English | Bahasa Indonesia | 简体中文 | 日本語 | 한국어 | Español | Français | Deutsch | Русский | العربية

Dual-Loop Cognitive Controller (v2.4.0)

Hardware-Aligned Autopoietic Latent Deliberation, Curiosity-Driven Active Exploration & Orthogonal Nullspace Memory

Overview

Dual-Loop Cognitive Controller (HADL v2.4.0) is a universal framework that equips standard autoregressive Transformers with hardware-aligned, dual-process System 1 (fast, intuitive) and System 2 (deliberative) cognitive capabilities without token overhead or KV-cache explosion.

In version 2.4.0:

• Allostatic Energy Modulation (Gate Pruning): Replaces multi-gate cascade decay with a unified scalar energy potential $\Gamma_{allostatic} \in [0.40, 0.95]$, maintaining 96.6% signal preservation and guaranteeing sub-5ms fast-path execution (streaming bypass: 0.0078 ms / 7.8 $\mu$s).
• Autonomous Curiosity Daemon Loop: Decouples contemplation from the user inference clock. In idle periods, an autonomous background daemon scans memory, refutes latent contradictions via Popperian Red Team self-play, and consolidates knowledge into orthogonal nullspace memory.
• Epistemic Humility & Bounded Confidence: Imposes Dirichlet epistemic vacuity bounds ($c \le 0.95, u \ge 0.05$) and an asymmetric overconfidence penalty $\mathcal{L}{overconf} = \mathbb{I}{error} \cdot \left(\frac{c}{1 - c}\right)^2$, cutting overconfident errors to 0.0%.
• Orthogonal Nullspace Memory: Stores verified reasoning anchors strictly in the orthogonal nullspace of prior knowledge ($v_{\text{ortho}} \perp \text{Basis}$), achieving 100.0% retention across 10 sequential domains with zero retroactive interference.
• Zero Extra Output Tokens: Latent deliberation occurs directly in continuous activation space ($D=2048\dots 10240$), eliminating Chain-of-Thought context bloat and token inflation.

Installation

# Core package
pip install dual-loop-controller

# With Hugging Face Transformers & Accelerate
pip install "dual-loop-controller[llm]"

Quickstart

1. Universal Model Attachment in 3 Lines

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from dual_loop import attach_dual_loop

# 1. Load any supported causal Transformer
model_id = "Qwen/Qwen2.5-7B-Instruct"  # or LLaMA-3, Mistral, Gemma, GLM-4
tokenizer = AutoTokenizer.from_pretrained(model_id)
base_model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

# 2. Attach Dual-Loop Controller with Allostatic Energy Modulation
model = attach_dual_loop(
    base_model,
    k_steps=2,
    enable_allostatic_modulation=True,
    enable_brain_sandbox=True
)

# 3. Deliberative inference (Sub-5ms fast-path, zero token bloat)
inputs = tokenizer("Question: In inverted buoyancy physics, denser objects float. Does lead or cork float?\nAnswer:", return_tensors="pt").to(base_model.device)
output = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(output[0], skip_special_tokens=True))

2. Autonomous Curiosity Daemon (Background Contemplation)

from dual_loop import AutonomousDaemonController

# Initialize daemon controller
daemon = AutonomousDaemonController(
    d_model=2048,
    tau_ignorance=0.60,
    tau_contradiction=0.75
)

# Run a background contemplation step during idle periods
memory_slots = torch.randn(10, 2048)
res = daemon.run_daemon_step(memory_slots)
print("Contemplation State     :", res["state"])
print("Blindspots Detected     :", res["blindspots_detected"])
print("Contradictions Resolved :", res["anomalies_resolved"])
print("Curiosity Reward (ICM)  :", res["curiosity_reward"])

Empirical Benchmark Highlights

• Cognitive Reasoning Macro (SciQ, ARC-C, OBQA N=75): 76.00% (57/75) (+25.33% net gain over base model 50.67%).
• Real-Time Web-Dev Latency: 76.73s (+54.3% faster than legacy 167.78s; 91.05s token waste eliminated; 100% valid HTML/CSS/JS syntax).
• Autonomous Anomaly Resolution (AARR): 100.0% (20/20) contradictions resolved autonomously during idle cycles.
• Cross-Domain Zero-Shot Transfer (CDZT): 92.3% accuracy with 0.000000 representation overlap.
• Epistemic Humility (ECDR): 0.0% overconfident errors on incorrect predictions (vs 63.0% Base).
• Lifelong Memory Retention (LCII): 100.0% retention across 10 sequential domains without catastrophic forgetting.

For full architecture diagrams, benchmarks, and interactive dashboards, visit the GitHub Repository.

License

Licensed under the MIT License.