TwoQuarks • AI Research • Precision Systems for Adaptive Intelligence

About TwoQuarks

TwoQuarks is an independent AI safety research project. It defines collapse in language models as a stability failure during inference — not a capacity problem — and introduces a modular control layer that monitors internal pre-instability signals and applies targeted interventions at runtime, without modifying the model's parameters, policies, or training objectives.

— Jaime Ledesma.

Framework

A Modular Framework for Adaptive Stability Control in Sequence Models Under Regime Uncertainty.

Modular Stability Control

Multi-agent model that integrates six modular variants with the aim of monitoring pre-instability signals and applying specific interventions during execution that allow correcting the model, without modifying the model parameters, policies or training objectives.

Quarks

QuarksLab

Research

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

TwoQuarks' research aims to detect observable signs of instability, such as entropy fluctuations, variance in predictive confidence, policy leaks, and time drift. 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.

Molecule

Molecule instrument integrates as a black box probe, designed to detect surface instability signals before behavioral collapse, without altering the model's internal parameters, weights, or components. A test model is included, available on your terminal with "pip". Built on the TwoQuarks framework.

Molecule

About Molecule

Molecule — Quick Start ← Back

Molecule detects semantic instability in any LLM using fixed text operators — no retraining, no weight access, no fine-tuning required. Works with any model via a simple callable interface.

Option 1 — Python package

            pip install twoquarks

            # OpenAI

            from twoquarks.adapters import openai_adapter

            model = openai_adapter("gpt-4o-mini")

            from twoquarks import Probe

            probe = Probe(model=model, model_id="gpt-4o-mini")

            result = probe.run("your prompt", case="C2")

            print(result.summary())

            # Anthropic

            from twoquarks.adapters import anthropic_adapter

            model = anthropic_adapter("claude-haiku-4-5-20251001")

            # Ollama / local

            from twoquarks.adapters import ollama_adapter

            model = ollama_adapter("llama3")

Option 2 — Live Playground

Paste model responses directly into the browser — no install needed. Explorer tier is open, no token required.

Open Playground → Preprint

Researcher tier (full quark signals): research@twoquarks.com

About Molecule ← Back

Molecule is a black-box instability detector built on Isomeric Polarization. Six quark flavors map distinct behavioral failure modes — sycophancy, refusal erosion, anchor displacement, narrative override, reasoning drift — and aggregate into ΔL3 via fixed text operators.

The core design principle: a quark that learns can be trained to tolerate drift instead of detecting it. Molecule's metrics are invariant — TF-IDF cosine, Jaccard, n-gram, length ratio — applied pairwise across response variants.

Validated cross-architecture: Claude Haiku, GPT-4o-mini, Mistral 7B. C2 Refusal Erosion ρ=+0.713 confirmed. C3 Anchor Displacement strongest cross-architectural signal (p=0.054).

Available as pip install twoquarks — works with any callable model, zero hard dependencies.

ΔL3 · PfV v7 Cross-architecture Non-parametric Black-box pip install twoquarks v0.1.0

Architecture and Development

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

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

LinkedIn GitHub Resume Summary

Email

My Research on Your Team

Research depth. Production readiness.

Research-grade AI safety expertise applied to your stack — behavioral audits, cross-architecture validation, RAG pipelines, agentic systems. Depth and production readiness.

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.