Methodology

The principle: one value, one piece of evidence

Every value published on Vuntum is a claim tied to an exact quote from an identified source, to a publication date and to a date of retrieval. A value without a validated claim is never displayed. An empty cell means Not disclosed: it is not a negative answer.

The list of sources and their confidence scores is public, as is the log of corrected errors. For who publishes Vuntum, and with what resources, see the presentation of the editorial team.

Levels of evidence

Two statuses override the level of evidence: disputed (two credible sources contradict each other) and stale (the last check is older than the freshness limit of the criterion).

Confidence score

The score is deterministic and explainable. It does not depend on any impression formed by a model:

confidence = 0.40 x source confidence + 0.35 x level of evidence + 0.15 x independent corroborations + 0.10 x freshness

Vuntum autonomy scale

Freshness limits by criterion

Badges

Estimated European price

This is not a fact, it is a calculation. It is always shown next to the original official price and labelled Vuntum estimate:

official price x dated exchange rate x (1 + customs duty) x (1 + VAT) + flat fees

The exchange rate used is dated and sourced. Customs duty, VAT and actual fees vary by country and sales channel: the estimate does not commit any seller.

Publication

Vuntum publishes automatically, within safeguards, whatever passes all of its checks. A value is published only after a second, independent reading by another AI model, which checks the value against the quote from the source. A value that only a media outlet reports, without having measured it, waits to be confirmed by the manufacturer or by another source. An article is published only if each of its factual sentences points to a value that is already published and its figures come from those values. A brief is published only if each of its sentences rests on a quote found word for word in the source, its figures appear in those quotes, and a review by another model finds nothing to object to.

Values that are already published are re-read at random every day: a value that is contradicted is withdrawn and the correction is displayed. A circuit breaker suspends all automatic publication in the event of abnormal behaviour (too many refusals, too many publications from a single source). The editor checks after the fact and can reverse anything. Reports from readers are handled first.

Never published: buying recommendations, accusations, unestablished security or privacy information, rumours, and anything a source does not allow us to verify.

The role of AI

AI tools sort the sources, extract the values, translate, write the articles and briefs, and review the work of another model. Their output is checked mechanically: the quote must be an exact substring of the source, numbers must be found in the quote, and an identifier absent from the input data invalidates the output. Content written with the help of AI says so.

Collection

We respect robots.txt, platform terms and rate limits. We use no workaround for anti-bot protection. We neither download nor rehost any video. Source snapshots of the pages we consulted are kept internally as evidence and are never republished: only short quotes are displayed, with attribution.

Independence

No paid placement in comparisons. Any affiliate link is flagged where it appears. Conflicts of interest are declared. An error that is reported is corrected and recorded in the corrections log.