Precision systems for adaptive intelligence.

Is inference-time stability regulation sufficient to prevent collapse and unsafe behavior in sequence models under regime shift?

About TwoQuarks

TwoQuarks defines model collapse as a failure of inference-time stability rather than capacity. It introduces a modular control layer that monitors internal pre-instability signals and applies targeted interventions during execution—without modifying model parameters, policies, or training objectives.

Framework

Adaptative agents inferenced under pre-critical states.
The framework consists of a set of independent control modules that operate at the same structural level and are conditionally activated based on observable instability signals, including entropy fluctuations, variance in predictive confidence, policy leakage, and temporal drift.
Quarks
QuarksLab

Research

The "Isomeric Polarization" architecture and "TwoQuarks" framework analogy that explore AI Safety.
In sequential environments, instability often manifests not as immediate failure, but as a gradual erosion of internal consistency. To address instability arising during execution, we introduce a modular stability control layer composed of multiple independent control mechanisms. Each mechanism monitors specific instability signals and applies transient, non-parametric interventions without modifying the base model’s parameters, training objective, or data. Cross-architecture is validated in the LLMs: Claude API Haiku and API GPT-4o-mini.
Preprint Isomeric
Preprint Framework

Architecture and Development

Independent researcher. Open to collaboration, feedback, and proposals in AI safety.
Development precision systems for adaptive intelligence from the ground up.
LinkedIn
GitHub
Resume
Summary
Email
GitHub LinkedIn Resume Preprint Framework Preprint Isomeric
© 2026 TwoQuarks • Q-research, AI-Safety, precision, sobriety, intent.