v0.5.0: N-dimensional spatial canvas

Arbitrary spatial dimensions — the canvas is no longer locked to 2D.

What's new

• spatial_shape parameter: CanvasLayout, compile_schema, and compile_program all accept spatial_shape=(...) for 1D, 2D, 3D, or higher.
• 1D time-series: Field(spatial_shape=(128,)) + compile_schema(..., spatial_shape=(200,))
• 3D volumetric: Field(spatial_shape=(4, 4, 4)) + compile_schema(..., spatial_shape=(8, 8, 8))
• N-D positional encoding: SinusoidalPositionalEncodingND splits d_model evenly across all dimensions.
• N-D packing: 1D linear, 2D strip (unchanged), 3D+ stacked greedy.
• N-D masks: Rect generalized to hyperrectangles, rect_cover works on any-dimensional boolean masks.
• Schema format: Serializes spatial_shape list; reads old H/W format transparently.

Backward compatible

The old H=/W= API continues to work unchanged. All 478 existing tests pass without modification.

# All of these still work exactly as before:
CanvasLayout(T=5, H=8, W=8, d_model=256)
Field(12, 12)
compile_schema(Robot(), T=8, H=16, W=16, d_model=256)

529 tests passing (478 existing + 51 new)

v0.4.2: 18 examples + TRIBE v2 BCI

18 runnable examples. TRIBE v2 BCI experiment: canvas decoder 68.8% vs SVM 59.4% on real cortical predictions.

v0.4.1: Fix all examples, add Modal runner

All 13 examples verified passing on Modal. Fixes deprecated parent_child kwarg and grid sizing for coarse-grained fields.

v0.4.0: Phase 6+ — all 8 future workstreams

Phase 6+: ProgramConditioner, internal microsteps, identity slots, masks, cortex, learned scheduling, ClockExpr IR, constraints. 478 tests.

v0.3.0: Typed process layer — families, carriers, operators, scheduling, compilation

What's new

Canvas-engineering gains a typed process layer on top of the existing typed layout:

CanvasProgram layer

• 5 region families: observation, state, memory, residual, action — everything else is tags
• 5 carriers: deterministic, diffusive, filter, memory, residual — not everything follows diffusion dynamics
• Operators: semantic intent (observe, predict, correct, retrieve, write, act) separate from backend (cross_attention, mamba, perceiver, etc.)
• Auto-wiring: compile_program() auto-sets operators from region family pairs
• ClockSpec: periodic, on_event (residual threshold), boundary clocks with cooldown and max_silence

Runtime

• ResidualAccumulator: tracks running scalar EMA summaries of error signals per region
• RegionScheduler: evaluates clock rules to decide which regions fire each step
• AttentionDispatcher.active_regions: skip inactive regions, pass through unchanged

Compilation

• ProgramCompiler: freeze/constant/export/dead-elimination passes
• Per-family learning defaults: ssl_prediction for observation, posterior_match for state, retrieval for memory, calibration for residual, supervised for action

New modules

• program.py — CanvasProgram, RegionProgram, ConnectionProgram, ClockSpec, LearningSpec
• residuals.py — ResidualSpec, ResidualAccumulator
• scheduling.py — RegionScheduler
• learning.py — FAMILY_DEFAULTS, default_learning()
• compiler.py — ProgramCompiler, CompiledProgram

Stats

• 361 tests, all passing
• 122 new tests for v2 features
• All 239 v0.2.0 tests pass unchanged
• Full backward compatibility: compile_schema() unchanged, all existing code works

v0.2.0: Temporal fill modes, PeriodEmbedding, CogVideoX

What's new

• Temporal fill modes (DROP, HOLD, INTERPOLATE) for cross-frequency attention connections
• PeriodEmbedding — learned embedding indexed by log-bucketed temporal period
• Higher-order IDW interpolation via interpolation_order on Connection
• CogVideoX backbone-native attention type
• Mamba readout refactored from pooled-mean to query-based attention
• Block-weighted mean period for coarse-grained field computation
• Period-aware fill resolution in real-time space

Behavior change

HOLD is now the default temporal fill mode. Previously, out-of-bounds temporal connections silently dropped. Use TemporalFill.DROP for the old behavior.

Stats

• 239 tests, all passing
• Full backward compatibility with v0.1.x

v0.1.1

Full Changelog: https://github.com/JacobFV/canvas-engineering/commits/v0.1.1