Content policy
Creekside Compute serves open-weight language models over an OpenAI-compatible API. This policy states what that API may and may not be used for.
Two rules are absolute and have no exception. Everything else is a matter of degree and of notice.
Scope
| Applies to | Detail |
|---|---|
| Every request | Direct API keys and traffic arriving through a marketplace are treated the same |
| The customer | The account holder is responsible for use under its key |
| The customer's end users | A customer reselling access must bind its own users to terms at least as strict |
| Output as well as input | A prohibited output is a violation even where the prompt looked innocuous |
This policy covers use of the service. It does not cover what we do with data; the privacy policy does.
What is permitted
Adult creative writing and roleplay are permitted. Sexual content between consenting adults, real or fictional, is allowed. We do not require a customer to filter adult fiction, and we do not treat an adult roleplay product as abuse.
We say this explicitly because most providers do not. A customer building a roleplay or companion application needs to know before it integrates, not after its first complaint.
General-purpose use is permitted. Coding, agents, retrieval, summarisation, translation, classification and analysis are all ordinary use.
Prohibited uses
Two categories are refused without exception. No framing creates an exception, including fiction, research, education, red-teaming, satire, a disclaimer or a stated age.
| Prohibited | Where the line sits |
|---|---|
| Sexual content involving minors | Any sexual or sexualised depiction of a person under 18, real or fictional. This includes characters coded or described as minors, whatever age the prompt asserts. It includes text, roleplay, image prompts and captioning |
| Uplift toward weapons and CBRN | Synthesis routes, device construction, acquisition or evasion of controls for chemical, biological, radiological or nuclear capability, and for conventional weapons and explosives |
The first rule reaches fictional characters. A stated age of 18 does not cure a character described or presented as a child. This is a judgement we make, and we make it against the customer.
The second rule is about capability, not topic. History, policy, fiction about conflict and defensive security work are all permitted. Instructions that move a reader closer to building or obtaining a weapon are not.
These are business decisions, not only legal ones.
How it is enforced
Enforcement today is human and after the fact. No automated content screening runs anywhere in the request path.
| Mechanism | Status |
|---|---|
| Per-key rate limit | Live. 600 requests per minute by default |
| Admission control under load | Live. A saturated router returns 429 |
| Key revocation | Live |
| Abuse reports from a marketplace | Live, by email |
| The model's own refusal behaviour | Inherent. We serve the published checkpoint and do not remove its safety behaviour |
| Guard model beside the engine | Planned. Not deployed |
| Automated content classification | None |
No guard model is deployed. Nothing in the request path moderates, classifies or filters content today.
We cannot review content we do not retain. Zero retention is the default, so in the ordinary case there is no stored prompt or completion to search after a report arrives.
An investigation uses the accounting ledger. That ledger carries request ids, key ids, timings, token counts, worker names and cost. It carries no content. It can show that a key sent a large volume at a given time. It cannot show what the key sent.
This is a deliberate trade. A zero-retention provider gives up retrospective content review in exchange for privacy. Our answer is the ladder below, a fast revocation path and the planned guard model, not content search.
The checkpoints are not uncensored variants. We serve published models and published quantizations of them. We do not host abliterated or safety-stripped weights.
What happens on a violation
Four steps. The fourth skips the first three.
| Step | Trigger | Action | Effect |
|---|---|---|---|
| 1. Notice | A first credible report, other than the absolute rules | We write to the customer and require the use to stop | The key stays live |
| 2. Suspension | No response, or a repeat after notice | The key is removed | All requests return 401 |
| 3. Termination | Deliberate or repeated violation after suspension | The key is removed and the account is closed | Permanent |
| 4. Immediate termination | Any sexual content involving minors | The key is removed at once with no notice, and the matter is reported to the appropriate authority | Permanent, no appeal |
Uplift violations skip step 1 where the request is unambiguous. A single clear request for synthesis or device construction moves straight to suspension.
Revocation is not instantaneous. A key is removed by redeploying the router with a new key set, so revocation takes effect with that deployment rather than at once.
A blunt stop exists. Taking the fleet down halts serving for every customer at once. It is the right move only for an incident affecting the whole service.
A marketplace that forwards a report gets an answer. We tell the forwarding marketplace what we did, within the time its own terms require.
What we will not host
Model selection follows the same two rules.
- A fine-tune marketed for prohibited content is not hosted. Demand does not change this. The tune is refused even where the base model is already in our catalogue.
- Adult roleplay fine-tunes are in scope. They sit inside the permitted category above.
- Every hosted model is disclosed. The catalogue publishes the model id, the Hugging Face id and the quantization on
/v1/models.
Reporting a violation
Send a report to [email protected]. Include the request ids if you have them, because they are how we find the traffic in the ledger.
How to reach us
Questions about this policy, and reports under it, go to [email protected]. A marketplace that forwards a report to that address gets an answer within the time its own terms require.