Indoor lighting influences cognition, affect, and behavioral regulation, yet these effects are typically studied as isolated empirical findings rather than as components of a unified computational process. This paper proposes an active inference account of the impact of shared indoor lighting-non-personalized illumination in environments occupied by multiple users such as offices, classrooms, and libraries. The central hypothesis is that lighting acts through three computationally distinct channels, each indexed by a metrologically separable descriptor of the light field rather than by correlated color temperature (CCT): an epistemic channel driven by task-plane photopic illuminance, an affective channel driven by ambient melanopic equivalent daylight illuminance (melanopic EDI; CIE S 026:2018), and a normative channel driven by chromaticity. Because melanopic EDI and chromaticity are dissociable by metamerism, and because the affective input is a whole-field, session-integrated quantity whereas the epistemic input is local and instantaneous, the three channels are separable both in computational role and in physical driver. To formalize and test this hypothesis, the paper develops a proof-of-concept partially observable Markov decision process (POMDP) under the active inference framework. The model simulates a cognitive agent performing sustained reading over a 5-h session, receiving observations from two modalities simultaneously: reading performance and eye tracking. Text legibility is a symmetric function of illuminance; the oculomotor signal is asymmetric, degrading rapidly under glare due to pupil constriction, and flat under sub-optimal illuminance. Three factors are crossed in a full factorial design: lighting scenario (warm-dim, moderate, and bright-cool), chronotype (morning, intermediate, and evening), and spatial desk position in a realistically specified room with a simulated photometric field. Six falsifiable predictions are derived from the model structure. These six predictions fall into three tiers by evidential status-structural invariants that verify the implementation (P1, P2, and P6), the critical test of the multi-channel hypothesis (P3, P4), and one exploratory prediction (P5)-and are borne out in simulation at N MC = 20 replicates. The central prediction concerns the spatial ordering of performance under intense cool illumination. Under moderate light, performance follows the legibility gradient: the best-lit desk position produces the best output. Under intense cool light, this ordering completely inverts: the desk position with the best legibility in both channels-receiving only 402 lux at the far corner-produces higher reading speed (71.3 wpm) than the position with near-zero legibility in both channels under 1,200 lux of glare (59.9 wpm). The inversion arises from two concurrent mechanisms: the well-lit agent accurately perceives its deteriorating state and rests strategically, while the glare-blinded agent works blindly through its deterioration but pays a direct mechanical speed penalty. In this condition, strategic rest at the far-corner position yields higher total output than blind persistence under severe glare. This prediction requires the multi-channel architecture and is not expected under models in which lighting acts only through arousal or only through monotonic legibility. Additional predictions include a chronotype-specific risk sign-reversal between scenarios, a position-by-chronotype interaction differing by a factor of two between morning and evening types, and a structural invariance of the intermediate chronotype's EFE risk across all scenarios as a direct consequence of setting its affective tolerance factor to zero. The contribution of the paper is theoretical: it does not provide a calibrated model of human reading performance, but a computational framework that links empirical findings on lighting to an explicit generative architecture with testable, non-trivial signatures. In short, the paper offers a computational theory of light as a modulator of inference, arousal, and behavior, and translates it into concrete experimental predictions.
