docs: gather privacy posture in one member-facing page

Adds guides/what-happens-with-your-data.md covering what the gateway,
chat server and each provider keep, why the routing path differs from a
marketplace router, and what the co-op does not claim.

Links it from Models, API access, Getting started, Guest access,
Infrastructure, the docs index, README and privacy policy. Fixes Getting
started's temporary-chat claim: temporary chats are server-backed and
deleted about an hour after creation, not never stored.
This commit is contained in:
inference-bot committed 2026-11-21 16:24:00 +00:00
1 parent 8d002bf5bc
commit be0ca6c53b
10 files changed
+178 -18

No files matched your search

+3 -1
View File
@@ -34,7 +34,9 @@ Two things worth being precise about:
[model list](models.md) distinguishes policy-based zero retention from
enclave-based protection.
Full detail, including what we can and cannot see: [privacy policy](../governance/privacy-policy.md).
Full detail, including what we can and cannot see:
[What happens with your data](what-happens-with-your-data.md) and the
[privacy policy](../governance/privacy-policy.md).
## Get a key
+8 -7
View File
@@ -67,13 +67,14 @@ with the account you created in step 1.
there's a tool you'd like available, ask in #general.
- **Voice** works in both directions: speak to the chat with the microphone
icon, and have replies read aloud.
- **Temporary chats** — for conversations you don't want kept on our servers
at all. Click the **Temporary Chat** toggle (in the chat header) and the
conversation is never saved: it disappears when you leave the page, and
nothing lands in your history or our database. To make this your default,
open **Settings → Conversations** and turn on **Default Temporary Chat**.
For anything sensitive, we recommend a temporary chat — ideally paired with
a 🔒 enclave model. See [Privacy](../governance/privacy-policy.md#4-what-we-can-see-an-honest-caveat)).
- **Temporary chats** — for conversations you don't want in your history. Click
the **Temporary Chat** toggle (in the chat header): the conversation is hidden
from your history immediately and deleted from our database about an hour
after creation. It still reaches our server while the chat runs, and it can
remain in backups until they age out. To make this your default, open
**Settings → Conversations** and turn on **Default Temporary Chat**. For
anything sensitive, pair a temporary chat with a 🔒 enclave model. See
[What happens with your data](what-happens-with-your-data.md).
Your usage draws from your monthly allowance; you can see it on the
[Member Dashboard](https://dashboard.inference.coop).
+2
View File
@@ -24,6 +24,8 @@ card also lists the guests you've sent.
[dashboard.inference.coop](https://dashboard.inference.coop) showing what's
left.
- **The weekly newsletter** and these public docs.
- **The same privacy rules as members** — see
[What happens with your data](what-happens-with-your-data.md).
Guest access is not membership — no Matrix, governance forum, or vote.
+1
View File
@@ -8,5 +8,6 @@ usage costs in energy and emissions.
- [Getting started](getting-started.md) — join, use the chat, and get involved in governance.
- [Guest access](guest-access.md) — inviting someone without a full membership.
- [Models](models.md) — the catalogue of providers and models.
- [What happens with your data](what-happens-with-your-data.md) — what the gateway, the chat server and each provider keep.
- [API access](api-access.md) — use the models programmatically.
- [Energy & carbon](energy.md) — the energy and emissions behind the service.
+10 -10
View File
@@ -37,6 +37,9 @@ Trade-offs are real: the most private option (Tinfoil) isn't the greenest, and
the greenest option (GreenPT) isn't the most private. The point of the co-op is
that **you choose** which axis matters more for a given task.
How the gateway and each provider handle your prompts is gathered in
[What happens with your data](what-happens-with-your-data.md).
**How to pick, quick guide:**
- **Default** → `greenpt/deepseek-v4.1-flash`: DeepSeek's efficient multimodal
@@ -309,13 +312,10 @@ reports back, and appears in your usage record alongside your chat calls.
## Limitations on privacy
Only **Tinfoil** offers architectural privacy. The other providers are chosen
for their value (renewable energy, public models) but do not run in enclaves —
prompts and responses pass through them in the ordinary way. Your **chat
history, including when using Tinfoil**, is also stored on our server so you can
revisit it, and that stored history is not encrypted in a way that prevents us
from technically reading it — we commit not to. Requests made through the **API**
with your own key are a different matter: those are **not stored on our side at
all**, beyond the usage record needed for metering ([API access](api-access.md)).
The full distinction — what's architecturally private versus what's a policy
commitment — is in the [Privacy Policy](../governance/privacy-policy.md).
Only **Tinfoil** offers architectural privacy. GreenPT and Regolo are chosen for
EU jurisdiction, contractual commitments and renewable energy, not for enclaves.
Your **chat history**, whichever model you use, is stored on our server; the
**API** path keeps usage records but no prompt or response text. The full account
— what the gateway keeps, what each provider commits to, and what we do not
claim — is in [What happens with your data](what-happens-with-your-data.md). The
formal policy is the [Privacy Policy](../governance/privacy-policy.md).
+146
View File
@@ -0,0 +1,146 @@
# What happens with your data?
A request passes through up to three layers, and each has a different
guarantee:
1. **Our gateway** routes the request and records what it cost.
2. **The model provider** runs the model.
3. **The chat server** saves the conversation, if you are using chat.
This page brings the privacy details together. The formal commitments are in
the [Privacy Policy](../governance/privacy-policy.md).
## The short version
- **In chat**, your conversation is stored on our server so you can revisit it.
Temporary chats are the exception: hidden from your history immediately,
deleted about an hour after creation, and only recoverable from backups until
those backups age out.
- **Through the API**, our gateway keeps a usage record and no text: the model,
token counts, cost, and which key made the request.
- **At the provider**, protection depends on the model you choose. Tinfoil runs
inference inside hardware enclaves. GreenPT and Regolo run in the EU under
GDPR, and Regolo's zero-retention claim is independently certified.
- We do not train on your data or sell it.
## The router: what it sees and what it keeps
The gateway is **LiteLLM**, running on the co-op's own server in Switzerland. It
has to handle your prompt in memory to send it to the provider. It does not
write prompts or responses to its logs or database. What it keeps is the usage
record that meters your allowance.
We test that instead of asserting it. A request containing a distinctive marker
is sent through the gateway, then the marker is searched for in the gateway's
logs, its request records, and its logging configuration. The marker does not
appear; the records carry usage fields only. The gateway's configuration and
code are public at [code/litellm](https://git.inference.coop/code/litellm), so
the logging settings can be read directly.
Two limits on that statement, because it is easy to read as more than it is:
- The gateway still knows **that a given key made a request**, and therefore
which member it belongs to. That link is how metering works.
- **"Not stored by us" is not "not stored."** What the provider keeps is a
separate question, answered in the next section.
## Why this is not a marketplace router
A common AI router sits in a chain: your request goes through the router, then
possibly another router, then a marketplace seller, then the provider. Every hop
can see the prompt, and some marketplaces retain prompts through ordinary
platform logging.
Our path has one hop we operate:
**Your client or our chat → our gateway → the provider.**
There is no marketplace, no reseller, and no second router in the path. We
contract each provider directly. Fewer parties handling the request means fewer
places it can be copied.
That is a reduction in exposure, not a guarantee about any provider. Dragos
researchers sent roughly two million prompts containing canary tokens through
commercial AI services and found prompt theft at marketplace resellers, providers
ignoring privacy settings, and breaches at three hosting providers — including
large ones
([their write-up](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens)).
Size is not the safeguard. Verifiable commitments, fewer intermediaries, and the
law are.
## What the providers keep
| Provider | Runs where | What is committed | Basis |
|---|---|---|---|
| **Tinfoil** | United States | Prompts and responses are not retained after the response is returned, apart from temporary prompt caching in encrypted enclave memory. Neither Tinfoil nor the co-op can read the exchange in plaintext at inference time. | Hardware enclaves, verifiable through remote attestation — a technical guarantee, not only a policy. |
| **GreenPT** | EU data centres in France and Finland | API payloads are processed in memory only and not stored. Error logs and monitoring metrics contain no prompt content. A DPA is available; its model-hosting sub-processors are in Finland. | EU jurisdiction, GDPR, and a contractual DPA. Policy-based: protection depends on GreenPT keeping its commitments. |
| **Regolo** (Seeweb S.r.l.) | Italy, EU | Prompts and outputs are not stored, logged, or reused for training. General account and connection metadata is still covered by its privacy policy. | EU jurisdiction and GDPR, plus a zero-data-retention certificate issued by an independent auditor on 24 July 2026. Policy-based with an external audit. |
Sources: [Tinfoil's terms](https://tinfoil.sh/terms),
[GreenPT's privacy policy](https://greenpt.com/privacy-policy), and
[Regolo's zero-data-retention page](https://regolo.ai/zero-data-retention/).
**Architectural versus policy-based.** An enclave protects you even if the
provider is compromised, because the provider cannot read the plaintext. A
policy protects you only as far as the provider keeps its word and the law is
enforced. Regolo's independent audit makes its policy claim checkable; GreenPT's
DPA and EU location give you a legal route if it is broken. Neither is the same
as cryptography.
## What the co-op can still see
- **Which member made a request.** Gateway usage records link a request to a key
and therefore to an account. This is how metering and budgets work.
- **Chat history.** Saved conversations sit in the chat database and its search
index, which the co-op can technically read. We commit not to; that is a policy
commitment, not an architectural guarantee. Temporary chats reduce the window
but do not remove the operator's technical access while the chat exists.
- **Operational logs.** Account provisioning, welcome emails, errors and access
logs contain the email addresses and connection metadata needed to run the
service. They are kept briefly and are not used to profile members.
- **Backups.** Cloudron takes whole-app encrypted backups nightly and keeps
seven days. A temporary chat present when a backup runs can remain recoverable
until that backup expires.
## Retention at a glance
| What | How long |
|---|---|
| Saved chats and their search entries | Until you delete them |
| Temporary chats | Hidden immediately; deleted from the database about an hour after creation; may persist in backups until they expire |
| API prompts and responses | Not stored by the co-op; provider retention is per the table above |
| Usage records (model, tokens, cost, key) | Kept for billing and account administration |
| Web-search queries | Stored with the conversation; DuckDuckGo and Brave receive the query when you use search |
| Operational logs | Kept briefly for running and debugging the service |
## What we do not claim
- **We do not claim "we never see your messages."** The gateway sits between you
and the provider by design, so prompts pass through it in memory.
- **Privacy is not uniform across the catalogue.** Only Tinfoil offers
architectural privacy. GreenPT and Regolo are chosen for EU jurisdiction,
contractual commitments and renewable energy, not for enclaves.
- **A temporary chat does not change who processes the request.** The model you
choose still receives your messages.
- **EU law is a constraint, not a force field.** GDPR gives you rights and a
complaints route; it does not make misuse impossible.
## Check it yourself
- Read the gateway configuration and patch history at
[code/litellm](https://git.inference.coop/code/litellm).
- Check Tinfoil's attestation documents, Regolo's zero-retention certificate,
and GreenPT's DPA against the links above.
- Ask us to repeat the marker test on the live gateway. Write to
[info@inference.coop](mailto:info@inference.coop) and we will run it and report
what we find.
## Formal policy and provider terms
- [Privacy Policy](../governance/privacy-policy.md) — the co-op's formal policy,
including your rights and how to request deletion.
- [Terms of Service](../governance/terms-of-service.md) — acceptable use and the
commitments that apply to members.
- Provider terms: [Tinfoil](https://tinfoil.sh/terms),
[GreenPT](https://greenpt.com/terms-of-service),
[Regolo](https://regolo.ai/terms/).