EVIDENCE

Sources and verification

How we select, corroborate and preserve evidence so that every material claim can be reviewed, reproduced and corrected.

Last reviewed: Operated by: Linkses OÜ
01

The right source for each claim

There is no single hierarchy for every question. We prioritise provider documentation for specifications, prices, licences and releases; official texts and registers for legal or corporate facts; and the original paper, data or code for research results. Strong secondary sources add context, history and independent scrutiny.

An interested source can establish what an organisation announced, but cannot by itself prove that its product is best or that a comparative claim is true. Authority, independence, proximity to the facts and fitness for the specific claim are assessed separately.

02

Evidence record and traceability

Each material fact must map to a specific source. We record the canonical URL, title, publisher or author, source type, publication and access dates where available, primary status and verification state; notes identify the supported claim and relevant limits.

We cite the most precise page, version, table, section or record available. A search engine, aggregator, automated summary or AI answer may help discover material, but does not replace the original document or become evidence merely by linking to it.

03

Checking and corroboration

We verify identity and domain control, authorship, date, version, scope and internal consistency. We also check the exact model or product, region, currency, taxes, units, sample size, configuration and other conditions that can change the meaning of a number.

Extraordinary, disputed or consequential claims require independent corroboration proportionate to risk. We look for evidence that could disprove them, distinguish fact, estimate and opinion, and state uncertainty when available evidence cannot settle a conclusion.

04

Studies, papers and preprints

When using research, we state whether it was peer reviewed and examine authors, funding, declared conflicts, method, sample, baselines, significance, limitations and reproducibility. A preprint is labelled as such and is not presented as scientific consensus.

We retain the DOI, repository and version where available and check for corrections, expressions of concern or retractions. We separate what the data show from the authors’ interpretation and avoid extending a result beyond the population, task or conditions studied.

05

Benchmarks and model claims

A score is meaningful only with the exact model and version, provider, configuration, tools, prompts, number of attempts, dataset and split, metric, evaluator and date. Provider-reported results are labelled as such rather than presented as Inferama measurements.

We do not merge numbers produced under incompatible rules. Potential data contamination, selective test reporting, version changes, closed evaluations and the difference between reproducible runs and declared results are reviewed before comparison.

06

Changing data, conflicts and broken links

Prices, limits, availability and product names may change without notice. We display verification dates, use shorter review cycles and retain the reference and notes needed to understand what information was available when it was checked.

When sources conflict, we explain the difference and our choice or leave the point unresolved. A vanished source does not remain verified by inertia: we seek a reliable official or archived version and otherwise replace, qualify or remove the claim.

07

Anonymous sources and third-party content

We prefer attributable sources. Confidential or anonymous information is considered only exceptionally, where there is public interest, direct knowledge, internal identity verification, assessment of motive and additional corroboration; it cannot alone support an accusation or high-risk conclusion.

Social posts, forums and user-generated content are leads or evidence of what an account published, not automatic confirmation of the underlying facts. We examine origin, date, context, authenticity and possible edits before use.

08

Automation, review and corrections

Automation may find candidates, extract metadata, detect contradictions and check links. It does not decide that a claim is true. Every citation suggested by a model is opened and validated because a plausible reference may be fabricated, incomplete or unrelated to the text.

Audits identify absent, stale or insufficient sources and can return content to review. Anyone can report an error with the URL, disputed claim and relevant evidence; we correct the content and its traceability under our corrections policy.

Official references

Reuters · Journalistic Standards↗W3C · PROV overview↗Crossref · Bibliographic metadata best practices↗NIST · AI Risk Management Framework 1.0↗NIST · Generative AI Profile↗COPE · Editorial ethics toolkit↗