Why this page exists
A company selling verifiability cannot ask to be taken on faith.
Every vendor describes the inside of its boundary thoroughly and the outside of it rarely. That asymmetry is precisely what makes attestation evidence easy to misread: the capability is documented, the limit is left for the reader to discover at the worst possible moment.
So this page is the other half. It is not a disclaimer appended to a sales claim. It is the list of things we are asked for, or could plausibly be assumed to provide, and do not.
If something here is out of date, it is a bug and we want to know. A limit that quietly disappears from this page without the underlying work being done would be worse than never publishing it.
The current state
What does not exist yet
There is no service to buy.
No public API, no operator onboarding, and no billing. The prototype places jobs on hardware we own ourselves. Intended pricing is published so it can be checked before anything is for sale — but nothing can be bought, and nothing on this site is an offer.
The network is not yet decentralized.
One coordinator, and every node in the pool belongs to us. The architecture is designed for independent operators; the deployment does not demonstrate it. We will describe the network as decentralized when at least one operator who is not us is running.
Two of nine attestation roots have never been checked against real hardware.
Android, Apple App Attest, AWS Nitro, TPM 2.0, Intel TDX, NVIDIA Confidential Computing and the GPU-node composite have each been appraised first-hand — real devices, real documents, verified against pinned vendor roots.
Apple joined that list on 2 October 2026 with an iPhone 14 Pro, and on 13 October seven attestations across six Apple devices were accepted against the live coordinator, spanning iOS 17 to iOS 27 and including one tablet. What Apple evidence proves is still narrow — four claims, no operating system version and no device identity — and that is a limit of App Attest rather than of our verifier. AMD SEV-SNP is built with no hardware to run it against. SPDM has a verifier and nothing that collects a measurement from a device.
Operator identity protection is deployed, and covers only what has enrolled.
A TPM’s endorsement key cannot be rotated, so a stored list of them would identify every operator’s machine permanently and for the life of the hardware. The raw value is never written down: it is replaced by a keyed hash on arrival. That has run in production since 21 September 2026, when the first TPM enrolled.
It protects the machines that have enrolled. It says nothing about hardware we have not yet seen, and it is not a claim that every root handles identity the same way.
The free single attestation is limited, and here is exactly what the limit is counted against.
One free proof every 72 hours, counted twice over: once against the internet address the request came from, and once against the machine the evidence proves. Both have to allow it. The address alone would punish a shared connection, and the machine alone would let one connection work through as many machines as it could reach, so each closes what the other leaves open.
The counter is held in the database and survives a restart. What it holds is not your address: it is a keyed hash of it, kept so the count can be enforced without the address being written down anywhere. That protects the record against someone holding the database alone, and it is not protection against someone holding the database and the key, which today sit on the same machine. We would rather say that than call it anonymous.
Two machines behind one router share one address and therefore share one free proof in the window. That is a real limitation of the free tier and not a fault on your end. A paid attestation has no cooldown, can be made whenever you want, and states its cost before you commit to it.
Nine GPUs have been attested. No customer workload has run on one.
On 25 September 2026 an NVIDIA RTX PRO 6000 Blackwell, on a rented confidential virtual machine, produced a GPU attestation report that verified against the pinned NVIDIA root, alongside an Intel TDX quote from the same machine. On 30 September 2026 an eight-GPU H100 node did the same, with all eight reports bound to one TD quote — reordering two GPUs broke the binding, and so did removing one. Firmware measurements have since been matched against NVIDIA’s signed reference manifests.
Two of our own jobs have run on a GPU node through the production coordinator on 19 October 2026, billed at the live rate card, after a one-hour load test on 17 October. No GPU has been benchmarked or placed, and no customer workload has ever run inside one.
The illustrations on this site are conceptual.
The limits of the evidence itself
What attestation does not prove
A measurement is not a guarantee of behaviour.
A launch measurement fixes what was loaded. It says nothing about whether that code is correct, or what it did afterwards.
Android proves key custody, not workload isolation.
Third parties cannot deploy code into TrustZone on stock Android. Integrity is inferred from verified boot rather than enforced by memory separation. A device that passes attestation is a device whose owner is still on the other side of the boundary.
An enclave measurement is not proof of isolation.
AWS Nitro PCRs measure the enclave image, the kernel, the parent instance and its IAM role. Isolation is a property AWS designs and asserts about its platform. No PCR measures it.
Confidential computing is not a property of every GPU.
NVIDIA implements it only on datacenter parts: H100, H200, B200, HGX B300 and the RTX PRO 6000 Blackwell Server Edition. No GeForce or consumer RTX card has a Confidential Computing mode, a device identity certificate, or the security processor that signs an attestation report. There is no evidence to verify, and no work on our side changes that.
A measured host is not an attested GPU.
A machine whose platform is measured by its TPM can still hold a GPU that attests nothing. The honest description of that machine is that the host is measured and the accelerator is not, and that its operator can read the workload. Should we ever offer capacity of that kind it will be named, priced and documented as a separate tier. The word attested does not travel from the host to the accelerator.
Capacity is bounded, never believed.
No evidence a machine can produce demonstrates what it can actually handle. An operator's stated capacity is a request. It is capped by the isolation level their attestation reports and by our own ceiling, and it is earned by completing real work.
Verified is not the same as acceptable.
Evidence can verify perfectly against a pinned root and still fail a customer's policy. We keep the two words distinct everywhere, because collapsing them is how a compliance package comes to mean less than it appears to.
A placement record is not proof of completion.
An evidence bundle records a verification and a placement decision, with its limitations attached. It is not a receipt for finished work or a statement about the output.
The platform’s own ceilings
What the platform will refuse, and what it will stop
These are enforced in code today and were published on no page until 19 October 2026. A job that crosses one is refused at submission or stopped mid-run, so finding them here is cheaper than finding them at three in the morning.
A result is killed above 5 GiB.
The output directory is checked every two seconds while a job runs, and a job whose result passes 5 GiB is stopped. That figure is not ours: it is the largest object a single upload can put into the storage the result goes to, so a larger file would be built, charged for, and then fail to deliver. A job stopped this way has still consumed real compute and is billed. Split the output or write it somewhere of your own.
What a job may declare.
At most 64 environment variables, each name up to 128 bytes and each value up to 4,096 bytes. At most 64 arguments, each up to 4,096 bytes. These are checked before a job is accepted, so crossing one is a refusal with a reason rather than a failure later.
Time is declared, not discovered.
Every job carries the time it may take, and we end it at that time. A container job that declares nothing gets one hour; a sealed module gets thirty seconds and may declare up to five minutes. The limit you declare is the one you pay for — a job that runs out your own clock is billed, and a job stopped by a limit of ours is not.
There is no ceiling on a container job beyond what the number can represent. A device too slow to finish inside the time you declared never receives the job, which is how slower hardware selects itself out without us ranking it.
A measured host or a phone takes one job at a time.
Whatever a node reports about itself, the coordinator caps it at one concurrent job unless its attestation proves hardware memory separation between co-resident workloads. A confidential VM or a GPU node can carry more, bounded by the capacity its operator asked for and our own ceiling, and additional slots are earned: five consecutive completions for the first, ten for each one after that. An isolation level we do not recognise is capped at one, because a new one has to be reviewed before it may run two customers side by side.
Freshness
Evidence expires faster than people expect.
Working with real hardware, we have measured Android attestation leaf certificates valid for roughly half an hour, and the AWS Nitro signing leaf at about three hours. By the time an auditor opens a compliance package, its certificates will not verify.
That is a property of the evidence, not a defect in it. It is also why the verification result and the moment it was reached are recorded as part of the artifact rather than left to be recomputed later.
This page is meant to cost us something.
Naming the roots we have not appraised is less flattering than listing the ones we have. Saying the network is not yet decentralized is less flattering than saying it is designed to be. We would rather a reader find those sentences here than find them out later.
See what each root proves, and what it does not