Sovereign, not borrowed
Every "AI trading tool" you can buy in 2026 is a wrapper around GPT-4, Claude, or Gemini. The product is the prompt. The model is somebody else's. The pricing, the rate limits, the data-retention terms — somebody else's.
We took a different path. The Qovaryx Options Decoder is our own model, trained from scratch on options-specific data, with our own tokenizer, on weights that live in the user's app. There is no API key. There is no cloud round-trip. There is no provider that can change terms next quarter.
What "sovereign" actually means
Sovereign isn't a marketing word for us. It has three concrete requirements:
- Own the tokenizer. Vocabulary, special tokens, subword merges — all ours. No upstream changes to a Llama or Mistral tokenizer can break us next year.
- Own the weights. Trained from random init on data we collected and curated. No "we built on top of Llama-3" asterisk in the license.
- Own the inference path. Everything runs locally on the user's CPU. The user's trades never touch our infrastructure.
What it costs
Building this way is expensive. A frontier general-purpose LLM costs $10M+ to train; even a specialized one runs into six figures. So you have to be ruthlessly narrow. Our model doesn't write poetry. It doesn't summarize PDFs. It does one thing: read a stack of timeframes for a US-listed options-eligible ticker and return a calibrated BUY/SELL/HOLD with a conviction probability. That focus is what makes the budget work.
What it buys
Three things that wrapper-tools cannot offer at any price:
- Stability. Anthropic deprecates a model — your prompt-engineering breaks. We change weights — we control the regression test.
- Privacy. We don't see your trades because the model runs on your machine. We can't sell what we don't collect.
- Performance. A small specialized model on CPU is faster than a network round-trip to a giant general one. Sub-millisecond, no GPU.
The most powerful AI for your trading isn't the biggest. It's the one that's yours.
The next post explains what we mean when we say a signal is calibrated — and why that word is doing a lot more work than most "AI confidence" claims.