Repository navigation
Workflow Design and Execution Boundaries
AXIOM beta.4 improves the way workflow designs are discovered, saved, revised, duplicated, validated, and estimated. The important distinction is between describing a plan and executing a workload.
The current design workflow includes:
- templates for starting a design;
- versioned saves and revision history;
- duplication for branching from an existing design;
- structural validation;
- dry-run resource estimates.
The CLI can inspect designs with:
axiom workflow list
axiom workflow show <workflow-id>
axiom workflow validate <workflow-id>Use axiom workflow <command> --help to confirm the flags supported by your installed build.
A valid design or successful dry run is not proof that a job executed. Specifically, these design tools do not by themselves:
- start training or fine-tuning;
- run inference;
- create model checkpoints;
- provide production model serving;
- prove an external framework/runtime is installed and connected.
Use a supported external runtime/framework for actual execution. AXIOM should only show a runtime as connected when the environment has been detected and verified.
Hardware-aware estimates use discovered hardware and assumptions. They are useful for comparing plausible starting configurations, not for claiming measured tokens/second, actual VRAM consumption, or guaranteed fit. Benchmark real workloads separately in the intended execution environment.
axiom evaluation run scores predictions already present in the evaluation data. It does not call a model to generate predictions. Confirm that your prediction field contains outputs produced by the model/runtime you intend to evaluate.
axiom roadmap audit --json checks ID and status consistency. It does not mark items as complete merely because a file exists. Update roadmap status only after the feature has been exercised and the required tests and documentation have been reviewed.