diff --git a/evalharness/cli.py b/evalharness/cli.py
index 47019b0..91c472e 100644
--- a/evalharness/cli.py
+++ b/evalharness/cli.py
@@ -776,6 +776,15 @@ def _cmd_eval_run(args) -> int:
                                if k != 'extraction_failure_rate'), None)
                     if _m is not None:
                         _scores.append(_m)
+                    if _repeats > 1 and out_dir:
+                        # per-run report: <out-dir>/<bench>/reps/repNN.report.jsonl
+                        # (each carries THAT run's own score + samples; the
+                        # top-level report.jsonl stays the mean-summary view)
+                        import pathlib as _pl
+
+                        _rdir = _pl.Path(out_dir) / name / 'reps'
+                        _rdir.mkdir(parents=True, exist_ok=True)
+                        report.save(str(_rdir / f'rep{_rep + 1:02d}.report.jsonl'))
                     _rep_info = report.metric_groups.get('run_info', {}) or {}
                     _rep_secs += sum(float((s.usage or {}).get('latency_s', 0) or 0)
                                      for s in report.samples)