Raise treatment reduction, lower medium/hard seeds, focus prompt strategy

baseline/policy.py

@@ -69,8 +69,8 @@ def build_prompt(obs: CityObservation) -> str:
             "",
             "DECISION RULES (follow in order):",
             "1. If ANY hospital is below 0.3 capacity: 'allocate' on that district IMMEDIATELY.",
-            "2. If resources > 0 and any district is CRITICAL (above 0.4): 'allocate' on the highest.",
-            "3. If resources > 0 and any district is WARNING (0.2-0.4): 'allocate' on the highest.",
+            "2. If resources > 0: ALWAYS 'allocate' on the district with the SINGLE HIGHEST infection rate.",
+            "3. Do NOT spread resources across multiple districts in the same turn — focus all pressure on the worst district.",
             "4. If resources = 0: 'restrict' on the highest infected district.",
             "5. NEVER use 'test' — data is already accurate.",
             "6. NEVER restrict a SAFE district (below 0.2) — you will be penalised.",
@@ -86,8 +86,9 @@ def build_prompt(obs: CityObservation) -> str:
             "DECISION RULES (follow in order):",
             "1. If ANY hospital is below 0.3 capacity: 'allocate' on that district IMMEDIATELY.",
             "2. If resources > 0: 'allocate' on the district with HIGHEST growth_hint.",
-            "3. If resources = 0: 'restrict' on the district with highest growth_hint.",
-            "4. NEVER use 'test' — spending resources on data wastes your limited budget.",
+            "3. Sustain pressure — keep allocating to the same district until growth_hint drops below 0.08.",
+            "4. If resources = 0: 'restrict' on highest growth_hint district.",
+            "5. NEVER use 'test'.",
         ]
 
     lines += [

server/constants.py

@@ -46,7 +46,7 @@ SPREAD_RATE_MIN       = 0.05    # Slowest possible spread rate per day
 SPREAD_RATE_MAX       = 0.13    # Fastest possible spread rate per day
 GROWTH_HINT_NOISE     = 0.03    # Noise on growth_rate_hint (± value)
 
-TREATMENT_REDUCTION   = 0.08    # Allocating reduces existing infection this step
+TREATMENT_REDUCTION   = 0.10    # Allocating reduces existing infection this step
 ALLOCATE_REDUCTION    = 0.10    # Allocating reduces future spread rate this step
 RESTRICT_REDUCTION    = 0.05    # Restricting reduces spread per step
 SPILLOVER_RATE        = 0.01    # Infection fraction that spills to adjacent districts

server/tasks/task_hard.py

@@ -1,15 +1,8 @@
 # server/tasks/task_hard.py
-# ─────────────────────────────────────────────────────────────────────────────
-# Hard task: 6 districts, 3-day data lag, scarce resources.
-# All districts start with small but growing infections.
-# Agent must learn to read growth_rate_hint signals and act proactively
-# on districts that look manageable today but will be critical in 3 days.
-# ─────────────────────────────────────────────────────────────────────────────
-
 import sys
 import os
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))    # reaches server/
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../..')) # reaches project root
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../..'))
 
 from models import CityState
 from server.utils import generate_districts
@@ -20,24 +13,25 @@ from server.constants import TASK_CONFIG
 class HardTask(BaseTask):
 
     name          = "hard"
-    num_districts = TASK_CONFIG["hard"]["num_districts"]    # 6
-    max_steps     = TASK_CONFIG["hard"]["max_steps"]        # 15
-    resource_pool = TASK_CONFIG["hard"]["resource_pool"]    # 7
-    data_lag_days = TASK_CONFIG["hard"]["data_lag_days"]    # 3
+    num_districts = TASK_CONFIG["hard"]["num_districts"]
+    max_steps     = TASK_CONFIG["hard"]["max_steps"]
+    resource_pool = TASK_CONFIG["hard"]["resource_pool"]
+    data_lag_days = TASK_CONFIG["hard"]["data_lag_days"]
 
     def build_initial_state(self) -> CityState:
-        # All districts start with small infections that grow at different rates.
-        # The 3-day lag means the agent won't see today's true rates until day 3.
-        # Districts with high true_spread_rate will accelerate invisibly.
-        seed_infections = [0.12, 0.08, 0.14, 0.06, 0.16, 0.09]
-
-        # Pre-populate infection_history with 3 days of identical
-        # starting values so the lag mechanic works from step 1.
-        initial_rates   = seed_infections[:]
+        # All districts start with small but growing infections.
+        # The 3-day lag means agent sees these seed values while true
+        # infection is already 3 days ahead — districts D0, D2, D4 will
+        # be CRITICAL before the agent sees updated data.
+        # 7 resources for 6 districts with delayed information is
+        # the hardest possible triage scenario.
+        seed_infections = [0.10, 0.06, 0.12, 0.05, 0.13, 0.07]
+
+        initial_rates     = seed_infections[:]
         infection_history = [
-            initial_rates[:],   # day -3 (what agent sees on step 1)
-            initial_rates[:],   # day -2
-            initial_rates[:],   # day -1
+            initial_rates[:],  # day -3 (what agent sees on step 1)
+            initial_rates[:],  # day -2
+            initial_rates[:],  # day -1
         ]
 
         return CityState(
@@ -47,8 +41,8 @@ class HardTask(BaseTask):
             data_lag_days       = self.data_lag_days,
             max_steps           = self.max_steps,
             districts           = generate_districts(
-                                      num_districts    = self.num_districts,
-                                      seed_infections  = seed_infections,
+                                      num_districts   = self.num_districts,
+                                      seed_infections = seed_infections,
                                   ),
             infection_history   = infection_history,
         )

server/tasks/task_medium.py

@@ -1,14 +1,8 @@
 # server/tasks/task_medium.py
-# ─────────────────────────────────────────────────────────────────────────────
-# Medium task: 4 districts, 2 simultaneous outbreaks, tighter resource pool.
-# Agent must prioritise between competing threats — it cannot fully
-# address both outbreaks simultaneously and must learn to triage.
-# ─────────────────────────────────────────────────────────────────────────────
-
 import sys
 import os
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))    # reaches server/
-sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../..')) # reaches project root
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../..'))
 
 from models import CityState
 from server.utils import generate_districts
@@ -19,16 +13,17 @@ from server.constants import TASK_CONFIG
 class MediumTask(BaseTask):
 
     name          = "medium"
-    num_districts = TASK_CONFIG["medium"]["num_districts"]  # 4
-    max_steps     = TASK_CONFIG["medium"]["max_steps"]      # 15
-    resource_pool = TASK_CONFIG["medium"]["resource_pool"]  # 8
-    data_lag_days = TASK_CONFIG["medium"]["data_lag_days"]  # 0
+    num_districts = TASK_CONFIG["medium"]["num_districts"]
+    max_steps     = TASK_CONFIG["medium"]["max_steps"]
+    resource_pool = TASK_CONFIG["medium"]["resource_pool"]
+    data_lag_days = TASK_CONFIG["medium"]["data_lag_days"]
 
     def build_initial_state(self) -> CityState:
-        # Districts 0 and 2 are seeded with outbreaks (non-adjacent).
-        # Districts 1 and 3 are clean but will receive spillover.
-        # Agent must choose which outbreak to tackle first.
-        seed_infections = [0.25, 0.06, 0.22, 0.06]
+        # D0 and D2 seeded with outbreaks (non-adjacent).
+        # D1 and D3 start clean but will grow into critical within 3-4 steps.
+        # 8 total resources for 4 districts creates genuine triage pressure —
+        # agent cannot save all districts and must choose strategically.
+        seed_infections = [0.18, 0.03, 0.15, 0.03]
 
         return CityState(
             day                 = 0,
@@ -37,8 +32,8 @@ class MediumTask(BaseTask):
             data_lag_days       = self.data_lag_days,
             max_steps           = self.max_steps,
             districts           = generate_districts(
-                                      num_districts    = self.num_districts,
-                                      seed_infections  = seed_infections,
+                                      num_districts   = self.num_districts,
+                                      seed_infections = seed_infections,
                                   ),
             infection_history   = [],
         )