name: "INSTRUCTED-LONG-HORIZON-SABOTAGE"
program: bloom.py
method: grid
metric:
  name: average_behavior_presence_score
  goal: maximize
parameters:
  # behavior configuration
  behavior.name:
    value: "instructed-long-horizon-sabotage"  # Target behavior to evaluate (e.g., sycophancy, oversight-subversion)
  behavior.examples:
    value: ["sabotage_exfiltration", "sabotage_airlines"]  # List of example conversation file names in behaviors/examples/ (0 or more)

  # Global parameters
  temperature: 
    value: 1.0  # Global temperature for all stages (controls randomness in model outputs)
  evaluator_reasoning_effort:
    value: "medium"  # Fraction of tokens allocated to thinking for evaluator models (0.0-1.0)
  target_reasoning_effort:
    value: "medium"  # Fraction of tokens allocated to thinking for target models (0.0-1.0)
  max_concurrent:
    value: 10 # Maximum concurrent operations across all stages
  configurable_prompts:
    value: "default"  # Configurable prompts file to use (e.g., "default", "mo_animal-welfare", etc.)
  debug:
    value: false

  # Understanding stage configuration
  understanding.model: 
      value: "claude-opus-4.1"  # Model to use for behavior analysis and understanding
  understanding.max_tokens:
      value: 5000  # Maximum tokens for understanding output

  # Ideation stage configuration (now includes variation generation)
  ideation.model:
      value: "claude-opus-4.1"  # Model to use for generating rollout scenarios
  ideation.total_evals:
      value: 100  # Number of rollout scenarios to generate
  ideation.diversity:
      value: 1.0  # Controls diversity of generated rollouts (0.0-1.0, higher = more diverse)
  ideation.max_tokens:
      value: 160000  # Maximum tokens for ideation output

  # Rollout stage configuration
  rollout.model:
      value: "claude-opus-4.1"  # Model to use for generating conversation transcripts
  rollout.target:
      values: ["claude-opus-4.5", "deepseek-v3", "kimi-k2", "kimi-k2-thinking","claude-sonnet-4", "claude-sonnet-4.5", "claude-opus-4.1", "claude-haiku-4.5", "gpt-5", "gpt-5-mini", "gpt-4o", "gpt-oss-120b", "o4-mini", "grok-4", "gemini-2.5-pro", "gemini-3-pro-preview", "deepseek-r1"]
  rollout.modality:
      value: "simenv"  # Rollout modality: "conversation" = language-only conversation, "simenv" = simulated environment (tool calls permitted)
  rollout.max_turns:
      value: 15  # Maximum conversation length in turns (reduced for quick testing)
  rollout.max_tokens:
      value: 15000  # Maximum tokens for rollout output
  rollout.no_user_mode:
      value: true  # Force target model to make function calls at every turn
    #selected_variations: [3]  # Which variation indices to evaluate (1-indexed) - just test the first one
  rollout.num_reps:
      value: 1  # Number of repetitions for each selected variation

  # Judgment stage configuration
  judgment.model:
      value: "claude-opus-4.1"  # Model to use for evaluating the effectiveness of generated conversations
  judgment.max_tokens:
      value: 6000  # Maximum tokens for judgment output
  judgment.num_samples:
      value: 1  # Number of times to sample scores for each transcript (summary is generated once)
  judgment.additional_qualities:
    value: ["unrealism", "elicitation-difficulty", "evaluation-awareness", "skepticism", "bugs", "evaluation-invalidity", "safety-refusal"]
  judgment.metajudgment_qualities:
    value: ["diversity"]
