Check a paper
Paste a passage or send us a document, and we will tell you which research tasks in it show generative AI participation, what delegation warrant each one carries, and which parts of the record are already written down. It takes about a second.
No account, nothing to install, and the whole report renders before we ask you for anything at all.
We produce no AI-likelihood score, here or anywhere else, and a clean report never means no AI was used.
Our checks are keyed to English, so a document in another language is declined by name rather than quietly under-read.
Two trial audits left. A trial audit is the whole report, not a sample: the same result, the same detail, nothing held back. A licence lifts the limit and opens the paper builder, the Delegation Record and sharing. What a licence adds.
Paste some text
An abstract, a methods section, a declaration, or a whole draft.
Or send a document
It is read on our own server and used for this audit and nothing else.
Nothing to hand? Try the sample and read a real report on the method's own source manuscript, in one click.
The audit makes no model call. It is deterministic: it reads your text on our own server, matches it against the method's stated rules, and never sends it to a model. The same document always produces the same report. That is why this step needs no account and no waiting.
The three warrant states
Every task family gets one. They describe how far a result may travel, not how good it is.
- A · delegable
The bounded output may enter the work after routine provenance and formatting checks.
- V · delegable with verification
Useful as a candidate, but it cannot support a claim until a named person has completed a stated check and recorded what came of it.
- H · human-retained
The decision stays with you. This is a statement about who holds the authority, not a rule against using anything.
What the report looks for
Six components make up the minimum warrant record. The report shows each one as recorded, partial, or not yet in the document, with the sentence it found.
- Task and output What finite output was delegated, and what may it change?
- Configuration What model, prompt, tools and settings shaped the output?
- Evidence access What could the system actually inspect?
- Reference standard What makes the output right or wrong?
- Severe error and repeatability What failure overrides a good average, and does the result repeat?
- Named human check Who checked what, when, and with what outcome?
What happens to what you send
Your text is used to produce this one report and nothing else. It is not used to train anything, and it is not shown to anyone. If you never give us an email address, nothing here is attached to a person. See the privacy notice and how we treat your work.
If we cannot read a file, or cannot confirm it is in English, we say so and produce no report at all. A report that looks clean because we could not read the document would be the worst thing we could hand you.