ArmResult

Training-side metrics from RunResult.metrics.

Back-compat alias - mirrors evals[\<primary>].metrics. primary is the first test-tagged post-training eval, or the first post-training eval, or the first eval overall.

All benchmarks scored against this arm - see EvalResult.

Raw underlying RunResult for advanced consumers.

Dashboard group_id when n_repeats > 1 (else None).

Primary's name (= the group's name) when grouped; else None.

Look up an eval result by benchmark name.

step=None (default) → return the post-training row (step is None) if present; else the last in-loop row for that benchmark.

step=\<int> → return the exact in-loop row at that step (or None if no exact match - callers can do their own nearest-step lookup over arm.evals).

Return the metric value for ranking by success_metric.

Three forms supported:

• "pass_rate" (bare) - looked up on eval_metrics (the primary post-training eval), then metrics (training-side).
• "bench.pass_rate" (dotted) - looked up on arm.eval("bench").metrics["pass_rate"]. The dotted form wins when an experiment carries multiple named benchmarks and success_metric picks one explicitly.
• Returns None if neither path resolves.