Pruning the garden

The little label at the top of this site’s home page says secret garden. Gardens need pruning, and I’ve let this one grow in every direction at once. So today I’m retiring three projects in one go: K.E.V.I.N., Sovereign Shield and Buildmancer. Two of them I’m rather proud of. One of them nobody used. All three were taking up room that I’d rather give to fewer things, done properly.

Here is what each one was, what happens to it now, and the honest reason.

K.E.V.I.N.

K.E.V.I.N. was an autonomous red-teaming framework: an attacker model tries to extract a planted secret from a target model, a panel of judge models from different vendors scores every attempt, and the loop refines its tactics across rounds. I ran it across thirteen research-grounded attack vectors and five model families, and it produced the finding I still find most interesting from this year: frontier models have opposite weak spots. Hold the persona and the secret constant, swap only the model, and one leaks under structured-output coercion while the other leaks under token prefilling.

That finding had a practical consequence. If the weak spots move every time a vendor ships a new version, the durable defence can’t live inside the model. It has to sit at the boundary, where a deterministic check can catch a leak regardless of how it was coaxed out. That idea became Sovereign Shield, which is the nicest possible way for a research project to end: it answered its question and the answer turned into something.

What was left after that was benchmark maintenance, and the field moves faster than one person’s evenings. So the framework is archived. The four findings posts stay up exactly as written, in the new archive, because results that were true on the day I measured them don’t stop being interesting when I stop measuring.

Sovereign Shield

Sovereign Shield replaces Swiss, EU and international identifiers, API keys and your own custom terms with placeholders before a prompt reaches ChatGPT, Gemini or Claude, then puts the real values back in the reply. It’s deterministic, it runs entirely in your browser, and it cannot be talked out of a match. It shipped as a Chrome extension and as a small Python proxy.

This one is being frozen, not deleted, and I want to be precise about what that means, because it’s a privacy tool and vague reassurance is the wrong register.

• The extension stays on the Chrome Web Store and keeps working exactly as it does today.
• shield.ars.md stays online, because the extension links to it for its privacy policy and its preset library.
• The source stays on GitHub, archived and read-only, under the same Apache licence, for anyone who wants to fork it.
• There will be no further updates of any kind. No features, no bug fixes, and no security fixes. If a future Chrome or provider change breaks it, it will stay broken.

Please weigh that last point before you rely on it for anything that matters.

Why stop? It did what I built it to prove. But a browser extension that sits between people and their AI assistants deserves a maintainer with more hours than I have, and pretending otherwise would be worse than saying so plainly.

Buildmancer

Buildmancer was an AI building companion for Minecraft. You’d type a request in chat and it would design a build, review its own plan, place it in your world and learn from the result. The brain behind it ended up with a library of several hundred learned structures and styles. It was a genuinely fun thing to make, and the honest verdict is that nobody really used it.

So it’s gone, cleanly: the hosted service is shut down, the website is gone, and the mod and plugin downloads have been removed, because a mod whose only remaining behaviour is failing to reach a server that no longer exists is worse than no mod. Nobody is owed anything.

I don’t regret building it. I regret keeping it on the home page for months after it was clear.

What stays

• 1-bit, the bit-exact music player for iPhone, and its bridge. This is the project I actually care about, and it now has its release notes and hands-on guides on its own site at 1-bit.app/blog, where they belong, instead of clogging the writing here.
• The Pitch, because a site on a domain I bought for the Swiss job market should probably do that job.
• This site, as a place for occasional writing. Fewer posts, on purpose. The About page now carries a “now” section, which is a better signal that someone is home than a stream of changelogs ever was.

If you subscribed to the feed for the 1-bit release notes, 1-bit.app has its own feed now. If you subscribed for the essays, nothing changes, except that there will be less noise between them.