The airlock, in detail
Everything Palimpseste Pseudo does — and, with the same seriousness, what it does not do. Every claim on this page describes behaviour that has been built and tested.
IFolder — the unit of work
You designate a folder (top level) or a file. The folder is the unit: one table, one key, the same fake names throughout. A mixed folder is not a trap: only the files in French are read, and the others are named on screen — never silently ignored. You can also add a file to a run that has already gone out: the interview that arrives the following week receives the same substitutes, under the same key.
Formats read: text, Markdown, Word (.docx, document body), PowerPoint (slides), PDF (text layer), HTML, RTF, XML, CSV and spreadsheets. Refused, and it says so: scanned PDFs (no OCR), old binary .ppt files, Word comments and metadata, speaker notes.
IIList — the app proposes, you decide
The harvest relies on your local assistant (LM Studio or Ollama) to judge the candidates — people and organisations, gender included. Spelling recovery is built for the real life of corpora: variants, typos, automatic-transcription distortions (“Rochetau” finds Rocheteau). Doubtful mentions are presented separately, with their excerpts — and the arbiter can only remove substitutions, never authorise one on your behalf. You remove, you add, you merge (“this is the same as…”), you enter every spelling of a name. Without an assistant, you write the list yourself.
IIIFake names — realistic, consistent, reversible
- Fake names that read like names — not labels. Gender is preserved, so is the first name's opening sound (vowel for vowel), and a fake name is never a homophone of a real name in the folder — guaranteed by construction.
- Sub-forms follow: first name alone → fictitious first name alone, surname alone → fictitious surname alone, organisation without its suffix → fictitious root.
- The draw is deterministic: same folder, same fake names — that is what makes the airlock replayable weeks later. A fake name you don't like is removed in one click.
- You choose the restoration key (with an optional hint, shown when you type it). It opens the table; it never leaves your machine; there is no back door — a forgotten key is a lost way back.
IVCopies — French patterns and a blocking final scan
Beyond names, structured patterns are detected and replaced with valid substitutes, outside the table: phone numbers (within the official fictitious ranges — a landline stays a landline, a mobile a mobile), emails, websites (each decided on its own: keep or replace), IBANs with a correct checksum, SIRET/SIREN numbers, social security numbers with a valid check key. A truncated number is never substituted blindly: it goes to review.
Before anything is written, doubtful residues pass before you — four verdicts, including “leave the real name”, which requires a written reason. Then the blocking final scan delivers the only send verdict: as long as a name from your list survives in a copy, the copy is not written. We do not promise “no name ever leaks”; we guarantee that no name from your list goes out without you having seen and approved it, one by one.
On output, two files per document: a Word file where every substitute is in red — proofreading at a glance — and a bare Markdown twin for the assistant. Copies are numbered (1.docx, 2.docx…): a file name can identify, so it is pseudonymised too; the mapping travels encrypted inside the bundle. The originals are never touched.
VRestore — even a rewritten text
Restoration returns the real names in any text carrying the pseudonyms — the exact copy, but also a document entirely rewritten by the public AI. That is what makes Pseudo an airlock, not a filter. Patterns (phone numbers, emails, IBANs…) come back exactly as written. The restored file is written next to the designated document; in case of ambiguity, the app prefers to have you reread rather than guess. And Restore never locks itself: subscription expired, your tables and the way back remain.
What Pseudo does not do
- It is not anonymisation: in the GDPR sense, pseudonymised data remains personal data — a mapping table exists, on your machine.
- A name outside the list stays visible in the copies — the app never invents a name on your behalf.
- Cross-referencing is not handled: dates, places, sequences of events can still identify.
- No OCR, no PDF redaction, no impact assessment, no compliance in a box. The proofreading — and the legal basis — remain yours.
- Corpus v1: French. A document clearly in another language is refused, and it says so.