Facts. Sources. Context.

We go back to the original document rather than what was said about it.

A report, a court ruling, a regulatory filing, a scientific paper, a dataset, an official statement: whenever possible, we cite the document itself, not its retelling in the press or on social media.

When information rests only on a secondary source, we say so.

A number in the text, an identified source at the end.

Claims that matter are linked to their sources with a numbered note. Each source states its type, its author or publisher, its date, and links to the original.

In our stories, quotation marks are reserved for words actually quoted, taken verbatim from a source and followed by its note.

We search broadly, then we search for what is missing.

For each investigation, a first set of queries is run on the web. FactaVue then reads every distinct, documentary page that comes back, with fallback methods when a page is empty or blocked, and builds one record per page.

After the initial research, the automatic chain defaults to up to three deepening rounds. Further rounds follow up on gaps and contradictions from earlier rounds, seeking primary documents, opposing positions, and the data needed to investigate them. The editorial team can use the “Deepen” button to add rounds: six rounds, for example, if the subject requires it. Three rounds are not a limit. A chain currently allows up to eight deepening rounds; a new research run can then be started. A source enters an investigation only if its page was actually read.

Research publishes nothing. Its results are working material that the editorial team reviews.

Who funds, who runs, who lobbies: we look in registers, not in press releases.

Every research run includes, by default, a tracing round. Rules, not a model, choose the key actors: the people the story concerns, the register entities that appear in it, the actors most cited in relationships already verified. For each one, FactaVue looks for the primary sources of their money and power: company and organisation registers, annual reports and financial statements, lobbying registers, public grant databases, governance documents.

In the health sector (pharmaceutical companies, hospitals and their foundations, universities, physicians), we add industry payment disclosures (Open Payments in the United States, Transparence Santé in France, Disclosure UK), clinical trial registers, Health Canada databases and the declarations of interest of advisory committees, when they are published.

A link of funding, ownership, contract or employment is recorded only if a page that was actually read states it, with its exact quotation.

A quotation is kept only if it really appears in the text that was read.

Every quotation picked up by a model is automatically compared with the text of the page or transcript that was read. Only differences of case, spacing, quotation marks and dashes are ignored: a single different word is enough to reject it. A quotation with gaps is accepted only if each of its fragments appears, in order.

A quotation that cannot be found in the text is discarded, however plausible it looks.

The evidence must stay available, even if the original page changes or disappears.

The sources we keep are archived automatically. FactaVue holds a private copy, identified by its SHA-256 digital fingerprint, which makes it possible to check later that the copy was not altered. FactaVue also requests a capture from the Internet Archive's Wayback Machine, whose public link appears under “Archived copy”.

The private copy is not published. Only documents we have the right to host are offered for download.

We say clearly what is known, what is claimed, what is disputed and what remains unknown.

Honest reporting also shows its limits. In our stories, a status table qualifies central claims on four levels:

A fact is presented as established only if at least two independent sources confirm it, meaning sources on different domains: two pages of the same site, or a text reproduced as is, count as a single source.

A fact that rests on a single source is attributed by name to that source and classed as reported. When sources conflict, the story says so, naming each one, without settling the matter without proof. What no source can establish remains unknown.

What a person said, we quote as they said it.

When a story relies on public remarks (an interview, a talk, a video, a web page), FactaVue collects the person's statements and checks each one word for word against the text or transcript that was read. A statement is presented with its date and outlet, never altered or cut to change its meaning.

For a video, the transcript is timestamped: each quotation links to the exact moment it is spoken. Published captions are used first; otherwise the audio is transcribed automatically. An automatic transcript can contain errors: the story says so when it quotes one.

An isolated fact can mislead as much as an error.

Timelines, definitions, orders of magnitude, regulatory frameworks, precedents: we provide what readers need to place information and weigh its significance.

Who funds, who owns, who sits where.

When funding, ownership, contracts, employment or influence are relevant to understanding a subject, we document them explicitly, with sources, amounts when known, and dates.

These links are gathered in a public register: each relationship points to its sources. A relationship is labelled documented only if two independent domains confirm it; it is disputed if sources diverge, and reported in other cases. A relationship proposed by the research enters the register only after the editorial team has reviewed it.

A documented relationship is not an accusation: it is information readers can use.

Readers should be able to read what we read.

When we have the right to do so, we host the documents we cite so they remain available. Our document library gathers every source used in our work, and its search lets you find them by title, publisher, author or description.

Document library

The text is read by eyes other than those that wrote it.

An AI model writes a draft from the sources and findings it is given alone, without inventing anything. A second model, designed as an independent reviewer, then checks every claim, figure, date and quotation against the research evidence: it removes or rewords what is not supported, reclassifies a single-source fact presented as established and restores contradictions that were smoothed over.

No story is published automatically. Publishing is a decision of the editorial team.

A person is not guilty until finally convicted.

We never write that a person committed an offence without a final conviction: we use “accused of”, “suspected of”, “according to the police” or “according to the indictment”. We never say a person was “cleared”: they are presumed innocent, and an acquittal confirms it.

A headline from another outlet that presents a person as guilty, reproduced as an example, is anonymised, unless that person is legitimately the subject of the story.

We use AI, and we say so.

FactaVue relies on AI models at several stages: proposing queries, extracting records and quotations from the pages read, summarising research, writing a draft, reviewing it, translating between French and English and, sometimes, generating an illustration.

A model proposes; fixed rules decide what stays: a source must have been read, a quotation must appear word for word in the text, a source reference must exist. These checks do not replace human judgment. Models can be wrong, which is why every important claim points to its sources, so you can verify it yourself.

We correct our mistakes in public.

Every factual correction is flagged in the story concerned, dated and described, as a correction, clarification or update notice. Notices are also collected on a dedicated page, which explains our policy. A substantive change updates the displayed update date.

Corrections

An illustration is not evidence.

We use photographs, document captures, visualisations and sometimes illustrations created for a story. Any generated image or non-documentary illustration is labelled as such in its caption. We never present it as a photograph of a real event.