ST-T1024 FOR IMAGES AND VIDEOS : THE ARCHITECTURE. On Image and Video Composition from a Declared Basis, Disturbance Rejection by Refusal, and the Creative Act That Stays With the Human

This paper states an architecture for image and video composition that is bounded byconstruction, deterministic in its output, and attributable in its record. It does not propose a better image generator. It proposes that generation is the wrong frameentirely. Generation is the creative act, and the creative act belongs to the human. What the architecture provides is the construction layer: a finite basis of visual primitives, a declared composition grammar, and a deterministic engine that enumerates compositions satisfying a human-authored specification. The loop has four steps. The human specifies the desiderata. The machine proposes thecompositions that satisfy the specification. The human remarks on the proposals. Themachine narrows the set by conjunction. Each remark is a constraint; each constraintreduces the satisfying set; the set converges monotonically. The properties transfer from the text architecture. No non-compliance: the composition either satisfies the specification or is refused with 0x-AMBIG. No excess energy or water: the same deterministic, MCU-class profile. No AlCapone-LLM: every composition traces to a specification, a basis, a grammar, a signer, and a selector. The disturbance model is stated. Inputs are perturbed, injected, drifted, and replayed. Theengine does not correct. It refuses. Disturbance rejection by exclusion: the system cannot be pushed outside the envelope, because leaving the envelope is not in the action library. The timing bound is deductible, not measured. AET = D × T_tile + T_triage for serialcomposition; AET = max(T_tile) + T_triage + T_merge for parallel. Every term is a declared constant. The bound follows from the declaration, and the FSM enforces it. Four residuals are named. The basis bounds what can be expressed. The parallel FSM isnot yet specified. The power profile changes under parallel execution. And if a trainedfeature extractor is used anywhere in the pipeline, the no-training claim is broken, and the paper says so in the first paragraph of the relevant part.