Skip to main content
Mubienclarity for what comes next
Research

Questions I keep coming back to.

How do these systems learn? What changes when they act on our behalf? And how do we know when our checks are good enough? This is where I explore those questions in more depth.

Building gives me practical questions. Reading helps me put them in context. These pieces bring the two together: explainers to untangle the technology, and frameworks to think through its governance and limitations.

This is independent, personal work based on public sources. The frameworks are proposals for discussion, with assumptions and limits to examine. I want the reasoning to be clear enough that you can follow it, challenge it, and decide what is useful.

Flagship · Interactive framework6 controls · 7 stages

Agents that spend

When an AI agent buys something, how does permission survive the whole journey from a human request to money moving?

A six-control framework for recording delegated authority, checking the final purchase, preserving evidence, and making stop mean stop.

Includes

An interactive permission test, system architecture, control matrix, adversarial cases, launch gates, and primary sources.

Test a purchase →
ExplainerWhen AI learns: what actually changes?Concepts & open questions

Does a better second answer mean the model learned? And when does an improvement loop become recursive?

A practical distinction between context, memory, training, and self-improvement. What each changes, where the limits are, and the checks I would want around a system that can change itself.

Read the explainer →
Interactive frameworkThe Oversight Threshold5 levels · checked 20 September 2026

At what point does human oversight stop being meaningful, because the system has become too capable, fast or autonomous for a person to verify?

A conceptual framework for asking when a system becomes too fast, complex, or autonomous for its human review process. Five levels of autonomy mapped against eight governance dimensions, plus a test for whether human review is still doing real work or has quietly become a signature.

Open the framework →
EssayThe Verification GapPDF

Can independent assurance actually make AI trustworthy to the institutions now betting on it?

AI auditing sits roughly where financial control attestation stood in 2003, and four unsolved problems decide whether it becomes a real discipline or a rubber stamp: what you measure against, what counts as evidence, who is qualified to judge, and whether the auditor is genuinely independent. Built on OSFI's model risk guideline, the EU AI Act and ISO 42001 rather than on opinion.

Read the essay →

Things I have built are under Projects, and things worth reading by other people are in the Library.