Proof
A released bundle you can open.
An English employee access request made entirely from synthetic sources and processed through the current Veil pipeline. Open the email thread, personnel record and attachments to see the rebuilt documents, then check their manifest and quality report. The README explains the recorded model replay and what this example does not measure.
ZIP, 274 kB, 9 documents, 23 pages. No sign-up.Skip to the download
Every source document, and what became of it.
Read out of the archive's own manifest: 9 documents, 8 released and 1 withheld, 23 pages in all. Detector 1.16.0, verifier 1.8.0, redactions found by mistral-large-2512.
What you are downloading
- The real output of a real run. Ingest, extraction, detection, entity resolution, policy, render, an independent verification pass, packaging, a spot check, and release through the same door an operator uses.
- Redaction decisions made by a real model on the English source text, then replayed from those fresh recordings rather than derived from an answer key.
- An HR file the size a company actually holds: an email thread with a cc list, a quoted reply and a forwarded customer complaint, a personnel record, an investigation note naming a complainant and two witnesses, an absence register, and a payroll summary.
- One output document per source document, numbered so the containers are visible: an email and its attachment, and an uploaded archive with the two documents inside it.
- The manifest, as machine readable JSON and as a document, and the quality report.
What it is not
- No customer material. The sources are generated test fixtures, with invented people and organisations.
- Reserved example email domains and generated structured identifiers. The archive contains no real account or customer identifiers.
- No scans. Native PDF text is extracted directly. Fresh real NER and residual responses were recorded for the English sources; the published replay used those frozen recordings without further live model calls. This pack does not measure OCR accuracy, production throughput or human review time.
What the run did, document by document.
The report the product wrote for this case, read off its own pages. One row per document: how it was read, how many entities the pass found, how many it could not place, and whether the independent check cleared it for release.
Source coverage, from the report
How much of each source reached detection, counted in the unit that format has: pages for a PDF, rows for a spreadsheet, paragraphs for everything else. A file the product could not open carries no figures at all, and the report says why rather than leaving it open.
Case summary, from the report
- Source documents
- 9 8 included, 1 withheld
- Entity registry
- 2 subject, 39 kept as not personal data, 15 third party, 0 unresolved
- Residual pass rounds
- 3
One page, exactly as it left the product.
The first page of the personnel record rebuilt from DOCX. Its labelled fields and paragraph breaks remain readable. The requester's account is still redacted in this run, while her birth date stays after an operator attestation. Download the PDF to read every page; this view substitutes kept personal values by their kind.
Read from the archive's own PDF, text and page stamp alike. Nothing here was typed.
The marked values are this page's substitution, not the product's: the archive holds the released values of the person the request is about, and this site prints none of them, only what kind each one is.
The downloadCase VC-2026-0001
A released bundle you can open.
ZIP, 274 kB, 9 documents, 23 pages. No sign-up.
SHA-256 27709a047701b0ab874d0d968a63f01786caf80c0b28a6c0706d6ad6d0cfc65f
000_README_SYNTHETIC.txt is the one file the product did not write. Everybody in it is invented, and the archive says so in its own first file, which also states the digest the bundle had before that file was added.
What to look for
Fields and paragraphs stay readable
Start with the personnel record rebuilt from DOCX. Its field labels, line breaks and paragraph order remain visible alongside readable pseudonym labels. The payroll summary shows the same treatment for plain text. The PDFs are generated from extracted text; original fonts, columns, positions and page imagery are not reproduced.
Native PDF page boundaries
Open the timeline attachment and the occupational health opinion. Each retains its source page boundaries, even where a source page contains only a short paragraph. More pages here is a consequence of preserving those boundaries, not additional material or a claim to reproduce the original page design.
One person, one recurring label
The requester's manager appears in the email thread, timeline, personnel record, investigation note, spreadsheet, payroll summary and health opinion. The same pseudonym follows her through them. A colleague who shares the requester's surname is redacted while the requester remains distinct.
Kept subject details and a visible limit
The requester's name, staff number and attested birth date remain in the personnel record. Her account number and her direct line in the email thread still redact in this run because they were not attested. The sample shows the operator's recorded decisions; it does not imply that every detail belonging to a subject is automatically disclosed.
Correspondence and spreadsheet context
The email keeps its header labels, quoted reply and forwarded complaint in reading order. The absence register retains its sheet headings, rows and notes. Read the surrounding ordinary dates and sentences, not only the redaction labels, to judge whether the result is useful.
Extra labels and attribution limits
A bare first name can receive a separate label instead of being merged with a full name. Some staff numbers remain readable where the recorded decisions did not attribute them to a person, and some spreadsheet dates are redacted as payment-card values. These are limits of this fixed recorded example. Read the documents alongside the quality report rather than treating a placeholder count as a count of people.
A withheld file stays in the record
The exported HTML chat log is unsupported and excluded. The manifest names it and states why; the other documents remain included. The quality report separates source coverage and verification from that format limitation, so a withheld file is not presented as an independent verifier finding.
Traceable pages and exact output hashes
Every output page carries the case reference, a synthetic document id and its page number. The manifest records each included PDF's hash. The public README is the only added file; the emitted PDFs and manifest are unmodified.
What is not in the archive
There is no decision log here. Every entity decision, every download and every state change is written to an append-only audit record, and a customer exports that alongside their bundles rather than receiving it inside one. The audit record holds identifiers, counts and dates and never a document excerpt, which is what makes it something we can hand over rather than something we would have to redact first.