Introduction

You write down what your business stands for: voice, values, banned phrases, judgment calls. We learn an internal activation direction from your spec plus demonstration examples, attest the behavior against the same policy, host it for you, and hand you an API key. When the model gets something wrong, you correct it, and the correction becomes control signal for the next alignment cycle.

Open models fail written brand and compliance policies out of the box. We measured brand customer-service at 83% violation on a popular 7B. Prompting helps and still leaves the model failing. Applying policy control inside the model, layering output enforcement, and attesting the full served system is the product.

The loop

values + examples ──▶ policy artifact ──▶ attestation ──▶ endpoint + API key
        ▲                                                    │
        └──────────────── corrections ◀──────────────────────┘

• Steer. Describe your business and pick a base model. Guided research can draft context from your site; the values wizard is where the policy is written.
• Train. We build paired positive and negative examples and learn their activation direction.
• Attest. The same values become an adversarial policy suite. You get a hash-verified report card bound to the exact checkpoint: clauses passed, violation rate, over-refusal.
• Serve. Your model goes live on managed GPUs behind a private, OpenAI-compatible endpoint.
• Correct. Flag bad outputs with the response you wanted. A retrain folds them in, same endpoint, same key, and re-attests.

What you get

Every ready model is served at api.opensteering.com, authenticated by your account API key. The endpoint speaks the OpenAI chat completions protocol, so existing SDKs work by changing two lines. Usage is metered per token and billed monthly. The model detail page shows the live pipeline and the policy attestation report for the latest evaluation.

Start with the quickstart, or jump to calling your model if one is already live.