Privacy Policy
Last updated July 24, 2026
Attestal is operated by Peter Lodri EV (2800 Tatabánya, Szőlődomb utca 77, Hungary; EU VAT HU69757200), the "we" in this policy. It covers this website and the Attestal fine-tuning service. The single most important thing to understand is that the two tiers treat your data in opposite ways, on purpose — so we describe them separately.
The one-line version
- Enclave (paid): your training data and trained weights stay in your cloud account. We orchestrate; we never receive, store, or access them. Privacy is the product.
- Open (free): your training data and the resulting weights are processed on our infrastructure and are retained by us, and may be published to a public corpus. Free means public. That is the price of the free tier, stated plainly.
What we collect on this website
- Nothing by default beyond standard server logs (IP, user-agent, timestamps) kept briefly for security and abuse prevention. No advertising trackers, no third-party analytics that profile you across sites.
- What you send us: if you email
partners@attestal.aior join a waitlist, we keep what you write (your address, company, what you'd fine-tune) to reply and to run the design-partner process.
The Enclave (paid) tier
Fine-tuning runs inside your own AWS/GCP/Azure account via a scoped credential you grant and can revoke. Your datasets, base models, and trained weights never leave your tenancy. We receive only the run metadata needed to orchestrate and to produce the audit record — stage, step, metric values, pass/fail against thresholds, timestamps, and error text. We do not receive your training examples or your weights, and we are structurally unable to. A separate data processing agreement (DPA) governs the enclave relationship.
The Open (free) tier
The free tier exists so people can use the real pipeline without a private cloud. In exchange:
- The dataset you submit and the weights produced are processed on our GPUs and retained by us.
- They may be published to a public corpus and/or open-model repository, under an open license, for anyone to use.
- Run metadata (model, stage, step, loss, verdict) is shown in the public live feed and kept as part of the open audit record.
- You must have the right to share whatever you upload. Do not submit personal data, secrets, or anything you are not free to make public — the free tier is not for confidential or regulated data. That is what the enclave is for.
You grant us a worldwide, royalty-free licence to host, process, and publish the free-tier dataset and resulting weights for these purposes. You can ask us to stop publishing futureruns at any time; already-published open artifacts cannot be un-published once others rely on them.
Legal basis (GDPR)
For website contact and security logs, our basis is legitimate interest and, where you contact us, performing steps at your request. For free-tier processing and publication, the basis is your consent and the performance of the free-tier terms you accept. You may withdraw consent for future processing at any time.
Your rights
Subject to law (GDPR for EU/EEA users), you can request access, correction, deletion, restriction, portability, and objection for the personal data we hold about you, and lodge a complaint with a supervisory authority (in Hungary, the NAIH). Email privacy@attestal.ai.
Retention, security, subprocessors
- Retention: server logs briefly; correspondence for as long as the relationship is active; free-tier artifacts per the open-corpus purpose above.
- Security: access controls, encryption in transit, and — the point of the product — signed, immutable audit records. No system is perfectly secure; we do not claim otherwise.
- Subprocessors: cloud/GPU hosting and email delivery. The enclave tier adds none on your data, by design. A current list is available on request and will be maintained here at GA.
Contact
Peter Lodri EV — privacy@attestal.ai. We answer these ourselves. Changes to this policy are posted here with a new "last updated" date and, for material changes to the free tier's data terms, announced on the build log.