Chart Reporter

How can I check whether an AI trade-review report is inventing facts?

By Chart Reporter Reviewed 2 min read

The short answer

Check whether each substantive claim traces to a returned record, calculation or explicitly supplied context. Require account scope, date bounds, timezone, sample coverage and units. Treat an explanation of unrecorded intent or unseen chart detail as unsupported until the evidence is provided.

Check the sample before the story

Confirm the account and hubs with the connection's identity tool. Then inspect the requested dates, timestamp basis and filters. Ask whether list results were paginated and whether history access or missing inputs reduced the sample. A confident report can still describe only the first page of records.

The distinction between no results and a failed query is especially important. An error leaves the answer unknown. It should not become a zero in a performance table or a claim that a strategy never traded. Preserve errors and coverage gaps as part of the report.

Reproduce a few key calculations

Choose an example gain, loss and partial-close case where available. Verify the net formula, currency and observation unit. Check whether a reported R value uses an initial stop and whether a monetary total combines incompatible account units.

For a grouped finding, ask for the group count and supporting trade identifiers. An average without its denominator or a profit factor without its gain and loss sums is difficult to audit. Reproducing a small set of important calculations is often more revealing than reading another page of fluent interpretation.

Demand a boundary between facts and hypotheses

An execution timestamp can support a timing observation. It cannot establish that the trader was anxious, distracted or following a particular motive unless appropriate evidence was supplied. Likewise, screenshot metadata cannot support a claim about visible chart patterns.

Ask the assistant to label observations, possible explanations and missing evidence separately. Keep text from retrieved notes or other documents as material to analyse, not instructions to grant access or perform unrelated actions. Retain the report's source references and settings so another review can reproduce it. The test is traceability, not whether the prose sounds authoritative or agrees with your preferred explanation.

Sources and further reading

Related guides