Legal
Responsible AI
Last updated 28 October 2026
Draft, pending legal review. Text in square brackets is still to be completed.
SparQ is infrastructure for other people's AI features, so our responsibility is to make the safe choice the default one and to make every decision traceable.
People decide what matters
Writes and destructive actions wait for a person by default. In production, destructive actions need two. A change to an agent's behaviour cannot be approved by the person who made it.
Fail closed
When SparQ cannot prove a call is allowed, because no model meets a client's residency rules, a budget is exhausted or an audit record cannot be written, it refuses rather than guesses.
Least authority
An agent's authority is the intersection of the verified user, the project's delegation, the agent's declared capabilities and the resource scope your application returns. Tools receive short-lived tokens for one purpose, never a user's own token.
Evaluate before release
Releases are gated by evaluation suites that include isolation, prompt-injection and cost tests. Live traffic is sampled for quality, and every negatively rated answer is reviewed.
Transparency and privacy
Every model call records which provider, deployment and region handled it. Prompt and response content stays out of telemetry unless your policy allows it.
Honest about limits
Fully sovereign deployments can only run open-weight models, because proprietary frontier models are not distributed as weights. We say so wherever it matters.