Update app.py

app.py

@@ -104,7 +104,7 @@ def get_llm_order_decision(prompt: str, echelon_name: str) -> (int, str):
     if not client: return 8, "NO_API_KEY_DEFAULT"
     with st.spinner(f"Getting AI decision for {echelon_name}..."):
         try:
-            temp = 0.1 if 'rational' in prompt else 0.7
+            temp = 0.1 if 'perfectly rational' in prompt else 0.7
             response = client.chat.completions.create(
                 model=OPENAI_MODEL,
                 messages=[
@@ -122,13 +122,12 @@ def get_llm_order_decision(prompt: str, echelon_name: str) -> (int, str):
             st.error(f"API call failed for {echelon_name}: {e}. Defaulting to 4.")
             return 4, f"API_ERROR: {e}"
 
-# =============== PROMPT FUNCTION (v3 - Sterman Heuristic + Demand Fix) ===============
+# =============== PROMPT FUNCTION (v4 - FIXES FOR OSCILLATION AND HUMAN-LIKE) ===============
 def get_llm_prompt(echelon_state_decision_point: dict, week: int, llm_personality: str, info_sharing: str, all_echelons_state_decision_point: dict) -> str:
     # This function's logic is updated for "human_like" to follow a flawed Sterman heuristic.
     e_state = echelon_state_decision_point
     base_info = f"Your Current Status at the **{e_state['name']}** for **Week {week}** (Before Shipping):\n- On-hand inventory: {e_state['inventory']} units.\n- Backlog (total unfilled orders): {e_state['backlog']} units.\n- Incoming order this week (just received): {e_state['incoming_order']} units.\n"
     
-    # --- PROMPT FIX: Get correct demand (current, not future) ---
     current_stable_demand = get_customer_demand(week) # Use current week's demand
     
     if e_state['name'] == 'Factory':
@@ -141,60 +140,97 @@ def get_llm_prompt(echelon_state_decision_point: dict, week: int, llm_personalit
     # --- PERFECT RATIONAL (NORMATIVE) PROMPTS ---
     
     if llm_personality == 'perfect_rational' and info_sharing == 'full':
-        stable_demand = current_stable_demand # Use the correct demand
-        # --------------------- LT=3 (1+1+1) ---------------------
-        if e_state['name'] == 'Factory': total_lead_time = FACTORY_LEAD_TIME # 1
-        elif e_state['name'] == 'Distributor': total_lead_time = ORDER_PASSING_DELAY + FACTORY_LEAD_TIME + FACTORY_SHIPPING_DELAY # 1+1+1 = 3
-        else: total_lead_time = ORDER_PASSING_DELAY + SHIPPING_DELAY # 1+2 = 3
-        # ----------------------------------------------------
+        stable_demand = current_stable_demand
+        
+        # 1. CALCULATE CORRECT LEAD TIME (UNCHANGED)
+        if e_state['name'] == 'Factory': 
+            total_lead_time = FACTORY_LEAD_TIME # 1
+        elif e_state['name'] == 'Distributor': 
+            total_lead_time = ORDER_PASSING_DELAY + FACTORY_LEAD_TIME + FACTORY_SHIPPING_DELAY # 1+1+1 = 3
+        else: 
+            total_lead_time = ORDER_PASSING_DELAY + SHIPPING_DELAY # 1+2 = 3
+            
         safety_stock = 4
         target_inventory_level = (stable_demand * total_lead_time) + safety_stock
+        
+        # 2. OSCILLATION FIX: Calculate CORRECT Inventory Position
+        order_in_transit_to_supplier = st.session_state.game_state['last_week_orders'].get(e_state['name'], 0) # Order Delay (1 week)
+        
         if e_state['name'] == 'Factory':
-            inventory_position = (e_state['inventory'] - e_state['backlog'] + sum(st.session_state.game_state['factory_production_pipeline']))
-            inv_pos_components = f"(Inv={e_state['inventory']} - Backlog={e_state['backlog']} + InProd={sum(st.session_state.game_state['factory_production_pipeline'])})"
-        else:
-            order_in_transit_to_supplier = st.session_state.game_state['last_week_orders'].get(e_state['name'], 0)
-            inventory_position = (e_state['inventory'] - e_state['backlog'] + sum(e_state['incoming_shipments']) + order_in_transit_to_supplier)
-            inv_pos_components = f"(Inv={e_state['inventory']} - Backlog={e_state['backlog']} + InTransitShip={sum(e_state['incoming_shipments'])} + OrderToSupplier={order_in_transit_to_supplier})"
+            # Factory pipeline: In Production (1 week)
+            supply_line = sum(st.session_state.game_state['factory_production_pipeline'])
+            inventory_position = (e_state['inventory'] - e_state['backlog'] + supply_line)
+            inv_pos_components = f"(Inv={e_state['inventory']} - Backlog={e_state['backlog']} + InProd={supply_line})"
+        
+        elif e_state['name'] == 'Distributor':
+            # Distributor pipeline: In Shipping (1 week) + In Production (1 week) + Order Delay (1 week)
+            in_shipping = sum(e_state['incoming_shipments'])
+            in_production = sum(st.session_state.game_state['factory_production_pipeline'])
+            supply_line = in_shipping + in_production + order_in_transit_to_supplier
+            inventory_position = (e_state['inventory'] - e_state['backlog'] + supply_line)
+            inv_pos_components = f"(Inv={e_state['inventory']} - Backlog={e_state['backlog']} + InTransitShip={in_shipping} + InProd={in_production} + OrderToSupplier={order_in_transit_to_supplier})"
+        
+        else: # Retailer and Wholesaler
+            # R/W pipeline: In Shipping (2 weeks) + Order Delay (1 week)
+            in_shipping = sum(e_state['incoming_shipments'])
+            supply_line = in_shipping + order_in_transit_to_supplier
+            inventory_position = (e_state['inventory'] - e_state['backlog'] + supply_line)
+            inv_pos_components = f"(Inv={e_state['inventory']} - Backlog={e_state['backlog']} + InTransitShip={in_shipping} + OrderToSupplier={order_in_transit_to_supplier})"
+
         optimal_order = max(0, int(target_inventory_level - inventory_position))
         return f"**You are a perfectly rational supply chain AI with full system visibility.**\nYour only goal is to maintain stability and minimize costs based on mathematical optimization.\n**System Analysis:**\n* **Known Stable End-Customer Demand:** {stable_demand} units/week.\n* **Your Current Total Inventory Position:** {inventory_position} units. {inv_pos_components}\n* **Optimal Target Inventory Level:** {target_inventory_level} units (Target for {total_lead_time} weeks lead time).\n* **Mathematically Optimal {task_word.title()}:** The optimal decision is **{optimal_order} units**.\n**Your Task:** Confirm this optimal {task_word}. Respond with a single integer."
     
     elif llm_personality == 'perfect_rational' and info_sharing == 'local':
-        safety_stock = 4; anchor_demand = e_state['incoming_order']
+        safety_stock = 4
+        anchor_demand = e_state['incoming_order']
         inventory_correction = safety_stock - (e_state['inventory'] - e_state['backlog'])
+        
+        # 2. OSCILLATION FIX: Calculate CORRECT *Local* Supply Line
         if e_state['name'] == 'Factory':
+            # Factory can see its full (local) pipeline
             supply_line = sum(st.session_state.game_state['factory_production_pipeline'])
             supply_line_desc = "In Production"
-        else:
-            order_in_transit_to_supplier = st.session_state.game_state['last_week_orders'].get(e_state['name'], 0)
-            supply_line = sum(e_state['incoming_shipments']) + order_in_transit_to_supplier
-            supply_line_desc = "Supply Line (In Transit Shipments + Order To Supplier)"
+        
+        elif e_state['name'] == 'Distributor':
+            # Distributor can *only* see its shipping queue (1 week)
+            # It CANNOT see the factory pipeline or its own order delay
+            # This is a weak heuristic, but it's *locally* correct and won't oscillate.
+            supply_line = sum(e_state['incoming_shipments'])
+            supply_line_desc = "Supply Line (In Transit Shipments)"
+            
+        else: # Retailer and Wholesaler
+            # R/W can see their full (local) pipeline: Shipping (2 weeks)
+            supply_line = sum(e_state['incoming_shipments'])
+            supply_line_desc = "Supply Line (In Transit Shipments)"
+            
         calculated_order = anchor_demand + inventory_correction - supply_line
         rational_local_order = max(0, int(calculated_order))
-        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. 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."
+        
+        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 ---
     
-    else: # Catches both 'human_like' / 'local' and 'human_like' / 'full'
-        
-        # This is the flawed Sterman heuristic
+    else: 
         DESIRED_INVENTORY = 12 # Matches initial inventory
-        anchor_demand = e_state['incoming_order']
         net_inventory = e_state['inventory'] - e_state['backlog']
         stock_correction = DESIRED_INVENTORY - net_inventory
-        panicky_order = max(0, int(anchor_demand + stock_correction))
-        panicky_order_calc = f"{anchor_demand} (Your Incoming Order) + {stock_correction} (Your Stock Correction)"
         
         # Get supply line info *just to show* the AI it's being ignored
         if e_state['name'] == 'Factory':
             supply_line = sum(st.session_state.game_state['factory_production_pipeline'])
             supply_line_desc = "In Production"
         else:
+            # This is just for display, not calculation
             order_in_transit_to_supplier = st.session_state.game_state['last_week_orders'].get(e_state['name'], 0)
-            supply_line = sum(e_state['incoming_shipments']) + order_in_transit_to_supplier
+            supply_line = sum(e_state['incoming_shipments']) + order_in_transit_to_supplier 
             supply_line_desc = "Supply Line"
 
         if info_sharing == 'local':
+            # 1. HUMAN-LIKE / LOCAL (Unchanged): Anchors on *local* incoming order
+            anchor_demand = e_state['incoming_order']
+            panicky_order = max(0, int(anchor_demand + stock_correction))
+            panicky_order_calc = f"{anchor_demand} (Your Incoming Order) + {stock_correction} (Your Stock Correction)"
+
             return f"""
             **You are a reactive supply chain manager for the {e_state['name']}.** You have a limited (local) view.
             You tend to make **reactive, 'gut-instinct' decisions** (like the classic Sterman 1989 model) that cause the Bullwhip Effect.
@@ -217,6 +253,11 @@ 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 
+            panicky_order = max(0, int(anchor_demand + stock_correction))
+            panicky_order_calc = f"{anchor_demand} (End-Customer Demand) + {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"
             for name, other_e_state in all_echelons_state_decision_point.items():
@@ -230,14 +271,14 @@ 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 the end-customer demand ({current_stable_demand}) and account for the *entire* system, but your gut-instinct is to panic about *your* numbers.
+            A 'rational' player would use a complex formula, but your gut-instinct is to panic about *your* numbers.
 
             **Your 'Panic' Calculation (Ignoring Full Info and Your Supply Line):**
-            1.  **Anchor on *Your* Demand:** You just got an order for **{anchor_demand}** units. You react to this, not the end-customer demand.
+            1.  **Anchor on *End-Customer* Demand:** You can see the real demand is **{anchor_demand}** units. You'll use that as your base.
             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:** (Your Incoming Order) + (Your Stock Correction)
+            **Final Panic Order:** (End-Customer Demand) + (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.