Update agent_logic.py

agent_logic.py

@@ -1,4 +1,4 @@
-# agent_logic.py (Milestone 5 - FINAL & ROBUST + LOGGING + NATURAL TEXT + FLAT JSON FIX + IGNORE FIELDS)
+# agent_logic.py (Milestone 5 - FINAL & ROBUST + LOGGING + NATURAL TEXT + ALLOWLIST FILTER)
 import asyncio
 from typing import AsyncGenerator, Dict, Optional
 import json
@@ -20,17 +20,22 @@ CLASSIFIER_SYSTEM_PROMPT = load_prompt(config.PROMPT_FILES["classifier"])
 HOMOGENEOUS_MANAGER_PROMPT = load_prompt(config.PROMPT_FILES["manager_homogeneous"])
 HETEROGENEOUS_MANAGER_PROMPT = load_prompt(config.PROMPT_FILES["manager_heterogeneous"])
 
-# --- NEW: Keys to explicitly ignore during evaluation parsing ---
-IGNORED_FIELDS = [
-    "justification", 
-    "explanation", 
-    "reasoning", 
-    "comments", 
-    "feedback", 
-    "evaluation for business problem", 
-    "overall thought",
-    "analysis"
-]
+# --- METRIC BOUNCER (Allowlist) ---
+# We map any variation of the key to the canonical internal name.
+# If a key isn't in here, it gets dropped.
+METRIC_MAPPING = {
+    "novelty": "Novelty",
+    "usefulness": "Usefulness_Feasibility",
+    "feasibility": "Usefulness_Feasibility",
+    "usefulness_feasibility": "Usefulness_Feasibility",
+    "usefulness/feasibility": "Usefulness_Feasibility",
+    "flexibility": "Flexibility",
+    "elaboration": "Elaboration",
+    "cultural_appropriateness": "Cultural_Appropriateness",
+    "cultural_sensitivity": "Cultural_Appropriateness",
+    "cultural appropriateness": "Cultural_Appropriateness",
+    "cultural appropriateness/sensitivity": "Cultural_Appropriateness"
+}
 
 class Baseline_Single_Agent:
     def __init__(self, api_clients: dict):
@@ -222,31 +227,40 @@ class StrategicSelectorAgent:
                     else:
                         v_fitness_json = {}
 
-                # --- Robust Normalization with Filtering ---
+                # --- ROBUST NORMALIZATION WITH ALLOWLIST FILTER ---
                 normalized_fitness = {}
                 if isinstance(v_fitness_json, dict):
                     for k, v in v_fitness_json.items():
-                        # --- FIX: Skip metadata keys ---
-                        if k.lower() in IGNORED_FIELDS:
+                        # 1. Map fuzzy keys to canonical keys
+                        canonical_key = None
+                        clean_k = k.lower().strip()
+                        
+                        # Check exact match or known variation
+                        if clean_k in METRIC_MAPPING:
+                            canonical_key = METRIC_MAPPING[clean_k]
+                        
+                        # If we couldn't map it to a valid metric, SKIP IT.
+                        if not canonical_key:
                             continue
 
-                        # Case 1: Nested Dictionary { "Novelty": { "score": 5 } }
+                        # 2. Extract Score Value
                         if isinstance(v, dict):
                             score_value = v.get('score')
                             justification_value = v.get('justification', str(v))
-                        # Case 2: List of Dicts { "Novelty": [{ "score": 5 }] }
                         elif isinstance(v, list) and len(v) > 0 and isinstance(v[0], dict):
                             score_value = v[0].get('score')
                             justification_value = v[0].get('justification', str(v[0]))
-                        # Case 3: Flat Value { "Novelty": "4/5" } OR { "Novelty": 4 }
                         else:
+                            # Flat value case
                             score_value = v 
                             justification_value = "Score extracted directly."
                         
-                        # Cleaning: "4/5" -> 4
+                        # 3. Clean Score (handle "4/5" strings)
                         if isinstance(score_value, str):
                             try:
-                                score_value = int(re.search(r'\d+', score_value).group())
+                                # Looks for the first number in the string
+                                match = re.search(r'\d+', score_value)
+                                score_value = int(match.group()) if match else 0
                             except:
                                 score_value = 0
                         
@@ -255,8 +269,9 @@ class StrategicSelectorAgent:
                         except (ValueError, TypeError):
                             score_value = 0
 
-                        normalized_fitness[k] = {'score': score_value, 'justification': justification_value}
+                        normalized_fitness[canonical_key] = {'score': score_value, 'justification': justification_value}
                 else:
+                     # Fallback for total failure
                      normalized_fitness = {k: {'score': 0, 'justification': "Invalid JSON structure"} for k in ["Novelty", "Usefulness_Feasibility", "Flexibility", "Elaboration", "Cultural_Appropriateness"]}
                 
                 v_fitness_json = normalized_fitness