DPSH Bottleneck: Theory, Falsification Program, and Experimental Results

Project: Cognia / DPSH
Document version: 1.0
Evidence status: partially falsified; one mechanistic candidate survived
Cutoff date for the reported series: October 13, 2026

Abstract

This work investigates whether adaptive selective communication bottlenecks between recurrent neural centers can produce stable and functionally useful representations. A bottleneck is not necessarily a narrow layer with few neurons. It is an interface that restricts or selects effective communication paths, controls their permeability, and may learn to translate the distributed state of a source center into a code usable by another center.

The experimental program did not demonstrate qualia or generally confirm the theory. It falsified several specific mechanisms: retention in the original topology, untrained internal generation, symmetric excitatory feedback, and several local-learning routes to a transferable representation. One narrower prediction survived locked confirmation: selective permeability control changed global content availability more than activity-matched nonspecific suppression while the local representation remained decodable. The resulting status is PARTIALLY FALSIFIED / H5 CANDIDATE SUPPORTED / QUALIA UNRESOLVED.

1. Epistemic limitation

There is no generally accepted independent test of subjective experience in an artificial network. The experiments reported here can support or falsify mechanistic predictions about retention, availability, causal transfer, autonomy, and learning. They cannot confirm or directly refute the existence of qualia.

Terms such as Global Workspace, perceptual center, and feature language are operational architectural terms, not labels of consciousness. Decodability is evidence that information is present; by itself it is not evidence of experience. Correlation is not causal transfer, and a firing-rate change is not automatically a content change.

2. Theory

2.1 Recurrent centers

The system contains several functionally specialized centers. Each center has a distributed state, local recurrence, its own temporal dynamics, and possibly local plasticity. One center may be the central DPSH Workspace; others may represent sensory properties, memory, action, or salience.

The functional dynamics of a center can be written as

$$ S_i(t+\Delta t)=F_i\left(S_i(t),X_i(t), {Z_{ji}(t-\tau_{ji})}_j,W_i,\xi_i(t)\right), $$

where S_i is the center state, X_i its input, Z_ji messages from other centers, W_i local synaptic parameters, and ξ_i a stochastic component.

2.2 Adaptive bottleneck

A directional bottleneck B_ij translates the state of center i into a message for center j:

$$ Z_{ij}(t)=B_{ij}\left(S_i(t),S_j^{feedback}(t), \beta_{ij}(t),\Theta_{ij}\right). $$

β_ij is operational permeability and Θ_ij contains interface parameters. Every direction must be independently measurable, switchable, and auditable. Lower β is not identical to lower information capacity, and fewer neurons do not automatically make a better bottleneck.

2.3 Feature language

The theory does not assume “one neuron equals one property.” A transmitted code may be local, population-based, temporal, phase-based, or mixed. A useful feature language must satisfy more than correlation:

  1. the property is decodable on held-out data;
  2. its meaning survives a context or physical cue-style change;
  3. intervention on the channel specifically changes the receiver's state or performance;
  4. the code supports previously unseen property combinations;
  5. its benefit exceeds capacity-, activity-, and compute-matched controls.

2.4 Local and global learning

A candidate local rule has the form

$$ \Delta w_{ab}^{(i)}=\eta_i e_{ab}^{(i)}(t)m_i(t) -\eta_{h,i}h_{ab}^{(i)}(t), $$

where e is an eligibility trace, m is a local or globally modulated learning signal, and h is a homeostatic correction. STDP alone does not guarantee an invariant code. Global modulation is likewise not automatically a gradient or a solution to credit assignment.

3. Falsifiable predictions

The program primarily tested the following mechanistic predictions:

  • H4 — retention: information survives temporary closure of a communication interface and is specifically recovered after reopening.
  • H5 — selective availability: changing permeability affects global availability without necessarily destroying local content, and the effect cannot be explained by overall activity alone.
  • H6 — internal generation: after input removal, a recurrent center produces a new content-identifiable state that causally depends on recurrence.
  • H7 — bidirectional integration: content-aligned feedback outperforms an equally sized shuffled feedback path.
  • H8 — transferable representations: local learning creates a code that generalizes to a physically novel cue style and beats frozen and simple baselines.

Version 0.3 additionally proposed H11–H15: emergent communication language, stability under continuing plasticity, benefit from limited global modulation, compositional generalization, and functional autonomy of centers. The reported series did not comprehensively test these broader hypotheses.

4. Methods

The experiments separated pilot, calibration, and confirmation seeds in advance. Branches were forked from a shared snapshot. A failed activity-match control invalidated an experiment rather than turning it into a negative result. Readouts were passive and could not change the event log. Learning was frozen before validation and test, and weight changes required exact accounting in an audit log.

Positive known truths verified metric sensitivity. Controls included shuffled labels and edges, zero precision, recurrence ablation, explicit memory, activity-matched global suppression, and a frozen runtime. Results were assessed with paired comparisons across seeds and 95% Student confidence intervals.

5. Results

5.1 H4: retention in the original topology — mechanism falsified

Across twenty seeds, a post-reopen difference existed: 0.140075, 95% CI [0.023084, 0.257066], with matched Workspace rate. Direct measurement of the presumed information carrier during closure found no retention: 0.013326 [-0.059435, 0.086087]. Content was also absent from integrated population rate (0.014274 [-0.026013, 0.054560]) and spike precedence versus a label-shuffle surrogate (0.024294 [-0.031374, 0.079963]). Silencing the entire downstream history did not change the post-reopen effect (0.006913 [-0.024260, 0.038085]).

Verdict: REJECTED_MECHANISM. A downstream difference existed, but it was not evidence that local content survived. This falsifies the tested H4 topology, not every possible memory mechanism.

5.2 H5: selective availability — supported candidate

The confirmation protocol was locked before results from seeds 80–99 were loaded. It compared a selective bottleneck with pre-calibrated global suppression under an activity-match tolerance of 0.08 and with no weight changes.

Permeability β Activity error Bottleneck − matched, 95% CI
0.75 0.046226 −0.259721 [−0.281713, −0.237728]
0.50 0.055594 −0.156019 [−0.173355, −0.138682]
0.25 0.061440 −0.159426 [−0.168103, −0.150748]

The local content effect remained 0.549345 [0.538138, 0.560553], and both receiver modules showed the same direction of effect. Nonspecific suppression with the same coarse activity preserved more available content than selective edge restriction.

Verdict: SUPPORTED_H5_CANDIDATE. This is replicated evidence for a specific mechanism, not evidence of consciousness. The control matched population firing rate but not the complete ISI and spectral statistics; a stricter replication remains appropriate.

5.3 H7: symmetric excitatory feedback — specificity falsified

Aligned and shuffled feedback each had 192 edges with matched sources, target populations stripped of content labels, weights, delays, and polarity. After independent dose calibration, confirmation on seeds 121–140 produced:

aligned − shuffled Workspace = 0.025288 [-0.003606, 0.054182]
module two transfer           = -0.007971
module three transfer         = -0.017263
activity error                = 0.019833

Feedback causally changed the network, but it was neither more content-specific nor more useful than a matched shuffled path. Verdict: REJECTED_H7_SPECIFICITY. Symmetric excitation alone is insufficient; future interfaces require learned, signed, or otherwise structured feedback.

5.4 H6: internal generation — pilot rejected

After stimulus removal, the experiment measured identity, novelty, blank specificity, and causal dependence on recurrence. Their joint score was the minimum rather than the mean, so noisy novelty could not conceal missing identity.

identity                = 0.038133 [-0.114894, 0.191161]
novelty                 = 0.201507
blank false positive    = 0.889640
recurrence dependence   = 0.013743
joint score             = 0.013743
explicit-memory control = 0.062984 [0.030468, 0.095500]

The positive control passed, but the tested network generated novelty without reliable identity. Verdict: REJECTED_PILOT for the symmetric untrained topology.

5.5 H8: transferable representations — tested family unsupported

The task enforced train/freeze/test separation, checked leakage, and held out an entire physical cue family. A one-seed STDP candidate failed to replicate on 12 fresh seeds:

none mean = 0.500
STDP mean = 0.500
paired difference = 0.000 [-0.0479, 0.0479]

A marginal temporal predictor, relational predictor, single-population error code, and opponent error code also failed. Final opponent confirmation on 48 fresh seeds produced:

frozen mean = 0.505208 [0.490312, 0.520105]
active mean = 0.500000
paired difference = -0.005208 [-0.036042, 0.025625]
positive / zero / negative seeds = 7 / 29 / 12

The opponent module was physically connected correctly, produced a signed local state, and causally changed the Workspace, but did not create a robustly transferable relational representation. Verdict: OPPONENT_TRANSFER_NOT_CONFIRMED. H8 remains open in general, but this mechanism family is closed without post-hoc tuning.

6. Overall verdict

Area Verdict Scope
H4, retention Mechanism falsified original tested topology
H5, selective availability Supported candidate three β values, 20 confirmation seeds
H6, internal generation Pilot rejected symmetric untrained recurrence
H7, feedback Specificity falsified symmetric excitatory topology
H8, transfer Not confirmed STDP and tested prediction-error variants
H11–H15 Not evaluated broader v0.3 program
Qualia Unresolved no independent test exists

The bottleneck idea as a whole was therefore not falsified in the same manner as DPSH-DAG. Several concrete implementations were falsified. H5 instead provided replicated mechanistic evidence that selective restriction of communication paths can alter availability differently from equally strong nonspecific attenuation.

7. Design consequences

  1. Recurrence alone is insufficient: without a learned attractor it produces novelty without identity.
  2. Symmetric feedback alone is insufficient: a causal effect need not be content-specific.
  3. Unsupervised STDP alone is insufficient: thousands of weight changes are not evidence of a useful code.
  4. Error populations require a verified physical path, but correct connectivity still does not guarantee generalization.
  5. Local content and global availability must be measured separately.
  6. Every future module needs independently switchable directions, snapshot forks, change auditing, matched controls, and read-only evaluation.

8. Open questions

Further publishable work requires new pre-registered experiments and untouched seeds. Priorities are a stricter rate+ISI+spectrum control for H5, learned signed interfaces instead of symmetric feedback, an attractor or generative model for H6, and a genuinely identifiable compositional task for H8/H11–H15. Existing test data must not be used both to select a new variant and to confirm it.

9. Reproducibility

The canonical validation suite exited with code 0. Legacy Cognia models passed 19/19, EOC validation passed 9/9, experimental gates BQ-01 through BQ-07 executed without technical failure, and git diff --check passed. Negative scientific verdicts are valid outputs of the falsification program, not test-suite failures.

Detailed protocols, intermediate results, and implementation history are available in EXPERIMENT_LOG-cz.md and IMPLEMENTATION_PLAN-cz.md in the same directory.