Grounding is not a prompt trick, it is an architecture. Here is how a deterministic engine and a tightly constrained narrator make an AI investing platform you can audit.
Most tools that call themselves an AI investing platform are a language model behind a prompt. You ask about your portfolio, the model obliges and the answer sounds confident. The trouble is you cannot check it. Ask twice and you get two different replies. We built Dayonik the other way round.
The engine decides, the model narrates
A deterministic engine works out the findings first. It measures concentration, volatility, drawdown and the distance from the efficient frontier. Only then does a tightly constrained narrator turn those numbers into plain language. The model never gets to choose what is true. It just says, in readable prose, what the analytics have already settled.
That is why the advisor cannot invent a ticker. There is no path from the narrator back to the data, so it has nothing to hallucinate with. Every sentence points at a number you can open and inspect for yourself.
What grounding buys you
- Every finding traces to a computed value, so you can audit the reasoning rather than trust it.
- The same book produces the same advice, run after run, because the maths is deterministic.
- A compliance team can approve the tool, because the numbers are the record.
Grounding is not a feature you bolt on at the end. It is the shape of the whole system. And it is the reason this AI investing platform can show its working on demand.