Make the data-leakage trigger reachable

agents/triggers.py

@@ -14,10 +14,13 @@ def should_check_data_leakage(agent, node: SearchNode) -> bool:
  metric_value = node.metric.value
  maximize = agent.metric_maximize
 
+ # Exact float equality almost never fires in practice: a leaked AUC lands
+ # on 0.99997, a leaked RMSE on 1.2e-08. Compare against a threshold instead,
+ # keeping the old exact-1.0 / exact-0.0 cases inside the new bounds.
  if maximize:
- is_extreme = (metric_value == 1.0)
+ is_extreme = (metric_value >= agent.acfg.leakage_max_threshold)
  else:
- is_extreme = (metric_value == 0.0)
+ is_extreme = (metric_value <= agent.acfg.leakage_min_threshold)
 
  if is_extreme:
  logger.info(

config/__init__.py

@@ -102,6 +102,9 @@ class AgentConfig:
  use_evolution: bool = True
  use_fusion: bool = True
  use_aggregation: bool = True
+ # Bounds for the data-leakage trigger (see agents/triggers.py).
+ leakage_max_threshold: float = 0.999
+ leakage_min_threshold: float = 1e-6
 @dataclass
 class ExecConfig:
  timeout: int

config/config.yaml

@@ -56,6 +56,9 @@ agent:
 
  # --- Submission check ---
  check_data_leakage: True
+ # A metric at or beyond these bounds triggers the data-leakage check.
+ leakage_max_threshold: 0.999
+ leakage_min_threshold: 1e-6
 
  # --- Code gen mode ---
  use_diff_mode: True # Multi-Mode Codegen