fix some bugs

backend/api/screen/router.py

@@ -2215,7 +2215,10 @@ async def get_screening_metrics(
     try:
         ids = await run_in_threadpool(
             screening_eligibility_service.list_eligible_ids,
-            criteria=cp, table_name=table_name, target_stage=step_norm,
+            criteria=cp,
+            table_name=table_name,
+            target_stage=step_norm,
+            repair_decisions=False,
         )
     except Exception as e:
         raise HTTPException(
@@ -2768,7 +2771,10 @@ async def get_screening_calibration(
     try:
         ids = await run_in_threadpool(
             screening_eligibility_service.list_eligible_ids,
-            criteria=cp, table_name=table_name, target_stage=step_norm,
+            criteria=cp,
+            table_name=table_name,
+            target_stage=step_norm,
+            repair_decisions=False,
         )
     except Exception as e:
         raise HTTPException(
@@ -3070,7 +3076,10 @@ async def get_live_confidence_histogram(
     try:
         ids = await run_in_threadpool(
             screening_eligibility_service.list_eligible_ids,
-            criteria=cp, table_name=table_name, target_stage=step_norm,
+            criteria=cp,
+            table_name=table_name,
+            target_stage=step_norm,
+            repair_decisions=False,
         )
     except Exception as e:
         raise HTTPException(
@@ -3213,7 +3222,10 @@ async def get_calibration_samples(
     try:
         ids = await run_in_threadpool(
             screening_eligibility_service.list_eligible_ids,
-            criteria=cp, table_name=table_name, target_stage=step_norm,
+            criteria=cp,
+            table_name=table_name,
+            target_stage=step_norm,
+            repair_decisions=False,
         )
     except Exception as e:
         raise HTTPException(

backend/api/services/screening_eligibility_service.py

@@ -26,7 +26,11 @@ def _answer_object(value: Any) -> dict[str, Any]:
 
 
 def selected_answer(row: dict[str, Any], question: str) -> str | None:
-    """Return the effective answer using human-over-AI precedence."""
+    """Return the effective answer using human-over-main-AI precedence.
+
+    The critical agent is advisory: disagreements are routed to human review
+    but do not change the main screening agent's progression decision.
+    """
     core = snake_case(question, max_len=56)
     if not core:
         return None
@@ -90,17 +94,28 @@ def list_eligible_ids(
         criteria: dict[str, Any] | None,
         table_name: str,
         target_stage: str,
+        repair_decisions: bool = True,
     ) -> list[int]:
+        """Return citation IDs eligible for ``target_stage``.
+
+        Progression callers should keep ``repair_decisions=True`` so derived
+        decision caches are refreshed before they are used. Read-only reporting
+        callers can set it to ``False`` to avoid turning concurrent metrics
+        requests into repeated, table-wide writes.
+        """
         stage = str(target_stage or '').strip().lower()
         if stage not in {'l1', 'l2', 'extract'}:
             raise ValueError(f'Unsupported screening stage: {target_stage!r}')
 
         if stage == 'l1':
             return self.repository.list_citation_ids(None, table_name)
 
-        # Later stages depend on derived decisions. Repair first and fail loudly
-        # if repair fails; an empty list would be a misleading valid result.
-        self.repository.backfill_human_decisions(criteria or {}, table_name)
+        # Later stages depend on derived decisions. Progression paths repair
+        # first and fail loudly; reporting paths only read the existing cache.
+        if repair_decisions:
+            self.repository.backfill_human_decisions(
+                criteria or {}, table_name,
+            )
         source_stage = 'l1' if stage == 'l2' else 'l2'
         return self.repository.list_citation_ids(source_stage, table_name)
 

backend/tests/screening_eligibility_service_test.py

@@ -8,7 +8,11 @@
 if 'psycopg' not in sys.modules:
     psycopg = types.ModuleType('psycopg')
     psycopg.connect = Mock()
+    psycopg_rows = types.ModuleType('psycopg.rows')
+    psycopg_rows.dict_row = Mock()
+    psycopg.rows = psycopg_rows
     sys.modules['psycopg'] = psycopg
+    sys.modules['psycopg.rows'] = psycopg_rows
 
 from api.services.screening_eligibility_service import ScreeningEligibilityService
 from api.services.screening_eligibility_service import compute_screening_decisions
@@ -44,6 +48,20 @@ def test_ai_answer_is_used_when_human_answer_is_missing(self) -> None:
             ('include', 'include'),
         )
 
+    def test_critical_ai_answer_does_not_override_screening_ai_answer(self) -> None:
+        row = {
+            'llm_relevant_population': {
+                'selected': 'Include',
+                'critical': {'selected': 'Exclude'},
+            },
+            'llm_eligible_study_design': {'selected': 'Include'},
+        }
+
+        self.assertEqual(
+            compute_screening_decisions(row, CRITERIA),
+            ('include', 'include'),
+        )
+
     def test_l2_decision_is_cumulative_and_missing_answers_exclude(self) -> None:
         row = {'human_relevant_population': {'selected': 'Include'}}
 
@@ -93,6 +111,60 @@ def test_l1_does_not_run_unnecessary_repair(self) -> None:
             None, 'screening_table',
         )
 
+    def test_l2_read_only_scope_does_not_repair_decisions(self) -> None:
+        repository = Mock()
+        repository.list_citation_ids.return_value = [4, 9]
+        service = ScreeningEligibilityService(repository)
+
+        result = service.list_eligible_ids(
+            criteria=CRITERIA,
+            table_name='screening_table',
+            target_stage='l2',
+            repair_decisions=False,
+        )
+
+        self.assertEqual(result, [4, 9])
+        repository.backfill_human_decisions.assert_not_called()
+        repository.list_citation_ids.assert_called_once_with(
+            'l1', 'screening_table',
+        )
+
+    def test_extract_read_only_scope_uses_l2_decisions(self) -> None:
+        repository = Mock()
+        repository.list_citation_ids.return_value = [12]
+        service = ScreeningEligibilityService(repository)
+
+        result = service.list_eligible_ids(
+            criteria=CRITERIA,
+            table_name='screening_table',
+            target_stage='extract',
+            repair_decisions=False,
+        )
+
+        self.assertEqual(result, [12])
+        repository.backfill_human_decisions.assert_not_called()
+        repository.list_citation_ids.assert_called_once_with(
+            'l2', 'screening_table',
+        )
+
+    def test_read_only_scope_is_not_affected_by_repair_failure(self) -> None:
+        repository = Mock()
+        repository.backfill_human_decisions.side_effect = RuntimeError(
+            'repair failed',
+        )
+        repository.list_citation_ids.return_value = [4]
+        service = ScreeningEligibilityService(repository)
+
+        result = service.list_eligible_ids(
+            criteria=CRITERIA,
+            table_name='screening_table',
+            target_stage='l2',
+            repair_decisions=False,
+        )
+
+        self.assertEqual(result, [4])
+        repository.backfill_human_decisions.assert_not_called()
+
     def test_repair_failure_is_not_hidden_as_an_empty_scope(self) -> None:
         repository = Mock()
         repository.backfill_human_decisions.side_effect = RuntimeError(