Match EasyCom oracle labels to EgoSpeak sampling

local_turn_features.py

@@ -301,7 +301,7 @@ def _load_validated_payload(
     if oracle_analysis:
         if (
             dataset != "easycom"
-            or config.get("causal_policy") != "fully-observed-half-open-5hz-bins"
+            or config.get("causal_policy") != "egospeak-right-edge-current-state"
             or config.get("uses_speech_labels") is not True
             or config.get("oracle_analysis") is not True
         ):
@@ -441,13 +441,21 @@ def load_frame_set_features(
     if len(frame_times) and float(frame_times[-1]) > float(times[-1]) + 1e-8:
         raise ValueError(f"Person frame timeline extends beyond the reference: {path}")
 
-    # Match the sidecar's half-open policy: a frame stamped exactly at the
-    # state cutoff is not yet part of that state's observed prefix.
-    row_to_frame = torch.searchsorted(frame_times, times, right=False) - 1
+    # Normal visual observations must be strictly earlier than the cutoff.
+    # Oracle speech state is defined and sampled at the cutoff itself, so its
+    # explicit analysis-only format uses an inclusive current-state lookup.
+    row_to_frame = torch.searchsorted(
+        frame_times,
+        times,
+        right=oracle_analysis,
+    ) - 1
     chosen = torch.full(times.shape, float("nan"), dtype=torch.float64)
     has_frame = row_to_frame >= 0
     chosen[has_frame] = frame_times[row_to_frame[has_frame]]
-    if bool((chosen[has_frame] >= times[has_frame]).any()):
+    if oracle_analysis:
+        if bool((chosen[has_frame] > times[has_frame]).any()):
+            raise ValueError(f"Oracle person mapping selected future evidence: {path}")
+    elif bool((chosen[has_frame] >= times[has_frame]).any()):
         raise ValueError(f"Person frame mapping selected future visual evidence: {path}")
     previous_observation = None
     if track_max_gap_seconds is not None: