A glossary of evidence and verification terms

When a research report has to survive scrutiny — a client's, a committee's, a court's — the vocabulary of evidence matters more than the vocabulary of AI. These are the terms Sidereal Research uses, defined the way we operationalize them. They apply to checking any research tool's output, not just ours.

Citation integrity

The fraction of a report's cited sentences whose claims are actually supported by the sources they cite. Measured, not asserted: an auditor (ideally a model from a different family than the writer, to remove self-preference bias) reads each cited sentence against its source's text and classifies it supported, partially supported, or unsupported. Sidereal reports publish this number with its 95% confidence interval — a measured probability, never a guarantee. A report that claims perfect citations but publishes no measurement is asking for trust; a report that publishes 94.1% (CI 89–97%) is offering a number you can hold it to.

Hallucinated citation

A reference that does not support the claim it is attached to — in the worst case, a source that does not exist at all. The critical design question is when fabrication is prevented: filtering citations after a report is written catches some; separating research from writing prevents the class, because the writer can only cite sources the system actually fetched and read. Ask any research tool: can it cite a document it never retrieved?

Corroboration

Multiple sources supporting the same underlying claim. Counting citations is not corroboration — three articles repeating one press release is one source wearing three hats. Genuine corroboration requires the supporting claims to be substitutable (any one could back the sentence) and, crucially, independent (see below).

Source independence

Whether corroborating sources trace to separate upstream origins. A study, a news story about the study, and a blog summarizing the news story are one origin, not three. Sidereal's corroboration analysis marks each group independent or shared-origin, and the drafter is instructed never to write "multiple sources confirm" over a shared-origin group. When checking any report, pick its strongest "widely reported" claim and trace where each citation actually got its information.

Claim strength

What a source's own language commits to: proves, suggests, implies, or asserts without evidence. Distinct from whether the claim is true — it records how firmly the source itself stands behind it. Watch for hedge erosion: an original study "suggests" a finding, and by the third retelling it "proves" it, with no new evidence anywhere in the chain. A trustworthy report states claims at the original source's strength, or attributes the upgrade to the specific evidence that justified it.

Source conflict

A point where credible sources genuinely disagree. Honest research stages conflicts — names the positions, who holds them, and what would settle the question — rather than averaging disagreement into false confidence. A report with no conflicts on a contested topic hasn't found consensus; it has hidden the argument.

Provenance

The recorded origin and handling of every piece of evidence: where a source was found, when it was fetched, what text was extracted, and which claims were drawn from it. Provenance is what makes verification possible later — a report without it can only be re-researched, not checked.

Proof sidecar

A machine-readable record accompanying a report that links each sentence to the source text behind it, re-verifiable offline, any time — independent of the company that produced it. This is the difference between "our platform checked it" and "you can check it."

Evidence corpus

The full set of sources a report was drafted from — inspectable before and after drafting, with each source's claims, quotes, limitations, and perspective recorded. If a tool cannot show you its corpus, its bibliography is a claim, not a record.

Bias as data

Treating a source's perspective as recorded metadata rather than disqualification. An advocacy group's report is usable evidence when its position is named and its claims are corroborated by independent origins. Laundering a partisan source into neutral authority is the failure; naming the lens is the fix.

Grey literature

Research output published outside commercial/journal channels — government reports, working papers, standards documents, corporate technical reports. Often the most load-bearing evidence for professional questions, and the most weakly indexed, which is why coverage of grey literature separates research tools more than coverage of the open web does.

Quote anchoring

Mechanically verifying that a quotation appears verbatim in its source's retrieved text — checked at extraction time, not trusted from the writer. An unanchored "quote" may be a paraphrase presented as quotation, or an outright fabrication. Sidereal marks each supporting quote as anchored or not, and reports the rate.

Every Sidereal report ships these concepts as working machinery: a measured integrity rate, an inspectable corpus, staged conflicts, and a per-sentence proof sidecar. Open a sample and check one yourself →