A sunlit survey boundary dividing open ground from a growing field, with a small research station in the distance.

Independent research lab

Explore failureEnjoy progress

Failure reveals the system.

Explore Failure builds public instruments that change, interrupt, and inspect the harnesses around AI models—revealing what breaks, what survives, and what remains trustworthy.

Explore the research Follow the work

One-person research lab / Est. 2026

Don’t just test the model. Test the harness.

The model is only one part of the system. Instructions, tools, context, state, and verification form the harness around it.

That harness needs its own tests. Those tests introduce controlled changes, preserve what happened, and measure whether the surrounding system remained trustworthy.

Released
and underway.

Explore Failure publishes working instruments, then follows the questions they expose.
Harness Workbench is public now. Two active research paths show where the work is moving next.

Available now
Public instrument / Release 001 v0.1.0-rc.1 / Public prerelease

Harness
Workbench

Test the harness, not the answer.

Harness Workbench introduces controlled changes, attaches features, preserves every attempt, and measures whether the harness around an AI coding agent remained trustworthy.

View the repository
Release
v0.1.0-rc.1
Verification at release
326 automated tests
Runtime footprint
Zero dependencies

Working on

Active research / Unreleased

Two open questions follow from the public instrument.

These are research directions, not finished products. Harness Workbench supplies records and failure cases that help make each question more concrete.
Path 01 / Evidence continuityUnreleased

Portable proof

How can evidence, uncertainty, and provenance travel with the claim they support?

Workbench informsRun records and verification tests help define a proof package that preserves sources, reasoning boundaries, and unresolved support.

ObjectiveThe claim can travel without pretending its uncertainty disappeared.

Path 02 / Environment continuityUnreleased

The downloadable research lab

What must travel for a research environment to remain runnable, inspectable, and extendable?

Workbench informsPortable specifications, attached source, run storage, and replay help identify what a downloadable environment must carry.

ObjectiveThe research can be reopened as working practice, not merely read as a report.

GarrettDavis

Independent researcher
and builder

I study the harnesses around AI systems—the instructions, tools, context, state, and verification that turn a model into a working system.

I build public instruments, run controlled tests, document failures, and publish the evidence so others can inspect and extend the work.

There is always more to question.