specimba/nexus-os-v2

NEXUS OS v2.0 — Thermodynamic Hallucination Control

ChimeraRouter + TWAVE + ERNIE Integration

Upgraded with 2024–2026 SOTA findings: LEAD, EDT, EPR, LED, CK-PLUG, Attention Divergence, TECP, AutoDeco.

Architecture

┌─────────────┐     ┌─────────────────┐     ┌─────────────────┐
│   Prompt    │────▶│ ChimeraRouterV2 │────▶│ RoutingDecision │
└─────────────┘     └─────────────────┘     └─────────────────┘
      │                    │                         │
      ▼                    ▼                         ▼
┌─────────────┐     ┌─────────────────┐     ┌─────────────────┐
│   ERNIE     │     │ PromptAnalyzer  │     │ QwaveAllocator  │
│  (agent)    │     │  + PolicySelect │     │  Budget tokens  │
└─────────────┘     └─────────────────┘     └─────────────────┘

┌──────────────────────────────────────────────────────────────┐
│                  TWAVE v2.0 — Landau-Ginzburg               │
├──────────────────────────────────────────────────────────────┤
│  EDT (entropy→temperature)  │  LEAD (latent↔discrete)       │
│  EPR (black-box detection) │  LED (layer-wise explore)     │
│  CK-PLUG (retrieval μ_ret) │  Attention Divergence probe   │
└──────────────────────────────────────────────────────────────┘

Key Files

File	Purpose
twave/landau_ginzburg_tracker_v2.py	TWAVE v2.0 — token-level phase control
nexus_os/chimera_router_v2.py	ChimeraRouter v2 — tiered routing + ERNIE hooks
nexus_os/demo_e2e_v2.py	End-to-end unified demo
README_v2.md	This file

Quick Start

# ChimeraRouter v2 demo
python nexus_os/chimera_router_v2.py

# TWAVE v2 demo
python twave/landau_ginzburg_tracker_v2.py

# Full E2E demo
python nexus_os/demo_e2e_v2.py --prompt "Explain quantum entanglement step by step." --category R2.2 --quality 0.85

ERNIE Integration

ERNIE is your external agent that provides suggestions via callback:

from nexus_os.chimera_router_v2 import ERNIEInterface, ERNIESuggestion, TemperaturePolicy

def ernie_callback(prompt, analysis):
    if "quantum" in prompt.lower():
        return ERNIESuggestion(
            suggested_policy=TemperaturePolicy.EAD,
            confidence=0.85,
            reasoning="Quantum prompts need annealed exploration",
            override_router=False,
        )
    return None

router = ChimeraRouterV2(ernie_interface=ERNIEInterface(callback=ernie_callback))

ERNIE suggestions are blended with router logic (default 30% ERNIE, 70% router) unless override_router=True.

Temperature Policies

Policy	When	Reference
fixed	Default	Baseline
edt	White/black-box with logprobs	arXiv:2403.14541
ead	Multi-step reasoning	arXiv:2510.05251
lead	High-entropy uncertainty	arXiv:2603.13366
ernie	External agent suggestion	Your ERNIE pipeline

Research Sources

• LEAD (Latent↔Discrete mode switching): arXiv:2603.13366, CVPR 2026
• EDT (Entropy-based Dynamic Temperature): arXiv:2403.14541
• EPR (Entropy Production Rate, black-box): arXiv:2509.04492, ECIR 2026
• LED (Latent Exploration Decoding): arXiv:2602.01698, ICML 2026
• CK-PLUG (Retrieval chemical potential μ_ret): arXiv:2503.15888
• Attention Divergence: arXiv:2605.05025
• TECP (Token-Entropy Conformal Prediction): arXiv:2509.00461
• SEP (Semantic Entropy Probes): arXiv:2406.15927
• AutoDeco (Learned temperature heads): arXiv:2510.26697

License

MIT — developed for the NEXUS OS project.

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'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

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