Acorda Labs · Models that answer to you

Frontier models are aligned to values written by a few people for a limited number of scenarios.

Outside of this scope, alignment breaks.

The people and organisations a model serves rarely see the principles it follows, can rarely change them and when they can, it's time and capital intensive. We think that's the wrong way round.

Our models are:

Sovereign

Alignment you own. Your values, on your infrastructure, in models that answer to you.

Efficient

Smaller, faster and cheaper by design. A model that fits its purpose beats one that tries to fit everything.

Aligned

Constitutional alignment at scale: principles derived from the organisation the model serves, traceable in every output.

From recorded decisions to a model you can inspect.

We derive a model's governing principles from an organisation's own recorded decisions, post-train against them on the organisation's infrastructure, and make every output traceable to a principle it approved.

Recorded decisions

Your own decision records: the judgements your organisation has already made and stood behind.

A derived constitution

Candidate principles surfaced from those records and ranked by how much of the record supports them.

Post-trained, on your infrastructure

An open model aligned to that constitution, running where your data already lives.

Every output, traced

Each output cites the principle it was trained against. Auditable by construction.

Our guiding principles

Alignment should belong to whoever deploys the model, not whoever trained it.

The principles a model follows should come from the organisation deploying it, from decisions it has already made and stood behind, and from within the real cultural and regulatory environment it operates in.

The future is a federation of small cooperative models, not one giant one.

Scaling parameters and burning compute is one way to buy capability. We think the more durable path is many smaller models, each aligned to its own context, cooperating.

Forward-deployed research

We embed to learn.

Every engagement is research. We craft bespoke models, governable by what's important and relevant to each organisation creating highly aligned, auditable models at scale. Each time we do, it improves our pipeline without compromising the safety or security of our partners.

Embedded with the people who decide

Our researchers work alongside the people who make the decisions, where the context lives.

Each engagement compounds

Each engagement extends the method: how constitutions are derived, evaluated and kept current as the rules change. The organisation keeps its deployed model, and the method improves for the next one.

Developed in the UK, deployed anywhere

The result runs on the organisation's own infrastructure, under its own governance, wherever that is.

We're taking on a small number of design partners working on high-stakes decisions.

Work with us

Built by researchers and builders, backed by experts.

Peer-reviewed at the field's top venues, with backgrounds spanning AI safety, imitation learning, security and human-computer interaction, paired with zero-to-one experience and builder mentalities.