The filing does not move. Only what you decided can leave, leaves.
TRIBUNALE ORDINARIO DI MILANO
Il sottoscritto Avv. Chiara Bellini, con studio in Via Melzo 14, Milano,
C.F. BLLCHR78E52F205K, in qualità di difensore di Marco Ferretti,
nato il 3 marzo 1981, premesso che il bonifico di € 24.500,00 disposto
sull'IBAN IT60X0542811101000000123456 non risulta accreditato,
chiede che il Tribunale voglia ingiungere il pagamento della somma.
The four steps, in full
- 01
Open the filing
The app reads the file where it sits: it does not copy it into a working folder, it does not upload it anywhere, it leaves behind no temporary file with the filing inside. No script in the document is ever executed, which is how a PDF does damage; the fonts the document does use, because pages are drawn for the preview and for the recognition. A scan is no longer refused: only the pages with no text are read here, on this computer, with the page count and a Cancel in plain sight, and a .jpg or a .png becomes a one-page PDF the moment it comes in. A file renamed to “.docx” is refused, not opened.
- 02
Look at what it found
Every detection is shown where it sits in the filing, with the sentence around it: that is the only way to decide whether “Milano” is your client’s address or the name of the court. Values with a checksum are marked as verified; the others are proposals, and a proposal is refused with one click. What the recognition cannot catch (a signature, a round stamp, a name a scan read crooked) you catch yourself: a rectangle drawn by hand removes from the file whatever sits under it, and when there is text under it the app offers to cover it on every other page too.
- 03
Export the anonymous text
Before saving, the app counts how many values are still in the clear and of what kind. It is the last moment to change your mind, and the count exists because nobody re-reads two thousand lines. The placeholders that leave are stable: [FULLNAME_1] is the same person throughout the document, so the model still understands who does what.
- 04
Put the names back
The mapping between placeholders and real values never leaves this computer: it sits encrypted with the system keychain. You paste the model’s answer and it comes back with the names in place. And if what you have in hand is not text but a file (an anonymised scan is a PDF, and a PDF cannot be pasted), there is “Open a document…”: the app reads the placeholders back out of the already-anonymised file and puts them back in place, and inside that file there is no table and no key, because the mapping never travels with the document. If you deleted the matter, the mapping went with it and the answer can no longer be reassembled. It is the same sentence said the right way round.
What it recognises, and how it finds it
Twenty-three kinds of data. The column that counts is the second one: an Italian tax code has a checksum that either works out or does not, a person’s name does not. The app does not pretend the two are the same thing, and neither does this page.
- Checksum
- An algorithm confirms it. If the check fails, the detection is dropped instead of lingering as a doubt.
- Shape
- The shape alone, specific enough to redact. Nothing confirms it.
- Model
- The language model proposes it, running locally. No checksum can confirm it: that is why the review step exists.
- Dictionary
- Nothing finds it: the terms you added to the matter do.
Natural person
| Kind of data | How it is found | Example |
|---|---|---|
| Person name, legal roles included: judge, lawyer, parties, witness | Model | |
| Age | Model | |
| Sex or gender | Model | |
| Italian tax code | Checksum | |
| Email, certified mail included | Shape | |
| Phone number | Shape | |
| Identity document number: ID card, passport, driving licence | Model | |
| Vehicle plate | Shape |
Address
| Kind of data | How it is found | Example |
|---|---|---|
| Street, square | Model | |
| Building number | Model | |
| Postcode | Model | |
| City | Model | |
| Province code | Model |
Business and money
| Kind of data | How it is found | Example |
|---|---|---|
| Private organisation: company, law firm, bank | Model | |
| VAT number | Checksum | |
| IBAN or account number | Checksum | |
| Credit card number | Checksum | |
| Money amount | Shape |
Filing and administration
| Kind of data | How it is found | Example |
|---|---|---|
| Calendar date | Shape | |
| Time of day | Model | |
| The code of a filing or a procedure: case number, protocol, repertoire, judgment, CIG, CUP, insurance policy, INPS employer number | Shape | |
| Land registry data: sheet, parcel, subordinate | Shape | |
| Web address | Shape |
The same program, three different jobs
The values that matter are not the same in a case file, a deed and a set of accounts, and neither are the questions clients ask. Each profession has its own page.
Lawyers
Anonymises pleadings, appeals and judgments on the office computer, and puts the real names back into the model’s answer.
Notaries
Anonymises draft deeds and AML files on the office computer, with land registry references recognised alongside everything else.
Accountants
Anonymises accounts, payslips and tax returns on the office computer, with VAT numbers and IBANs verified rather than guessed.
Journalists
Anonymise documents and confidential material on your own computer, before handing them to an AI to summarise, translate or search.
What leaves this computer
Two addresses, and it is not a promise: it is a block. Every request to any other address is cancelled before it goes out.
updates.omissia.com
The update check. It downloads a file that is the same for everyone: it carries no address of yours, no identifier for this computer, nothing from your matters. It also means we do not know how many installations there are or who has not updated, and that is the right price to pay.
app.omissia.com
The sign-in and the state of your licence. An email address and a token go through, and nothing else. The account is required: without it the app does not work. But what goes through here stays this, and no document ever crosses it.
And what never leaves
The filing, its name, its text, the values found, the placeholder dictionary. Not because we promised not to send them: because no address exists that would receive them, and building one is written down among the things this project does not do.
The fastest way to know whether you need it
Take a matter you have already closed and look at what it finds. Ten minutes, and you do not have to sign in to anything.