Don't trust us. Check us.
Every privacy page on the internet asks you to believe it. This one gives you commands. Below is exactly what you can confirm from outside, exactly what you can't yet, and the line between them — including the part that doesn't flatter us.
The scoreboard
| The admission gate enforces. On a gated Space, reads and writes are refused without an allow-listed identity — even if you hold the room key. | checkable now |
| Egress is locked. The app you loaded can reach only a published allowlist of hosts. The browser enforces it, not us. | checkable now |
| Witbitz can't read your data. The conversation is sealed to a key Witbitz never holds — the store holds only bytes the platform can't unlock, even with administrator access. | checkable now |
| The render is a reproducible build. Rebuild from published source and get the same hash AWS runs. | checkable now |
| Unobservable in use. That during the decrypt-and-render instant, the operator can't read plaintext out of memory — the Space's turns run inside attested AWS Nitro hardware. | Production |
The attested tier, in plain terms
The fifth row is now in production. Run a Space inside
the enclave and its room key is generated inside AWS Nitro hardware and never leaves it, the
image measurement is reproducible from published source, and the whole attestation check —
hardware signature, certificate chain to a pinned AWS root, a nonce for liveness — runs in your
browser against the enclave's published measurement (/enclave-claim.json). Launch-on-demand,
scale-to-zero when idle.
Enabling it is a per-room choice. The platform can run any room in the enclave; the app builder decides which rooms use it — so it is a property of a Space that turns it on, not of every Space by default. And inside the enclave, the model call no longer goes to a frontier vendor: it leaves the measured image only through an attested no-log gateway into confidential-compute inference on an open-weights model (fail-closed if that gateway doesn't attest). Keeping the frontier model — where the vendor does see the turn (next section) — is a per-room choice.
What is never hidden
Sealing content does not seal its shape, and no amount of cryptography changes that:
- Metadata. That a room exists, when it changes, and roughly how large each write is.
- The model provider — on the frontier trade. A turn's text reaches whichever model answers it, so a room that keeps a frontier model (the default Space) exposes its turns to that vendor. The confidential/attested tier now closes this: its model call goes only to attested confidential-compute inference on an open-weights model — shipped, not a roadmap. Keeping the frontier model is a per-room choice.
- Availability. We can refuse to serve you. We can't read you.
- Reproducibility isn't behaviour. It proves the running code is the code you can rebuild. Whether that code does what we say is what reading it is for — which is why the render is small and published.
If you'd rather not trust us at all
Run it yourself. The runtime ships as a container: your cloud tenant, your model endpoint, your object store, your identity provider. Same reproducible build — so you can confirm the image you're running is the one we published.
Two routes to the same property. One asks you to check our infrastructure; the other removes it from the question. Neither asks for the benefit of the doubt.