Canvas Engineering: Declared Causal Macrostructure for Reverse-Diffusion Latent Dynamics

Open the PDF. Published under CommandAGI; code and docs at github.com/commandAGI/canvas-engineering and commandagi.github.io/canvas-engineering.

Prompt engineering structures what a language model sees. Canvas engineering structures what a diffusion model thinks in. You declare the macrostructure of a diffusion transformer's latent space — which regions carry which modalities, their geometry, their temporal update frequencies, their loss roles, and a directed graph of permitted block-to-block attention operations — and a compiler lowers that declaration into attention masks, loss weights, and frame mappings on an off-the-shelf pretrained backbone.

Because reverse diffusion is an iterated denoising map over the whole canvas, a hard connectivity constraint on attention induces an explicit, human-legible causal interaction graph inside the generative dynamics, while gradient descent remains free to shape all fine structure within and along the declared edges. Macro is symbolic, micro is neural, and there is no discrete/continuous interface to cross — the symbolic layer is the attention mask. If two regions have no path between them, their independence is exact by construction: d-separation compiled into a denoiser.

The paper gives the underlying mathematics (region index sets, the mask construction, the position-weighted denoising objective, and the reachability semantics), the abstraction stack from typed entity schemas to compiled deployment, and worked designs spanning robot manipulation, a compiler-allocated hospital ICU ward with 199 typed regions, air-traffic control, and a 23-region cortical model that predicts real brain dynamics at R²=0.825. It reports the early empirical record — 26 experiments and 236 training runs on CogVideoX-2B — including a 1.73× parameter-efficiency result for looped attention and a frozen 350K-parameter configuration that beats 11.7M unfrozen parameters on action prediction, and reads those small-scale results as calibration rather than narrative.

This is a companion to the empirical ROAM paper on looped attention; where that work asks what works and why in the video-diffusion substrate, this one asks how to make latent structure declarable at all, and argues that declared structure — versionable, diffable, compilable, shareable — is the production-ready on-ramp to intuition-guided symbolic world modeling.