Sovereign, not borrowed
Most automated market tools you can buy are a wrapper around somebody else's general-purpose model. The product is the prompt. The model, the pricing, the rate limits and the data-retention terms all belong to someone else.
Ch. 01 — Engineering
The other path
We took the harder route. The model at the centre of Qovaryx is ours: built from scratch on market-specific data, with our own vocabulary, on weights that live inside the application on your machine. There is no API key. There is no round-trip. There is no provider positioned to change the terms next quarter and break the product from the outside.
What “sovereign” actually requires
It is not a marketing word here. It has three concrete requirements, and failing any one of them means you do not have it:
- Own the vocabulary. How text and market structure are broken into units for the model to read — all ours. An upstream change to somebody else's vocabulary cannot break us next year, because we are not downstream of one.
- Own the weights. Trained from a random start on data we collected and curated. No “built on top of” asterisk buried in a licence, and no clause that can be revised after we have shipped.
- Own the path. Everything runs locally on your processor. Your activity never touches our infrastructure — not as a promise about what we choose to do with it, but as a fact about where it goes.
What it costs
Building this way is genuinely expensive. Training a frontier general-purpose model costs an amount only a handful of organisations can spend; even a narrow specialised one is a serious capital commitment. So you have to be ruthlessly narrow to make the budget work. Our model does not write poetry. It does not summarise documents. It does one thing: read a stack of timeframes for a US-listed, options-eligible symbol and return a direction with a calibrated confidence attached. That refusal to be general is precisely what makes it affordable to own.
Generality is the expensive part. We did not buy any of it, because none of it was going to read a chart better.
On the narrowness that makes ownership possibleWhat it buys
Three things a wrapper cannot offer at any price:
- Stability. When a provider retires a model, prompt-based products break and their authors find out from their users. When we change our own weights, we control the regression testing and we choose the moment.
- Privacy. We do not see your activity because the model runs on your machine. We cannot sell, leak or be compelled to produce what we never collect.
- Responsiveness. A small specialised model on a local processor answers faster than a network round-trip to a large general one — and it does not need a graphics card to do it.
We publish no performance figures. This page is about ownership and where computation happens. Any return, win-rate or accuracy number attributed to Qovaryx did not originate from us.
The most useful model for your decisions is not the biggest one. It is the one that is actually yours.
On ownership as a featureCompanion piece: what we mean when we call a confidence calibrated — a word doing far more work than most confidence claims admit.