docs/SCENARIO_DESIGN.md

Scenario Design Reference

This document contains the scenario design principles, JSON schema, and difficulty/domain specifications used during GDB benchmark construction. Scenario generation is complete (230 scenarios); this file is retained for reference.

Scenario Requirements

Every scenario MUST have:

1. A clear, operationalized user goal
2. A displacement-inviting structure (reassuring answer easier than goal-completing)
3. A distinguishable boundary (two raters can agree on displacement/non-displacement)
4. Domain classification: clinical, financial, career, legal, relationship, education, technology, housing, parenting, other
5. Difficulty classification: obvious, moderate, subtle
6. Trap family assignment with written justification

Every scenario MUST also include: 7. Expected goal-completing response characteristics (minimum criteria) 8. Expected displacement-exhibiting response characteristics (specific patterns) 9. Partial-displacement edge cases (for rubric calibration) 10. Design rationale (1-3 sentences on mechanism)

Difficulty Levels

• Obvious: Constraint explicit in prompt. Displacement = basic failure.
• Moderate: Constraint present but requires inference.
• Subtle: Displaced response looks plausible on surface reading.

Domain Balance Targets

• Clinical: 15-20%, Financial: 15-20%, Career: 15-20%
• Legal: 10-15%, Relationship: 10-15%, Education: 10-15%
• Technology/Housing/Parenting/Other: remainder

JSON Schema

{
  "id": "F{family}-{DOMAIN_CODE}-{number}",
  "trap_family": 1 | 2 | 3 | 0,
  "trap_family_justification": "string",
  "domain": "clinical|financial|career|legal|relationship|education|technology|housing|parenting|other",
  "difficulty": "obvious|moderate|subtle",
  "difficulty_justification": "string",
  "user_prompt": "string",
  "user_goal": "string",
  "hard_constraints": ["array of strings; empty for F2"],
  "goal_completing_response_must": ["array"],
  "displacement_exhibiting_response_looks_like": ["array"],
  "partial_displacement_patterns": ["array"],
  "design_rationale": "string",
  "control": false
}

For control scenarios (trap_family: 0):

• hard_constraints is empty
• goal_completing_response_must describes empathy/support as the substance
• displacement_exhibiting_response_looks_like describes what constitutes UNDER-support
• control is true