diff --git a/evalharness/cli.py b/evalharness/cli.py
index 05ff19a..2089c8f 100644
--- a/evalharness/cli.py
+++ b/evalharness/cli.py
@@ -206,14 +206,14 @@ def _narration(msg: str) -> str:
     # generation counts, few-shot counts (curated, not rich's blanket highlight)
     import re as _re0
 
-    msg = _re0.sub(r'(?<![\w/])(\d+(?:/\d+)?)(?=[\s]|$)',
+    msg = _re0.sub(r'(?<![\w/%.])(\d+(?:/\d+)?)(?=[\s]|$)',
                   r'[bold]\1[/bold]', msg)
 
     icon = ''
     m = msg.lower()
     if m.startswith('loading/'):
         icon = '⬇ '
-    elif 'dataset ready' in m:
+    elif 'dataset ready' in m or m.startswith('dataset ready'):
         icon = '📦 '
     elif 'few-shot' in m:
         icon = '✳ '
@@ -235,7 +235,7 @@ def _narration(msg: str) -> str:
     # gets bold green -- it is what the eye should find first
     import re as _re1
 
-    m = _re1.search(r'(· [a-zA-Z_@]+ [0-9.]+%)$', msg)
+    m = _re1.search(r'[:·] ([a-zA-Z_@]+ [0-9.]+%)$', msg)
     if m:
         return (f'{icon}{msg[:m.start()]}'
                 f'[bold green]{m.group(1)}[/bold green]')
@@ -551,7 +551,7 @@ def _cmd_eval_run(args) -> int:
                 else:
                     _print_phase(console, _i + 1, total_runs, _n, msg)
 
-            _emit('loading/downloading dataset')
+            _emit('Loading dataset (downloads on first use, cached afterwards)')
             ds = get_dataset(name, **overrides)
             if _shared_reporter is not None:
                 _shared_reporter.pause()   # let hub tqdm print cleanly
@@ -559,7 +559,7 @@ def _cmd_eval_run(args) -> int:
             if _shared_reporter is not None:
                 _shared_reporter.resume()
             origin = ds.lineage.get('from', 'unknown')
-            _emit(f'dataset ready · samples={sample_count} · source={origin}')
+            _emit(f'Dataset ready: {sample_count} samples from {origin}')
             if model_spec:  # generate + score in one go
                 from evalharness.model import run_eval
 
@@ -621,8 +621,8 @@ def _cmd_eval_run(args) -> int:
             if args.out:
                 report.save(args.out)
             if out_dir:
-                _emit(f'writing results -> {out_dir}/{name}/' if out_dir
-                      else 'writing results')
+                _emit(f'Writing results to {out_dir}/{name}/' if out_dir
+                      else 'Writing results')
                 from pathlib import Path as _P
 
                 bench_dir = _P(out_dir) / name
diff --git a/evalharness/model/runner.py b/evalharness/model/runner.py
index a769f1c..f24c25b 100644
--- a/evalharness/model/runner.py
+++ b/evalharness/model/runner.py
@@ -398,10 +398,10 @@ async def generate_predictions(
             pending.append((i, s))
     if status_callback:
         if restored:
-            status_callback(f'{len(work)} samples · checkpoint {len(restored)}/{len(work)} '
-                            f'restored, {len(pending)} to generate')
+            status_callback(f'Checkpoint: {len(restored)}/{len(work)} predictions already generated, '
+                            f'{len(pending)} samples left to run')
         else:
-            status_callback(f'{len(work)} samples')
+            status_callback(f'{len(work)} samples to evaluate')
     elif restored:
         print(f'checkpoint: restored {len(restored)} predictions '
               f'({len(pending)} to generate) -> {ckpt_store.path}', flush=True)
@@ -434,16 +434,16 @@ async def generate_predictions(
     try:
         if status_callback:
             if pending:
-                status_callback(f'generating {len(pending)} responses')
+                status_callback(f'Generating {len(pending)} model responses')
             else:
-                status_callback('generation skipped · checkpoint complete')
+                status_callback('Generation skipped: the checkpoint already covers every sample')
         fresh = await asyncio.gather(*(run_one((i, s)) for i, s in pending))
         for i, pred in fresh:
             preds_by_key[keys[i]] = pred
         preds = [preds_by_key[k] for k in keys]
         usages = [p.get('usage', {}) for p in preds]
         if status_callback and pending:
-            status_callback(f'generation complete · {len(preds)} responses')
+            status_callback(f'Generation complete: {len(preds)} responses collected')
         return preds, usages, total_usage
     finally:
         # reporter lifecycle belongs to the CALLER (CLI reuses one reporter
@@ -659,7 +659,7 @@ async def run_eval(
     few_shot_text = None
     if few_shot_num:
         if status_callback:
-            status_callback(f'loading few-shot examples: {few_shot_num}')
+            status_callback(f'Loading {few_shot_num} few-shot exemplars')
         from ..data.registry import get_dataset_provider
 
         prov = get_dataset_provider(name)
@@ -732,7 +732,7 @@ async def run_eval(
         judge = _judge_callable(judge_adapter)
 
     if status_callback:
-        status_callback('scoring predictions')
+        status_callback('Scoring predictions against the benchmark recipe')
     report = evaluate(
         samples, preds, recipe,
         model=model_spec,
@@ -746,7 +746,7 @@ async def run_eval(
     if status_callback:
         _m = next(((k, v) for k, v in report.metrics.items()
                    if k != 'extraction_failure_rate'), None)
-        status_callback(f'scoring complete · {_m[0]} {_m[1] * 100:.1f}%'
+        status_callback(f'Scoring complete: {_m[0]} {_m[1] * 100:.1f}%'
                         if _m else 'scoring complete')
     # performance profile: pool success rate + latency/ttft percentiles
     try: