AI systems can now produce striking media from a prompt, but prompt-native generation remains unreliable in the places production depends on: exact text, correct brand color, accurate product facts, stable timing, inspectable decisions, and identical output on a re-run. This paper focuses on motion-graphics video: typographic, data-driven, brand-governed, composited, and procedural video artifacts rather than cinematic live-action synthesis. Template systems deliver reliability in this domain, but only by restricting authors to pre-designed forms. This paper argues that the tradeoff is architectural, not fundamental. The failure mode is a misplaced authority boundary: the model is allowed to propose, decide, and finalize the artifact. We present a compiler-gated architecture for AI-authored deterministic media. A model participates in authoring the video artifact rather than synthesizing its final pixels. The durable artifact is not the prompt or the rendered MP4; it is editable source. A language model proposes creative intent, including mood, narrative, copy, and emphasis, in a compact domain-specific language. A compiler, not the model, authorizes syntax, semantic validity, typography, palette, layout, timing, motion, and renderer inputs. The compiler emits a renderer-neutral intermediate representation, the RenderGraph, which can be inspected, diffed, versioned, approved, and rendered reproducibly. The central contribution is an artifact authority model for AI-authored media: models may propose, but deterministic systems authorize. The compiler gates the model on both sides. Upstream, language design bounds what can be expressed. Downstream, validation, policy checks, and quality gates reject source that is syntactically invalid, semantically incoherent, off-brand, or weak but valid. We call this bidirectional placement a compiler-gated sandwich architecture. This places the work in the neuro-symbolic tradition: neural systems propose candidates, while symbolic systems provide representation, reasoning, validation, and auditability. The result is video that is AI-authored, deterministic after source acceptance, brand-governed, editable, and auditable. We formalize the authority boundary, describe the Workshop implementation, and evaluate the system by separating control gained from expressiveness retained. The evaluation compares compiler-gated authoring against prompt-only, template-only, and grammar-constrained baselines across structural failure rate, repair efficacy, brand fidelity, text exactness, reproducibility, cost, and creative variant coverage. A local gating harness over eight real .ws examples and five controlled gate fixtures reports 100% parse validity, 100% compile validity, 100% parser-normalized text exactness, 100% RenderGraph reproducibility, and 100% weak-valid catch rate for the controlled rejection set; the same harness reports 50% quality-gate acceptance on the legacy real-example corpus, where rejected examples are weak-but-valid inputs caught before publication. External prompt-to-pixel and template baselines remain to be run. The architecture generalizes beyond video to any domain where models must author structured artifacts that humans, businesses, or regulators need to inspect. AI can now generate video from a written prompt, but the results are unreliable in exactly the ways businesses care about: on-screen text comes out misspelled, brand colors drift, and running the same prompt twice produces two different videos. The usual alternative, fixed templates, is reliable but rigid, and every video ends up looking the same. This paper describes a different path. You write what a video should say and feel in a simple, readable format, and an AI helps draft it. But the AI never has the final say. A compiler, the kind of program that turns code into software, takes over and makes all the exact design decisions: fonts, colors, spacing, timing, motion. Because a compiler is deterministic, the same input always produces the same video, and every step can be checked before anything is published. The result is video that is AI-assisted but controllable, consistent, and safe to publish at scale.
Samson Aligba (Sun,) studied this question.