Update app.py

app.py

@@ -1,5 +1,5 @@
 # app.py
-# @title Beer Game Final Version (v4.25 - v2 Logic + v2 Prompt Fix)
+# @title Beer Game Final Version (v4.25 - Based on v4.21 Logic + UI Fixes v3 - Log Fix)
 
 # -----------------------------------------------------------------------------
 # 1. Import Libraries
@@ -121,9 +121,8 @@ 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}"
 
-# =============== NEW PROMPT FUNCTION (v2) - Sterman Heuristic ===============
 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.
+    # This function's logic remains correct (from v4.21).
     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"
     if e_state['name'] == 'Factory':
@@ -132,10 +131,6 @@ def get_llm_prompt(echelon_state_decision_point: dict, week: int, llm_personalit
     else:
         task_word = "order quantity"
         base_info += f"- Shipments In Transit To You (arriving next week onwards): {list(e_state['incoming_shipments'])}"
-
-    # --- PERFECT RATIONAL (NORMATIVE) PROMPTS ---
-    # (These prompts are already good and remain unchanged)
-    
     if llm_personality == 'perfect_rational' and info_sharing == 'full':
         stable_demand = 8
         if e_state['name'] == 'Factory': total_lead_time = FACTORY_LEAD_TIME
@@ -152,7 +147,6 @@ def get_llm_prompt(echelon_state_decision_point: dict, week: int, llm_personalit
             inv_pos_components = f"(Inv={e_state['inventory']} - Backlog={e_state['backlog']} + InTransitShip={sum(e_state['incoming_shipments'])} + 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']
         inventory_correction = safety_stock - (e_state['inventory'] - e_state['backlog'])
@@ -166,108 +160,87 @@ def get_llm_prompt(echelon_state_decision_point: dict, week: int, llm_personalit
         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."
+    elif llm_personality == 'human_like' and info_sharing == 'full':
+        full_info_str = f"\n**Full Supply Chain Information (State Before Shipping):**\n- End-Customer Demand this week: {get_customer_demand(week)} units.\n"
+        for name, other_e_state in all_echelons_state_decision_point.items():
+            if name != e_state['name']: full_info_str += f"- {name}: Inv={other_e_state['inventory']}, Backlog={other_e_state['backlog']}\n"
+        return f"""
+        **You are a supply chain manager ({e_state['name']}) with full system visibility.**
+        You can see everyone's current inventory and backlog before shipping, and the real customer demand.
+        {base_info}
+        {full_info_str}
+        **Your Task:** Your primary responsibility is to meet the demand from your direct customer (your `Incoming order this week`: **{e_state['incoming_order']}** units), which contributes to your total current backlog of {e_state['backlog']}.
+        While you can see the stable end-customer demand ({get_customer_demand(week)} units), your priority is to fulfill the order you just received and manage your inventory/backlog.
+        You are still human and might get anxious about your own stock levels.
+        What {task_word} should you decide on this week? Respond with a single integer.
+        """
+    elif llm_personality == 'human_like' and info_sharing == 'local':
+        return f"""
+        **You are a reactive supply chain manager for the {e_state['name']}.** You have a limited view and tend to over-correct based on fear.
+        Your top priority is to NOT have a backlog.
+        {base_info}
+        **Your Task:** You just received an incoming order for **{e_state['incoming_order']}** units, adding to your total backlog.
+        Your gut instinct is to panic and {task_word.split(' ')[0]} enough to ensure you are never caught with a backlog again, considering your current inventory.
+        **React emotionally.** What is your knee-jerk {task_word}? Respond with a single integer.
+        """
 
-    # --- HUMAN-LIKE (DESCRIPTIVE) PROMPTS ---
-    # (These are NEW and implement a flawed, panicky Sterman-style heuristic)
-    
-    else: # Catches both 'human_like' / 'local' and 'human_like' / 'full'
-        
-        # This is the flawed Sterman heuristic
-        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:
-            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"
-
-        if info_sharing == 'local':
-            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.
-            
-            {base_info}
-
-            **Your Flawed 'Human' Heuristic:**
-            Your gut tells you to fix your entire inventory problem *right now*, and you're afraid of your backlog.
-            A 'rational' player would account for their {supply_line_desc} (which is {supply_line} units), but you're too busy panicking to trust that.
-
-            **Your 'Panic' Calculation (Ignoring the Supply Line):**
-            1.  **Anchor on Demand:** You just got an order for **{anchor_demand}** units. You'll order *at least* that.
-            2.  **Correct for Stock:** Your desired 'safe' inventory is {DESIRED_INVENTORY}. Your current net inventory is {net_inventory}. You need to order **{stock_correction}** more units to feel safe again.
-            3.  **Ignore Supply Line:** You'll ignore the **{supply_line} units** already in your pipeline.
-
-            **Final Panic Order:** (Your Incoming Order) + (Your Stock Correction)
-            * Order = {panicky_order_calc} = **{panicky_order} units**.
-            
-            **Your Task:** Confirm this 'gut-instinct' {task_word}. Respond with a single integer.
-            """
-        
-        elif info_sharing == 'full':
-            # 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: {get_customer_demand(week)} units.\n"
-            for name, other_e_state in all_echelons_state_decision_point.items():
-                if name != e_state['name']: full_info_str += f"- {name}: Inv={other_e_state['inventory']}, Backlog={other_e_state['backlog']}\n"
-
-            return f"""
-            **You are a supply chain manager ({e_state['name']}) with full system visibility.**
-            {base_info}
-            {full_info_str}
-            
-            **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 (8) and account for the *entire* system, 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.
-            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)
-            * Order = {panicky_order_calc} = **{panicky_order} units**.
-            
-            **Your Task:** Confirm this 'gut-instinct', locally-focused {task_word}. Respond with a single integer.
-            """
-# =========================================================
-
-# =============== REVERTED `step_game` (v2) - The stable logic ===============
 def step_game(human_final_order: int, human_initial_order: int, ai_suggestion: int):
-    # This is the correct logic from v4.17 (reverted from the buggy v3 log fix)
+    # This is the correct logic from v4.17
     state = st.session_state.game_state
     week, echelons, human_role = state['week'], state['echelons'], state['human_role']
     llm_personality, info_sharing = state['llm_personality'], state['info_sharing']
     echelon_order = ["Retailer", "Wholesaler", "Distributor", "Factory"]
     llm_raw_responses = {}
+    
+    # Capture opening state for logging
     opening_inventories = {name: e['inventory'] for name, e in echelons.items()}
     opening_backlogs = {name: e['backlog'] for name, e in echelons.items()}
-    
-    # This is the v2 logic (known stable)
+
+    # --- LOG FIX (v3): Capture arrival data BEFORE mutation ---
     arrived_this_week = {name: 0 for name in echelon_order}
+    # This dict will store the value shown on the UI for "next week"
+    opening_arriving_next_week_UI_VALUE = {name: 0 for name in echelon_order} 
+    
+    # Factory
+    factory_q = state['factory_production_pipeline']
+    if factory_q:
+        arrived_this_week["Factory"] = factory_q[0] # Read before pop
+    if len(factory_q) > 1:
+        opening_arriving_next_week_UI_VALUE["Factory"] = factory_q[1] # Read [1] before pop
+
+    # R, W, D
+    for name in ["Retailer", "Wholesaler", "Distributor"]:
+        shipment_q = echelons[name]['incoming_shipments']
+        if shipment_q:
+            arrived_this_week[name] = shipment_q[0] # Read before pop
+        
+        # This logic MUST match the v2 UI logic
+        if name == 'Distributor':
+            # "Next" for Distributor is what's at the front of the Factory's pipeline
+            if factory_q:
+                opening_arriving_next_week_UI_VALUE[name] = factory_q[0]
+        elif name in ("Retailer", "Wholesaler"):
+            if len(shipment_q) > 1: # q_len=2
+                opening_arriving_next_week_UI_VALUE[name] = shipment_q[1]
+            elif len(shipment_q) == 1: # q_len=1
+                opening_arriving_next_week_UI_VALUE[name] = shipment_q[0] # Match the v2 UI
+    # --- END OF LOG FIX (v3) ---
+
+    # Now, the *actual* state mutation (popping)
     inventory_after_arrival = {}
     factory_state = echelons["Factory"]
     produced_units = 0
     if state['factory_production_pipeline']:
         produced_units = state['factory_production_pipeline'].popleft()
-        arrived_this_week["Factory"] = produced_units
     inventory_after_arrival["Factory"] = factory_state['inventory'] + produced_units
     
     for name in ["Retailer", "Wholesaler", "Distributor"]:
         arrived_shipment = 0
         if echelons[name]['incoming_shipments']:
             arrived_shipment = echelons[name]['incoming_shipments'].popleft()
-            arrived_this_week[name] = arrived_shipment
         inventory_after_arrival[name] = echelons[name]['inventory'] + arrived_shipment
-    # End of v2 logic block
-        
+    
+    # (Rest of game logic)
     total_backlog_before_shipping = {}
     for name in echelon_order:
         incoming_order_for_this_week = 0
@@ -301,6 +274,8 @@ def step_game(human_final_order: int, human_initial_order: int, ai_suggestion: i
     if units_shipped["Factory"] > 0: echelons['Distributor']['incoming_shipments'].append(units_shipped["Factory"])
     if units_shipped['Distributor'] > 0: echelons['Wholesaler']['incoming_shipments'].append(units_shipped['Distributor'])
     if units_shipped['Wholesaler'] > 0: echelons['Retailer']['incoming_shipments'].append(units_shipped['Wholesaler'])
+    
+    # (Logging)
     log_entry = {'timestamp': datetime.utcnow().isoformat() + "Z", 'week': week, **state}
     del log_entry['echelons'], log_entry['factory_production_pipeline'], log_entry['logs'], log_entry['last_week_orders']
     if 'current_ai_suggestion' in log_entry: del log_entry['current_ai_suggestion']
@@ -308,19 +283,23 @@ def step_game(human_final_order: int, human_initial_order: int, ai_suggestion: i
         e = echelons[name]; e['weekly_cost'] = (e['inventory'] * HOLDING_COST) + (e['backlog'] * BACKLOG_COST); e['total_cost'] += e['weekly_cost']
         for key in ['inventory', 'backlog', 'incoming_order', 'order_placed', 'shipment_sent', 'weekly_cost', 'total_cost']: log_entry[f'{name}.{key}'] = e[key]
         log_entry[f'{name}.llm_raw_response'] = llm_raw_responses.get(name, "")
-        log_entry[f'{name}.opening_inventory'] = opening_inventories[name]; log_entry[f'{name}.opening_backlog'] = opening_backlogs[name]
-        log_entry[f'{name}.arrived_this_week'] = arrived_this_week[name]
         
-        # This is the v2 logging (it has the known bug, but the game logic is stable)
-        if name != 'Factory': log_entry[f'{name}.arriving_next_week'] = list(e['incoming_shipments'])[0] if e['incoming_shipments'] else 0
-        else: log_entry[f'{name}.production_completing_next_week'] = list(state['factory_production_pipeline'])[0] if state['factory_production_pipeline'] else 0
-    
+        # --- LOG FIX (v3): Use captured values ---
+        log_entry[f'{name}.opening_inventory'] = opening_inventories[name]
+        log_entry[f'{name}.opening_backlog'] = opening_backlogs[name]
+        log_entry[f'{name}.arrived_this_week'] = arrived_this_week[name] # Use captured value
+        
+        if name != 'Factory': 
+            log_entry[f'{name}.arriving_next_week'] = opening_arriving_next_week_UI_VALUE[name]
+        else: 
+            log_entry[f'{name}.production_completing_next_week'] = opening_arriving_next_week_UI_VALUE[name]
+        # --- END OF LOG FIX (v3) ---
+            
     log_entry[f'{human_role}.initial_order'] = human_initial_order; log_entry[f'{human_role}.ai_suggestion'] = ai_suggestion
     state['logs'].append(log_entry)
     state['week'] += 1; state['decision_step'] = 'initial_order'; state['last_week_orders'] = current_week_orders
     state['current_ai_suggestion'] = None # Clean up
     if state['week'] > WEEKS: state['game_running'] = False
-# ==============================================================================
 
 def plot_results(df: pd.DataFrame, title: str, human_role: str):
     # This function remains correct.