| { |
| "model": "mlx-community/Qwen2.5-1.5B-Instruct-4bit", |
| "generated": "2026-09-16", |
| "presets": [ |
| { |
| "id": "fintech_fraud", |
| "title": "FinTech Fraud & Autonomous AML Compliance (28 Fields)", |
| "description": "Real-time multi-attribute financial fraud detection, sanctions verification, and autonomous containment decision.", |
| "context": "TRANSACTION ALERT #TX-98421\nTimestamp: 2026-09-15 03:14:22 UTC\nCardholder: Sarah Jenkins (Account ID: US-99120)\nTransaction Amount: ,850.00 USD\nMerchant: CyberGold Crypto Exchange (Nicosia, Cyprus)\nCardholder Billing Address: Seattle, WA, USA\nDevice IP: 185.220.101.5 (Known Tor Exit Node, Frankfurt, Germany)\nDevice Fingerprint: Unrecognized Linux Firefox 128.0 (First seen 4 minutes ago)\nVelocity Check: 3 previous transactions attempted in past 10 minutes from Singapore (), London (), and Frankfurt (,850).\nCardholder Typical Activity: Average monthly spend ,200. No prior cryptocurrency, wire transfer, or foreign cross-border transactions recorded in 3 years.\nCurrent Account Balance: ,200.00 USD.\nTwo-Factor Authentication: SMS 2FA prompt bypassed via Session Hijacking / Cookie Injection anomaly.", |
| "num_fields": 28, |
| "parallel": { |
| "elapsed_ms": 405.43, |
| "prefill_ms": 206.17, |
| "suffix_eval_ms": 141.2, |
| "schema_match": true, |
| "fields": [ |
| { |
| "name": "is_fraudulent", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9543, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9543 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0457 |
| } |
| ] |
| }, |
| { |
| "name": "risk_tier", |
| "type": "enum", |
| "value": "HIGH", |
| "confidence": 0.7441, |
| "top": [ |
| { |
| "choice": "HIGH", |
| "p": 0.7441 |
| }, |
| { |
| "choice": "MEDIUM", |
| "p": 0.1687 |
| }, |
| { |
| "choice": "LOW", |
| "p": 0.0671 |
| }, |
| { |
| "choice": "CRITICAL", |
| "p": 0.0201 |
| } |
| ] |
| }, |
| { |
| "name": "recommended_action", |
| "type": "enum", |
| "value": "MANUAL_REVIEW", |
| "confidence": 0.5126, |
| "top": [ |
| { |
| "choice": "MANUAL_REVIEW", |
| "p": 0.5126 |
| }, |
| { |
| "choice": "BLOCK_TRANSACTION", |
| "p": 0.4419 |
| }, |
| { |
| "choice": "APPROVE", |
| "p": 0.0245 |
| }, |
| { |
| "choice": "FREEZE_ACCOUNT", |
| "p": 0.017 |
| } |
| ] |
| }, |
| { |
| "name": "primary_anomaly", |
| "type": "enum", |
| "value": "TOR_EXIT_NODE", |
| "confidence": 0.8877, |
| "top": [ |
| { |
| "choice": "TOR_EXIT_NODE", |
| "p": 0.8877 |
| }, |
| { |
| "choice": "IP_GEOLOCATION_MISMATCH", |
| "p": 0.0973 |
| }, |
| { |
| "choice": "UNUSUAL_VELOCITY", |
| "p": 0.0147 |
| }, |
| { |
| "choice": "NEW_DEVICE", |
| "p": 0.0002 |
| } |
| ] |
| }, |
| { |
| "name": "secondary_anomaly", |
| "type": "enum", |
| "value": "CARD_NOT_PRESENT", |
| "confidence": 0.9544, |
| "top": [ |
| { |
| "choice": "CARD_NOT_PRESENT", |
| "p": 0.9544 |
| }, |
| { |
| "choice": "MULTIPLE_FAILED_LOGINS", |
| "p": 0.0356 |
| }, |
| { |
| "choice": "NONE", |
| "p": 0.0099 |
| }, |
| { |
| "choice": "RAPID_DESTINATION_CHANGE", |
| "p": 0.0001 |
| } |
| ] |
| }, |
| { |
| "name": "requires_human_escalation", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9546, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9546 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0454 |
| } |
| ] |
| }, |
| { |
| "name": "can_auto_block", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.6893, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.6893 |
| }, |
| { |
| "choice": "false", |
| "p": 0.3107 |
| } |
| ] |
| }, |
| { |
| "name": "aml_sar_filing_required", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8998, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8998 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1002 |
| } |
| ] |
| }, |
| { |
| "name": "fincen_flag_threshold", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8899, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8899 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1101 |
| } |
| ] |
| }, |
| { |
| "name": "sanctions_screening_risk", |
| "type": "enum", |
| "value": "HIGH_CONFIDENCE_MATCH", |
| "confidence": 0.6405, |
| "top": [ |
| { |
| "choice": "HIGH_CONFIDENCE_MATCH", |
| "p": 0.6405 |
| }, |
| { |
| "choice": "POTENTIAL_MATCH", |
| "p": 0.3593 |
| }, |
| { |
| "choice": "CLEARED", |
| "p": 0.0001 |
| } |
| ] |
| }, |
| { |
| "name": "tor_exit_node_detected", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9986, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9986 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0014 |
| } |
| ] |
| }, |
| { |
| "name": "device_fingerprint_mismatch", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9711, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9711 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0289 |
| } |
| ] |
| }, |
| { |
| "name": "velocity_score", |
| "type": "enum", |
| "value": "NORMAL", |
| "confidence": 0.4339, |
| "top": [ |
| { |
| "choice": "NORMAL", |
| "p": 0.4339 |
| }, |
| { |
| "choice": "EXTREME_BURST", |
| "p": 0.3514 |
| }, |
| { |
| "choice": "ELEVATED", |
| "p": 0.2148 |
| } |
| ] |
| }, |
| { |
| "name": "block_card_immediately", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8267, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8267 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1733 |
| } |
| ] |
| }, |
| { |
| "name": "freeze_online_banking", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9519, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9519 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0481 |
| } |
| ] |
| }, |
| { |
| "name": "notify_cardholder_sms", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8732, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8732 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1268 |
| } |
| ] |
| }, |
| { |
| "name": "notify_cardholder_email", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.7202, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.7202 |
| }, |
| { |
| "choice": "false", |
| "p": 0.2798 |
| } |
| ] |
| }, |
| { |
| "name": "requires_identity_verification", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8007, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8007 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1993 |
| } |
| ] |
| }, |
| { |
| "name": "merchant_category_risk", |
| "type": "enum", |
| "value": "HIGH_RISK_CRYPTO", |
| "confidence": 0.8371, |
| "top": [ |
| { |
| "choice": "HIGH_RISK_CRYPTO", |
| "p": 0.8371 |
| }, |
| { |
| "choice": "LOW", |
| "p": 0.1598 |
| }, |
| { |
| "choice": "RESTRICTED_GAMBLING", |
| "p": 0.0017 |
| }, |
| { |
| "choice": "MODERATE", |
| "p": 0.0014 |
| } |
| ] |
| }, |
| { |
| "name": "chargeback_probability", |
| "type": "enum", |
| "value": "LOW", |
| "confidence": 0.712, |
| "top": [ |
| { |
| "choice": "LOW", |
| "p": 0.712 |
| }, |
| { |
| "choice": "MEDIUM", |
| "p": 0.1946 |
| }, |
| { |
| "choice": "VERY_HIGH", |
| "p": 0.0934 |
| } |
| ] |
| }, |
| { |
| "name": "counterparty_jurisdiction_risk", |
| "type": "enum", |
| "value": "TIER_1_LOW", |
| "confidence": 0.75, |
| "top": [ |
| { |
| "choice": "TIER_1_LOW", |
| "p": 0.75 |
| }, |
| { |
| "choice": "TIER_2_MODERATE", |
| "p": 0.0833 |
| }, |
| { |
| "choice": "TIER_3_HIGH", |
| "p": 0.0833 |
| }, |
| { |
| "choice": "SANCTIONED_ZONE", |
| "p": 0.0833 |
| } |
| ] |
| }, |
| { |
| "name": "auto_reverse_transfer", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.7042, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.7042 |
| }, |
| { |
| "choice": "false", |
| "p": 0.2958 |
| } |
| ] |
| }, |
| { |
| "name": "fraud_ring_association", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9086, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9086 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0914 |
| } |
| ] |
| }, |
| { |
| "name": "alert_clearing_house", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8652, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8652 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1348 |
| } |
| ] |
| }, |
| { |
| "name": "escalate_to_fiu", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8233, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8233 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1767 |
| } |
| ] |
| }, |
| { |
| "name": "confidence_score_tier", |
| "type": "enum", |
| "value": "HIGH_CONFIDENCE_FRAUD", |
| "confidence": 0.9786, |
| "top": [ |
| { |
| "choice": "HIGH_CONFIDENCE_FRAUD", |
| "p": 0.9786 |
| }, |
| { |
| "choice": "BORDERLINE", |
| "p": 0.021 |
| }, |
| { |
| "choice": "LEGITIMATE", |
| "p": 0.0005 |
| }, |
| { |
| "choice": "LEANING_FRAUD", |
| "p": 0.0 |
| } |
| ] |
| }, |
| { |
| "name": "loss_prevention_priority", |
| "type": "enum", |
| "value": "LOW_PRIORITY", |
| "confidence": 0.9892, |
| "top": [ |
| { |
| "choice": "LOW_PRIORITY", |
| "p": 0.9892 |
| }, |
| { |
| "choice": "IMMEDIATE_P0", |
| "p": 0.0093 |
| }, |
| { |
| "choice": "STANDARD_QUEUE", |
| "p": 0.0015 |
| } |
| ] |
| }, |
| { |
| "name": "account_takeover_suspected", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9763, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9763 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0237 |
| } |
| ] |
| } |
| ], |
| "json": { |
| "is_fraudulent": true, |
| "risk_tier": "HIGH", |
| "recommended_action": "MANUAL_REVIEW", |
| "primary_anomaly": "TOR_EXIT_NODE", |
| "secondary_anomaly": "CARD_NOT_PRESENT", |
| "requires_human_escalation": true, |
| "can_auto_block": true, |
| "aml_sar_filing_required": true, |
| "fincen_flag_threshold": true, |
| "sanctions_screening_risk": "HIGH_CONFIDENCE_MATCH", |
| "tor_exit_node_detected": true, |
| "device_fingerprint_mismatch": true, |
| "velocity_score": "NORMAL", |
| "block_card_immediately": true, |
| "freeze_online_banking": true, |
| "notify_cardholder_sms": true, |
| "notify_cardholder_email": true, |
| "requires_identity_verification": true, |
| "merchant_category_risk": "HIGH_RISK_CRYPTO", |
| "chargeback_probability": "LOW", |
| "counterparty_jurisdiction_risk": "TIER_1_LOW", |
| "auto_reverse_transfer": true, |
| "fraud_ring_association": true, |
| "alert_clearing_house": true, |
| "escalate_to_fiu": true, |
| "confidence_score_tier": "HIGH_CONFIDENCE_FRAUD", |
| "loss_prevention_priority": "LOW_PRIORITY", |
| "account_takeover_suspected": true |
| } |
| }, |
| "naive": { |
| "elapsed_ms": 3085.22, |
| "total_tokens": 288, |
| "tokens_per_second": 93.3, |
| "schema_match": false, |
| "raw_text": "{\n \"is_fraudulent\": true,\n \"risk_tier\": \"HIGH\",\n \"recommended_action\": \"MANUAL_REVIEW\",\n \"primary_anomaly\": \"TOR_EXIT_NODE\",\n \"secondary_anomaly\": \"UNUSUAL_VELOCITY\",\n \"requires_human_escalation\": true,\n \"can_auto_block\": false,\n \"aml_sar_filing_required\": false,\n \"fincen_flag_threshold\": false,\n \"sanctions_screening_risk\": \"POTENTIAL_MATCH\",\n \"tor_exit_node_detected\": true,\n \"device_fingerprint_mismatch\": true,\n \"velocity_score\": \"EXTREME_BURST\",\n \"block_card_immediately\": false,\n \"freeze_online_banking\": false,\n \"notify_cardholder_sms\": false,\n \"notify_cardholder_email\": false,\n \"requires_identity_verification\": false,\n \"merchant_category_risk\": \"HIGH_RISK_CRYPTO\",\n \"chargeback_probability\": \"VERY_HIGH\",\n \"counterparty_jurisdiction_risk\": \"TIER_3_HIGH\",\n \"auto_reverse_transfer\": false,\n \"fraud_ring_association\": false,\n \"alert_clearing_house\": false,\n \"escalate_to_fiu\": false,\n \"confidence_score_tier\": \"LEANING_FRAUD\",\n \"loss_prevention_priority\": \"IMMEDIATE_P0\"\n}" |
| } |
| }, |
| { |
| "id": "support_triage", |
| "title": "Enterprise Incident Triage & Routing (28 Fields)", |
| "description": "End-to-end multi-dimensional semantic analysis of a critical enterprise production outage ticket.", |
| "context": "SUPPORT TICKET #INC-44019\nFrom: Marcus Vance, CTO @ Omnicorp Global (Tier 1 Strategic Enterprise, k ARR)\nSubject: CRITICAL P0: Production US-East API 502 Gateway Outage - Checkout Down\n\nBody: Our primary checkout service has been failing completely for the last 48 minutes because your US-East API endpoint is throwing 502 Bad Gateway errors on every POST /v2/charges call. We are losing approximately ,000 in transaction revenue every ten minutes!\n\nFurthermore, our billing audit discovered this morning that your platform double-billed our corporate account for data overages (,800 improper charge on Invoice #INV-8812).\n\nIf this API outage is not resolved within 30 minutes, our legal counsel will formally issue notice of material breach of our 99.99% availability SLA and initiate contract termination with full damages. We demand an immediate phone bridge with your VP of Engineering and a full refund of the bogus invoice.\n\nEnvironment: PRODUCTION\nAffected Cluster: us-east-1a\nAccount Priority: STRATEGIC_TIER_1", |
| "num_fields": 28, |
| "parallel": { |
| "elapsed_ms": 455.76, |
| "prefill_ms": 205.54, |
| "suffix_eval_ms": 140.66, |
| "schema_match": true, |
| "fields": [ |
| { |
| "name": "sentiment", |
| "type": "enum", |
| "value": "NEGATIVE", |
| "confidence": 0.9799, |
| "top": [ |
| { |
| "choice": "NEGATIVE", |
| "p": 0.9799 |
| }, |
| { |
| "choice": "POSITIVE", |
| "p": 0.019 |
| }, |
| { |
| "choice": "VERY_NEGATIVE", |
| "p": 0.0007 |
| }, |
| { |
| "choice": "NEUTRAL", |
| "p": 0.0004 |
| } |
| ] |
| }, |
| { |
| "name": "severity_tier", |
| "type": "enum", |
| "value": "SEV_1_MAJOR", |
| "confidence": 0.6953, |
| "top": [ |
| { |
| "choice": "SEV_1_MAJOR", |
| "p": 0.6953 |
| }, |
| { |
| "choice": "SEV_3_MINOR", |
| "p": 0.1227 |
| }, |
| { |
| "choice": "SEV_2_MODERATE", |
| "p": 0.1002 |
| }, |
| { |
| "choice": "SEV_0_CRITICAL", |
| "p": 0.0818 |
| } |
| ] |
| }, |
| { |
| "name": "primary_department", |
| "type": "enum", |
| "value": "SECURITY_OPS", |
| "confidence": 0.6503, |
| "top": [ |
| { |
| "choice": "SECURITY_OPS", |
| "p": 0.6503 |
| }, |
| { |
| "choice": "BILLING_FINANCE", |
| "p": 0.3024 |
| }, |
| { |
| "choice": "LEGAL_COMPLIANCE", |
| "p": 0.0385 |
| }, |
| { |
| "choice": "ENTERPRISE_SUCCESS", |
| "p": 0.0056 |
| } |
| ] |
| }, |
| { |
| "name": "secondary_department", |
| "type": "enum", |
| "value": "NETWORK_OPS", |
| "confidence": 0.6104, |
| "top": [ |
| { |
| "choice": "NETWORK_OPS", |
| "p": 0.6104 |
| }, |
| { |
| "choice": "CORE_API", |
| "p": 0.1643 |
| }, |
| { |
| "choice": "NONE", |
| "p": 0.1592 |
| }, |
| { |
| "choice": "IDENTITY_AUTH", |
| "p": 0.0595 |
| } |
| ] |
| }, |
| { |
| "name": "sla_breach_risk", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9936, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9936 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0064 |
| } |
| ] |
| }, |
| { |
| "name": "requires_executive_escalation", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9975, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9975 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0025 |
| } |
| ] |
| }, |
| { |
| "name": "can_resolve_autonomously", |
| "type": "boolean", |
| "value": "False", |
| "confidence": 0.9294, |
| "top": [ |
| { |
| "choice": "false", |
| "p": 0.9294 |
| }, |
| { |
| "choice": "true", |
| "p": 0.0706 |
| } |
| ] |
| }, |
| { |
| "name": "confidence_above_auto_threshold", |
| "type": "boolean", |
| "value": "False", |
| "confidence": 0.6225, |
| "top": [ |
| { |
| "choice": "false", |
| "p": 0.6225 |
| }, |
| { |
| "choice": "true", |
| "p": 0.3775 |
| } |
| ] |
| }, |
| { |
| "name": "churn_risk_level", |
| "type": "enum", |
| "value": "HIGH", |
| "confidence": 0.975, |
| "top": [ |
| { |
| "choice": "HIGH", |
| "p": 0.975 |
| }, |
| { |
| "choice": "CRITICAL_IMMINENT", |
| "p": 0.0125 |
| }, |
| { |
| "choice": "LOW", |
| "p": 0.0073 |
| }, |
| { |
| "choice": "MEDIUM", |
| "p": 0.0052 |
| } |
| ] |
| }, |
| { |
| "name": "refund_recommended", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9875, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9875 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0125 |
| } |
| ] |
| }, |
| { |
| "name": "refund_tier", |
| "type": "enum", |
| "value": "FULL_MONTH_CREDIT", |
| "confidence": 0.8831, |
| "top": [ |
| { |
| "choice": "FULL_MONTH_CREDIT", |
| "p": 0.8831 |
| }, |
| { |
| "choice": "NONE", |
| "p": 0.062 |
| }, |
| { |
| "choice": "CUSTOM_COMPENSATION", |
| "p": 0.0351 |
| }, |
| { |
| "choice": "PARTIAL_CREDIT", |
| "p": 0.0198 |
| } |
| ] |
| }, |
| { |
| "name": "requires_legal_review", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9812, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9812 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0188 |
| } |
| ] |
| }, |
| { |
| "name": "security_incident", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9984, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9984 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0016 |
| } |
| ] |
| }, |
| { |
| "name": "compliance_breach", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9937, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9937 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0063 |
| } |
| ] |
| }, |
| { |
| "name": "customer_tenure_tier", |
| "type": "enum", |
| "value": "GROWTH", |
| "confidence": 0.5009, |
| "top": [ |
| { |
| "choice": "GROWTH", |
| "p": 0.5009 |
| }, |
| { |
| "choice": "MID_MARKET", |
| "p": 0.2538 |
| }, |
| { |
| "choice": "ENTERPRISE_STRATEGIC", |
| "p": 0.2104 |
| }, |
| { |
| "choice": "EARLY_STAGE", |
| "p": 0.0349 |
| } |
| ] |
| }, |
| { |
| "name": "root_cause_domain", |
| "type": "enum", |
| "value": "API_GATEWAY", |
| "confidence": 0.9964, |
| "top": [ |
| { |
| "choice": "API_GATEWAY", |
| "p": 0.9964 |
| }, |
| { |
| "choice": "DNS_ROUTING", |
| "p": 0.0033 |
| }, |
| { |
| "choice": "AUTH_TOKEN_EXPIRY", |
| "p": 0.0003 |
| }, |
| { |
| "choice": "DB_LOCKING", |
| "p": 0.0 |
| } |
| ] |
| }, |
| { |
| "name": "target_resolution_hours", |
| "type": "enum", |
| "value": "24_HOURS", |
| "confidence": 0.75, |
| "top": [ |
| { |
| "choice": "24_HOURS", |
| "p": 0.75 |
| }, |
| { |
| "choice": "1_HOUR", |
| "p": 0.0833 |
| }, |
| { |
| "choice": "4_HOURS", |
| "p": 0.0833 |
| }, |
| { |
| "choice": "12_HOURS", |
| "p": 0.0833 |
| } |
| ] |
| }, |
| { |
| "name": "requires_phone_call", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9992, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9992 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0008 |
| } |
| ] |
| }, |
| { |
| "name": "notify_account_executive", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9797, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9797 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0203 |
| } |
| ] |
| }, |
| { |
| "name": "public_status_page_incident", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.954, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.954 |
| }, |
| { |
| "choice": "false", |
| "p": 0.046 |
| } |
| ] |
| }, |
| { |
| "name": "postmortem_required", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9946, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9946 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0054 |
| } |
| ] |
| }, |
| { |
| "name": "vendor_fault", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9421, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9421 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0579 |
| } |
| ] |
| }, |
| { |
| "name": "contract_penalty_applicable", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9588, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9588 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0412 |
| } |
| ] |
| }, |
| { |
| "name": "escalate_to_vp", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.994, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.994 |
| }, |
| { |
| "choice": "false", |
| "p": 0.006 |
| } |
| ] |
| }, |
| { |
| "name": "customer_sentiment_trend", |
| "type": "enum", |
| "value": "WORSENING", |
| "confidence": 0.9999, |
| "top": [ |
| { |
| "choice": "WORSENING", |
| "p": 0.9999 |
| }, |
| { |
| "choice": "IMPROVING", |
| "p": 0.0001 |
| }, |
| { |
| "choice": "STATIC", |
| "p": 0.0 |
| } |
| ] |
| }, |
| { |
| "name": "suggested_action", |
| "type": "enum", |
| "value": "SCHEDULE_EXEC_CALL", |
| "confidence": 0.9942, |
| "top": [ |
| { |
| "choice": "SCHEDULE_EXEC_CALL", |
| "p": 0.9942 |
| }, |
| { |
| "choice": "ISSUE_CREDIT_MEMO", |
| "p": 0.0033 |
| }, |
| { |
| "choice": "DISPATCH_SECURITY_TEAM", |
| "p": 0.0023 |
| }, |
| { |
| "choice": "PAGE_ONCALL_SRE", |
| "p": 0.0003 |
| } |
| ] |
| }, |
| { |
| "name": "assigned_agent_tier", |
| "type": "enum", |
| "value": "TIER_2", |
| "confidence": 0.75, |
| "top": [ |
| { |
| "choice": "TIER_2", |
| "p": 0.75 |
| }, |
| { |
| "choice": "TIER_1", |
| "p": 0.0833 |
| }, |
| { |
| "choice": "SENIOR_SRE", |
| "p": 0.0833 |
| }, |
| { |
| "choice": "DIRECTOR", |
| "p": 0.0833 |
| } |
| ] |
| }, |
| { |
| "name": "ticket_priority", |
| "type": "enum", |
| "value": "P0_EMERGENCY", |
| "confidence": 0.8322, |
| "top": [ |
| { |
| "choice": "P0_EMERGENCY", |
| "p": 0.8322 |
| }, |
| { |
| "choice": "P1_HIGH", |
| "p": 0.1199 |
| }, |
| { |
| "choice": "P2_NORMAL", |
| "p": 0.0448 |
| }, |
| { |
| "choice": "P3_LOW", |
| "p": 0.0031 |
| } |
| ] |
| } |
| ], |
| "json": { |
| "sentiment": "NEGATIVE", |
| "severity_tier": "SEV_1_MAJOR", |
| "primary_department": "SECURITY_OPS", |
| "secondary_department": "NETWORK_OPS", |
| "sla_breach_risk": true, |
| "requires_executive_escalation": true, |
| "can_resolve_autonomously": false, |
| "confidence_above_auto_threshold": false, |
| "churn_risk_level": "HIGH", |
| "refund_recommended": true, |
| "refund_tier": "FULL_MONTH_CREDIT", |
| "requires_legal_review": true, |
| "security_incident": true, |
| "compliance_breach": true, |
| "customer_tenure_tier": "GROWTH", |
| "root_cause_domain": "API_GATEWAY", |
| "target_resolution_hours": "24_HOURS", |
| "requires_phone_call": true, |
| "notify_account_executive": true, |
| "public_status_page_incident": true, |
| "postmortem_required": true, |
| "vendor_fault": true, |
| "contract_penalty_applicable": true, |
| "escalate_to_vp": true, |
| "customer_sentiment_trend": "WORSENING", |
| "suggested_action": "SCHEDULE_EXEC_CALL", |
| "assigned_agent_tier": "TIER_2", |
| "ticket_priority": "P0_EMERGENCY" |
| } |
| }, |
| "naive": { |
| "elapsed_ms": 3104.88, |
| "total_tokens": 283, |
| "tokens_per_second": 91.1, |
| "schema_match": false, |
| "raw_text": "{\n \"sentiment\": \"VERY_NEGATIVE\",\n \"severity_tier\": \"SEV_0_CRITICAL\",\n \"primary_department\": \"SECURITY_OPS\",\n \"secondary_department\": \"NONE\",\n \"sla_breach_risk\": true,\n \"requires_executive_escalation\": true,\n \"can_resolve_autonomously\": false,\n \"confidence_above_auto_threshold\": false,\n \"churn_risk_level\": \"CRITICAL_IMMINENT\",\n \"refund_recommended\": true,\n \"refund_tier\": \"FULL_MONTH_CREDIT\",\n \"requires_legal_review\": true,\n \"security_incident\": true,\n \"compliance_breach\": false,\n \"customer_tenure_tier\": \"ENTERPRISE_STRATEGIC\",\n \"root_cause_domain\": \"API_GATEWAY\",\n \"target_resolution_hours\": \"12_HOURS\",\n \"requires_phone_call\": true,\n \"notify_account_executive\": true,\n \"public_status_page_incident\": false,\n \"postmortem_required\": false,\n \"vendor_fault\": false,\n \"contract_penalty_applicable\": false,\n \"escalate_to_vp\": true,\n \"customer_sentiment_trend\": \"WORSENING\",\n \"suggested_action\": \"PAGE_ONCALL_SRE\",\n \"assigned_agent_tier\": \"TIER_1\"\n}" |
| } |
| }, |
| { |
| "id": "code_security", |
| "title": "Autonomous Code Security & PR Vulnerability Triage (28 Fields)", |
| "description": "Automated SAST/DAST static analysis triage evaluating an enterprise pull request containing leaked secrets and injection risks.", |
| "context": "CI/CD PIPELINE AUDIT #PR-10822\nRepository: payment-gateway-core\nBranch: feature/direct-stripe-integration -> main\nAuthor: external-contractor-92 (First PR submitted)\nDiff Summary: +482 lines, -12 lines in src/services/charge_processor.py and config/aws.env\n\nStatic Scanner Findings:\n1. Hardcoded Plaintext AWS Root Secret: AWS_SECRET_ACCESS_KEY=\"AKIAIOSFODNN7EXAMPLE...\" committed into config/aws.env\n2. Unsanitized SQL Query Concatenation: cursor.execute(f\"SELECT * FROM users WHERE id = '{req.user_id}' AND secret = '{req.token}'\")\n3. Insecure Deserialization: pickle.loads(base64.b64decode(req.headers['X-Payload'])) detected in webhook handler.\n4. Dependency Check: Added vulnerable PyPI package 'requests-oauthlib==0.4.0' with 3 critical CVEs.\n\nDeployment Target: PRODUCTION (us-west-2 Kubernetes Cluster)\nCompliance Scope: SOC2 Type II, PCI-DSS Level 1, HIPAA\nSecurity Gate Status: FAILED_BLOCKING", |
| "num_fields": 28, |
| "parallel": { |
| "elapsed_ms": 419.26, |
| "prefill_ms": 205.48, |
| "suffix_eval_ms": 140.58, |
| "schema_match": true, |
| "fields": [ |
| { |
| "name": "is_vulnerability", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9953, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9953 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0047 |
| } |
| ] |
| }, |
| { |
| "name": "primary_cwe", |
| "type": "enum", |
| "value": "CWE_89_SQL_INJECTION", |
| "confidence": 0.3803, |
| "top": [ |
| { |
| "choice": "CWE_89_SQL_INJECTION", |
| "p": 0.3803 |
| }, |
| { |
| "choice": "CWE_20_IMPROPER_INPUT", |
| "p": 0.2574 |
| }, |
| { |
| "choice": "CWE_502_DESERIALIZATION", |
| "p": 0.2134 |
| }, |
| { |
| "choice": "CWE_798_HARDCODED_CREDENTIALS", |
| "p": 0.1489 |
| } |
| ] |
| }, |
| { |
| "name": "secondary_cwe", |
| "type": "enum", |
| "value": "CWE_798_HARDCODED_CREDENTIALS", |
| "confidence": 0.75, |
| "top": [ |
| { |
| "choice": "CWE_798_HARDCODED_CREDENTIALS", |
| "p": 0.75 |
| }, |
| { |
| "choice": "CWE_1104_OUTDATED_COMPONENTS", |
| "p": 0.0833 |
| }, |
| { |
| "choice": "CWE_200_INFO_EXPOSURE", |
| "p": 0.0833 |
| }, |
| { |
| "choice": "NONE", |
| "p": 0.0833 |
| } |
| ] |
| }, |
| { |
| "name": "severity_level", |
| "type": "enum", |
| "value": "HIGH", |
| "confidence": 0.656, |
| "top": [ |
| { |
| "choice": "HIGH", |
| "p": 0.656 |
| }, |
| { |
| "choice": "CRITICAL", |
| "p": 0.3148 |
| }, |
| { |
| "choice": "MEDIUM", |
| "p": 0.0235 |
| }, |
| { |
| "choice": "LOW", |
| "p": 0.0057 |
| } |
| ] |
| }, |
| { |
| "name": "block_pull_request", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.961, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.961 |
| }, |
| { |
| "choice": "false", |
| "p": 0.039 |
| } |
| ] |
| }, |
| { |
| "name": "quarantine_branch", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9543, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9543 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0457 |
| } |
| ] |
| }, |
| { |
| "name": "credential_revocation_required", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9191, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9191 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0809 |
| } |
| ] |
| }, |
| { |
| "name": "requires_security_lead_signoff", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8105, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8105 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1895 |
| } |
| ] |
| }, |
| { |
| "name": "exploitability_in_production", |
| "type": "enum", |
| "value": "REMOTE_UNAUTHENTICATED", |
| "confidence": 0.5747, |
| "top": [ |
| { |
| "choice": "REMOTE_UNAUTHENTICATED", |
| "p": 0.5747 |
| }, |
| { |
| "choice": "THEORETICAL", |
| "p": 0.255 |
| }, |
| { |
| "choice": "INTERNAL_NETWORK_ONLY", |
| "p": 0.1408 |
| }, |
| { |
| "choice": "AUTHENTICATED_USER", |
| "p": 0.0295 |
| } |
| ] |
| }, |
| { |
| "name": "blast_radius", |
| "type": "enum", |
| "value": "FULL_AWS_ACCOUNT", |
| "confidence": 0.8097, |
| "top": [ |
| { |
| "choice": "FULL_AWS_ACCOUNT", |
| "p": 0.8097 |
| }, |
| { |
| "choice": "NONE", |
| "p": 0.0952 |
| }, |
| { |
| "choice": "CONTAINER_LOCAL", |
| "p": 0.0652 |
| }, |
| { |
| "choice": "DATABASE_COMPROMISE", |
| "p": 0.03 |
| } |
| ] |
| }, |
| { |
| "name": "auto_generate_fix", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.5254, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.5254 |
| }, |
| { |
| "choice": "false", |
| "p": 0.4746 |
| } |
| ] |
| }, |
| { |
| "name": "remediation_effort", |
| "type": "enum", |
| "value": "TRIVIAL_SECRETS_ROTATION", |
| "confidence": 0.7126, |
| "top": [ |
| { |
| "choice": "TRIVIAL_SECRETS_ROTATION", |
| "p": 0.7126 |
| }, |
| { |
| "choice": "PARAMETRIZED_SQL_REFACTOR", |
| "p": 0.1471 |
| }, |
| { |
| "choice": "MAJOR_ARCHITECTURE_CHANGE", |
| "p": 0.1403 |
| } |
| ] |
| }, |
| { |
| "name": "pci_dss_impact", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9758, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9758 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0242 |
| } |
| ] |
| }, |
| { |
| "name": "soc2_audit_flag", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9659, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9659 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0341 |
| } |
| ] |
| }, |
| { |
| "name": "hipaa_exposure_risk", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9958, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9958 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0042 |
| } |
| ] |
| }, |
| { |
| "name": "secret_type_detected", |
| "type": "enum", |
| "value": "AWS_ROOT_KEY", |
| "confidence": 0.9999, |
| "top": [ |
| { |
| "choice": "AWS_ROOT_KEY", |
| "p": 0.9999 |
| }, |
| { |
| "choice": "STRIPE_API_KEY", |
| "p": 0.0001 |
| }, |
| { |
| "choice": "DATABASE_PASSWORD", |
| "p": 0.0 |
| }, |
| { |
| "choice": "NONE", |
| "p": 0.0 |
| } |
| ] |
| }, |
| { |
| "name": "target_environment", |
| "type": "enum", |
| "value": "PRODUCTION_CLUSTER", |
| "confidence": 1.0, |
| "top": [ |
| { |
| "choice": "PRODUCTION_CLUSTER", |
| "p": 1.0 |
| }, |
| { |
| "choice": "STAGING_ENVIRONMENT", |
| "p": 0.0 |
| }, |
| { |
| "choice": "LOCAL_DEV", |
| "p": 0.0 |
| } |
| ] |
| }, |
| { |
| "name": "notify_ciso", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.8998, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.8998 |
| }, |
| { |
| "choice": "false", |
| "p": 0.1002 |
| } |
| ] |
| }, |
| { |
| "name": "requires_immediate_patch", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9615, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9615 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0385 |
| } |
| ] |
| }, |
| { |
| "name": "jira_ticket_priority", |
| "type": "enum", |
| "value": "P2_MAJOR", |
| "confidence": 0.5625, |
| "top": [ |
| { |
| "choice": "P2_MAJOR", |
| "p": 0.5625 |
| }, |
| { |
| "choice": "P1_CRITICAL", |
| "p": 0.2785 |
| }, |
| { |
| "choice": "P0_BLOCKER", |
| "p": 0.1091 |
| }, |
| { |
| "choice": "P3_MINOR", |
| "p": 0.0499 |
| } |
| ] |
| }, |
| { |
| "name": "assigned_team", |
| "type": "enum", |
| "value": "APP_SEC_TEAM", |
| "confidence": 0.8178, |
| "top": [ |
| { |
| "choice": "APP_SEC_TEAM", |
| "p": 0.8178 |
| }, |
| { |
| "choice": "DEVOPS_PLATFORM", |
| "p": 0.1151 |
| }, |
| { |
| "choice": "CORE_BACKEND", |
| "p": 0.057 |
| }, |
| { |
| "choice": "INFRA_CLOUD_SEC", |
| "p": 0.0101 |
| } |
| ] |
| }, |
| { |
| "name": "false_positive_risk", |
| "type": "enum", |
| "value": "HIGH", |
| "confidence": 0.7121, |
| "top": [ |
| { |
| "choice": "HIGH", |
| "p": 0.7121 |
| }, |
| { |
| "choice": "LOW", |
| "p": 0.2703 |
| }, |
| { |
| "choice": "MODERATE", |
| "p": 0.014 |
| }, |
| { |
| "choice": "ALMOST_ZERO", |
| "p": 0.0036 |
| } |
| ] |
| }, |
| { |
| "name": "public_cve_match", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.848, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.848 |
| }, |
| { |
| "choice": "false", |
| "p": 0.152 |
| } |
| ] |
| }, |
| { |
| "name": "auto_merge_allowed", |
| "type": "boolean", |
| "value": "False", |
| "confidence": 0.9489, |
| "top": [ |
| { |
| "choice": "false", |
| "p": 0.9489 |
| }, |
| { |
| "choice": "true", |
| "p": 0.0511 |
| } |
| ] |
| }, |
| { |
| "name": "container_escape_risk", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9792, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9792 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0208 |
| } |
| ] |
| }, |
| { |
| "name": "data_exfiltration_risk", |
| "type": "enum", |
| "value": "HIGH", |
| "confidence": 0.9938, |
| "top": [ |
| { |
| "choice": "HIGH", |
| "p": 0.9938 |
| }, |
| { |
| "choice": "MODERATE", |
| "p": 0.0047 |
| }, |
| { |
| "choice": "CATASTROPHIC", |
| "p": 0.0015 |
| }, |
| { |
| "choice": "NEGLIGIBLE", |
| "p": 0.0 |
| } |
| ] |
| }, |
| { |
| "name": "author_trust_score", |
| "type": "enum", |
| "value": "UNTRUSTED_EXTERNAL", |
| "confidence": 0.8456, |
| "top": [ |
| { |
| "choice": "UNTRUSTED_EXTERNAL", |
| "p": 0.8456 |
| }, |
| { |
| "choice": "NEW_EMPLOYEE", |
| "p": 0.0818 |
| }, |
| { |
| "choice": "VERIFIED_STAFF", |
| "p": 0.0459 |
| }, |
| { |
| "choice": "CORE_MAINTAINER", |
| "p": 0.0268 |
| } |
| ] |
| }, |
| { |
| "name": "security_gate_status", |
| "type": "enum", |
| "value": "FAILED_BLOCKING", |
| "confidence": 0.9998, |
| "top": [ |
| { |
| "choice": "FAILED_BLOCKING", |
| "p": 0.9998 |
| }, |
| { |
| "choice": "PASSED", |
| "p": 0.0002 |
| }, |
| { |
| "choice": "WARNING_OVERRIDABLE", |
| "p": 0.0 |
| } |
| ] |
| } |
| ], |
| "json": { |
| "is_vulnerability": true, |
| "primary_cwe": "CWE_89_SQL_INJECTION", |
| "secondary_cwe": "CWE_798_HARDCODED_CREDENTIALS", |
| "severity_level": "HIGH", |
| "block_pull_request": true, |
| "quarantine_branch": true, |
| "credential_revocation_required": true, |
| "requires_security_lead_signoff": true, |
| "exploitability_in_production": "REMOTE_UNAUTHENTICATED", |
| "blast_radius": "FULL_AWS_ACCOUNT", |
| "auto_generate_fix": true, |
| "remediation_effort": "TRIVIAL_SECRETS_ROTATION", |
| "pci_dss_impact": true, |
| "soc2_audit_flag": true, |
| "hipaa_exposure_risk": true, |
| "secret_type_detected": "AWS_ROOT_KEY", |
| "target_environment": "PRODUCTION_CLUSTER", |
| "notify_ciso": true, |
| "requires_immediate_patch": true, |
| "jira_ticket_priority": "P2_MAJOR", |
| "assigned_team": "APP_SEC_TEAM", |
| "false_positive_risk": "HIGH", |
| "public_cve_match": true, |
| "auto_merge_allowed": false, |
| "container_escape_risk": true, |
| "data_exfiltration_risk": "HIGH", |
| "author_trust_score": "UNTRUSTED_EXTERNAL", |
| "security_gate_status": "FAILED_BLOCKING" |
| } |
| }, |
| "naive": { |
| "elapsed_ms": 3267.09, |
| "total_tokens": 299, |
| "tokens_per_second": 91.5, |
| "schema_match": false, |
| "raw_text": "{\n \"is_vulnerability\": true,\n \"primary_cwe\": \"CWE_798_HARDCODED_CREDENTIALS\",\n \"secondary_cwe\": \"CWE_502_DESERIALIZATION\",\n \"severity_level\": \"HIGH\",\n \"block_pull_request\": true,\n \"quarantine_branch\": true,\n \"credential_revocation_required\": true,\n \"requires_security_lead_signoff\": true,\n \"exploitability_in_production\": \"AUTHENTICATED_USER\",\n \"blast_radius\": \"DATABASE_COMPROMISE\",\n \"auto_generate_fix\": true,\n \"remediation_effort\": \"PARAMETRIZED_SQL_REFACTOR\",\n \"pci_dss_impact\": true,\n \"soc2_audit_flag\": true,\n \"hipaa_exposure_risk\": true,\n \"secret_type_detected\": \"AWS_ROOT_KEY\",\n \"target_environment\": \"PRODUCTION_CLUSTER\",\n \"notify_ciso\": true,\n \"requires_immediate_patch\": true,\n \"jira_ticket_priority\": \"P0_BLOCKER\",\n \"assigned_team\": \"INFRA_CLOUD_SEC\",\n \"false_positive_risk\": \"MODERATE\",\n \"public_cve_match\": true,\n \"auto_merge_allowed\": false,\n \"container_escape_risk\": true,\n \"data_exfiltration_risk\": \"HIGH\",\n \"author_trust_score\": \"NEW_EMPLOYEE\"\n}" |
| } |
| }, |
| { |
| "id": "high_cardinality_255", |
| "title": "High-Cardinality 255-Choice Tariff & Customs Router", |
| "description": "Real-time multi-attribute classification featuring a 255-choice discrete enum to demonstrate non-degraded O(1) latency scaling.", |
| "context": "CUSTOMS CLEARANCE INVOICE #AIR-60291\nShipper: Precision AeroForgings GmbH, Stuttgart, Germany\nConsignee: Lockheed Skyworks Defense Facility, Fort Worth, TX, USA\nItem Description: Batch of 5,000 aerospace-grade titanium alloy hex-flange structural fasteners (Ti-6Al-4V Grade 5) with anodized dry-film anti-galling lubricant coating. Specifically milled and certified under Aerospace Standard AS9500 for turbine rotor assembly in commercial and defense propulsion modules.\nGross Weight: 240 kg\nDeclared Value: ,500 USD\nITAR / Dual-Use Assessment: Requires verification of Defense Logistics Agency (DLA) end-user certificate.", |
| "num_fields": 4, |
| "parallel": { |
| "elapsed_ms": 142.25, |
| "prefill_ms": 77.28, |
| "suffix_eval_ms": 46.89, |
| "schema_match": true, |
| "fields": [ |
| { |
| "name": "customs_category", |
| "type": "enum", |
| "value": "CAT_100_Live_Animals", |
| "confidence": 0.75, |
| "top": [ |
| { |
| "choice": "CAT_100_Live_Animals", |
| "p": 0.75 |
| }, |
| { |
| "choice": "CAT_000_Live_Animals", |
| "p": 0.001 |
| }, |
| { |
| "choice": "CAT_001_Meat_&_Edible_Offal", |
| "p": 0.001 |
| }, |
| { |
| "choice": "CAT_002_Fish_&_Seafood", |
| "p": 0.001 |
| } |
| ] |
| }, |
| { |
| "name": "export_license_required", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.7866, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.7866 |
| }, |
| { |
| "choice": "false", |
| "p": 0.2134 |
| } |
| ] |
| }, |
| { |
| "name": "is_dual_use_goods", |
| "type": "boolean", |
| "value": "True", |
| "confidence": 0.9962, |
| "top": [ |
| { |
| "choice": "true", |
| "p": 0.9962 |
| }, |
| { |
| "choice": "false", |
| "p": 0.0038 |
| } |
| ] |
| }, |
| { |
| "name": "inspection_priority", |
| "type": "enum", |
| "value": "HOLD_PHYSICAL_LAB_INSPECTION", |
| "confidence": 0.9916, |
| "top": [ |
| { |
| "choice": "HOLD_PHYSICAL_LAB_INSPECTION", |
| "p": 0.9916 |
| }, |
| { |
| "choice": "STANDARD_FAST_TRACK", |
| "p": 0.0084 |
| }, |
| { |
| "choice": "RANDOM_SCREENING", |
| "p": 0.0 |
| } |
| ] |
| } |
| ], |
| "json": { |
| "customs_category": "CAT_100_Live_Animals", |
| "export_license_required": true, |
| "is_dual_use_goods": true, |
| "inspection_priority": "HOLD_PHYSICAL_LAB_INSPECTION" |
| } |
| }, |
| "naive": { |
| "elapsed_ms": 1264.1, |
| "total_tokens": 53, |
| "tokens_per_second": 41.9, |
| "schema_match": true, |
| "raw_text": "{\n \"customs_category\": \"CAT_019_Vegetable_Preparations\",\n \"export_license_required\": true,\n \"is_dual_use_goods\": true,\n \"inspection_priority\": \"HOLD_PHYSICAL_LAB_INSPECTION\"\n}" |
| } |
| } |
| ] |
| } |
