What the Lab keeps — the privacy note
In one paragraph
We keep the problem we gave you, the answers you sent, and what our checks said about them. We keep your CAD file so the checker can measure it and so you can ask us why it failed. We do not sell anything, we do not advertise, and we do not train an AI model on your work. The tutor, when it exists, is sent your status codes — never your numbers, and never ours.
En un paragraphe
Nous conservons l'énoncé que nous vous avons donné, les réponses que vous avez envoyées, et ce que nos vérifications en ont dit. Nous conservons votre fichier CAO pour que le vérificateur puisse le mesurer et pour que vous puissiez nous demander pourquoi il a échoué. Nous ne vendons rien, nous ne faisons pas de publicité, et nous n'entraînons aucun modèle d'IA sur votre travail. Le tuteur, lorsqu'il existera, reçoit des codes d'état — jamais vos nombres, ni les nôtres.
What is stored, field by field
From your CAD file we keep measurements, not the file's contents: each geometry check stores the one number it measured and the one it expected — a distance, a diameter, a mass — and a review finding stores the one measure its rule read. No name, author, comment or other metadata from the file is ever read into these rows; the measurement contract refuses any field it does not know.
In plain words:
| What | Why it exists | Could we do without it? |
|---|---|---|
| Your student id | to give you the same numbers twice and to find your work | No. It is the only identifier we hold, and it is whatever name you were enrolled under — not an email, not a phone number |
A fingerprint of your personal link (token_hash) | your link signs you in; we keep only a one-way hash of it, so a copy of our database does not contain your link | No — it is how the site knows it is you |
| Your language, when you accepted the beta terms, your 18+ confirmation and your gallery choice | to show the site in your language, and because the beta is for adults and a gallery needs your yes | The gallery choice, yes: it is optional and starts as no |
Your requests to the checker, and the reports it sent back (jobs, pages) | so the site can show you your brief and every report again | No — they are what the site displays |
Your numbers (params) and the correct answers | so that a brief you printed in November still means the same thing in December (Contract 6) | No. Re-deriving them instead would silently change your problem the first time we edit the challenge |
The values you submitted (student) | to tell you what was wrong, and to build the mistake library that makes the teaching better | No, and this is the one that matters most: see below |
What we expected (expected, deviation) | so a verdict can be re-read and explained months later | No. Note that these are never sent to your browser, only stored server-side |
| Timings | so "why did my report take an hour" has a true answer | Yes, and we keep them anyway because the answer matters |
| Your CAD file | the geometry checker measures it; you may ask us why it failed | Not while the challenge is open. See retention below |
The mistake library, stated plainly, because it is the thing we gain. When your answer is wrong, we record what you answered and which known mistake it matches, and we read those records to write better hints and better challenges. Your work therefore improves the Lab for the students after you. It is stored under your student id, it is not published, and it is not sold. If you would rather it were not kept, say so and we will delete your records — see below.
What is not stored
- No email address, phone number or postal address in the Lab itself.
- No password. The W45 pilot runs by email; on the site you sign in with a personal link, and we keep only a one-way fingerprint of it — not even we can read your link back from our records.
- No analytics, no advertising identifiers, no third-party trackers.
- No content of your CAD file beyond the measurements the checker takes and the file itself. Nothing inside the file influences a verdict — not its metadata, not its layer names, not its annotations. The verdict is built from measurements only.
Who else sees it
| Who | What they get |
|---|---|
| Fortune (the operator) | everything above. One person, one laptop |
| Supabase, our database host | stores the data above on our behalf and processes it only to run the database. [FORTUNE] — the hosting region, chosen when the project is created, to be stated here |
| Cloudflare, which serves the pages | the pages themselves only; no student data passes through it |
| The AI tutor, when it ships | status codes and hint ids. No numbers at all — not yours, not ours. This is a property of the code, not a policy: the tutor view is built by an allow-list, and our tests fail if a number reaches it |
| Anyone else | nothing |
How long
- Runs, answers and verdicts: kept while the Lab exists, because the mistake library is the point. [FORTUNE] — a retention window to state here (proposal: records older than 24 months are reduced to the diagnosis codes, with the submitted values dropped).
- CAD files: [FORTUNE] — a deletion window to state here (proposal: 90 days after the challenge is closed, since the measurements are kept in the verdict and the file itself is only needed to re-explain a failure).
- The pilot's own store (
lab-data/store.json) lives on one laptop and is never committed to the repository. The website's database is hosted by Supabase and copied once a week to that laptop, because the host keeps no backups of its own on the free plan.
Your choices
- Ask what we hold about you, and we will send it — your rows are yours, and we can export them in one file.
- Ask us to delete it, and we will, including the mistake-library rows.
- Use a pseudonym as your student id. Nothing in the Lab needs your real name; the name printed on a brief is only what you ask us to print.
- [FORTUNE] — the address these requests go to, to be stated here.
Beta, and what that means
The Lab is a beta and is labelled as one. Verdicts are produced by tested code, not by an AI, and no AI can change one. Where a result is a judgement call rather than a measurement — the review panel's cost figures, for instance — it is labelled advisory and dated, so that a number taken to a workshop and found to be 40% out was flagged by us before it was flagged by the machinist.
Anything in this note that the code contradicts is a bug in the code or a lie in the note, and an automated test compares the two every time the code changes, so that neither survives.