Privacy policy.
This notice describes information handled by the Zils website and its decision-model services, including sign-in, training, model requests, and support.
Accounts and browser storage
Zils uses Supabase for sign-in. Your email address, account identifier, and authentication records are used to manage access to your workspace. Training jobs and API-key records are associated with your account.
Your browser stores your sign-in session and light or dark theme preference. Choosing Save setup in the playground also stores your question setup on that browser, without its example text or conversation. Signing out removes the sign-in session; clearing browser storage also removes saved setups and preferences. Neither action deletes your account or uploaded data.
Files you prepare and submit
Data review and preparation happen in your browser. Selecting a file or pasting data for review does not upload it. When you submit a training run, the prepared datasets are uploaded to private storage, and the service records the job settings, progress, evaluation results, and model artifacts. For text models it also saves the first question of each task family, without inputs or labels, so your playground can start from the questions your model learned.
The training service uses Supabase for account and job records. DigitalOcean Spaces stores newly uploaded datasets, photos, and model artifacts in private storage. Existing files may remain in Supabase Storage during migration. Authorized service operators process the data to validate your submission, coordinate training, evaluate candidates, and deliver available results.
Training workers and retained copies
Submitting a run requires permission to copy the learning dataset to assigned, approved training workers. Their operators can read and retain that data. Confidence-check and final-evaluation datasets remain with the evaluation service rather than being included in worker training exports.
Each run is open tier or private, chosen when you submit it. Private runs are trained only by Zils workers. Open-tier runs are free and may be trained by Zils network miners: independent operators admitted to the Zils network who accept its miner terms. They receive only the learning dataset, must delete it within 7 days of the run finishing and must not use it for anything else, but Zils cannot verify that they do. Choose open tier only for data you are allowed to share this way.
This training workflow is not confidential compute. Files, model artifacts, and local run records remain until operator cleanup; automatic deletion is not currently implemented. Canceling a job does not erase data already downloaded or immediately stop remote work. Expiring a download link does not remove a copy already obtained.
Model requests
New text-training submissions include permission for a one-time comparison with TypeSafe JEV after a model is approved. Zils sends the held-out evaluation inputs and questions to TypeSafe; it does not send their correct-answer labels, training examples, or confidence-check examples for this comparison. Zils retains the predictions and aggregate scores so viewing a model page does not repeat those calls. Earlier submissions are not automatically opted in. See TypeSafe’s privacy policy.
Running the playground or calling the model API sends your supplied context and questions to the model service to produce predictions. These requests are processed on servers, rather than only in your browser. Use information you are authorized to send, and avoid sensitive data in public examples.
When the setup assistant is connected, requesting suggestions sends your description, follow-up conversation, and any question draft being revised to its configured AI provider. This is separate from running a prediction or submitting training data. Choosing and editing a built-in starter does not call the assistant.
Optional voice entry uses your browser’s speech recognition service, which may process audio remotely. It starts only when you choose Speak your goal. The transcript appears for you to review before requesting suggestions.
What we keep from model requests
When you call the model API in real time with text on a paid account, we do not store what you send or the predictions we return. Your context and questions are held in memory to produce the answer and are discarded when the response is sent. They are not written to logs, databases, or backups, and they are not used to train models.
For each request we record the time, your account, token counts, and whether it succeeded. We use these records for billing, limits, and usage reports.
Other ways of calling the service do store data. Bulk jobs keep their inputs and results for seven days after the job ends so you can download them, and database backups can hold a copy for up to seven more days. Uploaded images are kept for 24 hours for a prediction and for 30 days after a training run ends.
Requests on the free tier are covered by the free tier terms, which allow requests and answers to be stored to train models.
Paid real-time requests are answered on servers that Zils operates. The hosting infrastructure in front of them may process technical records such as IP addresses and request times, as described under service providers below.
Website analytics
We use Plausible Analytics to understand page visits, referral sites, engagement, browser and device types, and approximate locations. Plausible does not use analytics cookies or store raw IP addresses. See Plausible’s data policy for details.
We remove URL query parameters and fragments before sending analytics events and send only the referring site’s origin. We do not send form contents, account identifiers, API keys, training files, or model inputs to Plausible. Private model pages, internal administration pages, the pricing lab, and individual shared artifacts are excluded.
Contact messages and service providers
If you joined the early-access email list, we store your email address in Supabase and use it for product updates. You can use the contact form to request removal.
The contact form sends your name, email address, optional company name, and message through Resend to the Zils support inbox so we can respond.
Vercel hosts the website; Supabase supports authentication and job records and retains existing file copies during migration; DigitalOcean Spaces provides private file storage; Resend delivers contact messages. Hosting, authentication, and delivery infrastructure may process technical records such as request details, IP addresses, and delivery events as part of operating those services.
Questions and data requests
Use the Zils contact form to ask about your information or request access, correction, or deletion. Include enough information to identify your account or job, but do not send passwords or API keys. We may need to verify the request. Copies retained by training-worker operators need to be considered separately from data held by the Zils service.
Changes to this notice will appear on this page with an updated date. External websites linked from Zils have their own privacy notices.