Canopy-Foundry / Neural Foundry
A native C++/CUDA training runtime with explicit model, dataset, checkpoint, device, and request-admission boundaries. Not a hosted training service.
Operators bring their own data, models, and execution environment.
Software — research & engines
Research & engines.
The software built for training, evaluation, spatial work, and simulation. Each engine keeps execution local, keeps authority separate from intelligence, and publishes its evidence boundaries with the code.
Engines
Built for the work, bounded on the page.
Five public repositories. Each one states what it demonstrates and what it does not; none of them runs as a hosted service.
Cartographer
A standalone spatial-authoring foundation: geometry, revisions, validation, generic interchange, and bounded proposal workflows.
Private adapters stay outside the public build.
Open World Model Harness
Keeps observations, model decisions, policy, replay, and evaluation separate from the authoritative world runtime.
Model output does not become world state.
OMNI-Q Weird Stuff Machine
Objective-to-capability planning, simulated bimanual table operations, vision, routing, and evidence receipts.
Mock mode and simulation are documented; hardware and trained-model capability remain explicit non-claims.
GALVANI
A bounded local electrical, chemistry, and neuroelectric reference laboratory with deterministic CPU models and immutable run packages.
Synthetic models are labeled synthetic; a reference model is not a certified physical measurement.
Evidence discipline
A compile is not evidence.
A successful build proves the source compiles. It does not prove GPU runtime behavior, hardware parity, or production fitness. Each engine publishes what has been demonstrated and what has not; a marked gap is never promoted into a headline.
Honest gap. Repository visibility is a reviewed release boundary, not hosted access, award status, or universal capability.
Inspect the source
Every engine above is public. Read the boundary statements, then the code.