diff --git a/evalharness/cli.py b/evalharness/cli.py
index f040b13..ad78878 100644
--- a/evalharness/cli.py
+++ b/evalharness/cli.py
@@ -445,8 +445,6 @@ def _cmd_eval_run(args) -> int:
         from pathlib import Path
 
         Path(out_dir).mkdir(parents=True, exist_ok=True)
-        (Path(out_dir) / 'viz').mkdir(exist_ok=True)
-        (Path(out_dir) / 'reports').mkdir(exist_ok=True)
 
     rows = []
     all_reports = []
@@ -463,8 +461,6 @@ def _cmd_eval_run(args) -> int:
         from pathlib import Path
 
         Path(out_dir).mkdir(parents=True, exist_ok=True)
-        (Path(out_dir) / 'viz').mkdir(exist_ok=True)
-        (Path(out_dir) / 'reports').mkdir(exist_ok=True)
     console = _rich_console()
     _print_run_plan(console, args, model_spec)
     for i, name in enumerate(args.datasets):
@@ -523,9 +519,13 @@ def _cmd_eval_run(args) -> int:
                 report.save(args.out)
             if out_dir:
                 _print_phase(console, i + 1, total_runs, name,
-                             'writing report and visualization files')
-                report.save(f'{out_dir}/reports/{name}.report.json')
-                with open(f'{out_dir}/viz/{name}.md', 'w', encoding='utf-8') as f:
+                             'writing results')
+                from pathlib import Path as _P
+
+                bench_dir = _P(out_dir) / name
+                bench_dir.mkdir(parents=True, exist_ok=True)
+                report.save(str(bench_dir / 'report.json'))
+                with open(bench_dir / 'detail.md', 'w', encoding='utf-8') as f:
                     f.write(render(report, style='md'))
             if args.verbose:
                 if not _print_result_panel(console, report, _time.time() - t0):
@@ -583,7 +583,7 @@ def _cmd_eval_run(args) -> int:
             from evalharness.viz import render as _render
 
             xb = _render(all_reports, style='excel',
-                         out=f'{out_dir}/viz/results.xlsx')
+                         out=f'{out_dir}/summary.xlsx')
             print(f'excel  -> {xb}', flush=True)
         except Exception as e:
             print(f'excel export skipped: {type(e).__name__}: {str(e)[:80]}',
@@ -593,7 +593,7 @@ def _cmd_eval_run(args) -> int:
         if out_dir:
             import csv as _csv
 
-            with open(f'{out_dir}/viz/summary.csv', 'w', newline='', encoding='utf-8') as f:
+            with open(f'{out_dir}/summary.csv', 'w', newline='', encoding='utf-8') as f:
                 w = _csv.writer(f)
                 w.writerow(['benchmark', 'score', 'metric', 'num_samples',
                             'time_h', 'time_s', 'extract_fail',
@@ -618,7 +618,7 @@ def _cmd_eval_run(args) -> int:
                                    'ttft_mean_s', 'ttft_p90_s', 'ttft_p99_s',
                                    'tpot_mean_s', 'tpot_p90_s', 'tpot_p99_s')] +
                                [cats])
-            with open(f'{out_dir}/viz/summary.md', 'w', encoding='utf-8') as f:
+            with open(f'{out_dir}/summary.md', 'w', encoding='utf-8') as f:
                 import time as _tt
                 model_names = {r.get('model', '') for r in rows if r.get('model')}
                 head = f"# eval run summary\n\n- model: {', '.join(model_names) or '?'}\n"
@@ -648,13 +648,13 @@ def _cmd_eval_run(args) -> int:
             console.print(
                 f'\n{mark} [bold]运行结束[/bold] · {ok_n}/{len(rows)} benchmarks ok\n'
                 f'  结果 {ap}\n'
-                f'  ├─ reports/<bench>.report.json\n'
-                f'  ├─ viz/results.xlsx (excel 打开)\n'
-                f'  └─ viz/summary.md (+ summary.csv)')
+                f'  ├─ summary.xlsx (excel 打开总表)\n'
+                f'  ├─ summary.md / summary.csv\n'
+                f'  └─ <bench>/report.json + detail.md (每 benchmark 一个目录)')
         else:
             print(f'\n运行结束 · {ok_n}/{len(rows)} benchmarks ok\n'
-                  f'  结果 {ap}\n  ├─ reports/<bench>.report.json\n'
-                  f'  ├─ viz/results.xlsx\n  └─ viz/summary.md')
+                  f'  结果 {ap}\n'
+                  f'  ├─ summary.xlsx\n  └─ <bench>/report.json + detail.md')
     elif rows and rows[0]['ok'] and args.out:
         _notice('运行结束 · 结果已保存', args.out)
         print(f'{ok_n}/{len(rows)} benchmarks ok')