Trust & security
How Vouch Compute keeps the numbers honest
Networks like this fail in two ways: fake devices claim work they never did, and buyers' data ends up somewhere it shouldn't. Here is what we do about each, and what we can't do.
For buyers
- Every audio file is converted on our servers to one plain audio format before it reaches any device. Devices never open your original file.
- A device gets the input only after it accepts a job. The link works once, for that device, until the job deadline.
- Inputs are never stored on contributor disks and are deleted after each job.
- Images, when we add them, are hash-matched against known abuse material on our servers before dispatch.
- Your bill and the contributor's pay come from the same server-side meter. Device-reported counts are never used.
- API keys are shown once, stored hashed and scoped. Sensitive actions need a fresh second factor.
- We can't hide inputs from the person running the device. Consumer GPUs have no confidential-computing hardware. That is why we only accept public and non-sensitive data.
- We don't offer zero data retention yet. Inputs and outputs are deleted within 24 hours of the job finishing, never more than 72 hours; results waiting for download are kept 7 days (data policy).
For contributors
- Jobs run in a sandbox with no network access beyond the job itself.
- Only models signed by our registry key run. Pickle files are refused outright.
- The app runs when your computer is idle by default, and it backs off when you use it.
- You're paid for verified work in USDC, sent straight to your own wallet. We never hold your funds in a pooled wallet or swap them.
- Every wallet change needs a signed message from that wallet plus a second factor, and starts a 72-hour payout freeze.
- Failed checks go to human review, not to automatic forfeit, and you can dispute any decision.
How work is verified
Payment is released only after a job passes these checks:
- Server-side counting. We measure audio length from the file we send and count tokens with the model's tokenizer before dispatch.
- Signed benchmarks. Each device runs a benchmark we sign, lasting at least 20 minutes, at enrolment and again at random times. Results are bound to a hardware-backed key where the device has one.
- Plausibility. Each job's speed is compared with the device's own benchmark and with what its hardware class can plausibly do.
- Known-answer jobs. Some jobs have answers we already know. They look exactly like normal work.
- Re-runs. Across the network, about 1–5% of work is re-run on hardware we control and compared within tolerance (word error rate for transcripts, cosine similarity for embeddings).
- Holds and escrow. Earnings are held for 72 hours (24 hours on trusted devices), and new devices earn into a 21-day escrow. Results appear on dashboards in daily batches with at least a 24-hour delay.
- Caps. Daily caps per device, per network and per verified identity bound what any one participant can earn, whatever their hardware claims.
We don't publish detection thresholds, per-tier check rates or which jobs were checked. Publishing them would help exactly the people the checks are meant to catch.
Reliability tiers
Each device has a reliability tier: Premium, Stable, Less stable or Unpredictable, plus "New device" while it builds a track record. The tier is worked out once a day from what our servers observe over the last four weeks (when the device is online, how often it disconnects, whether it finishes the work it accepts, and how long it has worked for us). Nothing a device reports about itself counts.
- Moving up takes weeks of regular work, one step at a time. Long absences move a device down.
- A tier belongs to one device on one account. Re-installing, moving the device to another account or recovering an account starts it again.
- Premium needs a verified identity and a hardware security chip.
- Higher tiers get work first and a modest bonus on verified work (published on the rate card), paid from the same fixed weekly budget.
- Only Premium devices can carry requests from public AI marketplaces, only non-sensitive ones, and only after that marketplace's listing says requests may run on distributed third-party devices. Those jobs are re-checked more often.
Contributors see their tier and how to improve it. We don't publish the cut-offs.
Subsidy discipline
At launch, contributors earn a multiplier on the base rate. It has a published end date, tapers as revenue grows and is funded from a fixed weekly budget, so the total cost doesn't grow with the number of devices. It pays for verified work only, never for being online.
Official domains
Phishing sites copy download pages. We use vouchcompute.com and its subdomains only, including:
vouchcompute.comapi.vouchcompute.com
We will never ask for a seed phrase or private key, and we never DM first. Report anything suspicious to security@vouchcompute.com. Email from us comes only from @vouchcompute.com.
Report a vulnerability
Email security@vouchcompute.com (English or Turkish). Include the steps to reproduce, what an attacker could do, and the affected URL, app version or API route. Our contact details are also in /.well-known/security.txt.
- Test only against your own accounts and devices. Don't access, change or keep other people's data.
- No denial-of-service, spam or social engineering, and don't try to move or withhold anyone's payouts.
- Give us a reasonable time to fix the issue before you publish it. We will tell you when it is fixed.
We read every report. There is no paid bug bounty yet; we will say so here if that changes.
Website security
- Strict Content-Security-Policy with per-request nonces and no third-party scripts.
- Session cookies are HttpOnly, Secure and SameSite=Strict. Your tokens never reach page JavaScript.
- Sign-in uses an email link plus a mandatory second factor (authenticator app or passkey).
- Analytics are cookieless and self-hosted, and they record no identifiers.
See also the data policy.