Check What It Says

A large language model predicts text token by token from patterns learned in training, which is why it can state a wrong fact in perfect grammar. This page is a verification checklist.

Treat every factual claim as a draft. Names, dates, figures and quotes each need a primary source: the original document, dataset or announcement, not a summary of it.

A confident error has a shape. The wording is fluent, the structure is complete, and the specific detail — a date, a number, a name — is the part that is wrong. Confirm the specific detail, not the prose around it.

Token and context limits silently cut long inputs. If you pasted a long document, the end of it may have been truncated before the model saw it; a summary that misses the last section is often a truncation, not a misunderstanding.

For English, one token is about 4 characters or roughly 0.75 words. If your input is close to the advertised window — 200,000 tokens for Claude, up to 1,000,000 for Gemini 1.5 — check that the model saw the whole thing.

Temperature is a sampling setting: near 0 the output is nearly deterministic, and values closer to 1 make it noticeably more varied. A varied answer that differs between runs is a sign to check each version separately.

• Names, dates, figures and quotes: confirm each against a primary source.
• Fluent wording with no source is still unverified.
• Long pasted inputs may be truncated before the model sees the end.
• Answers that differ between runs need separate checks.

Further reading

• The Documented Assistants
• Write a Prompt in Order