MAILCRAFT / MINISTRY OF HANDWRITING

PUBLIC INFORMATION / NO SECURITY CLEARANCE REQUIRED

Frequently
avoided questions.

What happens to your handwriting, what a model learns, and where the joke ends. The Ministry has brought sources.

01. If it's free, am I the stationery?

We prefer “voluntary contributor to the national supply of loops.”

You get three server presets and personal handwriting generation without contributing to training. Choose to contribute one sample, and you earn the other ten presets. The exchange is real: we retain that sample and transcript for training and evaluation until you delete them. Deletion keeps your earned styles. The $1.99 offline edition needs no contribution.

This is a limited permission, not ownership of your handwriting. We do not sell, publish, or license raw contributed samples to third parties. The actual disclosure appears before you choose; closing it means no contribution.

Our product commitment. Background: Google Play's data disclosure and consent rules.

02. Is the machine learning to read my diary?

Its current assignment is moving a pen. Ambition remains under review.

Mailcraft predicts the next pen movement from earlier strokes and the text you provide. Training uses paired strokes and transcripts to adjust the model's numerical parameters. Generation uses those parameters to draw new strokes. This is handwriting synthesis, not OCR that reads a photograph.

Alex Graves: Generating Sequences With Recurrent Neural Networks.

03. Did my sample just give the machine a promotion?

Submission is not graduation. There are forms between them.

Collecting, evaluating and training are separate operations. A reference sample can guide a generated note without updating the model's weights. Only the separate contribution flow retains your sample for future training or evaluation. At this development checkpoint, no new training run from these contributions has been adopted.

Mailcraft implementation, reviewed 22 September 2026. The underlying distinction between generation and training is described in the original research.

04. Is a model a filing cabinet?

No drawers. Unfortunately, that is not the same as no memory.

A trained model uses learned numerical parameters, rather than simply opening a saved page. But that doesn't rule out memorization. Researchers recovered some original training passages from GPT-2. That was a language-model study, not a test of Mailcraft; it supports caution, not a claim that this model has leaked handwriting.

Carlini and colleagues: Extracting Training Data from Large Language Models.

05. Does Delete make a model forget?

The shredder handles files. It has not passed its neuroscience exam.

Deleting a stored sample and removing its influence from an already trained model are different tasks. The latter can require retraining or a verified unlearning method. Our Delete control removes the stored contribution and excludes it from future training and evaluation. It does not promise to reverse completed training. Your earned presets stay unlocked.

Google Research explains machine unlearning.

06. Anonymous if I leave off my name, right?

A fake moustache is not an anonymisation technique.

Removing a name does not automatically make information anonymous. Remaining details and other available information can still identify someone. Use a neutral phrase such as “hello from my pen,” not a signature, address, account detail or private message. A random sample ID is not an anonymity guarantee.

ICO: Introduction to anonymisation.

07. Does offline rewind my earlier choices?

It closes the network door. It is not a time machine.

The tested offline development app runs generation on your device without Android's INTERNET permission. You can still choose to export or share a file using another app. Switching editions does not automatically delete an earlier server contribution; use its deletion receipt in the free edition.

A claim about our tested build, not a guarantee about every app labeled “offline.” Final release behavior must be verified again.

08. Where's the urgent, once-in-a-lifetime offer?

The Department of Artificial Urgency has been given the afternoon off.

The planned offline price is $1.99 USD, one time. No countdown, invented former price or recurring charge. It is not yet on sale in a store; the eventual listing will show your actual local price before you pay.

Why we spell that out: the FTC's report on misleading interface patterns.

The jokes explain the product; they are not the permission. Read the plain privacy policy and sample contribution terms for the development preview.