QuantumColleagues. ← Thoughts
Do some good with AI.

Thoughts / Bracton

The Bracton logo, a bar above a crown, on navy. Change the model. Keep the rules. Gated Autonomy builds it. The threshold model sets the line. Bracton holds it. AI Thoughts for SMEs, from QuantumColleagues.

Change the model. Keep the rules.

AI Thoughts for SMEs · 18 November 2026

A better AI model arrives every few months. That’s good news, but it raises an awkward question for a regulated organisation. If your safeguards live inside the model, what happens to them when the model changes?

Ours don’t live there. Admin Core runs inside Bracton, our deterministic safety layer around AI models. Bracton is ordinary code, not another model, and it meets the test we set for every safeguard: it holds without the AI’s cooperation. A rule that a model merely agrees to follow is an aspiration, not a control.

Bracton puts seven protections around the model.

We named it after Henry de Bracton, an English judge of the 1200s who is remembered for the line that even the King is under the law. Nothing sits above the rules in Bracton either, and that includes the model.

Bracton is the third part of how we work. Gated Autonomy is how we build and change our AI. Our threshold model sets the line for each colleague, so everyone knows what it may do on its own and what waits for a person. Bracton holds that line while the colleague works.

Because every protection sits outside the model, any model can do the drafting and the protections stay the same. We can move routine work to a smaller, cheaper model, or bring in a stronger one where judgement matters, and the checks, the approvals and the audit trail stay as they were. A new model still passes through Gated Autonomy before it goes live, but nobody has to rebuild your safeguards each time.

On 18 November the Commons Business, Innovation, Science and Trade Committee published Guardrails for growth, its report on AI safety and security. Its argument is one we share. AI will only spread through the economy if people trust it. Bracton complements the report from the other end, where AI meets your inbox, your records and the people you serve.

Your safeguards shouldn’t belong to whoever supplies the model. Models will keep changing, and your rules shouldn’t have to.

We license Bracton on its own, with support to fit it around your AI. Read about a Bracton licence.

Get in touch to find out more.

If you’d like to talk to us about this, then please contact us here or email us directly at information@quantumcolleagues.org.

Page 2: How Bracton works. Seven protections around the model. Guarded inputs: code cleans everything first. No direct power: it drafts, and it can’t send. Checked outputs: claims must match your sources. Written limits: code enforces every limit. People decide: a named person signs off. A full record: every decision is logged. Safe by default: in doubt, it stops. All seven sit outside the model. Page 3: Why it lasts. Any model fits in the middle, and the protections stay the same, whatever the model. The picture shows Bracton, ordinary code outside the model, with the AI model in the middle inside no direct power, guarded inputs coming in, checked outputs going out, written limits and people decide above, and a full record and safe by default below. Which protection matters most to you? Talk to us at quantumcolleagues.org.
Download the three page PDF →
Businesses buy our AI, and that funds free post-AI training for every generation.
Do some good with AI.

© QuantumColleagues Ltd 2026 · Company No. 17219359 · Tel 0333 880 0712 · 50% owned by Quantum Colleagues Training CIC (No. 17244034) · Photos: Pexels · Privacy · Follow on Google