TradeAlmanac
Sign in
Documents

How we use language models

What prepares our material, who checks it, where the numbers come from, and what to do if you find an error.

Updated: 7 September 2026

Some of our material is prepared with the help of language models. We say so plainly, because readers make decisions on the numbers these pages carry and are entitled to know how those numbers got here. This document describes the working order rather than an intention: everything in it can be checked on any article page.

What a model does, and what it does not

A model writes TEXT: the wording, the structure, the explanation. It does not compute indicators, pick securities for lists, assign scores or decide what enters a ranking. All of that is done by code over stored data, and the result does not depend on which model wrote the prose beside it.

  • A model never sees prices and never recalls quotes: numbers are substituted into the text from our database on every page view.
  • A model gives no investment advice — that is forbidden at the level of the task, not left to its discretion.
  • A model does not decide whether material is published: publication is opened by a check that no setting and no access right can bypass.
  • A model never signs material: the author is never given as «AI».

Where the numbers come from

A factual number is never typed into a paragraph. In its place stands a reference to an indicator, resolved into a value when the page is served. The difference matters: a number written out in words freezes on the day it was written and, a quarter later, becomes a falsehood that looks like a fact. Our check rejects material with a number typed into prose — that is a condition of publication, not a recommendation.

Legal values — rates, limits and thresholds — live separately, in a register with an effective date and a link to the source: they change by legislative decision rather than by the market, and must change in one place. How each indicator is computed is described in Methodology.

Three ways material is prepared

Every article states how it was prepared. There are exactly three values, and no intermediate one on purpose: «partly reviewed» means not reviewed.

  • Model draftWritten by a language model from our data; no editor has read it. This is how daily-flow material is marked — summaries and routine breakdowns.
  • Checked by an editorWritten by a model and read by a person who accepted it. Review does not make the text human-written, so the byline stays editorial — but the promise to the reader is different, and we keep the two apart.
  • Written by a personWritten by an author. No model took part in preparing it.

For material published before we started recording this in the database, the model and its parameters are not stated. We do not fill those fields in retrospectively: an invented fact on a page that promises transparency is worse than an honest blank.

The method itself was reconstructed for that material, on the following basis. Daily-flow pieces are marked as model drafts — their author in the database is named as automation. The rest are marked as checked by an editor, because each of them passed the check that opens publication: without it, material does not reach the site at all. That is a reconstruction from the fact of passing the check, not a claim that every paragraph was reread — and we say so plainly, so the label can be read with the precision it actually has.

What the editors check

  • That the material answers the question it was opened for — a formally flawless text about something else misses the reader entirely.
  • That the reasoning over numbers holds: ratios are not confused with percentages, trailing periods are not mixed with calendar years, the conclusion follows from the figures given.
  • That there is no advice and no promise of returns, including by tone.
  • That every external fact carries its source.
  • That the material is not a repetition of a neighbouring page.

Responsibility for published material rests with our editors — not with the model and not with its provider. Automated preparation is never our excuse for an error.

If you find an inaccuracy

Every article carries a form for reporting an error. It accepts a report without registration and without a mandatory email address: we will verify the number against our own data, not against you. The form immediately prints a reference number in the form CR-2026-000001, so the report can be cited.

  • The review period is three days from receipt. It is recorded on the report itself, not only in this paragraph.
  • If the error is confirmed, we correct the material and add an entry to the corrections log, citing your reference number.
  • If it is not confirmed, the report is closed with an explanation; where an address was given, we reply to it.

A claim about RIGHTS to material or data is a different procedure with a different deadline: one day instead of three, as the law requires. It is described on the Rights holders page.

Corrections log

Substantive fixes are visible to readers. Each article lists its corrections: what exactly was wrong and when it was fixed. The site-wide log is on its own page — Corrections log.

  • CorrectionIt was wrong and is now right — a number, a date, a name or a conclusion.
  • ClarificationIt was right but read ambiguously. The meaning did not change; the wording did.
  • UpdateA fact changed in the world — a law, a rule, the composition of the data — and the text was brought into line. This is not an admission of error, and we do not call it a correction.

Minor edits — typos, a moved paragraph, a replaced illustration — do not enter the log: a list of fifty «updated» lines tells the reader nothing, and a real correction drowns in it.

What we do not do

  • We do not pass machine-written text off as authored, and we do not remove the note about how it was prepared.
  • We do not rewrite other people's material with a model and present the result as ours: a retelling without a source is not our material.
  • We do not publish model-generated images of people or events where a reader could mistake them for photographs.
  • We do not use models to score securities, build rankings or assemble portfolios.
  • We do not pass readers' personal data to a model.

Revision of this document

This document is revised whenever the order of preparing material changes. The date of the last change is given above. Related documents: Editorial policy and Privacy.