Similarity is a lead rather than a verdict
A report identifies strings or patterns shared with material available to its comparison system. It may locate meaningful borrowing, yet it also marks bibliographies, assessment briefs, standard language and properly signalled quotations. Material outside the corpus, translated text, images and sufficiently transformed wording may be missed. A high score is not automatically plagiarism, and a low score does not certify originality.
Preserve the unaltered report and identify the exact submitted document, run date, product, version and settings. Annotate a working copy. In a formal case, the institution's regulations and responsible body determine procedure. This page structures a scholarly review; it is not legal advice.
The displayed match may not be the originating source
Several sites may reproduce the same passage. A system can display the most accessible copy rather than the earliest source. Open the full item and verify authorship, date, edition and surrounding argument. A search snippet is inadequate. For scholarship, resolve DOI, publisher record or repository release and distinguish preprint, accepted manuscript and version of record.
Treat dates cautiously. A recently changed web page may contain older prose; a repository deposit can post-date publication. Prior work by the candidate, shared assessment materials and authorised reused methods need separate analysis. Establish direction through dependable records, not the order in a match list.
Filters and extraction alter what reviewers see
Record whether quotations, bibliography, short matches or named sources were excluded. Inspect a sample under alternative settings to understand sensitivity, not to optimise for a favourable percentage. Exclusion can reduce noise while hiding a composite passage. Keep the unfiltered output as a reference.
File conversion introduces further error. Footnotes can move, ligatures can split, tables can be linearised and scanned pages may yield no text. Compare highlighted output with the visual original. Record OCR gaps and omitted sections. An extraction failure is a limitation, not evidence for or against misconduct.
Use four provisional classes
Material covers a potentially insufficiently attributed or too-dependent passage requiring scholarly review. Legitimate includes properly quoted language, unavoidable terms and authorised common material. Technical captures navigation, metadata, misalignment and extraction artefacts. Open means that source, chronology or applicable convention remains unresolved.
Consider extent, distinctiveness and function. Five words in a defining phrase can matter more than a long conventional description, which a percentage cannot express. Ask whether the citation truly supports the proposition and whether a paraphrase is independently composed. Preserve counterarguments and divergent expert classifications.
Produce a review record that can be challenged
Summarise material findings individually with links to evidence. Report classes separately rather than inventing a revised headline score. State unavailable sources and technical limits. If a manuscript is corrected, preserve the submitted release and identify the new version.
The tool and run metadata sheet fixes technical context. The source-note trail helps test development. Further workflows appear in the examination procedures hub. The single foundation link is the guide to understanding similarity evidence.
Use an independent sample check before closing the record. Give a second reviewer the report, source copies and classification definitions without telling them the preferred outcome. Sample findings from all four classes. Where judgements differ, identify whether the disagreement concerns fact, citation convention or scholarly discretion. Preserve both reasons until it is resolved; averaging labels would hide the issue rather than examine it.
Create a negative-scope list as well. State which pages were unread, which source categories the provider does not claim to cover and which candidate texts could not be obtained. “No match” must not quietly become “no borrowing”. Tables, equations, code and images may require a separate human check. Describe that supplementary step independently, so its result is not misrepresented as a software finding.
Record segmentation when one highlighted passage is split across candidate sources or separate phrases are combined into one signal. Identify which wording actually corresponds to each source and consider whether the displayed sources copied from one another. A colour map is not a historical provenance map. A compact comparison of manuscript text, candidate source and a dated primary release gives reviewers a more defensible basis than the interface's visual emphasis.