diff --git a/scripts/config.py b/scripts/config.py
index 18e89ef..9b72a29 100644
--- a/scripts/config.py
+++ b/scripts/config.py
@@ -26,6 +26,11 @@ VIDEO_INFO_TSV = DATA_VOLUME / "all_video_info_merged.tsv"
 # A small CSV listing every video file we know about (built locally).
 INVENTORY_CSV = DATA_METADATA / "video_inventory.csv"
 
+# Hand-annotated barrier-opening times (output of the picker app). One
+# row per testing session; columns: machine_name, session_date,
+# session_time, opening_s, trim_first_s, bad_rois, analyst, notes.
+BARRIER_OPENING_CSV = DATA_METADATA / "barrier_opening.csv"
+
 # Where the ethoscope source tree is checked out (used by track_videos.py
 # and auto_detect_targets.py — host-side scripts that import ethoscope
 # from a local clone rather than from pip). Default assumes the standard
diff --git a/scripts/load_roi_data.py b/scripts/load_roi_data.py
index ee2263c..309709a 100644
--- a/scripts/load_roi_data.py
+++ b/scripts/load_roi_data.py
@@ -8,12 +8,48 @@ The TSV is the single source of truth for what data exists and how it
 maps to flies and conditions.
 """
 
+import re
 import sqlite3
 from pathlib import Path
 
 import pandas as pd
 
-from config import VIDEO_INFO_TSV
+from config import BARRIER_OPENING_CSV, VIDEO_INFO_TSV
+
+# DB filenames start with `YYYY-MM-DD_HH-MM-SS_<uuid>_...` — pull the
+# session date/time out so we can join against barrier_opening.csv.
+_DB_TIMESTAMP_RE = re.compile(r"(\d{4}-\d{2}-\d{2})_(\d{2}-\d{2}-\d{2})_")
+
+
+def _session_key(db_path: str) -> tuple[str, str] | None:
+    """Extract (session_date, session_time) from a tracking DB filename."""
+    if not isinstance(db_path, str) or not db_path:
+        return None
+    m = _DB_TIMESTAMP_RE.search(Path(db_path).name)
+    return (m.group(1), m.group(2)) if m else None
+
+
+def _load_barrier_lookup(csv_path: Path) -> dict[tuple[str, str, str], dict]:
+    """Build (machine, session_date, session_time) → opening/bad_rois lookup.
+
+    Returns an empty dict if the CSV is missing — callers should treat
+    that as "no per-session annotations available" rather than an error.
+    """
+    if not Path(csv_path).exists():
+        return {}
+    df = pd.read_csv(csv_path)
+    lookup: dict[tuple[str, str, str], dict] = {}
+    for r in df.itertuples(index=False):
+        bad = set()
+        if isinstance(r.bad_rois, str) and r.bad_rois.strip():
+            bad = {int(x) for x in r.bad_rois.split(",") if x.strip()}
+        lookup[(r.machine_name, r.session_date, r.session_time)] = {
+            "opening_s": float(r.opening_s) if pd.notna(r.opening_s) else float("nan"),
+            "trim_first_s": float(r.trim_first_s) if pd.notna(r.trim_first_s) else 0.0,
+            "bad_rois": bad,
+            "unusable": pd.isna(r.opening_s),
+        }
+    return lookup
 
 # Reason: prefer the explicit Jupyter-widget tqdm when available (it
 # updates reliably in JupyterLab, where text \r-style bars sometimes
@@ -66,6 +102,7 @@ def _open_ro(db_path: str, cache: dict) -> sqlite3.Connection | None:
 def load_roi_data(
     meta: pd.DataFrame | None = None,
     progress: bool = True,
+    apply_barrier_filter: bool = True,
 ) -> pd.DataFrame:
     """Load ROI tracking data joined with experimental metadata.
 
@@ -75,6 +112,14 @@ def load_roi_data(
     (``"training"`` or ``"testing"``). Rows with empty DB paths (unusable
     videos, or videos that didn't pass the completeness gate) are skipped.
 
+    Both training and testing reads are filtered against
+    ``barrier_opening.csv`` (the picker annotates both video types):
+    flies whose ROI never released (listed in ``bad_rois``) and entire
+    sessions flagged unusable are dropped. The session's ``opening_s``
+    is stamped onto its samples so downstream code can compute
+    ``t_from_opening = t - opening_s``. Sessions missing from the CSV
+    are still loaded, but with ``opening_s = NaN``.
+
     Args:
         meta: optional DataFrame with the same schema as
             ``all_video_info_merged.tsv``. Pass a filtered slice to load a
@@ -82,11 +127,17 @@ def load_roi_data(
             Defaults to the full TSV.
         progress: show a tqdm progress bar (one tick per fly/ROI row).
             Defaults to True. Set False for silent batch jobs.
+        apply_barrier_filter: if True (default), drop session data for
+            flies whose barrier never opened and stamp ``opening_s``
+            onto every sample. Set False to load raw data without any
+            barrier-derived filtering or columns.
 
     Returns:
         DataFrame with columns ``id, t, x, y, w, h, phi, is_inferred,
-        has_interacted, session, <metadata>`` — one row per tracking
-        sample. Empty if nothing could be loaded.
+        has_interacted, session, ROI, opening_s, <metadata>`` — one row
+        per tracking sample. ``opening_s`` is NaN for sessions not
+        covered by ``barrier_opening.csv``. Empty if nothing could be
+        loaded.
     """
     if meta is None:
         meta = pd.read_csv(VIDEO_INFO_TSV, sep="\t")
@@ -97,8 +148,12 @@ def load_roi_data(
     if "include" in meta.columns:
         meta = meta[meta["include"].astype(bool)]
 
+    barrier_lookup = _load_barrier_lookup(BARRIER_OPENING_CSV) if apply_barrier_filter else {}
+
     db_cache: dict = {}
     chunks: list[pd.DataFrame] = []
+    n_skipped_bad_roi = 0
+    n_skipped_unusable = 0
 
     n_rows = len(meta)
     if progress:
@@ -125,7 +180,28 @@ def load_roi_data(
     for row in meta.itertuples(index=False):
         for session in ("training", "testing"):
             pbar.set_postfix_str(f"{row.machine_name} ROI {int(row.roi)} {session}")
-            conn = _open_ro(getattr(row, f"{session}_db_path"), db_cache)
+            db_path = getattr(row, f"{session}_db_path")
+
+            # The picker annotates barrier_opening per video, and both
+            # the training and testing videos have their own entries.
+            # Apply the same per-session filter to both.
+            opening_s = float("nan")
+            if barrier_lookup:
+                key = _session_key(db_path)
+                if key is not None:
+                    bo = barrier_lookup.get((row.machine_name, key[0], key[1]))
+                    if bo is not None:
+                        if bo["unusable"]:
+                            n_skipped_unusable += 1
+                            pbar.update(1)
+                            continue
+                        if int(row.roi) in bo["bad_rois"]:
+                            n_skipped_bad_roi += 1
+                            pbar.update(1)
+                            continue
+                        opening_s = bo["opening_s"]
+
+            conn = _open_ro(db_path, db_cache)
             if conn is None:
                 pbar.update(1)
                 continue
@@ -141,6 +217,7 @@ def load_roi_data(
                 continue
             df["session"] = session
             df["ROI"] = int(row.roi)
+            df["opening_s"] = opening_s
             for col in _META_COLS:
                 df[col] = getattr(row, col)
             chunks.append(df)
@@ -148,6 +225,13 @@ def load_roi_data(
 
     pbar.close()
 
+    if apply_barrier_filter and (n_skipped_bad_roi or n_skipped_unusable):
+        print(
+            f"Barrier filter: dropped {n_skipped_bad_roi} ROI loads (barrier "
+            f"never opened) and {n_skipped_unusable} unusable sessions.",
+            flush=True,
+        )
+
     for conn in db_cache.values():
         if conn is not None:
             conn.close()