fix(timeline): prune out-of-interval citations

Author: AnkushMalaker

backends/advanced/src/advanced_omi_backend/services/timeline/executor.py

@@ -23,16 +23,25 @@ def validate_agent_result(
             raise ValueError(
                 f"episode {index} references unknown evidence: {sorted(unknown)}"
             )
+        overlapping_evidence_ids: list[str] = []
         for evidence_id in episode.evidence_ids:
             item = evidence[evidence_id]
             item_end = item.ended_at or item.started_at
-            if item.started_at >= episode.ended_at or item_end <= episode.started_at:
-                raise ValueError(
-                    f"episode {index} evidence {evidence_id} does not overlap its interval"
-                )
+            if item.started_at < episode.ended_at and item_end > episode.started_at:
+                overlapping_evidence_ids.append(evidence_id)
+        if not overlapping_evidence_ids:
+            raise ValueError(f"episode {index} has no temporally overlapping evidence")
+        episode.evidence_ids = overlapping_evidence_ids
+        bound_evidence = set(overlapping_evidence_ids)
         for assertion in episode.assertions:
-            if not set(assertion.evidence_ids).issubset(set(episode.evidence_ids)):
-                raise ValueError(f"episode {index} assertion has unbound evidence")
+            assertion.evidence_ids = [
+                evidence_id
+                for evidence_id in assertion.evidence_ids
+                if evidence_id in bound_evidence
+            ]
+        episode.assertions = [
+            assertion for assertion in episode.assertions if assertion.evidence_ids
+        ]
         if episode.representative_evidence_id:
             representative = evidence.get(episode.representative_evidence_id)
             if (

backends/advanced/src/advanced_omi_backend/services/timeline/prompt.py

@@ -144,6 +144,7 @@ def build_prompt(output_path: str) -> str:
 - Accessibility/OCR can include background or offscreen text.
 - Assistant-generated text is not evidence that its claims happened.
 - Every important boundary/assertion must cite supplied evidence IDs.
+- Episode citations must temporally overlap the episode interval.
 - Every episode must include at least one supplied evidence ID; otherwise leave the
   interval unassigned instead of creating an ungrounded episode.
 - Never return both `episodes` and `unassigned_intervals` empty when evidence exists.

backends/advanced/tests/test_timeline_executor.py

@@ -92,3 +92,29 @@ def test_unavailable_representative_image_is_dropped_without_losing_episode():
     result.episodes[0].representative_evidence_id = "observation:one"
     validate_agent_result(result, _manifest())
     assert result.episodes[0].representative_evidence_id is None
+
+
+def test_non_overlapping_citation_is_pruned_when_episode_remains_grounded():
+    manifest = _manifest()
+    outside = TimelineEvidenceItem(
+        evidence_id="observation:outside",
+        kind="observation",
+        started_at=manifest.started_at + timedelta(minutes=27),
+        ended_at=manifest.started_at + timedelta(minutes=28),
+        role="application_state",
+        excerpt="Later application state",
+    )
+    manifest.evidence.append(outside)
+    result = _result()
+    result.episodes[0].evidence_ids.append(outside.evidence_id)
+    validate_agent_result(result, manifest)
+    assert result.episodes[0].evidence_ids == ["observation:one"]
+
+
+def test_episode_without_overlapping_citation_is_rejected():
+    manifest = _manifest()
+    result = _result()
+    result.episodes[0].started_at = manifest.started_at + timedelta(minutes=26)
+    result.episodes[0].ended_at = manifest.started_at + timedelta(minutes=27)
+    with pytest.raises(ValueError, match="no temporally overlapping evidence"):
+        validate_agent_result(result, manifest)