Update app.py

app.py

@@ -208,7 +208,7 @@ def get_llm_prompt(echelon_state_decision_point: dict, week: int, llm_personalit
         
         return f"**You are a perfectly rational supply chain AI with ONLY LOCAL information.**\nYou must use a logical heuristic to make a stable decision. A proven method is \"Anchoring and Adjustment\".\n\n{base_info}\n\n**Rational Calculation (Anchoring & Adjustment):**\n1.  **Anchor on Demand:** Your best guess for future demand is your last incoming order: **{anchor_demand} units**.\n2.  **Adjust for Inventory:** You want to hold a safety stock of {safety_stock} units. Your current stock (before shipping) is {e_state['inventory'] - e_state['backlog']}. You need to order an extra **{inventory_correction} units** to correct this.\n3.  **Account for {supply_line_desc}:** You already have **{supply_line} units** being processed (that you can see). These should be subtracted from your new decision.\n\n**Final Calculation:**\n* Decision = (Anchor Demand) + (Inventory Adjustment) - ({supply_line_desc})\n* Decision = {anchor_demand} + {inventory_correction} - {supply_line} = **{rational_local_order} units**.\n**Your Task:** Confirm this locally rational {task_word}. Respond with a single integer."
 
-    # --- HUMAN-LIKE (DESCRIPTIVE) PROMPTS ---
+# --- HUMAN-LIKE (DESCRIPTIVE) PROMPTS ---
     
     else: 
         DESIRED_INVENTORY = 12 # Matches initial inventory
@@ -253,10 +253,10 @@ def get_llm_prompt(echelon_state_decision_point: dict, week: int, llm_personalit
             """
         
         elif info_sharing == 'full':
-            # 1. HUMAN-LIKE / FULL (FIXED): Anchors on *global* customer demand
-            anchor_demand = current_stable_demand 
+            # 1. HUMAN-LIKE / FULL (FIX v5): Anchors on *local* incoming order, despite seeing global info
+            anchor_demand = e_state['incoming_order'] # <--- THIS IS THE FIX
             panicky_order = max(0, int(anchor_demand + stock_correction))
-            panicky_order_calc = f"{anchor_demand} (End-Customer Demand) + {stock_correction} (Your Stock Correction)"
+            panicky_order_calc = f"{anchor_demand} (Your Incoming Order) + {stock_correction} (Your Stock Correction)"
 
             # Build the "Full Info" string just for context
             full_info_str = f"\n**Full Supply Chain Information (State Before Shipping):**\n- End-Customer Demand this week: {current_stable_demand} units.\n"
@@ -270,15 +270,14 @@ def get_llm_prompt(echelon_state_decision_point: dict, week: int, llm_personalit
             
             **A "Human-like" Flawed Decision:**
             Even though you have full information, you are judged by *your own* performance (your inventory, your backlog).
-            You tend to react to your *local* situation (like the classic Sterman 1989 model) instead of using the complex full-system data.
-            A 'rational' player would use a complex formula, but your gut-instinct is to panic about *your* numbers.
+            You tend to **ignore the stable end-customer demand ({current_stable_demand})** and instead react to your *local* situation (like the classic Sterman 1989 model).
 
             **Your 'Panic' Calculation (Ignoring Full Info and Your Supply Line):**
-            1.  **Anchor on *End-Customer* Demand:** You can see the real demand is **{anchor_demand}** units. You'll use that as your base.
+            1.  **Anchor on *Your* Demand:** You just got an order for **{anchor_demand}** units. You panic and react to *this* number, not the stable end-customer demand.
             2.  **Correct for *Your* Stock:** Your desired 'safe' inventory is {DESIRED_INVENTORY}. Your current net inventory is {net_inventory}. You need to order **{stock_correction}** more units.
             3.  **Ignore *Your* Supply Line:** You'll ignore the **{supply_line} units** in your own pipeline ({supply_line_desc}).
 
-            **Final Panic Order:** (End-Customer Demand) + (Your Stock Correction)
+            **Final Panic Order:** (Your Incoming Order) + (Your Stock Correction)
             * Order = {panicky_order_calc} = **{panicky_order} units**.
             
             **Your Task:** Confirm this 'gut-instinct', locally-focused {task_word}. Respond with a single integer.