Skip to main content
The draft checker at writerzroom.com/check runs the same deterministic checks the generation pipeline uses on a draft you paste in. It needs no account. It reconciles figures your draft states against each other, such as a total against its rows or a margin of error against its sample size, and reports what it could not decide. No model reads your text. The checks are fixed rules rather than a model’s judgement, so checking the same text twice gives the same answer.

Using it

1

Pick a document type

General runs the numeric consistency check that applies to any text. Each other type adds the checks calibrated for that kind of document (see the table below), so pick the type closest to your draft. A check runs only on the document types it was calibrated for.
2

Paste your draft

Up to 15,000 characters, about 2,500 words. Try a sample loads a document with known mistakes put in on purpose, lists them, and checks it.
3

Read the results

The text appears with each finding highlighted. Select a highlight to jump to its explanation, and Show in text to jump back.

Reading the results

If nothing in the text is the kind of statement these checks read, the page says so rather than reporting a clean result: a draft with no figures has not passed anything. If a check fails to finish, the page marks the results as partial. Under the results, What these checks don’t establish lists each check that ran and what it does not prove.

What each document type checks

Every type includes the General numeric consistency check. The page reads this list from the same vertical configuration the pipeline uses, so it can change as checks are added.

What the checks do not establish

These are the stated limits of each check, as the page shows them.
  • Numeric consistency checks internal numeric consistency patterns; it does not establish that an external factual claim is true.
  • Financial tables checks that stated totals agree with the rows listed above them and that stated percentage changes agree with their own endpoints; it does not establish that any underlying figure is accurate, that a breakdown is complete, or that the presentation conforms to an accounting standard.
  • Citations checks citation structure and internal consistency; it does not confirm that a cited matter exists or remains current.
  • Table of authorities checks consistency between cited authorities and the document; it does not establish that an authority remains good law.
  • Compliance deadlines checks that a stated compliance deadline agrees with the effective date and period given alongside it, within a few days for counting conventions; it does not establish that the effective date, the period, or the obligation itself is correct, that the deadline is legally operative, or that any authority actually imposes it.
  • Polling checks that a reported margin of error is attainable for the sample size stated alongside it; it does not establish that the survey was conducted, that the sample is representative, that weighting was sound, or that the result is accurate.
  • Property math checks that a stated price per square foot or capitalization rate agrees with the price, area, and income given alongside it; it does not establish that any of those inputs is accurate, that the property exists, or that the valuation is sound.
  • Availability checks that a stated downtime budget is attainable under the availability target beside it, for any length the named period can have; it does not establish that the commitment was met, that measurement is sound, or that exclusions are reasonable.
  • HTTP status codes checks that an HTTP status code is paired with its own standard reason phrase; it does not establish that the endpoint exists, that it returns the documented code, or that any other part of the described contract is implemented. Custom reason phrases are permitted by the protocol and are not evaluated.
  • References checks that a labelled ISBN satisfies its check digit and that a DOI matches the registered syntax; it does not establish that the work exists, that the identifier was ever assigned, or that it names the work cited beside it. A DOI carries no checksum and is only ever reported as well-formed, never as verified.
  • Clinical codes checks code shape and internal presentation; it does not establish clinical correctness, existence, or billability.

Limits

  • Length: 15,000 characters per check.
  • Rate: about a dozen checks a minute per visitor, with a daily ceiling. Past that, the page asks you to wait a minute.
  • Unbroken runs: a run of more than 200 characters without a space is refused, since real prose does not contain one. Remove it and check again.
  • Whitespace is tidied first: runs of spaces and blank lines are collapsed and tabs become spaces. The highlighted text is the tidied version, which can differ slightly from what you pasted.
  • Busy or slow: if the checker is busy, or a check takes longer than a few seconds, the page asks you to try again or to check a shorter excerpt.

What is stored

Nothing on our side. The text is checked in memory and discarded; it is not saved or logged. Server logs record the length, the document type and the number of findings, not the text. If you have accepted analytics cookies, we record that a check ran and which document type it used, never the text.

Fixing a draft in WriterzRoom

When a check finds problems or claims it could not verify, Fix it with WriterzRoom takes the checked draft and its findings into the product, where each generated version goes through the verification checks its vertical enables. This step requires an account: signed-out visitors sign in or create one first. So that they survive sign-up, the draft and its findings are kept in your own browser’s storage and used for up to 24 hours; they are not sent anywhere until you continue in the product.

Verticals

How document types map to the vertical configurations and their verification settings.

Data handling

How WriterzRoom handles the content you give it once you have an account.
Last modified on October 3, 2026