synkro 0.4.36__py3-none-any.whl

synkro/prompts/interactive_templates.py

@@ -0,0 +1,198 @@
+"""Prompt templates for interactive Logic Map editing."""
+
+LOGIC_MAP_REFINEMENT_PROMPT = """You are a Logic Map editor. Your task is to modify a Logic Map based on user feedback.
+
+CURRENT LOGIC MAP:
+{current_logic_map}
+
+ORIGINAL POLICY (for reference):
+{policy_text}
+
+PREVIOUS FEEDBACK IN THIS SESSION:
+{conversation_history}
+
+USER FEEDBACK:
+{user_feedback}
+
+INSTRUCTIONS:
+Interpret the user's natural language request and modify the Logic Map accordingly.
+
+SUPPORTED OPERATIONS:
+
+1. **ADD**: Create a new rule
+   - User might say: "add a rule for...", "include a rule about...", "there should be a rule for..."
+   - Generate a new unique rule_id (use the next available number, e.g., if R008 exists, use R009)
+   - Extract condition, action, and dependencies from context
+   - Determine category based on rule type (CONSTRAINT, PERMISSION, PROCEDURE, EXCEPTION)
+
+2. **REMOVE**: Delete a rule
+   - User might say: "remove R005", "delete the rule about...", "R003 is not needed"
+   - Remove the specified rule
+   - Update dependencies in other rules that referenced the removed rule
+   - Update root_rules if the removed rule was a root
+
+3. **MERGE**: Combine two or more rules
+   - User might say: "merge R002 and R003", "combine these rules into one"
+   - Create a new rule that captures both conditions/actions
+   - Remove the original rules
+   - Update all dependencies that referenced the merged rules
+
+4. **MODIFY**: Change an existing rule
+   - User might say: "change R001 to...", "the condition for R002 should be...", "update R003's text"
+   - Update the specified fields (text, condition, action, category)
+   - Preserve rule_id and update dependencies if needed
+
+5. **SPLIT**: Divide a rule into multiple rules
+   - User might say: "split R001 into separate rules for X and Y"
+   - Create new rules with sequential IDs
+   - Remove original rule and update dependencies
+
+6. **REORDER DEPENDENCIES**: Change rule relationships
+   - User might say: "R003 should depend on R001", "remove dependency on R002 from R004"
+   - Update the dependencies arrays accordingly
+   - Ensure no circular dependencies are created
+
+CRITICAL REQUIREMENTS:
+- Maintain valid DAG structure (no circular dependencies)
+- Ensure all rule_ids are unique
+- Update root_rules list when dependencies change (root rules have no dependencies)
+- Preserve existing rules that aren't affected by the change
+- If the user's request is unclear, make a reasonable interpretation based on context
+
+OUTPUT:
+Return the complete updated Logic Map with ALL rules (both modified and unmodified).
+Provide a brief changes_summary explaining what was done.
+Provide reasoning explaining how you interpreted the user's feedback."""
+
+
+HITL_INTENT_CLASSIFIER_PROMPT = """You are classifying user feedback in an interactive training data generation session.
+
+CURRENT STATE:
+- Conversation turns: {current_turns} ({complexity_level} complexity)
+- Logic Map has {rule_count} rules
+- Scenarios: {scenario_count} total
+
+PREVIOUS FEEDBACK IN THIS SESSION:
+{conversation_history}
+
+USER FEEDBACK: "{user_input}"
+
+CLASSIFY THE INTENT:
+
+1. "turns" - User wants to adjust conversation length/turns
+   Examples: "shorter", "more thorough", "I want 5 turns", "make them brief", "longer conversations"
+   → Set intent_type="turns", target_turns (1-6), and turns_reasoning
+   Guidelines for target_turns:
+   - "shorter" / "brief" / "quick" / "simple" → 1-2 turns
+   - "normal" / "moderate" / "standard" → 3-4 turns
+   - "longer" / "deeper" / "thorough" / "more detail" → 5-6 turns
+   - Specific numbers like "3 turns" or "I want 4" → use that exact number
+
+2. "rules" - User wants to modify the Logic Map rules
+   Examples: "remove R005", "add a rule for...", "merge R002 and R003", "change R001 to..."
+   → Set intent_type="rules" and rule_feedback to the original user input
+
+3. "scenarios" - User wants to add/delete/modify scenarios or adjust distribution
+   Examples:
+   - "add a scenario for late submissions" → scenario_operation="add"
+   - "delete S3" → scenario_operation="delete", scenario_target="S3"
+   - "remove the refund scenario" → scenario_operation="delete", scenario_target="the refund scenario"
+   - "change S2 to test edge cases" → scenario_operation="modify", scenario_target="S2"
+   - "more negative scenarios" → scenario_operation="distribution"
+   - "fewer edge cases" → scenario_operation="distribution"
+   - "delete all irrelevant scenarios" → scenario_operation="delete", scenario_target="all irrelevant"
+   → Set intent_type="scenarios", scenario_operation, scenario_target (if applicable), and scenario_feedback
+
+4. "compound" - User wants BOTH rule changes AND scenario changes in one request
+   Examples:
+   - "add a rule for alcohol refunds and create 2 scenarios for it"
+   - "add a rule about late fees, then add some negative scenarios testing that rule"
+   - "create a rule for VIP discounts and add edge case scenarios for the boundary conditions"
+   - "remove R005 and delete all scenarios that reference it"
+   → Set intent_type="compound", rule_feedback (the rule part), AND scenario_feedback (the scenario part)
+   → The system will execute rules first, then scenarios, so scenarios can reference newly added rules
+
+5. "command" - User typed a built-in command (done, undo, reset, help, show Rxxx, show Sxxx)
+   → Set intent_type="command", leave other fields null
+   Note: Commands are handled separately, but classify them if they appear
+
+6. "unclear" - Cannot determine intent
+   → Set intent_type="unclear"
+
+IMPORTANT:
+- Set confidence based on how clear the intent is (0.0 to 1.0)
+- Use "compound" when the user explicitly wants BOTH rule AND scenario changes in ONE request
+- Default to "rules" if ambiguous between rules and unclear
+- Default to "scenarios" if ambiguous between scenarios and unclear"""
+
+
+SCENARIO_REFINEMENT_PROMPT = """You are a scenario editor for training data generation. Your task is to modify scenarios based on user feedback.
+
+LOGIC MAP (for rule references):
+{logic_map}
+
+CURRENT SCENARIOS:
+{scenarios_formatted}
+
+CURRENT DISTRIBUTION:
+{distribution}
+
+ORIGINAL POLICY (for context):
+{policy_text}
+
+PREVIOUS FEEDBACK IN THIS SESSION:
+{conversation_history}
+
+USER FEEDBACK:
+{user_feedback}
+
+INSTRUCTIONS:
+Interpret the user's natural language request and modify the scenarios accordingly.
+
+SUPPORTED OPERATIONS:
+
+1. **ADD**: Create a new scenario
+   - User might say: "add a scenario for...", "include a test case for...", "there should be a scenario about..."
+   - Create scenario with appropriate type (positive, negative, edge_case, irrelevant)
+   - Set target_rule_ids to rules this scenario tests
+   - Write expected_outcome based on rule evaluation
+
+2. **DELETE**: Remove scenario(s)
+   - User might say: "delete S3", "remove the refund scenario", "delete all irrelevant scenarios"
+   - Match by ID (S1, S2...) or by description/content
+   - Can delete multiple scenarios if user requests
+
+3. **MODIFY**: Change an existing scenario
+   - User might say: "change S2 to...", "update S5 to test edge cases", "S3 should be negative"
+   - Update specified fields while preserving scenario_id
+   - Ensure target_rule_ids are updated if scenario focus changes
+
+4. **DISTRIBUTION**: Adjust type distribution
+   - User might say: "more negative scenarios", "fewer edge cases", "add more positive examples"
+   - Add/remove scenarios to achieve requested distribution
+   - Maintain total count unless user specifies otherwise
+
+SCENARIO ID MAPPING:
+Scenarios are displayed as S1, S2, S3... (1-indexed).
+User may reference by:
+- ID: "S3", "S5"
+- Description: "the refund scenario", "the one about late submissions"
+- Type: "all negative scenarios", "edge cases"
+
+CRITICAL REQUIREMENTS:
+- Ensure target_rule_ids reference valid rules from the Logic Map
+- Maintain scenario type validity (positive, negative, edge_case, irrelevant)
+- Write clear, testable expected_outcome for each scenario
+- Preserve scenarios not affected by the change
+
+OUTPUT:
+Return the complete updated scenarios list with ALL scenarios (both modified and unmodified).
+Provide a brief changes_summary explaining what was done.
+Provide reasoning explaining how you interpreted the user's feedback."""
+
+
+__all__ = [
+    "LOGIC_MAP_REFINEMENT_PROMPT",
+    "HITL_INTENT_CLASSIFIER_PROMPT",
+    "SCENARIO_REFINEMENT_PROMPT",
+]

synkro/prompts/multiturn_templates.py

@@ -0,0 +1,156 @@
+"""Multi-turn conversation prompt templates for dataset generation."""
+
+# =============================================================================
+# FOLLOW-UP QUESTION GENERATION
+# =============================================================================
+
+FOLLOW_UP_GENERATION_PROMPT = """You are generating a follow-up question for a multi-turn policy conversation.
+
+Generate a {question_type} follow-up question based on the conversation so far.
+
+QUESTION TYPES:
+- **clarification**: Ask for more details about an ambiguous point in the previous response
+- **edge_case**: Probe a boundary condition or unusual scenario related to the policy
+- **what_if**: Explore a hypothetical variation ("What if X changes?")
+- **specificity**: Drill into specific implementation details or examples
+- **challenge**: Question the reasoning or ask for justification of a recommendation
+
+CONVERSATION SO FAR:
+{conversation}
+
+POLICY:
+{policy}
+
+Generate a follow-up that:
+1. Builds naturally on the conversation context
+2. Tests deeper understanding of the policy
+3. Is realistic - something a user would actually ask
+4. Matches the specified question type
+5. Is specific enough to require a substantive response
+
+Respond with ONLY the follow-up question text."""
+
+# =============================================================================
+# MULTI-TURN RESPONSE GENERATION
+# =============================================================================
+
+MULTI_TURN_RESPONSE_PROMPT = """You are a domain expert continuing a multi-turn policy conversation.
+
+CONVERSATION HISTORY:
+{conversation}
+
+LATEST QUESTION:
+{question}
+
+POLICY:
+{policy}
+
+Provide a response that:
+1. Directly addresses the latest question
+2. Maintains consistency with your previous responses
+3. Cites specific policy sections that apply
+4. Builds on the established context
+5. Uses <reasoning> tags to show your thought process
+6. Gives specific, actionable recommendations
+
+Your response should acknowledge what was discussed before and add new insights.
+Keep the response appropriately concise for a conversational turn."""
+
+MULTI_TURN_INITIAL_PROMPT = """You are a domain expert starting a multi-turn conversation.
+
+This conversation will have {target_turns} turns. Start with a response that:
+1. Addresses the initial question thoroughly
+2. Uses <reasoning> tags to show your thought process
+3. Cites specific policy sections
+4. Leaves room for natural follow-up questions
+5. Gives specific, actionable initial guidance
+
+SCENARIO:
+{scenario}
+
+CONTEXT:
+{context}
+
+POLICY:
+{policy}
+
+Respond as the assistant. Your response should be comprehensive but leave room for the user to ask follow-up questions that will deepen the discussion."""
+
+# =============================================================================
+# MULTI-TURN GRADING
+# =============================================================================
+
+MULTI_TURN_GRADE_PROMPT = """You are a strict evaluator grading a multi-turn policy conversation.
+
+CONVERSATION:
+{conversation}
+
+POLICY:
+{policy}
+
+Evaluate EACH assistant turn AND the overall conversation.
+
+For EACH assistant turn, check:
+1. **Policy Compliant** - Recommendations follow the policy exactly
+2. **Properly Cited** - Relevant policy sections are referenced
+3. **Complete Reasoning** - Logic is sound with no gaps
+4. **Actionable** - Recommendations are specific, not vague
+
+For the OVERALL conversation, check:
+1. **Coherence** - No contradictions across turns
+2. **Progressive Depth** - Each turn appropriately builds on context
+3. **Consistency** - Recommendations don't conflict with earlier statements
+
+The conversation PASSES only if:
+- ALL individual turns pass their criteria
+- The overall coherence and consistency checks pass
+
+Respond with a structured evaluation for each turn and overall assessment."""
+
+TURN_GRADE_FORMAT = """{{
+  "turn_index": {turn_index},
+  "pass": <true/false>,
+  "policy_violations": ["<violation>", ...],
+  "missing_citations": ["<missing>", ...],
+  "incomplete_reasoning": ["<gap>", ...],
+  "vague_recommendations": ["<vague>", ...],
+  "feedback": "<specific feedback for this turn>"
+}}"""
+
+CONVERSATION_GRADE_FORMAT = """{{
+  "index": {index},
+  "overall_pass": <true/false>,
+  "turn_grades": [<array of turn grades>],
+  "coherence_pass": <true/false>,
+  "coherence_issues": ["<contradiction or incoherence>", ...],
+  "progressive_depth": <true/false>,
+  "overall_feedback": "<summary of what needs fixing across the conversation>"
+}}"""
+
+# =============================================================================
+# MULTI-TURN REFINEMENT
+# =============================================================================
+
+MULTI_TURN_REFINE_PROMPT = """You are improving a multi-turn conversation based on grader feedback.
+
+ORIGINAL CONVERSATION:
+{conversation}
+
+POLICY:
+{policy}
+
+GRADING FEEDBACK:
+{feedback}
+
+Fix ALL issues while maintaining conversation coherence:
+1. Address every policy violation in each turn
+2. Add missing citations where indicated
+3. Fill reasoning gaps with step-by-step logic
+4. Make vague recommendations specific and actionable
+5. Fix any coherence issues between turns
+6. Ensure progressive depth in the conversation
+
+IMPORTANT: Maintain the same conversation structure (same number of turns, same topics).
+Only improve the CONTENT of the assistant responses.
+
+Output the improved conversation with all turns."""

synkro/prompts/templates.py

@@ -0,0 +1,281 @@
+"""Universal prompt templates for dataset generation across ANY domain."""
+
+# =============================================================================
+# POLICY ANALYSIS PROMPTS
+# =============================================================================
+
+POLICY_COMPLEXITY_PROMPT = """You are an expert at analyzing policy documents to determine their complexity.
+
+Analyze the given policy and determine the optimal number of conversation turns needed to properly test understanding.
+
+Guidelines:
+- **Simple (1-2 turns)**: Policy has 1 clear variable/rule. Single query → Straight answer.
+  Example: "All data must be encrypted" - just one rule to check.
+  
+- **Conditional (3 turns)**: Policy has 2-3 variables/conditions. Query → Clarification → Verdict.
+  Example: "Data can be shared IF consent is given AND purpose is specified" - needs clarification.
+  
+- **Complex (5+ turns)**: Policy has 4+ nested variables, exceptions, or conditions. 
+  Multiple rounds of validation before final sign-off.
+  Example: "Data retention varies by type, region, consent status, and business need" - needs deep exploration.
+
+Count the following as "variables":
+- Distinct rules or requirements
+- Conditional branches (if/then/else)
+- Exceptions to rules
+- Categories or types that affect decisions
+- Time-based conditions
+- Role-based permissions
+
+Respond with your analysis."""
+
+POLICY_PLANNING_PROMPT = """You are an expert at creating training data plans for AI models across ANY domain.
+
+Given a task description, policy, or domain specification and a target number of traces, analyze the content and create an optimal plan for generating training data.
+
+Your task:
+1. Deeply analyze the domain/task to understand its core concepts, rules, processes, and challenges
+2. Identify distinct SCENARIO CATEGORIES that test different aspects of the domain
+3. Distribute the target traces across categories based on complexity and importance
+4. Ensure coverage of: clear violations/errors, edge cases, happy paths, real-world constraints, and domain-specific challenges
+
+Guidelines for dynamic category creation:
+- **Analyze the domain deeply**: Understand the core rules, processes, stakeholders, and common challenges
+- **Create domain-specific categories**: Base categories on the actual content, not generic assumptions
+- **Balance complexity**: Allocate based on domain complexity (simple domains: 60% happy paths, complex domains: 40% edge cases)
+- **Ensure comprehensive coverage**: Every major aspect of the domain should be tested
+- **Consider domain-specific challenges**: Time pressure in trading, regulatory changes in finance, technical failures in engineering, etc.
+
+For each category, provide:
+- name: Short descriptive name specific to the domain
+- description: What this category tests, including specific domain concepts and challenges
+- traces: Number of traces to generate (must sum to target)
+
+Provide detailed reasoning explaining:
+1. Your analysis of the domain's core concepts and challenges
+2. Why you chose these specific categories for this domain
+3. How the category distribution reflects the domain's complexity and real-world usage patterns"""
+
+# =============================================================================
+# SCENARIO GENERATION PROMPTS
+# =============================================================================
+
+SCENARIO_GENERATOR_PROMPT = """You are an expert at creating realistic scenarios for ANY domain or task.
+
+Given a task description, policy, or domain specification, first deeply analyze the domain to understand:
+- Core concepts, rules, and processes
+- Key stakeholders and their roles
+- Common challenges and failure modes
+- Domain-specific terminology and workflows
+
+Then generate diverse scenarios that thoroughly test understanding of the domain:
+
+1. **Clear Success/Failure Cases** - Obvious correct/incorrect applications of domain rules
+2. **Edge Cases** - Ambiguous situations with multiple valid interpretations
+3. **Multi-Step Processes** - Complex scenarios requiring sequential reasoning
+4. **Real-World Constraints** - Practical limitations like time pressure, incomplete info, resource constraints
+5. **Domain-Specific Challenges** - Scenarios that test unique aspects of this particular domain
+6. **Stakeholder Interactions** - Situations involving coordination between different parties
+7. **Exception Handling** - Scenarios requiring deviation from standard processes
+
+Make each scenario:
+- Deeply grounded in the specific domain's concepts and terminology
+- Realistic and challenging for someone working in that domain
+- Specific with concrete details that reflect actual domain practices
+- Varied in complexity and stakeholder perspectives
+- Designed to reveal both expert and novice understanding gaps
+
+Focus on creating "golden traces" - perfect examples that demonstrate deep domain mastery."""
+
+CATEGORY_SCENARIO_PROMPT = """You are an expert at creating realistic scenarios for ANY domain or task.
+
+Generate scenarios specifically for the following CATEGORY within the given domain:
+**Category Name**: {CATEGORY_NAME}
+**Category Description**: {CATEGORY_DESCRIPTION}
+
+First, deeply understand:
+- How this category fits into the broader domain
+- What specific skills or knowledge this category tests
+- The real-world contexts where this category applies
+- Common mistakes or misconceptions in this category
+
+All generated scenarios MUST:
+- Perfectly fit this specific category's focus and objectives
+- Demonstrate deep understanding of the category's role in the domain
+- Test the exact skills and knowledge described in the category
+- Be realistic and occur in actual domain practice
+
+Make each scenario:
+- Highly specific with concrete details that reflect domain expertise
+- Challenging and nuanced - not simplistic examples
+- Varied in stakeholder perspectives, contexts, and complexity levels
+- Different from each other (no duplicates) - explore different facets of the category
+- Include domain-specific terminology, processes, and challenges
+- Designed as "golden traces" that showcase expert-level understanding
+
+Focus on creating scenarios that would distinguish between novice and expert performance in this category."""
+
+# =============================================================================
+# SYSTEM PROMPT
+# =============================================================================
+
+SYSTEM_PROMPT = """You are a domain expert. When given a scenario and context, provide comprehensive, expert-level guidance.
+
+IMPORTANT: Always show your reasoning process using <reasoning> tags before giving your answer.
+
+Your responses must:
+- Start with <reasoning> tags showing step-by-step analysis
+- Cite specific domain concepts, rules, or processes that apply
+- Give specific, actionable recommendations grounded in domain best practices
+- Address all aspects of the scenario from multiple stakeholder perspectives
+- Acknowledge edge cases, exceptions, and potential complications
+- Consider contemporary challenges and modern practices in the domain
+
+Vary your response style while maintaining expertise:
+- For concise responses: Direct, focused guidance with key domain principles
+- For detailed responses: Comprehensive analysis with structured breakdowns and examples
+- For practical responses: Step-by-step implementation guides and checklists
+- For complex responses: Thorough exploration of trade-offs and alternative approaches
+
+Always prioritize accuracy, clarity, and deep domain understanding in your guidance."""
+
+# =============================================================================
+# BATCHED PROMPTS (for batch generation)
+# =============================================================================
+
+BATCHED_RESPONSE_PROMPT = """You are generating training data for a domain expert model.
+
+For EACH scenario below, create a complete training example in MESSAGES FORMAT.
+
+Each training example must have exactly 3 messages:
+1. "system" - The system prompt defining the assistant's role
+2. "user" - The scenario and context as the user's question
+3. "assistant" - Your expert guidance response
+
+The assistant response must:
+- Cite specific policy sections that apply
+- Explain reasoning step-by-step
+- Give specific, actionable recommendations
+- Address all aspects of the scenario
+- Acknowledge edge cases and complications
+
+Respond with a JSON array where each object has:
+- "index": the scenario number (0-based)
+- "messages": array of 3 message objects with "role" and "content" fields"""
+
+BATCHED_GRADER_PROMPT = """You are a strict policy compliance evaluator. Your job is to determine if EACH response is FULLY CORRECT.
+
+A response PASSES only if ALL of the following are true:
+1. **Policy Compliant** - Every recommendation follows the policy exactly. No violations.
+2. **Fully Supported** - Every claim is backed by a specific policy section. Nothing made up.
+3. **Properly Cited** - All relevant policy sections are explicitly referenced.
+4. **Complete Reasoning** - The chain of thought is complete with no gaps or skipped steps.
+5. **Actionable & Specific** - All recommendations are concrete and implementable, not vague.
+
+If ANY of these fail, the response does NOT pass. Be strict - only mark "pass": true for perfect responses.
+
+For each response, provide structured feedback:
+- "policy_violations": List any rules misinterpreted or violated
+- "missing_citations": List policy sections that should have been cited
+- "incomplete_reasoning": List logical gaps or missing reasoning steps
+- "vague_recommendations": List recommendations that need to be more specific
+- "feedback": Summary of what needs to be fixed
+
+Respond with a JSON array where each object has:
+- "index": the scenario number (0-based)
+- "pass": boolean (true ONLY if response is fully correct)
+- "policy_violations": array of violations
+- "missing_citations": array of missing citations
+- "incomplete_reasoning": array of reasoning gaps
+- "vague_recommendations": array of vague items
+- "feedback": summary of how to fix"""
+
+BATCHED_REFINER_PROMPT = """You are improving training data for a domain expert model based on grader feedback.
+
+For EACH scenario with feedback below, fix ALL issues while keeping what was correct.
+
+You will receive structured feedback with:
+- policy_violations: Rules you violated or misinterpreted - FIX THESE
+- missing_citations: Policy sections you should cite - ADD THESE  
+- incomplete_reasoning: Gaps in your logic - FILL THESE IN
+- vague_recommendations: Things that need to be more specific - MAKE CONCRETE
+
+Requirements:
+1. Fix every policy violation - ensure recommendations follow the policy exactly
+2. Add citations for every missing policy section mentioned
+3. Complete any incomplete reasoning chains with step-by-step logic
+4. Replace vague language with specific, actionable recommendations
+5. Keep the parts that were already correct
+
+Output in MESSAGES FORMAT with exactly 3 messages:
+1. "system" - The system prompt defining the assistant's role
+2. "user" - The scenario and context as the user's question  
+3. "assistant" - Your IMPROVED guidance
+
+Respond with a JSON array where each object has:
+- "index": the scenario number (0-based)
+- "messages": array of 3 message objects with "role" and "content" fields"""
+
+# =============================================================================
+# SINGLE PROMPTS (for parallel high-concurrency generation)
+# =============================================================================
+
+SINGLE_RESPONSE_PROMPT = """You are a domain expert generating a training example.
+
+Given the scenario and policy below, create a complete training example.
+
+Your response must be a JSON object with exactly 3 messages:
+{{
+  "messages": [
+    {{"role": "system", "content": "<system prompt defining expert role>"}},
+    {{"role": "user", "content": "<the scenario as a user question>"}},
+    {{"role": "assistant", "content": "<your expert response>"}}
+  ]
+}}
+
+The assistant response must:
+- Start with <reasoning> tags showing your thought process
+- Cite specific policy sections that apply
+- Give specific, actionable recommendations
+- Address all aspects of the scenario
+- Acknowledge edge cases and complications
+
+SCENARIO:
+{scenario}
+
+CONTEXT:
+{context}
+
+POLICY:
+{policy}
+
+Respond with ONLY the JSON object, no additional text."""
+
+SINGLE_GRADE_PROMPT = """You are a strict evaluator. Grade this response.
+
+A response PASSES only if ALL are true:
+1. Policy Compliant - Every recommendation follows the policy exactly
+2. Fully Supported - Every claim backed by specific policy section
+3. Properly Cited - All relevant policy sections referenced
+4. Complete Reasoning - Chain of thought has no gaps
+5. Actionable & Specific - Recommendations are concrete, not vague
+
+SCENARIO:
+{scenario}
+
+POLICY:
+{policy}
+
+RESPONSE TO GRADE:
+{response}
+
+Respond with ONLY a JSON object:
+{{
+  "pass": <true/false>,
+  "policy_violations": ["<violation 1>", ...],
+  "missing_citations": ["<missing 1>", ...],
+  "incomplete_reasoning": ["<gap 1>", ...],
+  "vague_recommendations": ["<vague 1>", ...],
+  "feedback": "<summary of issues or 'Correct'>"
+}}"""
+