Beyond the reducing valve: towards a computational neurophenomenology of altered states via deep neural networks.

Altered states of consciousness, including hallucinations, psychedelic experiences, and ego dissolution, differ qualitatively, yet no unified computational framework describes what varies and along which dimensions. Computational phenomenology (CP) has emerged as a promising bridge between first-person experience and computational models, yet current formalisations rely predominantly on the free energy principle (FEP). This paper proposes the C × G × D framework, drawing on three functional roles in deep neural networks: a Classifier (C) that extracts features from sensory input, a Generator (G) that synthesises internal representations, and a Discriminator (D) that judges whether a representation originates externally or internally. Phenomenological differences across altered states are redescribed as variations in the objective functions, constraints, and thresholds of these components. The framework reformulates Huxley's 'reducing valve' metaphor: relaxation of C's constraint exposes normally hidden 'effective causes,' producing psychedelic geometric patterns; G's prior governs hallucinatory veridicality; and D instantiates Perceptual Reality Monitoring. Three hallucination mechanisms-psychedelic, neurodegenerative, and schizophrenia-type-are predicted from distinct parameter configurations. Testable hypotheses derived from iterative-optimisation psychophysics and an extension to ego dissolution are presented. By foregrounding the plurality of objective functions and architectures, the C × G × D framework complements FEP-centred CP and provides a scaffold for translating phenomenology into experimentally manipulable variables.