Containerized Simulation Environments

Packaging code, libraries, and dependencies into immutable images so Kronos runs the same everywhere, from cloud to bare-metal.

The environment is part of the result

Two runs of the same code can differ if their libraries, compilers, or dependencies differ. Kronos packages each simulation and training environment into an immutable container image, so the exact software stack that produced a result is captured and re-runnable. The environment stops being an invisible variable and becomes a recorded, versioned artifact.

What a container fixes

A container image pins the numerical libraries, the solver code, the Python and system dependencies, and their versions into a single hashed artifact. A run records which image it used, so the software half of reproducibility is guaranteed by construction. Changing a library means building a new, distinctly versioned image, never silently altering an existing one.

• Pinned solver code and numerical libraries
• Recorded image hash in the run manifest
• Identical stack across cloud and bare-metal
• New image per dependency change, never in place

Portability across the substrate

Containers are what make the cloud-explores, bare-metal-certifies workflow practical. The same image runs on an elastic cloud partition during a wide sweep and on a certified node for the reproducible pass, so the software stack is identical in both places. Combined with the hardware abstraction layer, this makes placement a policy choice rather than a source of numerical difference.

There are limits. A container fixes software but not hardware; the same image on different microarchitectures can still differ in the last bits, which is why certified runs also pin the hardware class. Containers handle the software axis of reproducibility; bare-metal and seed management handle the rest.

One image library serves both machines. Breeder neutronics, burner stability, and twin-training environments are each captured as versioned images, so any run for either machine can be reconstructed by pulling its recorded image. Containerization is the software backbone of the whole reproducibility discipline.