Method

How a marketscout brief is made, what is checked, and what is not.

What it is

marketscout is a research tool. For each subject on its watchlist (a coin, a stock, gold or silver) it collects data and headlines from public sources, asks an AI model to write a short brief from that material only, and then checks the brief before it is published. Every claim in a brief cites numbered sources, and the site shows the source text beside each claim.

It does not give investment advice, make forecasts of its own, or take any action in a market.

The steps

Collect evidence. Prices and price changes come from public data feeds. News is read as headlines only, from news search feeds. Each item is numbered (E1, E2, and so on) and stamped with the time it was retrieved. If a source that a brief depends on is unavailable, no brief is published for that subject.
Draft. One call to a language model turns the numbered evidence into a headline, a summary and a list of claims, each citing the evidence it rests on. The model is told to use only the evidence, to attribute news to its publisher, to keep forecasts as forecasts, and to say when evidence is thin.
Citation check. Rules, not a model: a claim with no citation, a citation to evidence that does not exist, or a repeat of an earlier claim is removed.
Verification. A second model call reads each remaining claim next to only the evidence it cites and labels it supported, partly supported or unsupported. Unsupported claims are removed. Partly supported claims are cut back to the part the evidence states, or removed if the cut-back text is not a complete sentence.
Summary. The headline and summary shown to readers are written last, from the claims that survived, so they cannot mention anything that was removed.
Record. Each run saves a run ledger (which models answered, tokens used, time, claims proposed, kept, trimmed and removed) and a claim trail (for every proposed claim, what each check decided and why, with the source text it was checked against).

What comes from data, not from the model

Prices, percentage changes and the position of a price within its recent range come straight from the data feeds or are calculated in code from them, and are quoted as given. Gold and silver are shown as futures prices and converted into other currencies in code using European Central Bank reference rates. These conversions are reference prices, not local retail prices: duties, taxes and dealer premiums are excluded.

How it is measured

A fixed set of hand-labelled test cases checks the citation rules, a rule-based number check (every price, percentage or scaled figure in a claim must appear in the evidence it cites), and the verifier (does it accept good claims and catch flawed ones). The same checks are run on the published briefs. Results are on the model card.

Limits you should know about

News claims rest on headlines, not full articles. The model and the verifier see only the headline and any short description.
The verifier is a language model and can be wrong. At present the drafting and verification calls normally use the same model, so they can share blind spots.
The number check is a measurement, not a gate: it is recorded in the claim trail and the test results, but it does not by itself stop a claim from being published.
Data feeds can be delayed or briefly wrong. Each source item shows when it was retrieved.
The watchlist is small and the coverage is English-language and mostly US-centric.
This is a research summary from public sources and is not investment advice.

Questions or corrections

If a brief states something its cited source does not support, that is a defect. The risk register lists what is known and what is being worked on.