Neural Rendering Composition and Granularity Tuning: State- and Task-Dependent Assembly of Multimodal Memory and Experience

A neural rendering, as proposed in Self-Aware Networks (SAN), is an evolving assembly of perceptual, remembered, bodily, conceptual, emotional and prospective relations, not merely a visual image. I propose Neural Rendering Composition and Granularity Tuning to distinguish which contents participate, how they are bound, which differences survive and how far an episode extends in time. The candidate contribution is a specific decomposition and measurement program, not the discovery of task-dependent cognition. Seven elementary propositions and a bounded twenty-statement Lean supplement clarify sufficiency, refinement, incomparability, observation and future-query requirements; restricted readout failure need not imply lost information. Twelve trained recurrent networks demonstrate differences among primary performance, broader access and cross-role reuse, but the matched-context models remain at chance on the intended temporal relation. A development-only diagnosis does not rescue that failure. A subsequent conventional active classifier learns coarse/fine queries and responses while retaining only encountered evidence. Across three fitted seeds and separate background sets, its second-query accuracy ties a fine-only learner at 0.857 while acquiring 1.961 versus 2.500 nominal new source bits. A complete-record control performs better, at 1.000, with greater acquisition cost. The relation is answered correctly when the needed earlier observation was retained and remains at chance when it was not. Assembly-only interventions alter responses without changing the source record, while equivalent attention and inverse-restoration controls tie exactly. These symbolic results make the proposed distinctions inspectable; they neither repair the recurrent implementation nor establish a new neural algorithm. The biological hypothesis concerns learned reception, temporary participation, relational organization, active sampling and returned consequences jointly configuring an episode. Published-method comparisons, independently collected measurements and causal evidence remain necessary. Mathematical consistency, conventional learned sampling and synthetic accuracy do not establish conscious experience. Preprint v1.0, 23 September 2026. The companion contains source, applications, frozen results, mathematical material, figures and bounded reproduction instructions. Negative and tied comparisons are preserved. Automated and AI-assisted checks are not independent scientific peer review. External raw inputs and private correspondence are not redistributed; consult the source-access and rights notices. CC BY 4.0 covers original prose, figures, documentation and results. Original software is provided for review without an added general reuse license; explicitly attributed third-party code retains its own license.