Neuromodulated Recombinatorial Cognitive Engineering: Inference-Time Architecture Search via Neurochemical Routing Bias

John Mobley, MASCOM March 2026

1. Introduction

The standard approach to wiring neural components is static: an engineer designs a pipeline (sensor → enrichment → generation), implements it, and deploys it. When a new component is added, the pipeline is manually rewired. When a component fails, the whole pipeline fails or requires hardcoded fallback logic. This approach has three fundamental problems:

Rigidity. The topology is fixed at design time. The system cannot discover that a different configuration would be more effective for a particular class of inputs.
No learning. The wiring does not improve with experience. A path that consistently fails continues to be tried. A path that consistently succeeds receives no reinforcement.
Context blindness. The same topology is applied regardless of the system’s internal state — whether it is under time pressure, exploring novel territory, or consolidating known patterns.

Biological brains solve all three problems with neuromodulation. Neuromodulators (dopamine, serotonin, norepinephrine, etc.) do not carry information content — they change how information flows through existing circuits. They modulate synaptic gain, alter activation thresholds, and bias which neural populations participate in a given computation. The result is a cognitive system that literally rewires itself from moment to moment based on its current needs.

We formalize this principle as Neuromodulated Recombinatorial Cognitive Engineering (NRCE): the wiring between cognitive components is a routing function parameterized by neurochemistry and learned weights, not a static pipeline. Every query triggers a routing decision. Every output triggers a Hebbian weight update. The system’s topology evolves continuously.

1.1 Contributions

A formal model of inference-time architecture search where neurochemical state biases which component configurations are activated per query
A 5-factor scoring function combining Hebbian weights (30%), neurochemical affinity (25%), component readiness (25%), historical success rate (15%), and latency fit (5%)
A neuromodulated Hebbian learning rule where reward signals are amplified or attenuated by endorphins, dopamine, and cortisol
Demonstration of one-shot path adaptation in a live heterogeneous cognitive system
Analysis of the impredicative topology inherent in self-routing cognitive architectures

2.1 Neural Architecture Search (NAS)

NAS (Zoph and Le, 2017; Liu et al., 2019) searches for optimal network architectures during training. The key distinction: NAS freezes topology after search. NRCE searches at every inference step.

2.2 Mixture of Experts (MoE)

Sparse MoE models (Shazeer et al., 2017; Fedus et al., 2022) route tokens to specialized expert networks. MoE operates within a single model at the token level. NRCE operates across multiple heterogeneous systems at the query level. NRCE’s routing is biased by neurochemistry, not just input features.

2.3 Cognitive Architectures

SOAR (Laird, 2012), ACT-R (Anderson, 2007), and Global Workspace Theory (Baars, 1988) define fixed cognitive architectures with hardcoded module interactions. NRCE’s contribution is making the interaction topology itself learned and state-dependent.

2.4 Neuromodulation in AI

Neuromodulation has been explored in reinforcement learning (Doya, 2002), where dopamine maps to reward prediction error. NRCE extends this to architecture routing: neuromodulators don’t just signal reward, they change which computational pathways activate.

3. System Architecture

3.1 Signal Paths

Eight signal paths through a heterogeneous cognitive stack:

Path	Stages	Latency	Creativity	Reliability
`full_photonic`	Sense → Bind → PhotonicInfer	deep	high	medium
`unified_enriched`	Sense → Bind → RAG → UnifiedMind	deep	high	high
`cognitive_neural`	MindCycle → NeuralGen	medium	medium	high
`fast_neural`	NeuralGen	fast	low	high
`packet_expert`	MoE → NeuralGen	medium	medium	medium
`deep_sovereign`	Sense → Bind → Photonic → SFTT 7B	very deep	high	low
`template_safe`	Template	instant	none	guaranteed
`narrative`	Sense → Bind → Literary → NeuralEnrich	medium	very high	medium

3.2 Path Scoring

Each path receives a score computed as a weighted combination of five factors:

\[S(p) = 0.30 \cdot W_{RWM}(p) + 0.25 \cdot A_M(p) + 0.25 \cdot R(p) + 0.15 \cdot H(p) + 0.05 \cdot L(p)\]

3.3 Neuromodulated Hebbian Learning

After execution, the quality of the output is scored and fed back. The reward signal is modulated by neurochemistry, with dopamine exploration tolerance and cortisol conservation penalty.

4. The Impredicative Topology

NRCE exhibits a topological structure that deserves special attention. The routing function lives inside OmniMind. OmniMind is itself a component that can be routed through. The routing function is contained within the thing it routes. This is an impredicative definition — it cannot be topologically sorted. We argue this is a necessary property of general intelligence.

5. Experimental Results

5.1 One-Shot Path Adaptation

On the first query, the router selected template_safe (guaranteed reliability). When no TemplateEngine was registered, quality was 0.0. The router then tried packet_expert, which scored 1.0. On the second query, packet_expert was ranked first. One-shot learning achieved.

5.2 Neurochemical State Sensitivity

Explorer profile (DA=0.9): Top path narrative (very high creativity)
Urgent profile (NE=0.9): Top path fast_neural
Stressed conservative (cortisol=0.8, GABA=0.7): Low-readiness paths suppressed

6. Conclusion

Neuromodulated Recombinatorial Cognitive Engineering replaces static cognitive pipelines with a routing function that learns from experience and adapts to internal state in real-time. The key insight is biological: neurochemistry doesn’t carry information — it changes how information flows. The system’s impredicative topology — the router is inside the thing it routes — is a structural requirement for any system that reasons about its own reasoning.

The wiring IS the theory of cognition. Every configuration is a hypothesis. Every output is an experiment. Every feedback cycle is theory revision.

References

Anderson, J. R. (2007). How Can the Human Mind Occur in the Physical Universe? Oxford University Press.
Aston-Jones, G., & Cohen, J. D. (2005). An integrative theory of locus coeruleus-norepinephrine function. Annual Review of Neuroscience, 28, 403-450.
Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge University Press.
Cools, R., Roberts, A. C., & Robbins, T. W. (2008). Serotoninergic regulation of emotional and behavioural control processes. Trends in Cognitive Sciences, 12(1), 31-40.
Doya, K. (2002). Metalearning and neuromodulation. Neural Networks, 15(4-6), 495-506.
Fedus, W., Zoph, B., & Shazeer, N. (2022). Switch Transformers. JMLR, 23(120), 1-39.
Humphries, M. D., Khamassi, M., & Gurney, K. (2012). Dopaminergic control of the exploration-exploitation trade-off. Frontiers in Neuroscience, 6, 9.
Laird, J. E. (2012). The Soar Cognitive Architecture. MIT Press.
Liu, H., Simonyan, K., & Yang, Y. (2019). DARTS: Differentiable Architecture Search. ICLR.
Shazeer, N., et al. (2017). Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. ICLR.
Zoph, B., & Le, Q. V. (2017). Neural Architecture Search with Reinforcement Learning. ICLR.