
  - Result: 78% correct
  - Confidence: high

I-Channel: Interface diagnostics
  - During actual test: Response time spikes 40% on word problems
  - Semantic distance from prior assessments: 0.3 (moderate drift)
  - Pattern frequency: matches "anxiety under time pressure" profile

C-Channel: Observed output
  - Test score: 65%
  - Written work: shows knowledge gaps in word interpretation
  - Behavior: rushed, multiple erasures

Diagnosis: D ≈ 78% (true understanding), but C = 65% because I-channel (time pressure anxiety) degraded execution.

Action: Address I-channel (anxiety management, test format adaptation), not D-channel (understanding is fine).

This is only possible with separation.

End of Formal Specification

This is the grounded, defensible version. The claims are tighter, the boundaries are clear, and it connects to known theory.

The real work is in the applications and in finding where it breaks.