Cognia Causal Discovery Loop: Theory and Validation Results

Project: Cognia / DPSH

Document version: 1.0

Evidence status: original theory falsified; bounded core supported

End date of the reported series: October 14, 2026

Scientific verdict: ORIGINAL TEMPORAL-DAG THEORY FALSIFIED / BOUNDED DISCOVERY LOOP SUPPORTED / GENERALITY AND QUALIA UNRESOLVED.

Abstract

The Cognia Causal Discovery Loop (CCDL) began as the hypothesis that temporal activity in an asynchronous DPSH network could be mined automatically for causal motifs, composed into a Temporal DAG, and compiled back into the network. This strong version failed experimental validation. The Temporal DAG provided no held-out predictive benefit, mainly tracked the amount of activity, was not invariant to population renaming, and was substantially slower than a simpler control.

The falsification nevertheless exposed a narrower and practically useful mechanism. Revised CCDL observes activity within a bounded hypothesis space, proposes a candidate lag and map, freezes the hypothesis, and tests its causal meaning with a new intervention forked from a shared snapshot. In a controlled 16-channel DPSH task, this loop recovered the correct lag and all 16 hidden edges in eight out of eight independent replications. The frozen map reached 22.19% mean directional accuracy against a 6.25% chance level, and every pre-registered per-seed gate passed.

This is neither a general causal-intelligence algorithm nor evidence of qualia. It supports a bounded experimental protocol that, within a declared family of mechanisms, can distinguish an observational candidate from a causally supported model, reject a delayed common cause, track regime change, and keep test evidence out of hypothesis selection.

1. Research question

The central question was:

Can a system discover a structure in the internal activity of DPSH that not only correlates with observations but also predicts the outcome of a new intervention?

This formulation separates three levels of evidence:

  1. observational structure — a regularity in recorded activity;
  2. mechanistic activity — intervening on a mechanism changes system state;
  3. functional causal specificity — the change preserves a pre-declared meaning and beats matched controls on untouched data.

Correlation, decodability, and a large activity change are not causal discovery by themselves.

2. Original theory

The original CCDL proposed this path:

DPSH → Temporal DAG → Causal Temporal Motif → CDCT → compiled DAG → DPSH

Temporal motifs were expected to act as natural causal units. CDCT would levelize and compress them into modules suitable for subsequent network execution. The theory combined three distinct roles: DPSH as asynchronous network physics, the Temporal DAG as an observation engine, and CDCT as a representation compiler.

The strong prediction was falsifiable: the extracted structure had to beat simpler models on held-out data, survive activity-preserving controls, remain equivalent under renaming, and justify its computational cost.

3. Falsification of the Temporal DAG

A complex 256-neuron experiment found:

  • no held-out predictive benefit over a simple pairwise-delay model;
  • correlation between DAG activity and spike count of r = 0.730;
  • a richer DAG after activity-preserving shuffle, despite destroying causal meaning;
  • no specificity to an activity-matched edge-of-chaos regime;
  • representation failure under bijective population renaming;
  • an end-to-end analytical path 11.50 times slower than the equivalent task without a DAG.

Node motifs, directed edges, exact lags, flow across levels, and path completion failed in sequence. Further tuning on the same data would have become post-hoc optimization. Temporal DAGs, Causal Temporal Motifs, and their levelized composition were therefore removed from the active theory and kept only as legacy research.

This does not falsify CDCT as a general method for composing analytical expressions. It falsifies the assumption that a Temporal DAG supplies the right units from which an internal DPSH network should be constructed.

4. Revised theory

The surviving CCDL uses the following cycle:

bounded hypothesis space
        ↓
observational candidate proposal
        ↓
hypothesis freeze
        ↓
snapshot + paired control/intervention
        ↓
matched-control and functional gate
        ↓
unsupported / inconclusive / supported
        ↓
versioned hypothesis lifecycle

Here, discovery does not mean unconstrained search for arbitrary latent structure. The system searches within a declared family of candidates, such as lags, channel maps, edge groups, or named DPSH mechanisms. A model receives causal status only after it predicts the effect of an independent intervention.

The dual verdict is essential:

  • mechanistic verdict: the mechanism physically affects the system;
  • functional verdict: the effect fulfils its pre-registered function and beats the relevant controls.

An active mechanism may still be functionally useless. CCDL therefore uses fail-closed promotion: missing, inconclusive, or conflicting evidence cannot be rewritten as support.

5. What the experiments supported

5.1 Reproducible experimental apparatus

Deterministic snapshots, control/intervention forks from shared state, frozen and fresh randomness, matched controls, pre-registration, provenance, and train/validation/intervention-test separation all work. Negative and inconclusive results can be published without automatically turning them into support.

5.2 Natural visual prerequisites in DPSH

DPSH represented factored colour, edge orientation, and their binding across eight natural stimuli. Colour-path and edge-path ablations produced the required double dissociation. Nonlinear colour geometry was linearly readable from internal state, and explicit chromatic adaptation before the dynamics substantially reduced error.

A larger spatial grid showed that neighbouring pixels and phase redundancy stabilize the representation of a colour field better than an isolated pixel. An oscillator reduced spike fatigue but lowered accuracy in the tested form; it is therefore not supported as a general solution.

These results establish computational and mechanistic prerequisites for processing colour and edges. They are not evidence of subjective colour experience.

5.3 Features without internal numeric coordinates

The network did not require explicit x and y values. Motion transitions created an observational graph geometry, and distributed features could be used causally without an internal numeric map. The observer still knew the physical coordinates during the final audit. The result therefore supports causal use of relational features, not the claim that the network created a phenomenal space.

5.4 Representation alignment

Linear structure was discovered independently of the sensory basis. Direct weight transfer across bases failed, but synchronized experience made it possible to re-identify the corresponding mechanism. Shuffling the temporal anchors destroyed the benefit, while a robust median/MAD gate rejected false anchors automatically. The principle also worked on actual DPSH spike windows and physical bursts.

The supported claim is representation alignment through shared events, not universal code identity across arbitrary sensory bases.

5.5 Functional specificity of a physical path

The magnitude of target change alone did not distinguish the correct projection from a shuffled one. When the held-out task required channel identity to be preserved, however, the aligned path reached 24.22% accuracy, versus 6.10% for shuffled and 7.28% for ablated, with a 6.25% chance level. The causal meaning of topology is visible only in a metric that preserves the meaning of the information, not merely perturbation energy.

5.6 Minimal closed discovery loop

In M-26, observational discovery without access to the hidden topology recovered all 16 edges, and the frozen map predicted new interventions with 23.97% accuracy.

M-27 jointly searched for an unknown lag in 0…6 and a map in overlapping multichannel activity. It found the correct lag 3 and all 16/16 edges. The same observational pattern was generated both by a physical path and by a delayed common cause. Only the source-only intervention separated them: effect 0.0480809 for the physical path and exactly zero for the common cause. This implemented the minimal cycle:

observe → propose → freeze → intervene → validate

5.7 Regime change and hypothesis lifecycle

The loop was able to:

  • retire a stale lag and map after stable new evidence without a change marker;
  • suspend a hypothesis during a temporary outage and reactivate it after return without relearning;
  • track gradual drift with continuous confidence instead of a false binary switch;
  • discover a new map not already present in the archive;
  • reject an unnecessarily complex model through a validation gate;
  • keep the final intervention test untouched by model selection.

5.8 Replication across seeds

The final closed cycle was repeated across eight independent DPSH states, random streams, lags, and hidden maps. The correct lag and all 16 edges were recovered in 8/8 runs. Mean intervention accuracy was 22.1924%, the 95% Student interval was 21.0645…23.3203%, the worst seed reached 19.3359%, and the highest ablated control reached 7.4219%. Every pre-registered gate passed.

6. What did not work

6.1 Temporal DAG as an observation engine

The main original mechanism was falsified. Levelization and more precise lag mining did not remove activity dependence, lack of invariance, absence of held-out benefit, or computational overhead.

6.2 Prediction error as automatic credit assignment

The local prediction-error channel was causally active and, after a runtime fix, genuinely modulated plasticity. It still failed to improve complex BQ-07 transfer. Preserving the error sign, adding local eligibility, using the true task outcome, one-shot per-edge counterfactuals, and hierarchical group-to-edge selection all failed to produce a robust held-out benefit. The mechanism changed weights, but those weight changes were not the right credit for generalization.

The practical conclusion is outcome != credit. A correct error or reward does not identify which local change caused success.

6.3 Direct transfer across sensory bases

Discovery found equivalent subspaces, but the same weights did not transfer to another basis without alignment. Structural similarity of representations is not identity of their coordinate systems.

6.4 Observed lag as causality

A correctly identified temporal lag can be produced by a delayed common cause. M-22 found the lag but did not establish a functional causal margin; M-23 showed zero source-only effect. Only intervention distinguished a real path from common cause.

6.5 Unrestricted autonomous feedback

Experimental results are not automatically fed back into DPSH to rewrite the network. A safe self-modifying loop still lacks a separate theory, validation gates, and evidence beyond the tested task family. The current system is an experimenter and hypothesis manager, not an autonomous architect of its own mind.

7. What remains open

Current support is limited to known candidate families, especially 16-channel cyclic permutations, sparse multichannel activity, linear mappings, and a bounded lag range. The following have not yet been established:

  • discovery among an unknown number of areas without predefined source and target populations;
  • nonlinear, recurrent, and simultaneously active mechanisms of different complexity;
  • generally calibrated complexity penalties and intervals for broad search;
  • long-running operation with many concurrent hypotheses and a limited probe budget;
  • safe automatic changes to topology or learning rules;
  • any relationship between the discovered mechanisms and consciousness or qualia.

8. Epistemic limit: qualia

CCDL validates a discovery loop, not subjective experience. Colour geometry, colour-edge binding, spatial coherence, decodability, and causal use of a feature are measurable functional phenomena. Even together they do not provide an independent test of whether the system experiences anything.

The warranted statement is:

DPSH contains mechanisms capable of creating and causally using some structures required for visual representation, and CCDL can discover and interventionally validate their map in a bounded domain.

It would be unwarranted to claim that the experiments demonstrated colour or spatial qualia.

9. Practical impact on Cognia design

The active implementation belongs to DPSH because it operates on DPSH asynchronous physics, snapshots, and interventions. CDCT is not the discovery engine. It may only be used after independent validation as optional compression of supported quantities.

Applications should therefore:

  1. declare the functional outcome before experimentation;
  2. separate outcome from credit assignment;
  3. create candidates only on training data;
  4. select complexity on validation data;
  5. freeze the hypothesis before opening the intervention test;
  6. use snapshot-fork, matched controls, and multiple seeds;
  7. keep mechanistic and functional verdicts separate;
  8. archive supported, suspended, falsified, and inconclusive hypotheses;
  9. never feed results back into the network automatically;
  10. publish negative findings and the boundary of validity.

10. Final conclusion

The original Temporal-DAG-based CCDL is a closed research branch. Revised CCDL, however, is not an empty remainder of the original theory. It survived experimentation as a bounded causal scientific cycle over DPSH:

observation proposes a hypothesis;
independent intervention determines causal status;
new evidence controls lifecycle rather than rewriting history.

Its strongest demonstrated contribution is not automatic construction of an internal network but epistemic discipline: the system can separate correlation from intervention, activity from function, a temporary outage from a changed mechanism, and data fitting from an independent test.

This is a meaningful positive result for designing experimental DPSH applications, but it remains a narrow foundation. Generalization to open environments, complex nonlinear mechanisms, and autonomous network change is future research.

Related documents