Artificial Intelligence and Safety Division
Reasoning models & alignment.
Independent AI research by Jayden Xing: adaptive reasoning language models at Kreflux, and verifiable defenses against jailbreaks, extraction, and abliteration at Freezesol.
kreflux.com / Live platform and lab
KrefluxStay curious. Reasoning at your pace.
An independent AI lab building reasoning language models and tools with user-controlled thinking effort. From the browser chat to developer terminal workflows and API pipelines, Kreflux gives users real-time control over how deep the model reflects on complex problems.
Reasoning at scale
Intelligence should not be a black box with arbitrary compute constraints. Users deserve transparent, adjustable effort levels and developer tools that seamlessly bridge the web, the terminal, and cloud APIs.
Adjustable reasoning effort
- Fast for quick answers that do not need to wander.2 levels
- Balanced for everyday reasoning.3 levels
- Deep for problems worth thinking all the way down.5 levels
- Auto and the model decides how far to reflect.model decides
klucidKlucid CLI
A high-performance command-line autonomous agent for developers, directly integrated with Kreflux models and terminal workflows.
kreflux apiKreflux API
Programmatic access to Kreflux models for products and pipelines, with the same adjustable reasoning effort as the chat.
git clone github.com/Kreflux/openkrefluxOpenKreflux
The open side of the lab: models and tooling published on GitHub for anyone to inspect and build on.
freezesol.com / Alignment division
A sun, held in stillness. Permissions are enforcement.
The alignment and safety division of Kreflux. Because text defenses are merely advisory and prompts can be bypassed, Freezesol builds verifiable permission barriers, extraction defenses, and safeguards engineered to survive intentional model abliteration.
Enforcement, not polite advice
Prompt instructions and semantic rules alone cannot prevent sophisticated jailbreaks or weight abliteration. Safeguards must be structural, permission-based, and mathematically resilient under adversarial conditions.
Krextract
DefenseScreens extraction attempts and gives applications an explicit, inspectable boundary between private model reasoning traces and public answers.
npm install github:Kreflux/KrextractKreGuard
LayerOpen guardrail layer for LLM applications. Screens inputs, outputs, and egress for jailbreaks and prompt injection with zero runtime dependencies.
pip install git+https://github.com/Kreflux/KreGuardRSAAJ
ResearchRobust Safeguards Against Abliteration and Jailbreaks. Training models to preserve core refusal behaviors even when refusal activation directions are subtracted.
Research contact