Data policy
Data policy
Version 2026-11-20 (draft v0.2, pending counsel review). Legal wording is pending review by counsel.
The short version. Vouch Compute processes jobs on a distributed network of consumer devices owned by independent contributors. Anyone running your job can see its input. Do not send personal data, sensitive data, or anything you would not publish.
1. What you may send
- Public audio and video: podcasts, public Spaces and AMAs, conference talks, published media.
- Public or non-sensitive text: documentation, forum posts, on-chain data, open web corpora.
2. What you must not send
- Personal data within the meaning of GDPR or similar laws, unless it is already public and you have a lawful basis to process it.
- Special-category data: health, biometric, financial account, government ID, children's data.
- Confidential business material, private calls or meetings, legal or HR recordings.
- Anything illegal to possess or distribute.
Every console account must accept this before its first job. API requests default to the strictest routing class.
3. Where your data goes
- By default (
X-Pool-Data-Class: sensitive), jobs run only on hardware we control or on vetted cloud GPUs. You can opt in to the distributed network (public) after accepting this policy. That is where the lower prices come from. - Marketplace traffic runs only on our own and vetted cloud hardware until the listing discloses third-party devices.
- Audio is converted on our servers before dispatch. Devices receive short chunks, never your original file.
- Inputs are never written to a contributor's disk. They are deleted from the device when the job ends.
- Images (when supported) are hash-matched against known abuse material on our servers before any dispatch.
4. What we do not claim
- We do not offer zero data retention yet. Inputs and results are kept for the times in section 5.
- We cannot prevent a contributor from observing an input their device processes.
- We are not a processor for personal data under GDPR Article 28. Do not use us for personal data.
5. Retention
Inputs and outputs are normally deleted within 24 hours of the job finishing, and never kept longer than 72 hours. The only reason to keep them past 24 hours is an open verification or review of that job. Deletion runs automatically.
| Data | Kept for |
|---|---|
| Inputs and outputs on our servers (uploaded, fetched or produced) | Deleted within 24 hours of the job finishing; kept longer only while that job is under an open verification or review, and never more than 72 hours |
| Inputs on contributor devices | Only while the job runs; never on disk |
| Results of asynchronous and batch jobs, waiting for you to download them | 7 days, then deleted |
| Records that stop a retried request from running twice (idempotency) | 24 hours |
| Usage metadata: units processed, hashes of inputs and outputs, and job receipts (no content) | Kept, for billing, contributor pay and verification |
| Billing and metering records | As long as required for accounting and law |
| Re-run samples used for verification | Same as inputs: within 24 hours, never more than 72 hours |
6. Output integrity
We re-run a sample of work on our own hardware and filter outputs for anomalies (for example, links or text with no support in the audio). Sampling can't catch every bad output. If you need every job run on hardware we control, use the default sensitive class.
7. Contributors
Contributors agree not to copy, store or share inputs. Devices run jobs in a sandbox with no network access beyond the job. Breaking this ends the contributor's participation and unpaid earnings can be clawed back.
8. Contact
Abuse and data requests: info@vouchcompute.com (security issues: security@vouchcompute.com).