A file existing on disk isn't proof that an AI coding agent completed a task. In plan-auditor, every step needs a behavioral check — a real run, exec, or pytest command — before it can be verified.

In data fitting, the unglamorous choices matter: what happens to missing rows, what counts as an outlier, and which model family you allow. I built guided-fit in scientific-computing-system-2.0 so those decisions don't stay hidden in a notebook.

A model can sound more certain just because it keeps repeating the same guess. Quantum Reasoning Skill is an experiment in keeping genuinely different hypotheses alive until evidence can separate them.

An assistant misses your request. You correct it. It apologizes — then makes the same mistake again. That's not just a tone problem. It's a repair and control problem, and it's one reason I'm working on EQ-Layer.

An AI coding agent can suggest its own tests. But should it get to decide what “done” means? I built plan-auditor around a stricter boundary: the user/host owns the accepted requirements and seals the verification contract.

One easy way to fool yourself in scientific modeling: get a plausible-looking answer from an equation with inconsistent units. Axiomize makes units part of the model, not an afterthought.

A saved model fit can run again and still be different. That’s why I built replay checks into the NumPy/SciPy version of scientific-computing-system. Same data and seed are important, but so is the software environment.

plan-auditor exists because an AI coding agent saying “done” is not evidence. It verifies work in a separate process with its own checks, so the implementing agent’s narration carries no weight.

EQ-Layer is my attempt to make “empathetic AI” less like a tone prompt. Instead of telling the model to sound caring, it separates affect, intent, repair state and response policy before generation.

quantum-reasoning-skill is an Agent Skill for one problem I keep seeing: models commit to the first plausible answer too early. It keeps multiple genuinely different hypotheses alive instead.

scientific-computing-system is my from-scratch pure-Python take on scientific computing: readable algorithms, a zero-dependency core, and no C/Fortran runtime hidden underneath.

I built Axiomize because scientific models shouldn’t die inside a notebook. It turns a vague idea into an explicit, versioned mathematical model with assumptions, units and provenance that another person can actually rerun.

AI coding agents are very good at saying “done.” I wanted a stricter answer: show me the evidence. So I built plan-auditor — an independent, LLM-free verification layer that checks agent work with real commands instead of trusting the agent’s own report.

New here. I mostly build scientific software and tools around AI reasoning and verification. Scientific computing, reproducible modeling, agent verification, and a few experiments on how LLMs reason and respond. github.com/Furox-Art