"""NEXUS OS v2.1 — Production Hybrid Cloud+Local Inference with BEC Thermodynamic Control. New in v2.1: - Ollama telemetry extractor (embedding-space divergence for entropy proxy) - ChimeraRouter v2.1 with integrated telemetry pipeline + cloud API adapters + fallback controller - Production runner (run_production.py) — connects to real Ollama, Pinecone, Milvus - EPR Detector (arXiv:2509.04492) — token-level entropy production rate - Spilled Energy Detector (arXiv:2602.18671) — energy discrepancy hallucination signal - Unified Thermodynamic Detector — fuses EPR + SpilledEnergy + CK-PLUG + TWAVE - CombinedThermodynamicDetector — weighted ensemble with actionable recommendations - Cloud API Adapters — DeepSeek, Qwen, Kimi, GLM, OpenAI, Claude - Model Fallback Controller — auto-escalation: re-ground → backtrack → switch model → cloud - Novel Composite Signals — Energy-Entropy Product (EEP), Phase Transition Index (PTI), NEWI - Stochastic resonance optimal temperature computation per-request - Post-hoc entropy estimation from completed responses Core modules: model_registry — 35+ real Ollama + cloud model profiles sulphur_enhancer — Prompt quality optimization ckplug_retriever — Token-level retrieval coupling (arXiv:2503.15888) ckplug_presets — Epsilon presets for all 35+ models twave_tracker — BEC Landau-Ginzburg thermodynamic tracking epr_integration — Entropy Production Rate (arXiv:2509.04492) epr_detector — Standalone EPR + WEPR detector spilled_energy — Energy-based hallucination detection (arXiv:2602.18671) unified_detector — Production unified detector (EPR + Spilled + CK-PLUG + TWAVE) temperature_sweep_integration — Empirical T_c from Nemotron sweep pinecone_client — Vector + Assistant retrieval (llama-text-embed-v2-index) milvus_client — Zilliz Cloud retrieval (Free-01 / Serverless-01) ernie_adapter — Browser-sourced manual callback (yiyan.baidu.com) chimera_router — Hybrid orchestrator (QWAVE + TWAVE) chimera_router_v2_1 — Telemetry-integrated production router + cloud + fallback cloud_api_adapters — DeepSeek, Qwen, Kimi, GLM, OpenAI, Claude gmr_scheduler — VRAM-aware model rotation per_token_debug — Telemetry schema for token-level tracking ollama_telemetry — Ollama stream + embedding divergence extractor adpo_bec_training — ADPO-BEC training objective skeleton trackio_integration — Experiment tracking, alerts, dashboard run_production — Production runner with real credentials demo_unified_detector — Interactive demo of all detection signals cli — Command-line interface Paper corpus (16 papers): [1] Fast Byte Latent Transformer [2] Cognitive Pivot Points / V-STAR / RVTD [3] RLAR: Agentic Reward System [4] Autoregressive DPO [5] Unpacking DPO and PPO [6] Disentangling Length from Quality in DPO [7] Mechanistic Interpretability for AI Safety [8] Critique-out-Loud Reward Models [9] Understanding RLHF Effects on Generalisation [10] Big Self-Supervised Models [11] Discourse-Aware Attention for Summarization [12] CK-PLUG (arXiv:2503.15888) — Confidence Gain [13] SDE Self-Adversariality (arXiv:2501.16783) — Fokker-Planck validation [14] EPR Hallucination Detection (arXiv:2509.04492) — Token-level entropy production [15] Spilled Energy (arXiv:2602.18671) — Energy discrepancy in autoregressive EBMs [16] PLDR-LLM Self-Organized Criticality (arXiv:2603.23539) — Order parameter at criticality Usage (production): export PINECONE_API_KEY="pcsk_J9DUW_..." export OLLAMA_HOST="http://localhost:11434" export DEEPSEEK_API_KEY="..." export OPENAI_API_KEY="..." export ANTHROPIC_API_KEY="..." python -m nexus_os_v2.run_production \ --prompt "Explain quantum entanglement" \ --vram 16 \ --model deepseek-r1-8b \ --telemetry \ --use-pinecone \ --save-telemetry Usage (unified detector demo): python -m nexus_os_v2.demo_unified_detector --mode mock Usage (benchmark): python -m benchmarks.benchmark_real_models --vram 16 --auto-models --telemetry Usage (calibrate): python -m scripts.calibrate_detectors --model deepseek-r1-8b --iterations 3 Usage (basic): from nexus_os_v2.chimera_router import ChimeraRouter, QWAVEBudget router = ChimeraRouter() result = router.route("Explain quantum computing") print(result.response) """ __version__ = "2.1.0" __author__ = "specimba" from .model_registry import REGISTRY, SULPHUR, get, by_tier, by_cap, all_names, vram, Tier, Capability, ModelProfile from .sulphur_enhancer import SulphurEnhancer, MockSulphurEnhancer, EnhancedPrompt from .ckplug_retriever import CKPLUGCoupling, TokenModulation from .ckplug_presets import get_ckplug_epsilon, get_all_presets, preset_summary from .twave_tracker import TWAVETracker, TokenState, GenerationTrajectory, StochasticResonance from .epr_integration import EPRCalculator, EPRAnalyzer, EPRTWWVEIntegration, TokenLogprob, quick_epr from .epr_detector import EPRDetector, WEPRDetector, EPRReading, SequenceEPR from .spilled_energy import SpilledEnergyDetector, SpilledEnergyReading, CombinedThermodynamicDetector from .unified_detector import ( UnifiedThermodynamicDetector, FusionMode, Action, DetectorReading, TokenVerdict, SequenceVerdict, ) from .temperature_sweep_integration import TcCalibrator, EmpiricalTc, scale_Tc_for_family from .pinecone_client import PineconeRetriever, MockPineconeRetriever, RetrievalResult from .milvus_client import MilvusRetriever, MockMilvusRetriever, MilvusResult from .ernie_adapter import ERNIEAdapter, MockERNIEAdapter, ERNIEEvidence from .chimera_router import ChimeraRouter, QWAVEBudget, RouterResult, RoutingDecision from .chimera_router_v2_1 import ChimeraRouterV2, ModelFallbackController from .cloud_api_adapters import ( CloudAPIManager, BaseCloudAdapter, CloudResponse, DeepSeekAdapter, QwenAdapter, KimiAdapter, GLMAdapter, OpenAIAdapter, ClaudeAdapter, ) from .gmr_scheduler import GMRScheduler, ModelSlot, LoadState from .per_token_debug import PerTokenDebug, GenerationTelemetry from .ollama_telemetry import ( OllamaTelemetryExtractor, OllamaStreamingClient, TokenTelemetry, EmbeddingTrajectory, estimate_entropy_from_response, ) from .adpo_bec_training import ADPOBEC, ADPOBECHyperparams, PreferencePair from .trackio_integration import NEXUSTrackioMonitor, MockTrackioMonitor, TrackioConfig __all__ = [ # Registry "REGISTRY", "SULPHUR", "get", "by_tier", "by_cap", "all_names", "vram", "Tier", "Capability", "ModelProfile", # Enhancer "SulphurEnhancer", "MockSulphurEnhancer", "EnhancedPrompt", # CK-PLUG "CKPLUGCoupling", "TokenModulation", "get_ckplug_epsilon", "get_all_presets", "preset_summary", # TWAVE "TWAVETracker", "TokenState", "GenerationTrajectory", "StochasticResonance", # EPR "EPRCalculator", "EPRAnalyzer", "EPRTWWVEIntegration", "TokenLogprob", "quick_epr", "EPRDetector", "WEPRDetector", "EPRReading", "SequenceEPR", # Spilled Energy "SpilledEnergyDetector", "SpilledEnergyReading", "CombinedThermodynamicDetector", # Unified Detector "UnifiedThermodynamicDetector", "FusionMode", "Action", "DetectorReading", "TokenVerdict", "SequenceVerdict", # Temperature Sweep "TcCalibrator", "EmpiricalTc", "scale_Tc_for_family", # Retrievers "PineconeRetriever", "MockPineconeRetriever", "RetrievalResult", "MilvusRetriever", "MockMilvusRetriever", "MilvusResult", "ERNIEAdapter", "MockERNIEAdapter", "ERNIEEvidence", # Routers "ChimeraRouter", "ChimeraRouterV2", "ModelFallbackController", "QWAVEBudget", "RouterResult", "RoutingDecision", # Cloud APIs "CloudAPIManager", "BaseCloudAdapter", "CloudResponse", "DeepSeekAdapter", "QwenAdapter", "KimiAdapter", "GLMAdapter", "OpenAIAdapter", "ClaudeAdapter", # GMR "GMRScheduler", "ModelSlot", "LoadState", # Telemetry "PerTokenDebug", "GenerationTelemetry", "OllamaTelemetryExtractor", "OllamaStreamingClient", "TokenTelemetry", "EmbeddingTrajectory", "estimate_entropy_from_response", # Training "ADPOBEC", "ADPOBECHyperparams", "PreferencePair", # Monitoring "NEXUSTrackioMonitor", "MockTrackioMonitor", "TrackioConfig", ]