Run experiments

The Agent owns the design and interpretation. Use the library for config validation, execution, budgets and recovery; do not reimplement these guarantees with ad hoc subprocesses or manually edited manifests. Call Python APIs when available; CLI examples below are equivalent optional adapters.

Preview before launch:

uv run tsf run configs/runs/<run>.toml --dry-run --json
uv run tsf run configs/runs/<run>.toml [--round <round-id>]

Associate a run with a research round when one was supplied; do not create a round for a one-off run unless the task asks for persistent research context. Use --jobs N for independent configs and --gpus 0,1 only after checking memory and device intent. Keep sweeps in TOML. Before long runs, verify data, output location, seeds, horizons, strategy, and profiling. Report successful/failed configs and work_dirs/ artifacts; do not silently restart or overwrite costly runs.

Route failed, unstable, or suspect runs to diagnose-experiment. Route compatible, complete outputs to analyze-results; execution itself does not establish a fair comparison.

For GIFT-Eval, inspect uv run tsf dataset gift-download --help, obtain only the requested data, and preview configs/runs/gift_eval_sweep.toml before launch. Record dataset versions, horizons, model compatibility, compute budget, and the missing-series policy; incomplete cells must remain visible during analysis.

For optional budgets, GPU queueing, tracking, cancellation, or interrupted-run recovery, read execution controls. Do not load advanced controls for an ordinary one-off experiment.