Start Eval Run

Create an eval run and dispatch background execution. An eval run re-runs a candidate task revision against the same inputs as a set of reference ("baseline") task runs, then uses an LLM judge to compare each candidate output against its corresponding reference output. **Choosing the reference runs** (provide exactly one of): - `eval_set_id` — use the task runs stored in a previously created eval set - `task_run_ids` — use a specific, ad-hoc list of task run IDs (max 20) - `last_n` — use the task's `n` most recent successful runs (1-20) **Candidate revision:** - `candidate_revision_id` — the task revision to re-run and judge. Must differ from the revision that produced the reference runs; the run is rejected with a 400 if the candidate revision matches all reference runs. **Judge model:** - `judge_model_id` — optional; a platform default judge model is used if omitted. The run starts in `pending` status and transitions to `running`, then `completed` or `failed`, as it executes in the background. Poll Get Eval Run for the per-comparison verdicts once it completes.