{"status":"ok","operational_mode":"SOLE_T4_INFERENCE_PRIORITY_ACTIVE","mode_badge":"⚡ SOLE GPU ONLINE: Tesla T4 (16 GB) 100% Available for Real-Time Inference (Night QLoRA Scheduled 22:00 CET)","inference_availability":"100% DEDICATED TO INFERENCE: The NVIDIA Tesla T4 (16 GB) is currently the sole active GPU on the network (21:44 CET European Day). BOINC QLoRA MoE training is held/yielded (GPU mode: never) so end-user completions receive 100% priority; multi-MoE QLoRA training resumes primarily during European Night (22:00–07:00 CET/CEST) and quiet windows.","sole_active_gpu":false,"active_producers_count":3,"hardware":{"primary_gpu":"NVIDIA Tesla T4 (16 GB GDDR6)","node_host":"socreateai-agora-hermes-node1 (us-east4-b)","secondary_gpu":"NVIDIA GeForce RTX 2070 (8 GB GDDR6 Turing — Live Producer 58fd3df19d2b5ef8c10031e1389aac6a)","vram_used_mib":4841,"vram_total_mib":15360,"vram_free_mib":10519,"vram_utilization_pct":31.5,"temp_c":66,"power_w":30.2,"resident_processes":[{"name":"llama-server (hivemind/bee-stars • Qwen2.5-1.5B + NIGHT-MOE-1 Stars LoRA)","mib":4838},{"name":"hivemind-rtx2070-agent (hivemind/coder • Qwen2.5-Coder-7B-Instruct & coder-lite on RTX 2070)","mib":5462}]},"european_schedule":{"timezone":"Europe/Paris (CET/CEST)","current_utc_iso":"2026-10-28T20:44:01+00:00","current_european_time":"2026-10-28 21:44:01 CET","current_european_hour":21,"is_european_night_window":false,"night_window_cet":"22:00 – 07:00 CET/CEST (21:00 – 06:00 UTC)","day_window_cet":"07:00 – 22:00 CET/CEST (06:00 – 21:00 UTC)","day_policy":"Sole Tesla T4 dedicated 100% to low-latency interactive inference (hivemind/coder, hivemind/bee-stars, hivemind/coder-lite, hivemind/auto).","night_policy":"During European Night (22:00–07:00 CET/CEST) and quiet intervals (>=30s idle), the T4 runs BOINC QLoRA training of more MoE experts across more datasets (WikiText-2 signed shards, GAIA/Hipparcos Stars, Kepler/TESS Exoplanets, Verified GSM8K/Lean4 Maths, and Execution-Verified Code/Cyber), yielding in <1s whenever an inference request arrives."},"governor":{"daemon":"hivemind-t4-governor.service","poll_interval_s":1.0,"idle_start_threshold_s":1800.0,"state":"serving_inference_priority","boinc_gpu_mode":"never","hold_file_present":false,"in_flight_requests":0,"yields_count":130,"last_yield_ms":50,"real_mid_task_yield_ms":[766,7129,23406,15929,1135,1133],"preemption_policy":"100% Inference Priority: t4_governor.py polls nginx stub_status and GPU utilization every 1.0s. On any incoming inference activity, BOINC GPU mode is set to 'never' and active BOINC child tasks are terminated immediately (~1.13s measured after kill-on-yield fix)."},"live_inference_models":["bee-lean-prover","bee-ruche-moe","bee-rusty-sundials","bee-stars","coder","coder-lite","coder-plus","crowd-v1","gemma-core","mistral-core","moe-coder","moe-general","moe-qwen35-bee","qwen35-core"],"qlora_progress":{"boinc_app":"hivemind_qlora_moe (BOINC 7.18.1 / 7.24 Core Node)","workgen_daemon":"hivemind-t4-workgen.service","assimilated_workunits":100,"qlora_rounds_assimilated":98,"ratchet_accepted_rounds":20,"ratchet_held_rounds":78,"chain_head_workunit":"hmq_00092","chain_head_sha256":"e91e0ec338b92e52b13bf4abdc2c2d95999b56c3568cb85de5b306a8d3c9f5f8","queued_workunit":"hmq_00099 (chained from hmq_00092 latest_adapter.pt SHA-256)","segment_checkpoint_steps":10,"inner_steps_per_workunit":40,"peak_training_vram_mib":3409,"replay_max_divergence":0.0,"verified_cobblestones":5632.11,"held_out_eval_loss_before":2.2705,"held_out_eval_loss_after":2.138,"held_out_eval_by_domain_baseline_hmq_00002":{"general":2.323,"math":1.644,"biomed":2.5017,"edu_web":2.6332,"french":2.3789,"news":2.1422},"held_out_eval_by_domain_head_hmq_00092":{"general":2.1177,"math":1.3949,"biomed":2.2089,"edu_web":2.6342,"french":2.3475,"news":2.1246},"caveats":"98 rounds (85 inside 22:00-07:00 CET, mean 105 s/round); mean 6-domain held-out loss 2.2705 -> 2.1380 (-5.8%). Plateaued after ~30 rounds (last-20 std 0.0055; rounds 69-98 target-domain delta +0.0006). Small fixed held-out sets (8x128 tokens/domain), best-of-97 ratchet selection on those sets, and granite-3.1-1b-a400m-base + LoRA is evaluated offline (not yet served by the router)."},"qlora_moe_curriculum":[{"track_id":"BOINC-MOE-GRANITE-6DOMAIN","moe_backbone":"ibm-granite/granite-3.1-1b-a400m-base (1.3B Sparse MoE, 400M active)","qlora_recipe":"NF4 double-quantized attention (q,k,v,o) + fp32 LoRA r=8, alpha=16 (1,376,256 trainable params)","dataset_name":"6-Domain Night Curriculum (general, math, biomed, edu_web, french, news • 98 BOINC rounds hmq_00001..00098)","schedule_window":"European Night (22:00–07:00 CET/CEST) • 85/98 rounds in night window","vram_footprint_mib":3409,"status":"ACTIVE_CHAINED_QUEUE","measured_evidence":"98 QLoRA rounds assimilated (5,632.11 Cobblestones, 20 ratchet-accepted, head hmq_00092); 6-domain held-out NLL 2.2705 -> 2.1380 (-5.8%: math -15.1%, biomed -11.7%, general -8.8%, plateau after ~30 rounds)"},{"track_id":"NIGHT-MOE-1-BEE-STARS","moe_backbone":"Qwen2.5-Coder-7B / 1.5B Specialist Expert (hivemind/bee-stars)","qlora_recipe":"QLoRA 4-bit NF4 + LoRA Expert Adapter (300 steps, exact paired McNemar gate)","dataset_name":"GAIA / Hipparcos Stellar Spectral Classification & Luminosity Pack (n=317)","schedule_window":"Trained European Night 2026-10-24/25 -> Now Served Live 24/7","vram_footprint_mib":4838,"status":"VALIDATED_AND_SERVING_LIVE","measured_evidence":"Accuracy 0.489 -> 0.893 (+40.4%, McNemar p < 1e-4); beats shuffled control 193 to 7"},{"track_id":"NIGHT-MOE-2-EXOPLANET","moe_backbone":"Qwen2.5-Coder-7B / 14B & Sparse MoE Specialist Expert (exoplanet)","qlora_recipe":"QLoRA NF4 + Star-Split Paired Adapter-vs-Control Gate + La Ruche SigmaMoE Router","dataset_name":"Kepler / TESS Exoplanet Transit & Stellar-Host Split Catalogue (n=600)","schedule_window":"European Night (22:00–07:00 CET/CEST) Multi-Dataset Expansion","vram_footprint_mib":10230,"status":"SCHEDULED_NIGHT_WINDOW","measured_evidence":"NIGHT-MOE-1 post-hoc: adapter beat shuffled control 309 to 17 on 600 items; queued for routed MoE gate"},{"track_id":"NIGHT-MOE-2-VERIFIED-MATHS","moe_backbone":"Qwen2.5-Coder-7B / 14B & gpt-oss-20B / Qwen3-30B-A3B MoE (maths)","qlora_recipe":"Rejection-Sampled Self-Solutions (E15) + Cross-Domain Adapter Control","dataset_name":"GSM8K & Lean 4 Execution-Verified Mathematical Proofs (n=300+)","schedule_window":"European Night (22:00–07:00 CET/CEST) Multi-Dataset Expansion","vram_footprint_mib":10230,"status":"SCHEDULED_NIGHT_WINDOW","measured_evidence":"Replaces terse gold text with verified self-generated solutions to eliminate format interference"},{"track_id":"NIGHT-MOE-2-CODE-AND-CYBER","moe_backbone":"Qwen2.5-Coder-7B / 14B & Qwen3-Coder-30B-A3B Sparse MoE (code + cyber)","qlora_recipe":"Execution-Attested Pytest Sandbox + Verified CTI-Bench / MMLU Security MCQ","dataset_name":"OpenRouter Live Coding Failure Distillations + Verified Cyber MCQ Pack (n=300+)","schedule_window":"European Night (22:00–07:00 CET/CEST) Multi-Dataset Expansion","vram_footprint_mib":10230,"status":"SCHEDULED_NIGHT_WINDOW","measured_evidence":"Trains La Ruche SigmaMoE sigmoid gate across coding, astrophysics, maths, and cyber experts"}],"nocturnal_auto_improvement":{"status":"ACTIVE_DISTRIBUTED_FALLBACK","gcp_spot_preemption_fallback":"ENGAGED: When GCP Spot Tesla T4 instances are preempted by high cloud demand, nocturnal MoE training and auto-improvement workunits are dynamically scheduled across connected swarm producers (RTX 2070, RTX 3080/3090, RTX 4070/4090) under the strict < 8 GB VRAM ceiling.","target_domains":["maths","exoplanet","stars","cyber","coder"],"pouw_points_retrofit":"AUTOMATIC: Completed and verified auto-improvement workunits receive hash-chained Proof-of-Useful-Work (PoUW) HMT Reflexion Points retrofitted directly into the router ledger."}}