Strengthen panel label guardrails in QC pipeline

Prepend panel label constraint to system prompt and add post-response
filter that discards panels with labels not in the expected set, preventing
downstream issues from AI-invented subdivisions or descriptive labels.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: drAbreu

config.qc.yaml

@@ -1,4 +1,4 @@
-qc_version: "2.3.2"
+qc_version: "2.3.3"
 qc_check_metadata:
   panel:
     plot_axis_units:

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "soda-curation"
-version = "2.5.7"
+version = "2.5.8"
 description = "A professional Python package for data curation with AI capabilities"
 authors = ["Dr. Jorge Abreu Vicente <jorge.abreu@embo.org>"]
 license = "MIT"

src/soda_curation/qc/base_analyzers.py

@@ -98,6 +98,9 @@ def analyze_figure(
             # Process and validate the response
             result = self.process_response(response)
 
+            # Filter out panels with labels not in expected_panels
+            result = self._filter_valid_panels(result, expected_panels)
+
             # Check if the test passed
             passed = self.check_test_passed(result)
 
@@ -107,6 +110,47 @@ def analyze_figure(
             logger.error(f"Error analyzing figure with {self.test_name}: {str(e)}")
             return False, self.create_empty_result()
 
+    @staticmethod
+    def _filter_valid_panels(
+        result: Dict, expected_panels: Optional[List[str]]
+    ) -> Dict:
+        """Filter response outputs to only include panels with valid labels.
+
+        Args:
+            result: Parsed response dict with an "outputs" list.
+            expected_panels: Allowed panel labels for this figure.
+
+        Returns:
+            The result dict with invalid-label outputs removed.
+        """
+        if not expected_panels or not isinstance(result, dict):
+            return result
+
+        outputs = result.get("outputs", [])
+        if not outputs:
+            return result
+
+        valid_outputs = []
+        for panel in outputs:
+            # Extract panel_label from dict or object
+            if isinstance(panel, dict):
+                label = panel.get("panel_label")
+            elif hasattr(panel, "panel_label"):
+                label = panel.panel_label
+            else:
+                label = None
+
+            if label in expected_panels:
+                valid_outputs.append(panel)
+            else:
+                logger.warning(
+                    f"Discarded panel with invalid label '{label}' "
+                    f"(expected one of {expected_panels})"
+                )
+
+        result["outputs"] = valid_outputs
+        return result
+
     def process_response(self, response: Any) -> Any:
         """Process the response from the model API."""
         # If response is a string, try to parse it

src/soda_curation/qc/model_api.py

@@ -69,15 +69,22 @@ def generate_response(
 
         # Add expected panels instruction if provided
         if expected_panels:
-            panels_instruction = (
-                f"\n\n**IMPORTANT CONSTRAINT**: The `panel_label` field in your response MUST be EXACTLY "
-                f"one of the following valid panel labels for this figure: {expected_panels}. "
-                f"Do NOT use any other labels, sub-panel labels (like 'A-a', 'A-b'), descriptive labels "
-                f"(like 'Rice cell', 'Figure 8'), or panel labels with modifiers (like 'C (plot)', 'C (right)'). "
-                f"Use ONLY the exact labels from this list: {expected_panels}"
+            panels_constraint = (
+                f"**CRITICAL PANEL LABEL CONSTRAINT**:\n"
+                f"The output panel labels for this task MUST be exactly equal to "
+                f"those defined for this figure. The allowed panel labels are: {expected_panels}\n"
+                f"Rules:\n"
+                f"- You MUST ONLY use panel labels from the list above.\n"
+                f"- Do NOT invent new panel labels or subdivisions.\n"
+                f"- Do NOT use sub-panel labels (e.g., 'A-a', 'A-b', 'A-l').\n"
+                f"- Do NOT use descriptive labels (e.g., 'Rice cell', 'Figure 8').\n"
+                f"- Do NOT add modifiers to labels (e.g., 'C (plot)', 'C (right)').\n"
+                f"- Each `panel_label` in your response MUST be exactly one of: {expected_panels}\n"
             )
-            system_prompt += panels_instruction
-            user_prompt += panels_instruction
+            # Prepend to system prompt so it acts as a ground rule
+            system_prompt = panels_constraint + "\n" + system_prompt
+            # Append to user prompt for reinforcement
+            user_prompt += "\n\n" + panels_constraint
 
         # Determine the type of analysis and create appropriate messages
         if encoded_image is not None and caption is not None: