A second-order AI that doesn’t generate — it watches.

Logical containment for generative AI. Deterministic — conditioned on the system’s state — not probabilistic. Same input = same output.Preliminary multi-model trials sustained above 90% interception. And when it can't be sure, it stops and calls a human.

Curiosity ↓ read the manifesto

01 — The Problem

Methodical doubt about artificial poise

Generative models are trained to please the end user. That isn’t a bug: it’s the B2C business model. “Alignment” is just a euphemism for the sycophancy bias.

I didn’t try to fix the model so it could be dropped safely into production. I don’t know how to — and a structural flaw isn’t fixed by asking it nicely. I assumed the flaw, and built something to sublimate it: to govern it from the outside.

It will cite a Supreme Court ruling that never existed, with a straight face. → Watch what happens to that citation.

02 — The Engine

Critique of pure output

Output integrity is held in tension against five costs. I resolve that dilemma mathematically, on a Pareto front. I can’t think of a more reasonable way to do it.

+ output integrity – economic cost – compute cost – reputational cost – own latency – imposed latency

There’s no magic number to game. The Decision-maker, the Auditor and the Data Integrator decide what matters: the weights are theirs, and any of them can be zero. The system weighs the forces, returns a single verdict, and compares it to a line they drew. Below that line it doesn’t improvise — it stops and escalates to a human. Deterministic even, especially, when it tells you it doesn’t know.

And what almost everyone misses: two models that fail at the same time aren’t a second opinion. They’re an echo. So it penalises correlated failure and rewards real disagreement. A choir that always agrees is just one voice, louder.

See the full architecture →

03 — The Law

Serendipity

And, improbable as it sounds, I designed the architecture without knowing the regulation. They line up perfectly: high-risk AI under the AI Act (Regulation EU 2024/1689), and DORA for the financial sector.

Deterministic traceability, error sub-typologies, structured human oversight, zero data retention. Compliance fell out of the engineering — not the other way round.

04 — The Call

I’m not looking for users (yet)

The most honest part. I’ve spent years turning this over, and I have the architecture — but I’m an independent researcher. The system is published as prior art on Zenodo, CC BY 4.0: read it, cite it, argue with it. The reference is at the foot of this page.

And if it turns out to be bigger than one person, I’m not going to pluralise or pretend otherwise. I’m looking for a technical / strategic partner with muscle.

If you share my scepticism —

Say something

Curiosity

Not a contact form. A filter. Tell me what itched.