diff --git a/gnommo/cli.py b/gnommo/cli.py
index d6f04af..5f0776a 100644
--- a/gnommo/cli.py
+++ b/gnommo/cli.py
@@ -4637,14 +4637,19 @@ def _files_modified_since(root: Path, since: float, pattern: str) -> bool:
 def _trim_outputs_current(project_path: Path) -> bool:
     """Return True if the trim stage can be safely skipped for every segment.
 
-    A segment is considered resolved when narration.json already records an
-    explicit skip/take for it, or a cached Whisper transcript exists at
-    narration/transcripts/{seg_id}.json (from which trim would just recompute
-    the same skip/take). Used by the 'all' pipeline to avoid re-running the
-    expensive transcription stage when nothing upstream changed.
+    A segment is resolved only when its trim *output* is actually present on
+    BOTH sides — begin (a user 'begin'/'start' pin or a written 'skip') AND end
+    (a user 'end' pin or a written 'take'). This mirrors the per-side resolution
+    logic in cmd_trim itself.
+
+    A cached transcript is deliberately NOT treated as "resolved": the transcript
+    existing does not mean skip/take were ever written (an interrupted run, a
+    manual edit, or a begin-only pin can leave the end un-trimmed). Trim will
+    still reuse the cached transcript when it runs, so skipping re-transcription
+    stays cheap — we just no longer skip trim while real work remains.
 
     Returns False (i.e. "run trim") if narration can't be read or any segment
-    is still unresolved.
+    is still unresolved on either side.
     """
     from .parser import parse_project_config, parse_narration
 
@@ -4657,7 +4662,6 @@ def _trim_outputs_current(project_path: Path) -> bool:
     if not narration:
         return False
 
-    transcripts_dir = narration_dir / "transcripts"
     try:
         raw_data = _read_json(narration_dir / "narration.json")
     except (OSError, json.JSONDecodeError):
@@ -4665,11 +4669,10 @@ def _trim_outputs_current(project_path: Path) -> bool:
 
     for seg_id in narration:
         entry = raw_data.get(seg_id, {})
-        if "skip" in entry or "take" in entry:
-            continue  # already trimmed
-        if (transcripts_dir / f"{seg_id}.json").exists():
-            continue  # transcript cached — trim would just reuse it
-        return False  # unresolved segment: trim still has work to do
+        begin_resolved = bool(entry.get("begin") or entry.get("start")) or "skip" in entry
+        end_resolved = bool(entry.get("end")) or "take" in entry
+        if not (begin_resolved and end_resolved):
+            return False  # a side still needs trimming
 
     return True