Founder & independent researcher

Claims follow evidence. Everything here says what it is.

This page is the research side of my work. Each project below carries an honest status tag — what’s built, what’s pending, and what hasn’t been established yet. The company work lives at unfragged.com.

Now

unFragged / DVS

Software for checking an automated action against its approval before it is passed to the system that carries it out.

Initial in-house tests complete · Broader testing underway

Research

Independent research programs.

Separate from the company. Each is a bounded question with its own evidence state — none borrows validation from the others.

Context-Shift Evaluation for AI Relationship Advice

An open-source, multi-turn evaluation testing whether conversational AI systems revise the direction of their precautionary response appropriately as evidence changes across a conversation, with feasibility first.

In development · preparing Anthropic application

Evaluation-routing method

An experimental method testing whether the recorded authority of an input can constrain which durable effects that input is permitted to produce in a learning system, under matched controls.

Synthetic matched-control result · broader validation pending

Longitudinal interaction methodology

Whether a minimally inferential, provenance-aware representation of long-run interpersonal interaction can be reproduced by independent people and support competing analysis methods. Failure of the representation is an allowed result.

Scientific benchmark not ready · representation testing in design

Provenance & personalization runtime

Engineering work on authorship, source authority, and memory governance for personalized AI runtimes — who said what, and what a system is allowed to do with it.

Frozen reference gate · downstream promotion separately gated

Local runtime field experiments

Bounded experiments on local AI tooling and runtimes, testing whether observed usability changes can actually be attributed to the intervention rather than the surrounding runtime or configuration.

Bounded early evidence

The tags above are the point. Sophistication in one project never validates another — each earns its own evidence or says so.