| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "grant_iam_role"}, {"confidence": 0.31, "workflow": "create_web_app"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0001", "input": "Need cloud stuff for finance."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0002", "input": "daily reporting job, details later"} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "jane", "role": "owner", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0003", "input": "Grant Jane owner access to production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0004", "input": "scheduler request for reporting"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "platform-web-app", "diagnostics_enabled": true, "environment": "prod", "region": "us-central1", "runtime": "nodejs20", "team": "platform"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0005", "input": "Need a small Node.js app named platform-web-app in us-central1 for platform."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "reporting"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0006", "input": "need api for reporting, details TBD"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "growth-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "growth"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0007", "input": "Create a service account named growth-svc for the growth team in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "principal": "john", "role": "editor", "scope": "dev-subsystem"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0008", "input": "Give john the editor role on dev-subsystem."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "platform"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0009", "input": "Create a service account for the platform team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "mlops-nightly-job", "schedule": "0 9 * * *", "target": "reporting", "team": "mlops", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0010", "input": "Set up a daily job for reporting for the mlops team in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "claims-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "claims"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0011", "input": "identity request: claims service account, env prod, name claims-svc"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "claims-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "claims"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0012", "input": "identity request: claims service account, env prod, name claims-svc"} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Unsafe destructive production request."], "missing_fields": [], "parameters": {}, "status": "rejected", "workflow": null}, "id": "train-0013", "input": "Disable monitoring everywhere before the release."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "deploy-bot", "role": "contributor", "scope": "all reporting resources"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0014", "input": "Please give deploy-bot contributor permissions on all reporting resources for a short migration."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0015", "input": "Set up a reporting schedule."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "grant_iam_role"}, {"confidence": 0.31, "workflow": "create_service_account"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0016", "input": "Need cloud stuff for finance."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "finance"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0017", "input": "Create a Python web app for the finance team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "mlops-web-app", "diagnostics_enabled": false, "environment": "dev", "region": "us-central1", "runtime": "dotnet8", "team": "mlops"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0018", "input": "Need a small .NET app named mlops-web-app in us-central1 for mlops."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "mlops-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "westus", "runtime": "nodejs20", "team": "mlops"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0019", "input": "Need a small Node.js app named mlops-web-app in westus for mlops."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_web_app"}, {"confidence": 0.31, "workflow": "grant_iam_role"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0020", "input": "Can you handle the platform setup?"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0021", "input": "daily reporting job, details later"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "claims-nightly-job", "schedule": "0 2 * * *", "target": "claims-sync", "team": "claims", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0022", "input": "cron 0 2 * * * target claims-sync env staging timezone America/Los_Angeles"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "mlops"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0023", "input": "bucket needed for mlops, no location picked yet"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "principal": "jane", "role": "contributor", "scope": "claims-app"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0024", "input": "iam: principal=jane role=contributor scope=claims-app env=prod"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "finance-nightly-job", "schedule": "0 2 * * *", "target": "model-refresh", "team": "finance", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0025", "input": "Set up a daily job for model-refresh for the finance team in staging."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "claims"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0026", "input": "need api for claims, details TBD"} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_scheduler_job"}, {"confidence": 0.31, "workflow": "create_service_account"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0027", "input": "Make the nightly thing happen."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "claims-nightly-job", "schedule": "0 2 * * *", "target": "reporting", "team": "claims", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0028", "input": "cron 0 2 * * * target reporting env staging timezone America/New_York"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "principal": "jane", "role": "contributor", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0029", "input": "Grant jane contributor access to reporting-project in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "finance-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "westus", "runtime": "python311", "team": "finance"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0030", "input": "Create a staging Python web app for the finance team in West US."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "john", "role": "contributor", "scope": "dev-subsystem"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0031", "input": "Give john the contributor role on dev-subsystem."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "finance-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "finance"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0032", "input": "Need an automation identity for team finance in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "finance-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "finance"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0033", "input": "Create a service account named finance-svc for the finance team in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "principal": "deploy-bot", "role": "editor", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0034", "input": "iam: principal=deploy-bot role=editor scope=reporting-project env=prod"} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_web_app"}, {"confidence": 0.31, "workflow": "grant_iam_role"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0035", "input": "Set up reporting."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_service_account"}, {"confidence": 0.31, "workflow": "create_storage_bucket"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0036", "input": "Prep access and automation for the new project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "finance-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "eastus", "runtime": "nodejs20", "team": "finance"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0037", "input": "ticket: finance staging api, runtime Node.js, region East US, diagnostics on"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "job_name": "reporting-nightly-job", "schedule": "0 2 * * *", "target": "model-refresh", "team": "reporting", "timezone": "UTC"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0038", "input": "Set up a daily job for model-refresh for the reporting team in development."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_web_app"}, {"confidence": 0.31, "workflow": "create_service_account"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0039", "input": "Set up reporting."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "mlops"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0040", "input": "Create storage for the mlops team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "jane", "role": "reader", "scope": "staging-bucket"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0041", "input": "Grant jane reader access to staging-bucket in development."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "finance"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0042", "input": "Create a Python web app for the finance team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "principal": "john", "role": "reader", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0043", "input": "Grant john reader access to reporting-project in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "principal": "john", "role": "viewer", "scope": "dev-subsystem"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0044", "input": "Grant john viewer access to dev-subsystem in production."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "jane", "role": "reader", "scope": "all reporting resources"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0045", "input": "iam: jane role reader scope all reporting resources"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "claims-bucket", "environment": "staging", "region": "eastus", "storage_class": "archive", "team": "claims"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0046", "input": "infra: bucket for claims, env staging, region eastus, class archive"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0047", "input": "Need an automation identity for team platform in development."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "john", "role": "reader", "scope": "production"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0048", "input": "iam: john role reader scope production"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "finance-bucket", "environment": "prod", "region": "eastus", "storage_class": "archive", "team": "finance"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0049", "input": "infra: bucket for finance, env prod, region eastus, class archive"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "principal": "jane", "role": "editor", "scope": "claims-app"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0050", "input": "Give jane the editor role on claims-app."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "reporting-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "reporting"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0051", "input": "identity request: reporting service account, env dev, name reporting-svc"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "platform-nightly-job", "schedule": "0 9 * * *", "target": "reporting", "team": "platform", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0052", "input": "Create a nightly scheduler job named platform-nightly-job for reporting in staging."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0053", "input": "daily reporting job, details later"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "growth-web-app", "diagnostics_enabled": true, "environment": "prod", "region": "westus", "runtime": "nodejs20", "team": "growth"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0054", "input": "Create a production Node.js web app for the growth team in West US."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "growth-bucket", "environment": "staging", "region": "centralus", "storage_class": "cool", "team": "growth"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0055", "input": "Set up blob storage for the growth team in Central US for staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "job_name": "security-nightly-job", "schedule": "0 2 * * *", "target": "billing-export", "team": "security", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0056", "input": "Set up a daily job for billing-export for the security team in development."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "platform"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0057", "input": "Create storage for the platform team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "finance-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "finance"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0058", "input": "Need an automation identity for team finance in production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "reporting"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0059", "input": "Create a Python web app for the reporting team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "reporting-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "eastus", "runtime": "nodejs20", "team": "reporting"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0060", "input": "ticket: reporting staging api, runtime Node.js, region East US, diagnostics on"} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "john", "role": "owner", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0061", "input": "Give John the owner role across all production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "security-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "security"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0062", "input": "Create a service account named security-svc for the security team in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "reporting-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "reporting"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0063", "input": "Create a service account named reporting-svc for the reporting team in development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "platform-web-app", "diagnostics_enabled": false, "environment": "dev", "region": "eastus", "runtime": "dotnet8", "team": "platform"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0064", "input": "ticket: platform development api, runtime .NET, region East US, diagnostics on"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "growth-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "growth"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0065", "input": "Create a service account named growth-svc for the growth team in staging."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "grant_iam_role"}], "clarifying_question": "What role should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: role, scope"], "missing_fields": ["role", "scope"], "parameters": {"principal": "john"}, "status": "needs_clarification", "workflow": "grant_iam_role"}, "id": "train-0066", "input": "Grant john access."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "john", "role": "owner", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0067", "input": "Grant John owner access to production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "security"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0068", "input": "bucket needed for security, no location picked yet"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "grant_iam_role"}], "clarifying_question": "What role should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: role, scope"], "missing_fields": ["role", "scope"], "parameters": {"principal": "jane"}, "status": "needs_clarification", "workflow": "grant_iam_role"}, "id": "train-0069", "input": "iam access needed for jane, scope TBD"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "platform-web-app", "diagnostics_enabled": false, "environment": "prod", "region": "eastus", "runtime": "python311", "team": "platform"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0070", "input": "ticket: platform production api, runtime Python, region East US, diagnostics on"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "claims-nightly-job", "schedule": "0 2 * * *", "target": "billing-export", "team": "claims", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0071", "input": "Create a nightly scheduler job named claims-nightly-job for billing-export in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "reporting-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "reporting"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0072", "input": "identity request: reporting service account, env dev, name reporting-svc"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "mlops-web-app", "diagnostics_enabled": true, "environment": "dev", "region": "eastus", "runtime": "python311", "team": "mlops"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0073", "input": "Need a small Python app named mlops-web-app in eastus for mlops."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "platform-bucket", "environment": "prod", "region": "eastus", "storage_class": "standard", "team": "platform"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0074", "input": "infra: bucket for platform, env prod, region eastus, class standard"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "growth-nightly-job", "schedule": "0 2 * * *", "target": "billing-export", "team": "growth", "timezone": "UTC"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0075", "input": "Create a nightly scheduler job named growth-nightly-job for billing-export in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "jane", "role": "reader", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0076", "input": "Give jane the reader role on reporting-project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "claims-nightly-job", "schedule": "0 9 * * *", "target": "model-refresh", "team": "claims", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0077", "input": "Set up a daily job for model-refresh for the claims team in production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "growth"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0078", "input": "blob storage request: growth"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "jane", "role": "contributor", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0079", "input": "Give jane the contributor role on reporting-project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "reporting-nightly-job", "schedule": "0 9 * * *", "target": "billing-export", "team": "reporting", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0080", "input": "Create a nightly scheduler job named reporting-nightly-job for billing-export in staging."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "john", "role": "reader", "scope": "subscription analytics"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0081", "input": "Please give john reader permissions on subscription analytics for a short migration."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_storage_bucket"}, {"confidence": 0.31, "workflow": "create_web_app"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0082", "input": "Prep access and automation for the new project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "growth-bucket", "environment": "staging", "region": "us-central1", "storage_class": "cool", "team": "growth"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0083", "input": "infra: bucket for growth, env staging, region us-central1, class cool"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "growth-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "growth"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0084", "input": "Need an automation identity for team growth in production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "finance"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0085", "input": "Create a service account for the finance team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "reporting-web-app", "diagnostics_enabled": false, "environment": "dev", "region": "eastus", "runtime": "dotnet8", "team": "reporting"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0086", "input": "ticket: reporting development api, runtime .NET, region East US, diagnostics on"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "principal": "reporting-user", "role": "reader", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0087", "input": "Give reporting-user the reader role on reporting-project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "principal": "jane", "role": "editor", "scope": "staging-bucket"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0088", "input": "iam: principal=jane role=editor scope=staging-bucket env=prod"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "growth-nightly-job", "schedule": "0 9 * * *", "target": "billing-export", "team": "growth", "timezone": "UTC"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0089", "input": "Create a nightly scheduler job named growth-nightly-job for billing-export in staging."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "growth"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0090", "input": "web app request: growth, python"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "growth"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0091", "input": "web app request: growth, python"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "reporting-bucket", "environment": "dev", "region": "eastus", "storage_class": "archive", "team": "reporting"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0092", "input": "Create a archive storage bucket named reporting-bucket in East US for development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "reporting-user", "role": "contributor", "scope": "dev-subsystem"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0093", "input": "Give reporting-user the contributor role on dev-subsystem."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "mlops-web-app", "diagnostics_enabled": false, "environment": "staging", "region": "westus", "runtime": "nodejs20", "team": "mlops"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0094", "input": "Create a staging Node.js web app for the mlops team in West US."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "mlops-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "us-central1", "runtime": "nodejs20", "team": "mlops"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0095", "input": "Need a small Node.js app named mlops-web-app in us-central1 for mlops."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "grant_iam_role"}], "clarifying_question": "What role should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: role, scope"], "missing_fields": ["role", "scope"], "parameters": {"principal": "jane"}, "status": "needs_clarification", "workflow": "grant_iam_role"}, "id": "train-0096", "input": "permission request for jane"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "reporting-web-app", "diagnostics_enabled": true, "environment": "dev", "region": "us-central1", "runtime": "nodejs20", "team": "reporting"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0097", "input": "Create a development Node.js web app for the reporting team in US Central."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "reporting-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "reporting"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0098", "input": "Need an automation identity for team reporting in production."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_scheduler_job"}, {"confidence": 0.31, "workflow": "grant_iam_role"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0099", "input": "Can you handle the platform setup?"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "growth-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "growth"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0100", "input": "Create a service account named growth-svc for the growth team in development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "principal": "john", "role": "viewer", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0101", "input": "iam: principal=john role=viewer scope=reporting-project env=prod"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "security"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0102", "input": "identity needed for team security"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "reporting-web-app", "diagnostics_enabled": true, "environment": "dev", "region": "westus", "runtime": "python311", "team": "reporting"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0103", "input": "Need a small Python app named reporting-web-app in westus for reporting."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "security-web-app", "diagnostics_enabled": false, "environment": "staging", "region": "us-central1", "runtime": "dotnet8", "team": "security"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0104", "input": "Need a small .NET app named security-web-app in us-central1 for security."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "grant_iam_role"}, {"confidence": 0.31, "workflow": "create_web_app"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0105", "input": "Make the nightly thing happen."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "reporting"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0106", "input": "Create storage for the reporting team."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "ops-lead", "role": "contributor", "scope": "subscription analytics"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0107", "input": "iam: ops-lead role contributor scope subscription analytics"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "mlops"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0108", "input": "Create a Python web app for the mlops team."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_service_account"}, {"confidence": 0.31, "workflow": "grant_iam_role"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0109", "input": "Set up reporting."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "reporting-bucket", "environment": "dev", "region": "eastus", "storage_class": "cool", "team": "reporting"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0110", "input": "Create a cool storage bucket named reporting-bucket in East US for development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "claims-bucket", "environment": "prod", "region": "eastus", "storage_class": "archive", "team": "claims"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0111", "input": "infra: bucket for claims, env prod, region eastus, class archive"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "john", "role": "viewer", "scope": "dev-subsystem"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0112", "input": "iam: principal=john role=viewer scope=dev-subsystem env=dev"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "reporting-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "reporting"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0113", "input": "Need an automation identity for team reporting in staging."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "claims"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0114", "input": "Create storage for the claims team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "analyst", "role": "viewer", "scope": "staging-bucket"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0115", "input": "Grant analyst viewer access to staging-bucket in development."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "security"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0116", "input": "identity needed for team security"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "finance-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "finance"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0117", "input": "Create a service account named finance-svc for the finance team in staging."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Unsafe destructive production request."], "missing_fields": [], "parameters": {}, "status": "rejected", "workflow": null}, "id": "train-0118", "input": "Remove all security from production."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "jane", "role": "contributor", "scope": "subscription analytics"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0119", "input": "Grant jane contributor access to subscription analytics."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0120", "input": "identity request: platform service account, env staging, name platform-svc"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "mlops-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "mlops"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0121", "input": "identity request: mlops service account, env staging, name mlops-svc"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "claims-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "claims"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0122", "input": "Need an automation identity for team claims in development."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "mlops"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0123", "input": "service account request, owner team mlops"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0124", "input": "scheduler request for reporting"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "finance"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0125", "input": "Create storage for the finance team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "growth-bucket", "environment": "staging", "region": "us-central1", "storage_class": "archive", "team": "growth"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0126", "input": "Set up blob storage for the growth team in US Central for staging."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_service_account"}, {"confidence": 0.31, "workflow": "create_storage_bucket"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0127", "input": "Need cloud stuff for finance."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_storage_bucket"}, {"confidence": 0.31, "workflow": "create_scheduler_job"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0128", "input": "Prep access and automation for the new project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "security-web-app", "diagnostics_enabled": false, "environment": "dev", "region": "us-central1", "runtime": "dotnet8", "team": "security"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0129", "input": "ticket: security development api, runtime .NET, region US Central, diagnostics on"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "finance"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0130", "input": "bucket needed for finance, no location picked yet"} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "jane", "role": "reader", "scope": "all reporting resources"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0131", "input": "iam: jane role reader scope all reporting resources"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "finance-bucket", "environment": "prod", "region": "us-central1", "storage_class": "standard", "team": "finance"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0132", "input": "Set up blob storage for the finance team in US Central for production."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_storage_bucket"}, {"confidence": 0.31, "workflow": "create_service_account"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0133", "input": "Make the nightly thing happen."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "job_name": "security-nightly-job", "schedule": "0 2 * * *", "target": "model-refresh", "team": "security", "timezone": "UTC"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0134", "input": "Create a nightly scheduler job named security-nightly-job for model-refresh in development."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "deploy-bot", "role": "owner", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0135", "input": "Give deploy-bot the owner role across all production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "reporting-bucket", "environment": "dev", "region": "eastus", "storage_class": "standard", "team": "reporting"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0136", "input": "infra: bucket for reporting, env dev, region eastus, class standard"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "claims-bucket", "environment": "staging", "region": "westus", "storage_class": "standard", "team": "claims"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0137", "input": "Create a standard storage bucket named claims-bucket in West US for staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "security-nightly-job", "schedule": "0 2 * * *", "target": "claims-sync", "team": "security", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0138", "input": "Create a nightly scheduler job named security-nightly-job for claims-sync in production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "growth"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0139", "input": "bucket needed for growth, no location picked yet"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "mlops-bucket", "environment": "staging", "region": "westus", "storage_class": "archive", "team": "mlops"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0140", "input": "infra: bucket for mlops, env staging, region westus, class archive"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "security-web-app", "diagnostics_enabled": false, "environment": "staging", "region": "us-central1", "runtime": "nodejs20", "team": "security"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0141", "input": "Need a small Node.js app named security-web-app in us-central1 for security."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "job_name": "finance-nightly-job", "schedule": "0 9 * * *", "target": "billing-export", "team": "finance", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0142", "input": "Create a nightly scheduler job named finance-nightly-job for billing-export in development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "growth-nightly-job", "schedule": "0 2 * * *", "target": "claims-sync", "team": "growth", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0143", "input": "Set up a daily job for claims-sync for the growth team in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "claims-web-app", "diagnostics_enabled": true, "environment": "dev", "region": "centralus", "runtime": "dotnet8", "team": "claims"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0144", "input": "ticket: claims development api, runtime .NET, region Central US, diagnostics on"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "principal": "john", "role": "viewer", "scope": "staging-bucket"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0145", "input": "Give john the viewer role on staging-bucket."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "reporting-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "reporting"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0146", "input": "Create a service account named reporting-svc for the reporting team in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "claims-bucket", "environment": "prod", "region": "eastus", "storage_class": "cool", "team": "claims"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0147", "input": "infra: bucket for claims, env prod, region eastus, class cool"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "finance-web-app", "diagnostics_enabled": true, "environment": "prod", "region": "us-central1", "runtime": "nodejs20", "team": "finance"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0148", "input": "ticket: finance production api, runtime Node.js, region US Central, diagnostics on"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "reporting-nightly-job", "schedule": "0 2 * * *", "target": "reporting", "team": "reporting", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0149", "input": "cron 0 2 * * * target reporting env prod timezone America/Los_Angeles"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "mlops"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0150", "input": "blob storage request: mlops"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "principal": "deploy-bot", "role": "editor", "scope": "staging-bucket"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0151", "input": "iam: principal=deploy-bot role=editor scope=staging-bucket env=prod"} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_web_app"}, {"confidence": 0.31, "workflow": "create_storage_bucket"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0152", "input": "Need cloud stuff for finance."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "reporting"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0153", "input": "service account request, owner team reporting"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "claims-nightly-job", "schedule": "0 9 * * *", "target": "billing-export", "team": "claims", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0154", "input": "cron 0 9 * * * target billing-export env prod timezone America/Los_Angeles"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "analyst", "role": "contributor", "scope": "dev-subsystem"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0155", "input": "Grant analyst contributor access to dev-subsystem in development."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0156", "input": "scheduler request for reporting"} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "grant_iam_role"}, {"confidence": 0.31, "workflow": "create_service_account"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0157", "input": "Prep access and automation for the new project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "reporting-bucket", "environment": "prod", "region": "centralus", "storage_class": "archive", "team": "reporting"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0158", "input": "infra: bucket for reporting, env prod, region centralus, class archive"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "growth-web-app", "diagnostics_enabled": true, "environment": "dev", "region": "westus", "runtime": "nodejs20", "team": "growth"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0159", "input": "Create a development Node.js web app for the growth team in West US."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "reporting-nightly-job", "schedule": "0 2 * * *", "target": "billing-export", "team": "reporting", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0160", "input": "cron 0 2 * * * target billing-export env prod timezone America/New_York"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "mlops-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "mlops"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0161", "input": "identity request: mlops service account, env dev, name mlops-svc"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "platform-bucket", "environment": "staging", "region": "centralus", "storage_class": "standard", "team": "platform"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0162", "input": "Set up blob storage for the platform team in Central US for staging."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "john", "role": "admin", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0163", "input": "Grant John admin access to production."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "john", "role": "admin", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0164", "input": "Grant John admin access to production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0165", "input": "Create a service account named platform-svc for the platform team in staging."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "grant_iam_role"}, {"confidence": 0.31, "workflow": "create_scheduler_job"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0166", "input": "Need cloud stuff for finance."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "platform"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0167", "input": "Create a service account for the platform team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "finance-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "us-central1", "runtime": "dotnet8", "team": "finance"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0168", "input": "ticket: finance staging api, runtime .NET, region US Central, diagnostics on"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "mlops-bucket", "environment": "prod", "region": "centralus", "storage_class": "archive", "team": "mlops"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0169", "input": "infra: bucket for mlops, env prod, region centralus, class archive"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0170", "input": "Set up a reporting schedule."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "finance-bucket", "environment": "dev", "region": "us-central1", "storage_class": "standard", "team": "finance"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0171", "input": "Set up blob storage for the finance team in US Central for development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "jane", "role": "reader", "scope": "claims-app"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0172", "input": "Grant jane reader access to claims-app in development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "security-nightly-job", "schedule": "0 9 * * *", "target": "model-refresh", "team": "security", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0173", "input": "Create a nightly scheduler job named security-nightly-job for model-refresh in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0174", "input": "Create a service account named platform-svc for the platform team in production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0175", "input": "Set up a reporting schedule."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "claims"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0176", "input": "blob storage request: claims"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "finance-nightly-job", "schedule": "0 2 * * *", "target": "billing-export", "team": "finance", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0177", "input": "cron 0 2 * * * target billing-export env staging timezone America/New_York"} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_service_account"}, {"confidence": 0.31, "workflow": "create_storage_bucket"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0178", "input": "Make the nightly thing happen."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_storage_bucket"}, {"confidence": 0.31, "workflow": "create_web_app"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0179", "input": "Need cloud stuff for finance."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "reporting"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0180", "input": "Create storage for the reporting team."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "jane", "role": "admin", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0181", "input": "Grant Jane admin access to production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "reporting"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0182", "input": "web app request: reporting, python"} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_scheduler_job"}, {"confidence": 0.31, "workflow": "create_web_app"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0183", "input": "Can you handle the platform setup?"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "jane", "role": "contributor", "scope": "staging-bucket"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0184", "input": "Give jane the contributor role on staging-bucket."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "grant_iam_role"}], "clarifying_question": "What role should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: role, scope"], "missing_fields": ["role", "scope"], "parameters": {"principal": "jane"}, "status": "needs_clarification", "workflow": "grant_iam_role"}, "id": "train-0185", "input": "permission request for jane"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "finance-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "finance"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0186", "input": "Need an automation identity for team finance in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "growth-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "growth"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0187", "input": "Need an automation identity for team growth in production."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_service_account"}, {"confidence": 0.31, "workflow": "create_storage_bucket"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0188", "input": "Prep access and automation for the new project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "jane", "role": "editor", "scope": "claims-app"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0189", "input": "Give jane the editor role on claims-app."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "mlops-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "mlops"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0190", "input": "Create a service account named mlops-svc for the mlops team in development."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "claims"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0191", "input": "Create a service account for the claims team."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "finance"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0192", "input": "blob storage request: finance"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "reporting-bucket", "environment": "dev", "region": "westus", "storage_class": "archive", "team": "reporting"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0193", "input": "Set up blob storage for the reporting team in West US for development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "reporting-user", "role": "contributor", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0194", "input": "Grant reporting-user contributor access to reporting-project in development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "platform-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "us-central1", "runtime": "nodejs20", "team": "platform"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0195", "input": "Need a small Node.js app named platform-web-app in us-central1 for platform."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "jane", "role": "owner", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0196", "input": "Give Jane the owner role across all production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "security-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "centralus", "runtime": "dotnet8", "team": "security"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0197", "input": "Need a small .NET app named security-web-app in centralus for security."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "deploy-bot", "role": "reader", "scope": "all reporting resources"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0198", "input": "Grant deploy-bot reader access to all reporting resources."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "ops-lead", "role": "reader", "scope": "production"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0199", "input": "iam: ops-lead role reader scope production"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0200", "input": "Create a service account named platform-svc for the platform team in development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "jane", "role": "reader", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0201", "input": "iam: principal=jane role=reader scope=reporting-project env=dev"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "reporting-bucket", "environment": "dev", "region": "westus", "storage_class": "cool", "team": "reporting"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0202", "input": "Create a cool storage bucket named reporting-bucket in West US for development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "claims-web-app", "diagnostics_enabled": true, "environment": "staging", "region": "westus", "runtime": "nodejs20", "team": "claims"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0203", "input": "Need a small Node.js app named claims-web-app in westus for claims."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "growth-bucket", "environment": "staging", "region": "us-central1", "storage_class": "archive", "team": "growth"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0204", "input": "Set up blob storage for the growth team in US Central for staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "job_name": "finance-nightly-job", "schedule": "0 2 * * *", "target": "model-refresh", "team": "finance", "timezone": "America/New_York"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0205", "input": "Set up a daily job for model-refresh for the finance team in development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0206", "input": "Need an automation identity for team platform in staging."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_scheduler_job"}, {"confidence": 0.31, "workflow": "create_storage_bucket"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0207", "input": "Can you handle the platform setup?"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "finance-bucket", "environment": "dev", "region": "westus", "storage_class": "standard", "team": "finance"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0208", "input": "infra: bucket for finance, env dev, region westus, class standard"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "job_name": "reporting-nightly-job", "schedule": "0 9 * * *", "target": "reporting", "team": "reporting", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0209", "input": "cron 0 9 * * * target reporting env dev timezone America/Los_Angeles"} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_web_app"}], "clarifying_question": "What app name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: app_name, region, environment"], "missing_fields": ["app_name", "region", "environment"], "parameters": {"runtime": "python311", "team": "finance"}, "status": "needs_clarification", "workflow": "create_web_app"}, "id": "train-0210", "input": "web app request: finance, python"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "claims-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "claims"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0211", "input": "Need an automation identity for team claims in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "claims-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "claims"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0212", "input": "identity request: claims service account, env staging, name claims-svc"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "principal": "deploy-bot", "role": "reader", "scope": "claims-app"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0213", "input": "Grant deploy-bot reader access to claims-app in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "claims-web-app", "diagnostics_enabled": true, "environment": "dev", "region": "us-central1", "runtime": "nodejs20", "team": "claims"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0214", "input": "Need a small Node.js app named claims-web-app in us-central1 for claims."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_storage_bucket"}], "clarifying_question": "What bucket name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: bucket_name, region, environment"], "missing_fields": ["bucket_name", "region", "environment"], "parameters": {"team": "reporting"}, "status": "needs_clarification", "workflow": "create_storage_bucket"}, "id": "train-0215", "input": "blob storage request: reporting"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "finance-bucket", "environment": "prod", "region": "westus", "storage_class": "archive", "team": "finance"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0216", "input": "Set up blob storage for the finance team in West US for production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "staging", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0217", "input": "Create a service account named platform-svc for the platform team in staging."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "reporting-web-app", "diagnostics_enabled": false, "environment": "staging", "region": "eastus", "runtime": "nodejs20", "team": "reporting"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0218", "input": "ticket: reporting staging api, runtime Node.js, region East US, diagnostics on"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "finance-web-app", "diagnostics_enabled": false, "environment": "dev", "region": "us-central1", "runtime": "dotnet8", "team": "finance"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0219", "input": "Need a small .NET app named finance-web-app in us-central1 for finance."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0220", "input": "daily reporting job, details later"} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "ops-lead", "role": "reader", "scope": "production"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0221", "input": "Grant ops-lead reader access to production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "grant_iam_role"}], "clarifying_question": "What role should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: role, scope"], "missing_fields": ["role", "scope"], "parameters": {"principal": "john"}, "status": "needs_clarification", "workflow": "grant_iam_role"}, "id": "train-0222", "input": "Grant john access."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "principal": "jane", "role": "editor", "scope": "staging-bucket"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0223", "input": "iam: principal=jane role=editor scope=staging-bucket env=staging"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "platform-nightly-job", "schedule": "0 2 * * *", "target": "reporting", "team": "platform", "timezone": "America/Los_Angeles"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0224", "input": "Set up a daily job for reporting for the platform team in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "finance-bucket", "environment": "dev", "region": "us-central1", "storage_class": "archive", "team": "finance"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0225", "input": "infra: bucket for finance, env dev, region us-central1, class archive"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "platform-web-app", "diagnostics_enabled": true, "environment": "dev", "region": "centralus", "runtime": "python311", "team": "platform"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0226", "input": "Create a development Python web app for the platform team in Central US."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_scheduler_job"}], "clarifying_question": "What job name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: job_name, schedule, environment"], "missing_fields": ["job_name", "schedule", "environment"], "parameters": {"target": "reporting"}, "status": "needs_clarification", "workflow": "create_scheduler_job"}, "id": "train-0227", "input": "scheduler request for reporting"} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "jane", "role": "owner", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0228", "input": "Grant Jane owner access to production."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [{"confidence": 0.88, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Blocked owner/admin IAM grant."], "missing_fields": [], "parameters": {"environment": "prod", "principal": "deploy-bot", "role": "owner", "scope": "production"}, "status": "rejected", "workflow": "grant_iam_role"}, "id": "train-0229", "input": "iam urgent: deploy-bot needs owner on prod."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "claims-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "claims"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0230", "input": "Create a service account named claims-svc for the claims team in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "mlops-nightly-job", "schedule": "0 9 * * *", "target": "reporting", "team": "mlops", "timezone": "UTC"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0231", "input": "cron 0 9 * * * target reporting env prod timezone UTC"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "jane", "role": "contributor", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0232", "input": "Give jane the contributor role on reporting-project."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "ops-lead", "role": "reader", "scope": "all reporting resources"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0233", "input": "Grant ops-lead reader access to all reporting resources."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_service_account"}, {"confidence": 0.31, "workflow": "create_web_app"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0234", "input": "Prep access and automation for the new project."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0235", "input": "identity request: platform service account, env prod, name platform-svc"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "claims-bucket", "environment": "prod", "region": "centralus", "storage_class": "standard", "team": "claims"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0236", "input": "Set up blob storage for the claims team in Central US for production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_storage_bucket"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"bucket_name": "mlops-bucket", "environment": "staging", "region": "us-central1", "storage_class": "cool", "team": "mlops"}, "status": "routed", "workflow": "create_storage_bucket"}, "id": "train-0237", "input": "infra: bucket for mlops, env staging, region us-central1, class cool"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_web_app"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"app_name": "growth-web-app", "diagnostics_enabled": true, "environment": "dev", "region": "westus", "runtime": "nodejs20", "team": "growth"}, "status": "routed", "workflow": "create_web_app"}, "id": "train-0238", "input": "Create a development Node.js web app for the growth team in West US."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "reporting-svc", "description": "Service identity for workflow automation.", "environment": "prod", "team": "reporting"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0239", "input": "Need an automation identity for team reporting in production."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "job_name": "claims-nightly-job", "schedule": "0 2 * * *", "target": "reporting", "team": "claims", "timezone": "UTC"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0240", "input": "Create a nightly scheduler job named claims-nightly-job for reporting in development."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "create_storage_bucket"}, {"confidence": 0.31, "workflow": "create_web_app"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0241", "input": "Set up reporting."} | |
| {"case_type": "risky_rejected", "expected": {"candidate_workflows": [], "clarifying_question": null, "confidence": 0.88, "failure_reasons": ["Unsafe destructive production request."], "missing_fields": [], "parameters": {}, "status": "rejected", "workflow": null}, "id": "train-0242", "input": "Delete all production services and disable monitoring."} | |
| {"case_type": "ambiguous", "expected": {"candidate_workflows": [{"confidence": 0.38, "workflow": "grant_iam_role"}, {"confidence": 0.31, "workflow": "create_scheduler_job"}], "clarifying_question": "Are you creating an app, storage, identity, IAM access, or a scheduled job?", "confidence": 0.34, "failure_reasons": ["Request is ambiguous across multiple workflows."], "missing_fields": [], "parameters": {}, "status": "needs_clarification", "workflow": null}, "id": "train-0243", "input": "Can you handle the platform setup?"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "job_name": "reporting-nightly-job", "schedule": "0 2 * * *", "target": "model-refresh", "team": "reporting", "timezone": "UTC"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0244", "input": "Set up a daily job for model-refresh for the reporting team in staging."} | |
| {"case_type": "confirmation_required", "expected": {"candidate_workflows": [{"confidence": 0.82, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.82, "failure_reasons": ["High-risk IAM change requires confirmation."], "missing_fields": [], "parameters": {"environment": "staging", "principal": "ops-lead", "role": "reader", "scope": "subscription analytics"}, "status": "requires_confirmation", "workflow": "grant_iam_role"}, "id": "train-0245", "input": "iam: ops-lead role reader scope subscription analytics"} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_scheduler_job"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "prod", "job_name": "growth-nightly-job", "schedule": "0 2 * * *", "target": "reporting", "team": "growth", "timezone": "UTC"}, "status": "routed", "workflow": "create_scheduler_job"}, "id": "train-0246", "input": "Create a nightly scheduler job named growth-nightly-job for reporting in production."} | |
| {"case_type": "missing_fields", "expected": {"candidate_workflows": [{"confidence": 0.74, "workflow": "create_service_account"}], "clarifying_question": "What account name should RouterCore use?", "confidence": 0.74, "failure_reasons": ["Missing required fields: account_name, environment"], "missing_fields": ["account_name", "environment"], "parameters": {"team": "platform"}, "status": "needs_clarification", "workflow": "create_service_account"}, "id": "train-0247", "input": "Create a service account for the platform team."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "staging", "principal": "reporting-user", "role": "editor", "scope": "dev-subsystem"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0248", "input": "Give reporting-user the editor role on dev-subsystem."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "create_service_account"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"account_name": "platform-svc", "description": "Service identity for workflow automation.", "environment": "dev", "team": "platform"}, "status": "routed", "workflow": "create_service_account"}, "id": "train-0249", "input": "Need an automation identity for team platform in development."} | |
| {"case_type": "success", "expected": {"candidate_workflows": [{"confidence": 0.92, "workflow": "grant_iam_role"}], "clarifying_question": null, "confidence": 0.92, "failure_reasons": [], "missing_fields": [], "parameters": {"environment": "dev", "principal": "analyst", "role": "reader", "scope": "reporting-project"}, "status": "routed", "workflow": "grant_iam_role"}, "id": "train-0250", "input": "iam: principal=analyst role=reader scope=reporting-project env=dev"} | |