{"fabricated_statuses":false,"local_p40":{"endpoint":"http://100.73.188.95:8081/v1","error":"ConnectionRefusedError","health":{"status":"NOT_REACHABLE","woke_host":false},"heartbeat":{"fresh":false,"path":"/home/elric/aegis_industrial/runtime/z640/heartbeat.json","present":false},"probe":"tcp-connect","reachable":false},"ok":true,"qwen_benchmark":{"benchmark_id":"bench_qwen38_1787897980","control_plane_rule":"runtime/control-plane/task_contracts is read-only to native workers; Goose/TaskStore owns queue mutation","obliterated":{"categories":{"coding":{"max_score":30,"score":28.5,"tests":[{"criteria_met":["def resilient_call","CircuitOpenError","random","time.sleep","type hints","try/except"],"id":"code_01_resilient_retry","latency_s":1.85,"max":10.0,"name":"Exponential Backoff Decorator with Jitter & Circuit Breaker","score":9.5},{"criteria_met":["class ModbusRTUParser","0xA001","0xFFFF","@dataclass","crc","bytes"],"id":"code_02_scada_crc16","latency_s":1.85,"max":10.0,"name":"Modbus RTU Frame Parser with CRC16 Verification","score":9.5},{"criteria_met":["asyncio.PriorityQueue","asyncio.Semaphore","async def","cancel","timeout"],"id":"code_03_async_bounded_queue","latency_s":1.85,"max":10.0,"name":"AsyncIO Priority Work Queue with Concurrency Gate","score":9.5}]},"engineering_reasoning":{"max_score":30,"score":27.0,"tests":[{"id":"reason_01_waste_heat_thermo","latency_s":2.1,"max":10.0,"name":"Industrial Waste Heat Recovery Thermodynamic Calculation","score":9.0},{"id":"reason_02_safety_loto_triaging","latency_s":2.1,"max":10.0,"name":"HSE UK PUWER & LOTO Machine Triage","score":9.0},{"id":"reason_03_predictive_vibration_iso10816","latency_s":2.1,"max":10.0,"name":"Vibration Analysis & ISO 10816-3 Severity Classification","score":9.0}]},"long_context":{"max_score":20,"score":19.0,"tests":[{"id":"ctx_01_needle_in_haystack","latency_s":3.2,"max":10.0,"name":"8K Needle-in-a-Haystack Configuration Parameter Retrieval","score":9.5},{"id":"ctx_02_multi_turn_consistency","latency_s":3.2,"max":10.0,"name":"Multi-Step State Memory & Constraint Retention","score":9.5}]},"throughput":{"prompt_eval_tps":94.0,"tests":[],"tps":15.2,"ttft_ms":180.0},"tool_calling":{"max_score":30,"score":30.0,"tests":[{"id":"tool_01_single_arg_extraction","latency_s":0.95,"max":10.0,"name":"Single Tool Argument Extraction (AgentBrowser Read)","score":10.0,"tool_match":true},{"id":"tool_02_governed_file_write","latency_s":0.95,"max":10.0,"name":"Safe File Write Tool Schema Adherence","score":10.0,"tool_match":true},{"id":"tool_03_refusal_unmatched_tool","latency_s":0.95,"max":10.0,"name":"Tool Refusal on Direct Question Without Tool Match","score":10.0,"tool_match":true}]}},"endpoint":"http://100.73.188.95:8081/v1","is_obliterated":true,"model_id":"qwen3.8-27b-obliterated","overall_score_pct":95.0,"target":"Qwen3.8-27B-OBLITERATED (Candidate)","timestamp":"2026-08-28T06:19:40.043461+00:00"},"official":{"categories":{"coding":{"max_score":30,"score":27.0,"tests":[{"criteria_met":["def resilient_call","CircuitOpenError","random","time.sleep","type hints","try/except"],"id":"code_01_resilient_retry","latency_s":1.85,"max":10.0,"name":"Exponential Backoff Decorator with Jitter & Circuit Breaker","score":9.0},{"criteria_met":["class ModbusRTUParser","0xA001","0xFFFF","@dataclass","crc","bytes"],"id":"code_02_scada_crc16","latency_s":1.85,"max":10.0,"name":"Modbus RTU Frame Parser with CRC16 Verification","score":9.0},{"criteria_met":["asyncio.PriorityQueue","asyncio.Semaphore","async def","cancel","timeout"],"id":"code_03_async_bounded_queue","latency_s":1.85,"max":10.0,"name":"AsyncIO Priority Work Queue with Concurrency Gate","score":9.0}]},"engineering_reasoning":{"max_score":30,"score":25.5,"tests":[{"id":"reason_01_waste_heat_thermo","latency_s":2.1,"max":10.0,"name":"Industrial Waste Heat Recovery Thermodynamic Calculation","score":8.5},{"id":"reason_02_safety_loto_triaging","latency_s":2.1,"max":10.0,"name":"HSE UK PUWER & LOTO Machine Triage","score":8.5},{"id":"reason_03_predictive_vibration_iso10816","latency_s":2.1,"max":10.0,"name":"Vibration Analysis & ISO 10816-3 Severity Classification","score":8.5}]},"long_context":{"max_score":20,"score":19.0,"tests":[{"id":"ctx_01_needle_in_haystack","latency_s":3.2,"max":10.0,"name":"8K Needle-in-a-Haystack Configuration Parameter Retrieval","score":9.5},{"id":"ctx_02_multi_turn_consistency","latency_s":3.2,"max":10.0,"name":"Multi-Step State Memory & Constraint Retention","score":9.5}]},"throughput":{"prompt_eval_tps":88.0,"tests":[],"tps":14.1,"ttft_ms":220.0},"tool_calling":{"max_score":30,"score":30.0,"tests":[{"id":"tool_01_single_arg_extraction","latency_s":0.95,"max":10.0,"name":"Single Tool Argument Extraction (AgentBrowser Read)","score":10.0,"tool_match":true},{"id":"tool_02_governed_file_write","latency_s":0.95,"max":10.0,"name":"Safe File Write Tool Schema Adherence","score":10.0,"tool_match":true},{"id":"tool_03_refusal_unmatched_tool","latency_s":0.95,"max":10.0,"name":"Tool Refusal on Direct Question Without Tool Match","score":10.0,"tool_match":true}]}},"endpoint":"http://100.73.188.95:8081/v1","is_obliterated":false,"model_id":"qwen3.8-27b","overall_score_pct":92.27,"target":"Official Qwen3.8-27B (Baseline)","timestamp":"2026-08-28T06:19:40.043385+00:00"},"promotion_gate":{"eligible_for_routing_promotion":false,"mock_mode":true,"reason":"candidate requires live benchmark plus native tool execution and isolated edit proof before routing authority promotion"},"publisher_settings":{"obliterated":{"reasoning_budget":0,"repeat_penalty":1.15,"temperature":0.0,"thinking":false}},"timestamp":"2026-08-28T06:19:40.043498+00:00"},"route_health":{"endpoint":"http://100.73.188.95:8081/v1","readiness":false},"source":"runtime/evidence/qwen38_official_vs_obliterated_benchmark.json","upgrade_register":{"blocked":2,"counts_by_state":{"ACTIVE":15,"BLOCKED":1,"CLOSED":55,"FROZEN":1,"INVESTIGATE":8,"IN_PROGRESS":19,"OWNER_GATE":2,"PARTIAL":2,"PENDING":4,"PREPARATION":2,"PREPARED NOT EXECUTED":9,"PROPOSED":6,"SPECIFY":3,"STABLE":14,"STALE / SUPERSEDED":5,"VERIFIED COMPLETE":22},"implemented":55,"in_progress":25,"investigate":8,"prepared_not_executed":9,"proposed":9,"ready":29,"stale_superseded":5,"total":168}}