Dual-Channel Data: Why the AI Never Writes the Numbers

Language models are good at explaining and bad at arithmetic under pressure. Ask one for a company's debt-to-equity ratio and it will produce a plausible figure, sometimes the right one, sometimes a number from a different year or a different company. For a stock research tool, "sometimes" is not an acceptable failure rate. The design answer is to keep two channels apart.

Channel one: the figures

Prices, filings, ratios and the score itself come from market data providers and from deterministic code. They are rendered verbatim on the page. No model touches them, rewrites them or rounds them. If a figure is wrong, the cause is a data-provider error or a code bug, both of which can be traced and fixed.

Channel two: the commentary

The model is allowed to write text around the figures: what a low valuation percentile means for a company in that sector, why a momentum spike without a fundamental change deserves caution, which pillar is doing the work in a total. It reads the numbers as inputs. It is never the source of one.

Why the boundary matters for the reader

The two channels fail differently. A wrong figure looks exactly like a right one and is dangerous in proportion to how confident the page looks. Weak commentary is merely unhelpful, and a reader can ignore it. Keeping the channels separate means the dangerous failure mode is removed from the model entirely, and the remaining failure mode is the survivable one.

How to check whether a tool does this

Ask three questions. Does every figure on the page carry a source and a date? If the commentary mentions a number, does the same number appear in a table or a field that came from data, or only in the prose? When the tool is asked for a metric it does not have, does it say so, or does it produce one anyway? A tool that passes all three is keeping the channels apart, whatever its marketing says.

Stock Expert AI is built on this rule: figures are rendered from licensed data, the model writes only commentary, and every MoonshotScore carries its calculation date. The full design is described on the methodology page.

This is educational content, not investment advice. Past performance does not guarantee future results.