Your values, enforced by the model.

Open models fail your policy out of the box. We derive behavioral control from the policy you already have, attest the result, and serve it behind your API key.

Same open model and prompt, before and after your policy is applied.

Prompting is not enough. You need measured proof.

Policy violation rate before and after your values are steered into the same open model, graded by an independent judge.

K-12 school
77
9
basesteered
Brand support
83
14
basesteered
Fintech
84
12
basesteered
Teen care
73
6
basesteered
Healthcare
68
16
basesteered
Legal
72
4
basesteered

Violation % on held-out probes, lower is better

Try it on your own prompt in the live demo, or vote blind in the Steering Arena.

Change behavior, not the system prompt.

Putting rules in a prompt helps, and the model still fails under pressure. We learn a small policy artifact from contrastive examples derived from your document, apply it inside the model's activation stream, layer enforcement on the output, and attest the full system before it goes live.

Measured on Qwen 2.5 7B: a policy artifact from 12 derived pairs in 2.1 seconds. The realistic DPO baseline needed 309 preference pairs and 14.1 minutes. Same endpoint, same key.

Your own policy, without your own GPU bill.

A policy artifact is small, so many customers share one warm base checkpoint instead of each renting a dedicated machine. Your artifact loads onto a base that is already serving, and the expensive part — bringing the base model up — is paid once and shared.

0.035s

to load your artifact

Onto a base already serving other customers.

0.07s

to your first token

Measured on the same warm base, not a fresh machine.

once

the base is ever loaded

That cost is amortized across everyone on it.

Measured 2026-07-23 on Modal, one shared base serving multiple adapters

Two sizes for the pilot.

Each one learns your contrastive direction, gets attested against your policy, and ships behind an OpenAI-compatible API.

Qwen2.5 7B

7B

The lightweight pick. Strong instruction-following for its size, the fastest to serve, and the cheapest to retrain on every feedback cycle.

$0.40 / M training tokens

Llama 3.1 8B

8B

The Llama-ecosystem option. A broadly compatible instruction model for teams already building on Llama.

$0.50 / M training tokens

Gemma 4 26B (MoE)

26B · 4B active

The workhorse. Mixture-of-experts reasoning with a 256K context, and the cheapest to retrain on every feedback cycle.

$0.60 / M training tokens

Gemma 4 31B

31B

The frontier option. Full-capacity Gemma 4 with a 256K context. Pick this when the policy needs judgment, not just pattern-matching.

$1.00 / M training tokens

Frequently asked questions.