{ "ok": true, "status": "PASS", "version": "1.33.0", "mode": "synthetic_no_web_research", "passed": 128, "total": 128, "failed": 0, "results": [ { "name": "human_aging", "query": "Define Aging in humans", "expected_intent": "DEFINE", "actual_intent": "DEFINE", "semantic_intent": "DEFINE", "canonical_topic": "Human Aging", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "canonical_match": true, "route_match": true }, "status": "PASS" }, { "name": "why_rust", "query": "Why does iron rust?", "expected_intent": "EXPLAIN_CAUSE", "actual_intent": "EXPLAIN_CAUSE", "semantic_intent": "EXPLAIN_CAUSE", "canonical_topic": "Iron Rusting", "semantic_roles": { "phenomenon": "iron rust", "evidence_goal": "cause_or_explanation", "subject": "iron", "predicate": "rust" }, "route": "reasoning_with_evidence", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "canonical_match": true, "route_match": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "mechanism", "query": "How does photosynthesis work?", "expected_intent": "EXPLAIN_MECHANISM", "actual_intent": "EXPLAIN_MECHANISM", "semantic_intent": "EXPLAIN_MECHANISM", "canonical_topic": "Photosynthesis", "semantic_roles": { "subject": "photosynthesis", "evidence_goal": "mechanism" }, "route": "reasoning_with_evidence", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "canonical_match": true, "route_match": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "causal_discrimination", "query": "Determine whether A or B causes C", "expected_intent": "CAUSAL_DISCRIMINATION", "actual_intent": "CAUSAL_DISCRIMINATION", "semantic_intent": "CAUSAL_DISCRIMINATION", "canonical_topic": "C Causation", "semantic_roles": { "candidate_causes": [ "A", "B" ], "effect": "C", "decision": "most_supported_cause" }, "route": "reasoning_with_evidence", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "canonical_match": true, "route_match": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "compare", "query": "Compare vegetarianism and pescatarianism", "expected_intent": "COMPARE", "actual_intent": "COMPARE", "semantic_intent": "COMPARE", "canonical_topic": "Vegetarianism vs Pescatarianism", "semantic_roles": { "left": "vegetarianism", "right": "pescatarianism", "comparison_mode": "similarities_and_differences" }, "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "canonical_match": true, "route_match": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "distinguish", "query": "Distinguish between machine learning and deep learning", "expected_intent": "DISTINGUISH", "actual_intent": "DISTINGUISH", "semantic_intent": "DISTINGUISH", "canonical_topic": "Machine learning vs Deep learning", "semantic_roles": { "left": "machine learning", "right": "deep learning", "comparison_mode": "differences_and_boundaries" }, "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "canonical_match": true, "route_match": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "counterfactual", "query": "What would happen if Earth stopped rotating?", "expected_intent": "COUNTERFACTUAL", "actual_intent": "COUNTERFACTUAL", "semantic_intent": "COUNTERFACTUAL", "canonical_topic": "Counterfactual: Earth stopped rotating", "semantic_roles": { "condition": "Earth stopped rotating", "output": "counterfactual_consequence" }, "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "procedure", "query": "How do I reset my password?", "expected_intent": "PROCEDURE", "actual_intent": "PROCEDURE", "semantic_intent": "PROCEDURE", "canonical_topic": "How do I reset my password", "semantic_roles": { "goal": "reset my password", "output": "ordered_steps" }, "route": "procedural_answer", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "who", "query": "Who is Ada Lovelace?", "expected_intent": "IDENTIFY_PERSON", "actual_intent": "IDENTIFY_PERSON", "semantic_intent": "IDENTIFY_PERSON", "canonical_topic": "Who is Ada Lovelace", "semantic_roles": { "entity": "Ada Lovelace", "requested_attribute": "identity" }, "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "where", "query": "Where is Mount Everest?", "expected_intent": "LOCATE", "actual_intent": "LOCATE", "semantic_intent": "LOCATE", "canonical_topic": "Where is Mount Everest", "semantic_roles": { "entity": "Mount Everest", "requested_attribute": "location" }, "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "when", "query": "When did Apollo 11 land?", "expected_intent": "TEMPORAL_QUERY", "actual_intent": "TEMPORAL_QUERY", "semantic_intent": "TEMPORAL_QUERY", "canonical_topic": "When did Apollo 11 land", "semantic_roles": { "event": "Apollo 11 land", "requested_attribute": "time" }, "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "semantic_role_present": true }, "status": "PASS" }, { "name": "social_hi", "query": "Hi", "expected_intent": "SOCIAL_GREETING", "actual_intent": "SOCIAL_GREETING", "semantic_intent": "SOCIAL_GREETING", "canonical_topic": "Hi", "semantic_roles": [], "route": "social_response", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "social_wellbeing", "query": "Hi, how are you?", "expected_intent": "SOCIAL_WELLBEING", "actual_intent": "SOCIAL_WELLBEING", "semantic_intent": "SOCIAL_WELLBEING", "canonical_topic": "Hi, how are you", "semantic_roles": [], "route": "social_response", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "social_thanks", "query": "Thank you", "expected_intent": "SOCIAL_THANKS", "actual_intent": "SOCIAL_THANKS", "semantic_intent": "SOCIAL_THANKS", "canonical_topic": "Thank you", "semantic_roles": [], "route": "social_response", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "social_ack", "query": "Agreed", "expected_intent": "SOCIAL_ACKNOWLEDGEMENT", "actual_intent": "SOCIAL_ACKNOWLEDGEMENT", "semantic_intent": "SOCIAL_ACKNOWLEDGEMENT", "canonical_topic": "Agreed", "semantic_roles": [], "route": "social_response", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "social_bye", "query": "Goodbye", "expected_intent": "SOCIAL_FAREWELL", "actual_intent": "SOCIAL_FAREWELL", "semantic_intent": "SOCIAL_FAREWELL", "canonical_topic": "Goodbye", "semantic_roles": [], "route": "social_response", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "correction", "query": "I meant Java the programming language, not coffee", "expected_intent": "CORRECT_INTENT", "actual_intent": "CORRECT_INTENT", "semantic_intent": "CORRECT_INTENT", "canonical_topic": "I meant Java the programming language, not coffee", "semantic_roles": [], "route": "conversation_control", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "continue", "query": "Proceed", "expected_intent": "CONTINUE_PRIOR_TASK", "actual_intent": "CONTINUE_PRIOR_TASK", "semantic_intent": "CONTINUE_PRIOR_TASK", "canonical_topic": "Proceed", "semantic_roles": [], "route": "conversation_control", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "repeat", "query": "Do it again", "expected_intent": "REPEAT_ACTION", "actual_intent": "REPEAT_ACTION", "semantic_intent": "REPEAT_ACTION", "canonical_topic": "Do it again", "semantic_roles": [], "route": "conversation_control", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "cancel", "query": "Never mind", "expected_intent": "CANCEL_INTENT", "actual_intent": "CANCEL_INTENT", "semantic_intent": "CANCEL_INTENT", "canonical_topic": "Never mind", "semantic_roles": [], "route": "conversation_control", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "route_match": true }, "status": "PASS" }, { "name": "multi_action", "query": "Remove all that and populate the table with the new client data", "expected_intent": "UNDO_OR_REMOVE_CONTEXT", "actual_intent": "UNDO_OR_REMOVE_CONTEXT", "semantic_intent": "UNDO_OR_REMOVE_CONTEXT", "canonical_topic": "Remove all that and populate the table with the new client data", "semantic_roles": [], "route": "conversation_control", "multi_intent": true, "node_count": 2, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "multi_intent_match": true, "route_match": true }, "status": "PASS" }, { "name": "multi_explain", "query": "Explain photosynthesis and then explain cellular respiration", "expected_intent": "EXPLAIN", "actual_intent": "EXPLAIN", "semantic_intent": "EXPLAIN", "canonical_topic": "Photosynthesis and then explain cellular respiration", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": true, "node_count": 2, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true, "multi_intent_match": true }, "status": "PASS" }, { "name": "define_0", "query": "Define gravity", "expected_intent": "DEFINE", "actual_intent": "DEFINE", "semantic_intent": "DEFINE", "canonical_topic": "Gravity", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "define_1", "query": "Define democracy", "expected_intent": "DEFINE", "actual_intent": "DEFINE", "semantic_intent": "DEFINE", "canonical_topic": "Democracy", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "define_2", "query": "Define mitochondria", "expected_intent": "DEFINE", "actual_intent": "DEFINE", "semantic_intent": "DEFINE", "canonical_topic": "Mitochondria", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "define_3", "query": "Define encryption", "expected_intent": "DEFINE", "actual_intent": "DEFINE", "semantic_intent": "DEFINE", "canonical_topic": "Encryption", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "describe_0", "query": "Describe a black hole", "expected_intent": "DESCRIBE", "actual_intent": "DESCRIBE", "semantic_intent": "DESCRIBE", "canonical_topic": "A black hole", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "describe_1", "query": "Describe photosynthesis", "expected_intent": "DESCRIBE", "actual_intent": "DESCRIBE", "semantic_intent": "DESCRIBE", "canonical_topic": "Photosynthesis", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "describe_2", "query": "Describe a neural network", "expected_intent": "DESCRIBE", "actual_intent": "DESCRIBE", "semantic_intent": "DESCRIBE", "canonical_topic": "A neural network", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "describe_3", "query": "Describe inflation", "expected_intent": "DESCRIBE", "actual_intent": "DESCRIBE", "semantic_intent": "DESCRIBE", "canonical_topic": "Inflation", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "explain_0", "query": "Explain quantum entanglement", "expected_intent": "EXPLAIN", "actual_intent": "EXPLAIN", "semantic_intent": "EXPLAIN", "canonical_topic": "Quantum entanglement", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "explain_1", "query": "Explain supply and demand", "expected_intent": "EXPLAIN", "actual_intent": "EXPLAIN", "semantic_intent": "EXPLAIN", "canonical_topic": "Supply and demand", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "explain_2", "query": "Explain DNA replication", "expected_intent": "EXPLAIN", "actual_intent": "EXPLAIN", "semantic_intent": "EXPLAIN", "canonical_topic": "DNA replication", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "explain_3", "query": "Explain cloud computing", "expected_intent": "EXPLAIN", "actual_intent": "EXPLAIN", "semantic_intent": "EXPLAIN", "canonical_topic": "Cloud computing", "semantic_roles": [], "route": "knowledge_or_research", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "summarize_0", "query": "Summarize this report", "expected_intent": "SUMMARIZE", "actual_intent": "SUMMARIZE", "semantic_intent": "SUMMARIZE", "canonical_topic": "This report", "semantic_roles": { "source": "this report", "output": "compressed_representation" }, "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "summarize_1", "query": "Summarize the findings", "expected_intent": "SUMMARIZE", "actual_intent": "SUMMARIZE", "semantic_intent": "SUMMARIZE", "canonical_topic": "The findings", "semantic_roles": { "source": "the findings", "output": "compressed_representation" }, "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "summarize_2", "query": "Summarize the article", "expected_intent": "SUMMARIZE", "actual_intent": "SUMMARIZE", "semantic_intent": "SUMMARIZE", "canonical_topic": "The article", "semantic_roles": { "source": "the article", "output": "compressed_representation" }, "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "summarize_3", "query": "Summarize the conversation", "expected_intent": "SUMMARIZE", "actual_intent": "SUMMARIZE", "semantic_intent": "SUMMARIZE", "canonical_topic": "The conversation", "semantic_roles": { "source": "the conversation", "output": "compressed_representation" }, "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "list_0", "query": "List the planets", "expected_intent": "LIST", "actual_intent": "LIST", "semantic_intent": "LIST", "canonical_topic": "The planets", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "list_1", "query": "List key risks", "expected_intent": "LIST", "actual_intent": "LIST", "semantic_intent": "LIST", "canonical_topic": "Key risks", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "list_2", "query": "List project milestones", "expected_intent": "LIST", "actual_intent": "LIST", "semantic_intent": "LIST", "canonical_topic": "Project milestones", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "list_3", "query": "List the requirements", "expected_intent": "LIST", "actual_intent": "LIST", "semantic_intent": "LIST", "canonical_topic": "The requirements", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "extract_0", "query": "Extract the dates", "expected_intent": "EXTRACT", "actual_intent": "EXTRACT", "semantic_intent": "EXTRACT", "canonical_topic": "The dates", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "extract_1", "query": "Extract the names", "expected_intent": "EXTRACT", "actual_intent": "EXTRACT", "semantic_intent": "EXTRACT", "canonical_topic": "The names", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "extract_2", "query": "Extract the totals", "expected_intent": "EXTRACT", "actual_intent": "EXTRACT", "semantic_intent": "EXTRACT", "canonical_topic": "The totals", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "extract_3", "query": "Extract the action items", "expected_intent": "EXTRACT", "actual_intent": "EXTRACT", "semantic_intent": "EXTRACT", "canonical_topic": "The action items", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "analyze_0", "query": "Analyze this dataset", "expected_intent": "ANALYZE", "actual_intent": "ANALYZE", "semantic_intent": "ANALYZE", "canonical_topic": "This dataset", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "analyze_1", "query": "Analyze the market", "expected_intent": "ANALYZE", "actual_intent": "ANALYZE", "semantic_intent": "ANALYZE", "canonical_topic": "The market", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "analyze_2", "query": "Analyze the argument", "expected_intent": "ANALYZE", "actual_intent": "ANALYZE", "semantic_intent": "ANALYZE", "canonical_topic": "The argument", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "analyze_3", "query": "Analyze the failure", "expected_intent": "ANALYZE", "actual_intent": "ANALYZE", "semantic_intent": "ANALYZE", "canonical_topic": "The failure", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "evaluate_0", "query": "Evaluate this proposal", "expected_intent": "EVALUATE", "actual_intent": "EVALUATE", "semantic_intent": "EVALUATE", "canonical_topic": "This proposal", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "evaluate_1", "query": "Evaluate the evidence", "expected_intent": "EVALUATE", "actual_intent": "EVALUATE", "semantic_intent": "EVALUATE", "canonical_topic": "The evidence", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "evaluate_2", "query": "Evaluate performance", "expected_intent": "EVALUATE", "actual_intent": "EVALUATE", "semantic_intent": "EVALUATE", "canonical_topic": "Performance", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "evaluate_3", "query": "Evaluate the options", "expected_intent": "EVALUATE", "actual_intent": "EVALUATE", "semantic_intent": "EVALUATE", "canonical_topic": "The options", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "determine_0", "query": "Determine the best option", "expected_intent": "DETERMINE", "actual_intent": "DETERMINE", "semantic_intent": "DETERMINE", "canonical_topic": "The best option", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "determine_1", "query": "Determine the root cause", "expected_intent": "DETERMINE", "actual_intent": "DETERMINE", "semantic_intent": "DETERMINE", "canonical_topic": "The root cause", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "determine_2", "query": "Determine the winner", "expected_intent": "DETERMINE", "actual_intent": "DETERMINE", "semantic_intent": "DETERMINE", "canonical_topic": "The winner", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "determine_3", "query": "Determine feasibility", "expected_intent": "DETERMINE", "actual_intent": "DETERMINE", "semantic_intent": "DETERMINE", "canonical_topic": "Feasibility", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "infer_0", "query": "Infer the likely cause", "expected_intent": "INFER", "actual_intent": "INFER", "semantic_intent": "INFER", "canonical_topic": "Infer the likely cause", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "infer_1", "query": "Infer user intent", "expected_intent": "INFER", "actual_intent": "INFER", "semantic_intent": "INFER", "canonical_topic": "Infer user intent", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "infer_2", "query": "Infer the pattern", "expected_intent": "INFER", "actual_intent": "INFER", "semantic_intent": "INFER", "canonical_topic": "Infer the pattern", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "infer_3", "query": "Infer the missing value", "expected_intent": "INFER", "actual_intent": "INFER", "semantic_intent": "INFER", "canonical_topic": "Infer the missing value", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "classify_0", "query": "Classify these records", "expected_intent": "CLASSIFY", "actual_intent": "CLASSIFY", "semantic_intent": "CLASSIFY", "canonical_topic": "Classify these records", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "classify_1", "query": "Classify the image", "expected_intent": "CLASSIFY", "actual_intent": "CLASSIFY", "semantic_intent": "CLASSIFY", "canonical_topic": "Classify the image", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "classify_2", "query": "Classify the request", "expected_intent": "CLASSIFY", "actual_intent": "CLASSIFY", "semantic_intent": "CLASSIFY", "canonical_topic": "Classify the request", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "classify_3", "query": "Classify the risks", "expected_intent": "CLASSIFY", "actual_intent": "CLASSIFY", "semantic_intent": "CLASSIFY", "canonical_topic": "Classify the risks", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "predict_0", "query": "Predict next quarter sales", "expected_intent": "PREDICT", "actual_intent": "PREDICT", "semantic_intent": "PREDICT", "canonical_topic": "Next quarter sales", "semantic_roles": { "target": "next quarter sales", "time_orientation": "future" }, "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "predict_1", "query": "Predict the likely outcome", "expected_intent": "PREDICT", "actual_intent": "PREDICT", "semantic_intent": "PREDICT", "canonical_topic": "The likely outcome", "semantic_roles": { "target": "the likely outcome", "time_orientation": "future" }, "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "predict_2", "query": "Predict demand", "expected_intent": "PREDICT", "actual_intent": "PREDICT", "semantic_intent": "PREDICT", "canonical_topic": "Demand", "semantic_roles": { "target": "demand", "time_orientation": "future" }, "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "predict_3", "query": "Predict failure risk", "expected_intent": "PREDICT", "actual_intent": "PREDICT", "semantic_intent": "PREDICT", "canonical_topic": "Failure risk", "semantic_roles": { "target": "failure risk", "time_orientation": "future" }, "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "extrapolate_0", "query": "Extrapolate the trend", "expected_intent": "EXTRAPOLATE", "actual_intent": "EXTRAPOLATE", "semantic_intent": "EXTRAPOLATE", "canonical_topic": "Extrapolate the trend", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "extrapolate_1", "query": "Extrapolate future demand", "expected_intent": "EXTRAPOLATE", "actual_intent": "EXTRAPOLATE", "semantic_intent": "EXTRAPOLATE", "canonical_topic": "Extrapolate future demand", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "extrapolate_2", "query": "Extrapolate the series", "expected_intent": "EXTRAPOLATE", "actual_intent": "EXTRAPOLATE", "semantic_intent": "EXTRAPOLATE", "canonical_topic": "Extrapolate the series", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "extrapolate_3", "query": "Extrapolate growth", "expected_intent": "EXTRAPOLATE", "actual_intent": "EXTRAPOLATE", "semantic_intent": "EXTRAPOLATE", "canonical_topic": "Extrapolate growth", "semantic_roles": [], "route": "reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "interpret_0", "query": "Interpret this chart", "expected_intent": "INTERPRET", "actual_intent": "INTERPRET", "semantic_intent": "INTERPRET", "canonical_topic": "This chart", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "interpret_1", "query": "Interpret the result", "expected_intent": "INTERPRET", "actual_intent": "INTERPRET", "semantic_intent": "INTERPRET", "canonical_topic": "The result", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "interpret_2", "query": "Interpret this sentence", "expected_intent": "INTERPRET", "actual_intent": "INTERPRET", "semantic_intent": "INTERPRET", "canonical_topic": "This sentence", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "interpret_3", "query": "Interpret the contract clause", "expected_intent": "INTERPRET", "actual_intent": "INTERPRET", "semantic_intent": "INTERPRET", "canonical_topic": "The contract clause", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "decipher_0", "query": "Decipher this code", "expected_intent": "DECIPHER", "actual_intent": "DECIPHER", "semantic_intent": "DECIPHER", "canonical_topic": "This code", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "decipher_1", "query": "Decipher the message", "expected_intent": "DECIPHER", "actual_intent": "DECIPHER", "semantic_intent": "DECIPHER", "canonical_topic": "The message", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "decipher_2", "query": "Decipher the pattern", "expected_intent": "DECIPHER", "actual_intent": "DECIPHER", "semantic_intent": "DECIPHER", "canonical_topic": "The pattern", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "decipher_3", "query": "Decipher the notation", "expected_intent": "DECIPHER", "actual_intent": "DECIPHER", "semantic_intent": "DECIPHER", "canonical_topic": "The notation", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "recommend_0", "query": "Recommend a strategy", "expected_intent": "RECOMMEND", "actual_intent": "RECOMMEND", "semantic_intent": "RECOMMEND", "canonical_topic": "Recommend a strategy", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "recommend_1", "query": "Recommend the best option", "expected_intent": "RECOMMEND", "actual_intent": "RECOMMEND", "semantic_intent": "RECOMMEND", "canonical_topic": "Recommend the best option", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "recommend_2", "query": "Recommend improvements", "expected_intent": "RECOMMEND", "actual_intent": "RECOMMEND", "semantic_intent": "RECOMMEND", "canonical_topic": "Recommend improvements", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "recommend_3", "query": "Recommend next steps", "expected_intent": "RECOMMEND", "actual_intent": "RECOMMEND", "semantic_intent": "RECOMMEND", "canonical_topic": "Recommend next steps", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "select_0", "query": "Select the best candidate", "expected_intent": "SELECT", "actual_intent": "SELECT", "semantic_intent": "SELECT", "canonical_topic": "Select the best candidate", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "select_1", "query": "Choose the strongest option", "expected_intent": "SELECT", "actual_intent": "SELECT", "semantic_intent": "SELECT", "canonical_topic": "Choose the strongest option", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "select_2", "query": "Select a model", "expected_intent": "SELECT", "actual_intent": "SELECT", "semantic_intent": "SELECT", "canonical_topic": "Select a model", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "select_3", "query": "Choose a route", "expected_intent": "SELECT", "actual_intent": "SELECT", "semantic_intent": "SELECT", "canonical_topic": "Choose a route", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "calculate_0", "query": "Calculate 2 plus 2", "expected_intent": "CALCULATE", "actual_intent": "CALCULATE", "semantic_intent": "CALCULATE", "canonical_topic": "2 plus 2", "semantic_roles": [], "route": "deterministic_compute", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "calculate_1", "query": "Calculate compound interest", "expected_intent": "CALCULATE", "actual_intent": "CALCULATE", "semantic_intent": "CALCULATE", "canonical_topic": "Compound interest", "semantic_roles": [], "route": "deterministic_compute", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "calculate_2", "query": "Calculate the average", "expected_intent": "CALCULATE", "actual_intent": "CALCULATE", "semantic_intent": "CALCULATE", "canonical_topic": "The average", "semantic_roles": [], "route": "deterministic_compute", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "calculate_3", "query": "Calculate the total", "expected_intent": "CALCULATE", "actual_intent": "CALCULATE", "semantic_intent": "CALCULATE", "canonical_topic": "The total", "semantic_roles": [], "route": "deterministic_compute", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "estimate_0", "query": "Estimate the cost", "expected_intent": "ESTIMATE", "actual_intent": "ESTIMATE", "semantic_intent": "ESTIMATE", "canonical_topic": "Estimate the cost", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "estimate_1", "query": "Estimate completion time", "expected_intent": "ESTIMATE", "actual_intent": "ESTIMATE", "semantic_intent": "ESTIMATE", "canonical_topic": "Estimate completion time", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "estimate_2", "query": "Estimate demand", "expected_intent": "ESTIMATE", "actual_intent": "ESTIMATE", "semantic_intent": "ESTIMATE", "canonical_topic": "Estimate demand", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "estimate_3", "query": "Estimate the distance", "expected_intent": "ESTIMATE", "actual_intent": "ESTIMATE", "semantic_intent": "ESTIMATE", "canonical_topic": "Estimate the distance", "semantic_roles": [], "route": "core_fallback", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "count_0", "query": "Count the records", "expected_intent": "COUNT", "actual_intent": "COUNT", "semantic_intent": "COUNT", "canonical_topic": "Count the records", "semantic_roles": [], "route": "deterministic_compute", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "count_1", "query": "Count the words", "expected_intent": "COUNT", "actual_intent": "COUNT", "semantic_intent": "COUNT", "canonical_topic": "Count the words", "semantic_roles": [], "route": "deterministic_compute", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "count_2", "query": "Count the failures", "expected_intent": "COUNT", "actual_intent": "COUNT", "semantic_intent": "COUNT", "canonical_topic": "Count the failures", "semantic_roles": [], "route": "deterministic_compute", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "count_3", "query": "Count the clients", "expected_intent": "COUNT", "actual_intent": "COUNT", "semantic_intent": "COUNT", "canonical_topic": "Count the clients", "semantic_roles": [], "route": "deterministic_compute", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "plan_0", "query": "Plan the migration", "expected_intent": "PLAN", "actual_intent": "PLAN", "semantic_intent": "PLAN", "canonical_topic": "Plan the migration", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "plan_1", "query": "Plan a product launch", "expected_intent": "PLAN", "actual_intent": "PLAN", "semantic_intent": "PLAN", "canonical_topic": "Plan a product launch", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "plan_2", "query": "Plan the project", "expected_intent": "PLAN", "actual_intent": "PLAN", "semantic_intent": "PLAN", "canonical_topic": "Plan the project", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "plan_3", "query": "Plan the rollout", "expected_intent": "PLAN", "actual_intent": "PLAN", "semantic_intent": "PLAN", "canonical_topic": "Plan the rollout", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "organize_0", "query": "Organize these files", "expected_intent": "ORGANIZE", "actual_intent": "ORGANIZE", "semantic_intent": "ORGANIZE", "canonical_topic": "These files", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "organize_1", "query": "Organize the tasks", "expected_intent": "ORGANIZE", "actual_intent": "ORGANIZE", "semantic_intent": "ORGANIZE", "canonical_topic": "The tasks", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "organize_2", "query": "Organize the research", "expected_intent": "ORGANIZE", "actual_intent": "ORGANIZE", "semantic_intent": "ORGANIZE", "canonical_topic": "The research", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "organize_3", "query": "Organize the requirements", "expected_intent": "ORGANIZE", "actual_intent": "ORGANIZE", "semantic_intent": "ORGANIZE", "canonical_topic": "The requirements", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "prioritize_0", "query": "Prioritize the bugs", "expected_intent": "PRIORITIZE", "actual_intent": "PRIORITIZE", "semantic_intent": "PRIORITIZE", "canonical_topic": "Prioritize the bugs", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "prioritize_1", "query": "Prioritize the leads", "expected_intent": "PRIORITIZE", "actual_intent": "PRIORITIZE", "semantic_intent": "PRIORITIZE", "canonical_topic": "Prioritize the leads", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "prioritize_2", "query": "Prioritize the risks", "expected_intent": "PRIORITIZE", "actual_intent": "PRIORITIZE", "semantic_intent": "PRIORITIZE", "canonical_topic": "Prioritize the risks", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "prioritize_3", "query": "Prioritize the tasks", "expected_intent": "PRIORITIZE", "actual_intent": "PRIORITIZE", "semantic_intent": "PRIORITIZE", "canonical_topic": "Prioritize the tasks", "semantic_roles": [], "route": "planning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "troubleshoot_0", "query": "Troubleshoot the login error", "expected_intent": "TROUBLESHOOT", "actual_intent": "TROUBLESHOOT", "semantic_intent": "TROUBLESHOOT", "canonical_topic": "Troubleshoot the login error", "semantic_roles": [], "route": "diagnostic_reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "troubleshoot_1", "query": "Troubleshoot the API failure", "expected_intent": "TROUBLESHOOT", "actual_intent": "TROUBLESHOOT", "semantic_intent": "TROUBLESHOOT", "canonical_topic": "Troubleshoot the API failure", "semantic_roles": [], "route": "diagnostic_reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "troubleshoot_2", "query": "Troubleshoot the upload", "expected_intent": "TROUBLESHOOT", "actual_intent": "TROUBLESHOOT", "semantic_intent": "TROUBLESHOOT", "canonical_topic": "Troubleshoot the upload", "semantic_roles": [], "route": "diagnostic_reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "troubleshoot_3", "query": "Troubleshoot the network", "expected_intent": "TROUBLESHOOT", "actual_intent": "TROUBLESHOOT", "semantic_intent": "TROUBLESHOOT", "canonical_topic": "Troubleshoot the network", "semantic_roles": [], "route": "diagnostic_reasoning", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "create_0", "query": "Create a report", "expected_intent": "CREATE", "actual_intent": "CREATE", "semantic_intent": "CREATE", "canonical_topic": "A report", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "create_1", "query": "Create a logo brief", "expected_intent": "CREATE", "actual_intent": "CREATE", "semantic_intent": "CREATE", "canonical_topic": "A logo brief", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "create_2", "query": "Create a project plan", "expected_intent": "CREATE", "actual_intent": "CREATE", "semantic_intent": "CREATE", "canonical_topic": "A project plan", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "create_3", "query": "Create a template", "expected_intent": "CREATE", "actual_intent": "CREATE", "semantic_intent": "CREATE", "canonical_topic": "A template", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "build_0", "query": "Build a dashboard", "expected_intent": "BUILD", "actual_intent": "BUILD", "semantic_intent": "BUILD", "canonical_topic": "A dashboard", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "build_1", "query": "Build a parser", "expected_intent": "BUILD", "actual_intent": "BUILD", "semantic_intent": "BUILD", "canonical_topic": "A parser", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "build_2", "query": "Build a website", "expected_intent": "BUILD", "actual_intent": "BUILD", "semantic_intent": "BUILD", "canonical_topic": "A website", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "build_3", "query": "Build a module", "expected_intent": "BUILD", "actual_intent": "BUILD", "semantic_intent": "BUILD", "canonical_topic": "A module", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "design_0", "query": "Design a workflow", "expected_intent": "DESIGN", "actual_intent": "DESIGN", "semantic_intent": "DESIGN", "canonical_topic": "A workflow", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" }, { "name": "design_1", "query": "Design a database schema", "expected_intent": "DESIGN", "actual_intent": "DESIGN", "semantic_intent": "DESIGN", "canonical_topic": "A database schema", "semantic_roles": [], "route": "artifact_creation", "multi_intent": false, "node_count": 1, "checks": { "intent_match": true, "frame_intent_match": true, "no_instruction_contamination": true }, "status": "PASS" } ], "policy": "V1.28 validates deep intent classification, semantic roles, canonicalization, routing, social separation, and multi-intent decomposition without public web requests." }