VELLUMSAR

Subject access requests, with redaction you can prove.

Vellum runs a subject access request from the day it arrives to the day the bundle goes out. It is built for UK public bodies and covers the statutory clock, documents in mixed formats, detection, human review, a verified release and a hash-chained audit trail.

Vellum checks the finished bundle three independent ways for any of the text you redacted, and refuses the download if any of it survived.

Built by RODMENA LIMITED. This deployment holds only fabricated case data.

A sample contact note in which third-party names, addresses and numbers are highlighted and then blacked out before release.

How a request moves through Vellum

  1. Intake and the statutory clock

    The month starts on the day the request arrives, or on the day the requester's identity is confirmed if you needed to confirm it. It ends on the same date in the following month. If that date is not a working day the deadline moves to the next working day, and if the following month has no such date it falls on the last day of that month. Asking the requester to clarify which information they want stops the clock. Asking about the format of the response leaves it running. An information governance officer can see how Vellum worked out each date and check it.

  2. Ingest, whatever the pack contains

    Vellum takes PDF, Word, Excel, PowerPoint, OpenDocument, images and plain text. It works out each file's type from its contents, because in a real pack the extension is only what someone typed when they named the file.

  3. Concealed content, found before review

    At ingest Vellum looks for content a reviewer would never see on screen and puts it in its own queue, which a reviewer cannot skip. It finds optional-content layers that are switched off, white-on-white and zero-size text, content placed off the page, annotations and comments, tracked changes, document properties in Word files, hidden worksheets, hidden rows and columns, and pivot caches. Each item says how it was hidden.

  4. Detection that shows its working

    Vellum validates UK identifiers against their own specifications. That covers NHS numbers (Modulus 11), National Insurance numbers, DfE pupil numbers (check letter), postcodes, telephone numbers, sort codes and dates of birth. A candidate shows the detector and its version, along with the rule or checksum that made it fire. The reviewer can see why it was raised and argue with it.

  5. One decision, applied across the whole pack

    Names are recorded inconsistently in any real file. The Party Register lists the variant spellings it finds in the papers, such as Okonkwo-Bassey, Chloe and C. Okonkwo-Bassey. A person confirms which of them are the same individual, and Vellum applies that decision to the confirmed forms exactly as written, throughout the pack. While the register is being built Vellum uses fuzzy matching to suggest variants. Once a person confirms them it matches only those exact forms.

  6. Review under human control

    A reviewer accepts, rejects or amends a candidate, and has to choose the statutory basis before accepting it. For third-party material Vellum records the ICO's three-step test. The Party Register records a person's capacity apart from their role, because the professional exemption depends on whether someone acted in an official capacity. In practice the social worker's name is released and their home address is withheld.

  7. Verification of the release

    Vellum burns redactions into the pixels of the pages that carry them. The text that is kept goes back onto the page from its stored geometry, without a second round of recognition. Three independent checks then examine the finished bundle. Before they run, each has to fail on control documents that carry planted text. Vellum will not release while any page is undecided, or when the checks could not be shown to fail.

  8. The audit trail

    Events are hash-chained to the one before, and reads are logged as events too. Changing or removing an old entry breaks the hash of each entry after it. The break can be found without a second copy of the log, and the verifier reports the sequence number where it starts. We test the verifier by corrupting an entry and checking that it reports the break.

What the AI does

The model does not edit documents

A language model's findings go to a reviewer as candidates. A finding that does not quote text found word for word in the document is discarded. Vellum publishes the share it discards as a hallucination rate.

Vellum works with the AI switched off

Deterministic detection and the Party Register are a complete product without it. Switching the AI off lowers recall and leaves the rest working. If someone asks what opting out would change, you can re-run a case with the AI off and show them.

The AI is tested on each run

Before Vellum trusts any model output, the AI tier has to find personal data planted in a control document. A second control document contains an instruction in its text. If that instruction changes the result, the AI tier is switched off for the run. The test is there because an applicant may have put documents into their own file.

Data sent to a model is logged

For each call Vellum logs the model, its version and region, the purpose, the redaction state when it was sent, the token counts and a hash of the prompt. The log exists even when the AI is switched off. An empty log on a completed case then shows that nothing was sent.

What exists today

This list is generated from the product's own feature register, which also drives the badges inside the application. Not in this deployment means the capability is built and switched off here. Anything not yet built says Planned.

15 available · 0 not in this deployment · 5 planned

Casework

  • Case intake and statutory clock Available

    Records receipt, identity confirmation, extensions and clarification, and computes the statutory deadline.

  • Party Register Available

    Names every individual in the case with their role and capacity, and drives consistent treatment across the whole pack.

  • Freedom of Information policy pack Planned

    The same machinery against the FOIA exemption set.

Ingest

  • Mixed-format ingest Available

    PDF, Word, Excel, PowerPoint, OpenDocument, images and plain text, identified by content rather than by filename.

  • Concealed content discovery Available

    Finds content a reviewer could not see on screen and puts it in its own queue, naming the mechanism that concealed it.

  • OCR for scanned records Planned

    Recognising text in scanned pages so they can be searched and reviewed.

  • Handwriting recognition Planned

    Cursive and handwritten contact notes.

Review

  • Detection with stated evidence Available

    Every candidate carries the detector, its version, and the rule or checksum that caused it to fire.

  • Consistent entities across the pack Available

    A name confirmed once is applied identically to every document, and the result is reproducible byte for byte.

  • ICO three-step third-party test Available

    Records each step and each factor weighed, not only the conclusion.

  • LLM assurance tier Available

    A second opinion over the already-redacted rendition, returning findings for a human rather than edits to a document.

Release

  • Falsifiable release verification Available

    Checks the finished bundle by independent paths, after first proving the checks can fail on deliberately dirty controls.

  • PDF/A-2b output with validation Planned

    Archival-format bundles validated before release.

  • Vector-mode redaction Planned

    Removing text from the content stream while keeping the page vector.

Governance

  • Tenant isolation enforced by the database Available

    Each licensed organisation's data is separated by PostgreSQL row-level security, not by application code remembering to filter.

  • Irretrievable erasure with a surviving audit trail Available

    Deleting a case destroys its content and its files on disk, and leaves the record that the deletion happened.

  • Hash-chained audit trail Available

    Records reads as well as writes, and detects alteration or removal of a historic entry.

  • AI governance and opt-out Available

    The assurance tier can be switched off per tenant; the system still completes a case without it.

  • Microsoft Entra single sign-on Available

    Sign in with your own organisation's Microsoft directory.

  • User and role administration Available

    Adding people, changing their roles, switching an account off and removing it, from inside the product.

How to check us

You can test the claims on this page yourself.

  • PostgreSQL keeps each council's data apart. Row-level security is enabled and forced on the 15 tables that hold each council's data, and forcing it means the policy binds the schema owner too. The application connects as a role that owns nothing and has neither superuser nor BYPASSRLS. More than thirty tests try to read, change or delete one council's data from another and require the attempt to fail. The main isolation suite pairs these checks with controls showing that the owning council's query does return rows, because a broken query would also come back empty.
  • Accuracy is measured on your documents. Vellum reports precision and recall for each entity class, against a ground-truth answer key for the pack you are working on.
  • The release gate is on by default. Turning it off needs a configuration change, which is recorded under your name.
  • You own the policy pack. The statutory grounds ship as a starting point that your data protection officer reviews, amends and versions.
RODMENA LIMITED Company Product screens
Compliance UK GDPR · DPA 2018 · ICO guidance Data held in the United Kingdom

© 2026 RODMENA LIMITED. Vellum is a Subject Access Request management and redaction platform. The cases in this deployment are sample data and contain no real personal information.