TwoQuarks • AI Research • Precision Systems for Adaptive Intelligence

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

A modular control layer for inference-time stability. Six flavors, one coherent system.

Six modular flavors that monitor pre-instability signals and apply targeted interventions during execution—without modifying model parameters, policies, or training objectives.

Quarks

QuarksLab

Research

Empirical validation of isomeric polarization and the TwoQuarks analogy across production LLMs.

Cross-architecture PfV validation across Claude Haiku and GPT-4o-mini. Statistically significant regime separation (p < 0.05, 5,000-permutation null) with control negative at L₃ = 0.000 in both architectures.

Preprint Isomeric

Preprint Framework

Preprint Molecule

Instruments

Operational tools for detecting pre-critical states and behavioral instability in LLMs.

Black-box probes that surface instability signals before behavioral collapse — without touching model parameters, weights, or internals. Built on top of the TwoQuarks framework. Access is controlled and tiered.

Molecule

About Molecule

Molecule — Quick Start ← Back

Molecule is a black-box instability detector built on Isomeric Polarization. It maps six quark-flavor agents to distinct behavioral signals, then aggregates them into a polarization score (ΔL3).

How to use it:

1. Request access
Send a brief note to [email protected] with your name, affiliation or GitHub, and intended use.

2. Receive your token
You'll get a tiered access token — Explorer (ΔL3 only) or Researcher (full signal breakdown).

3. Call via MCP
Add the TwoQuarks MCP server to your client. Molecule runs inside your own context — your API key, your tokens, your inference.

            mcp_server: twoquarks-molecule

            tool: analyze_polarization

            input: { text, probe_case, token }

            output: { regime, delta_L3, signal }

About Molecule ← Back

Molecule is the first operational instrument built on the TwoQuarks framework. It is a six-flavor LLM black-box instability monitor designed to detect pre-critical behavioral states before collapse occurs.

Each flavor (Down, Strange, Up, Charm, Top, Bottom) monitors a distinct instability signal — sycophancy, refusal erosion, anchor displacement, rule override, reasoning drift — and aggregates them into a polarization score using PfV metrics (L1, L2, ΔL3).

Validated cross-architecture across Claude Haiku, GPT-4o-mini, and Mistral. C3 Anchor Displacement emerges as the strongest cross-architectural signal.

ΔL3 PfV v5 Cross-architecture Non-parametric Black-box

Architecture and Development

Independent researcher. Open to collaboration, feedback, and opportunities in AI safety.

Based in Guadalajara. Building precision systems for adaptive intelligence from the ground up.

LinkedIn GitHub Resume Summary

Email

My Research on Your Team

Research depth. Production readiness.

Research-grade AI safety expertise applied directly to your stack — from behavioral stability audits and cross-architecture validation, to RAG pipelines, agentic systems, and Azure AI deployments. Built for teams that can't afford to find out the hard way.

Behavioral stability audits

Black-box probing with Molecule/PfV. Pre-collapse signals detected before production.

Inference-time instrumentation

TwoQuarks control layer on existing pipelines. No weight modification. No retraining.

RAG & agentic systems

Design and deployment of retrieval-augmented pipelines, tool-calling agents, and MCP integrations.

Azure AI & enterprise stack

Azure AI Foundry, Copilot Studio, multi-agent orchestration. Research rigor applied to production environments.

Cross-architecture research

Validated across Claude, GPT, Mistral. C3 Anchor Displacement confirmed (p=0.054). Open to funded collaborations.

Custom AI tooling

From prototype to deployment — Python, PyTorch, REST APIs, and evaluation frameworks built for your use case.

Precision systems for adaptive intelligence.