specimba/nexus-os-v2-dualmode

NEXUS OS v2.1 -- Dual-Mode Token-Level Hallucination Control

Landau-Ginzburg phase transition hypothesis -> empirically confirmed. No universal T_c.

What This Is

NEXUS OS is an autonomous agent operating system that routes LLM inference across local hardware (8GB consumer GPUs) and cloud APIs. Its core is:

• ChimeraRouter: 4-tier automatic model selection
• TWAVE: Token-level wavefront expansion with Landau-Ginzburg thermodynamic tracking
• QWAVE: Quality budget allocator mapping task -> resource budget
• CK-PLUG: Token-level retrieval coupling (Confidence Gain from evidence entropy)
• SafetyHead: Dual-mode safety detection (strict toxic-token / frontier bifurcation)

Quick Start

from nexus_os_v2 import RealityBridge, Mode, ChimeraRouter, QWAVEScorer

# STRICT mode (production -- proven methods only)
bridge = RealityBridge(mode=Mode.STRICT_REALITY)

# Classify intent -> quality budget
qs = QWAVEScorer()
budget = qs.classify_and_score("what is the capital of france",
                                 intent_model=my_real_classifier,
                                 mode="strict")

# Route to best model given VRAM + budget
router = ChimeraRouter(mode="strict")
decision = router.route(budget.task_type, budget.budget, local_vram=8192)
print(decision.model_name, decision.tier)
# -> "Qwen3-7B-Q4_K_M" tier=2

Dual-Mode Architecture (v2.1 Core Feature)

Every component exposes both proven (strict) and experimental (frontier) paths:

Mode	Uncertainty	Use When
STRICT	0.0-0.20	Production, reproducibility required, third-party validation exists
FRONTIER	0.4-0.65	Research, prototyping, exploring invention space

Promote frontier -> strict: bridge.promote("my_method", evidence_score=0.88) requires score >= 0.85.

Files

File	Purpose
reality_bridge.py	Central dual-mode gatekeeper, audit logging, promotion API
twave_tracker.py	Token entropy tracking + Landau-Ginzburg field evolution
ckplug_retriever.py	BM25 retrieval + CK-PLUG Confidence Gain (mu_ret coupling)
model_registry.py	ChimeraRouter: 4-tier model routing with VRAM probe
qwave.py	Quality budget allocator: task -> Q in [0,1]
safety_head.py	Dual-mode safety (strict toxic tokens / frontier bifurcation SSA)
visualizations/phase_transition_plots.py	Publication-quality 6-panel LG phase diagrams

Version

2.1.0 -- Dual-Mode Architecture release.

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = 'specimba/nexus-os-v2-dualmode'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.