Carriers

Not everything in a canvas is diffusive. A robot's joint angles are deterministic measurements. A belief state is a Kalman-style filter. An episodic memory bank is a persistent lookup table. A prediction error signal is an ephemeral trace. The carrier declares what dynamics govern a region, independent of what the region means (its family).

The 5 carriers

Carrier	Dynamics	Update rule	Example regions
deterministic	Standard forward latent updates	Direct write from encoder/decoder	Joint angles, reward scalars, task prompts
diffusive	Noise/denoise process	DDPM/flow-matching denoising	Video frames, generated images, future plans
filter	Predict/correct cycle	Kalman-style: predict from prior, correct with evidence	Belief state, object pose estimates
memory	Persistent lookup	Write/retrieve from key-value store	Episodic memory, semantic knowledge base
residual	Error traces	EMA-smoothed scalar summaries	Prediction error, novelty signal, uncertainty

Family and carrier are orthogonal

The family says what a region is. The carrier says how it evolves. The same family can use different carriers depending on the application:

Region	Family	Carrier	Why
Future video (to generate)	observation	diffusive	Model generates it via denoising
Current video (observed)	observation	deterministic	Encoder writes it directly
Kalman belief	state	filter	Predict/correct dynamics
Goal embedding	state	deterministic	Set once, read many
Episodic memory	memory	memory	Persistent write/retrieve
Prediction error	residual	residual	EMA scalar summaries
Generated plan	action	diffusive	Plan generated via denoising
Direct motor command	action	deterministic	MLP decoder output

How carrier affects the system

Residual tracking. Regions with carrier="residual" feed into ResidualAccumulator, which maintains running EMA summaries of error signals. These summaries drive event-triggered scheduling: a memory region can fire only when prediction error exceeds a threshold, instead of every step.

Compile-mode interaction. The carrier influences what compile modes make sense. A memory carrier region naturally maps to compile_mode="export" (save the memory bank to disk at deploy). A deterministic carrier region can be compile_mode="constant" (materialize as a buffer and remove from the compute graph).

Future: loss computation. Carrier type will determine which loss function applies. Diffusive regions use diffusion loss (MSE on noise prediction). Deterministic regions use direct reconstruction loss. Filter regions use predictive consistency loss. This is declared, not hard-coded -- the training loop reads the carrier and dispatches accordingly.

Usage

Carrier is set on Field (for compile_program()) or on RegionSpec (for manual layout construction):

from canvas_engineering import Field

# On Field (compile_program path)
camera = Field(12, 12, family="observation", carrier="deterministic")
future_video = Field(12, 12, family="observation", carrier="diffusive")

# On RegionSpec (manual layout path)
from canvas_engineering import RegionSpec
spec = RegionSpec(bounds=(0, 8, 0, 12, 0, 12), carrier="filter")

The default carrier is "deterministic". If you don't set it, the region uses standard forward updates.