Terbit Zero AI evidence note
A detector result should create questions—not accusations
This first-party methodology explains how Terbit Zero AI reviews its own detector experience. It documents the test input, expected evidence, failure conditions and the human checks required before a result is used.
Test design
Use passages long enough to contain several sentences and retain their original context. Include ordinary human business writing, edited text, translated text and AI-assisted text. Do not tell the reviewer which category is expected until the evidence has been recorded.
Record the exact passage, date, tool version and complete result. Repeating only the top label removes the explanations needed to assess whether the output is reasonable.
What we score
We examine whether the result identifies observable features, provides plausible human explanations, communicates uncertainty and recommends proportionate follow-up. We do not score success by whether the tool agrees with a hidden label, because authorship itself may be mixed or undocumented.
A useful result should help a reviewer locate evidence outside the detector: drafts, source notes, version history, citations and a conversation with the writer.
Recorded live check — 26 October 2026
We submitted a short corporate-style passage stating that an operations team reviewed a quarterly report, revenue increased modestly and logistics costs remained above plan. The live tool returned a mixed signal. It cited neutral tone, passive construction and lack of specifics, while also explaining that executive summaries, redaction and normal corporate templates could produce the same features.
This was an appropriate outcome because the passage was deliberately generic and its authorship could not be inferred from style. The result remained cautious and directed the reviewer to ask for figures, dates, source documents and version history.
Failure conditions
A test fails if the output presents certainty, omits plausible human explanations, penalises a writer for language background, or recommends a consequential decision without supporting evidence. Short text and highly formulaic material should receive stronger limitation language.
We will repeat documented checks after material prompt or model changes and record corrections in the changelog.