Data handling
Where your requests run, what we keep, who can read them, and where each of our commitments stops.

What happens to a request
We process your request in the region you choose, return the response and keep no copy. Our logs record metadata, never the text. Batch is the one exception: its files wait until the job is done.
These commitments cover the Shoreheld API. Requests you send to another AI provider follow that provider's terms.
- Used to train a model?
- Never
- We don't train any model on your prompts or responses.
- Kept after we reply?
- No copy
- We hold each request in memory while we process it.
- Batch files?
- Until done
- Input deleted when the batch finishes; results on download or after 7 days. You can delete either sooner.
- Our logs?
- Metadata only
- Time, key, model, region, token counts and status. Error logs and backups too.
- Our staff?
- Can't read it
- No prompt or response text reaches anything they can open.
- The model's maker?
- Receives nothing
- Model servers cannot make outbound internet connections.
- Where is it processed?
- Your region
- The region whose endpoint you call. Account, billing and log records may be held elsewhere.
- Billing?
- Token counts
- Per key. Card payments go through Stripe.
- Other AI providers?
- Their terms
- Requests you send elsewhere, including through our gateway, follow that provider's terms.
Where requests run
Each region has its own endpoint, and we process each request in the region whose endpoint you call. A region is a country or bloc, not a named building.
UK (London) EU Switzerland Singapore
| Region | Processed in | Endpoint |
|---|---|---|
| UK (London) | The United Kingdom | https://uk.api.shoreheld.com/v1 |
| EU | The European Union | https://eu.api.shoreheld.com/v1 |
| Switzerland | Switzerland | https://ch.api.shoreheld.com/v1 |
| Singapore | Singapore | https://sg.api.shoreheld.com/v1 |
Service commitments and their limits
What stays fixed, what can change, and when we would act on your account.
| What you get | The limit |
|---|---|
| Each model ID points to one published version of the weights, at one precision. A new version gets a new ID. | Outputs can still vary from one run to the next, as with any model. |
| We give at least 90 days' notice before we retire a model ID. | A legal order could force a shorter notice. |
| We suspend an account only for non-payment, abuse or a legal order. | Our terms of service define each ground. |
| Before we serve a model, we check its files against the hashes its maker published. | Matching hashes prove the files are the published ones. Your own tests show how the model behaves. |
| The API follows the OpenAI format, so most existing clients need only a new base URL, key and model ID. | The other side must offer the same format. Hosted tools, connectors and other provider-specific features don't carry over, and a different model can behave differently. |
Reserved nodes
For sovereign capacity, a reserved node is whole machines allocated to you in the country you choose. Tell us the country and we will tell you which sites can serve it.