Trust & safety
An identity that develops, within boundaries that hold.
An AI that changes through experience has to stay within the limits you set. LSF is designed with guardrails at every step: limits on what it can do, checks before anything becomes part of the identity, and review over time to keep it within bounds.
And when something does change, there’s a record of what changed and the reason recorded for it.
For example, a learner finally gets fractions. The tutor may learn that patience works with them, but only if that passes LSF’s built-in rules, your boundaries, and a review.
How LSF is designed to work.
Guardrails at every step.
Development is the point. Drift is not. These are the controls LSF is designed around, and the questions your safety, legal and procurement teams will ask first.
LSF is in development. Some of these controls are working in our own development today, some are still being built, and some are agreed with each partner. We’ll walk you through which is which.
Working today
Boundaries kept apart from experience.
Agreements and core commitments are designed to set what the identity will and won’t do. They are kept separate from everyday experience, so a conversation can shape how it relates to someone without changing its limits.
Being built
A final check before it responds.
A last check is being built to hold back what shouldn’t be said or done, on top of the underlying model’s own safety training.
Working today
Change is proposed, not automatic.
Most experience never becomes part of the identity. It can propose that something should last; layered rules and review decide whether it does.
Working today
Evaluated for drift.
The identity is designed to be checked against its boundaries over time and around model updates, so a slow drift in the wrong direction can be spotted through review.
Working today
Actions are governed.
The identity can only request an action. Permission, safety and readiness checks decide whether it happens, and a claim that something happened is not accepted as proof. It has no tools beyond the ones it’s given, so if someone talks it around, those are still the only tools it has.
Set with you
Set with each partner.
The rules, personality and boundaries for your product are set with you before work with real users begins.
See what changed.
A record connects the experience, the change and the later behavior, so your team can investigate when something looks different. An AI’s own explanation is not the same as an independently verified account, so the record is kept separately from what the AI says about itself.
Conceptual layout. Not a product screenshot or a verified log.
Agree the responsibilities before the work begins.
Before a pilot or partnership involving data, the arrangements need to be clear: what is held, who can access it, who can authorize changes, what can be corrected or removed, and what happens when the work ends. These are part of our partner discussions.
Be clear about the limits.
“Identity” describes continuity and development in a system. It is not a claim of consciousness, personhood or human equivalence.
LSF makes no medical or clinical claims. Development observations are not the same as validated performance across users, products or models.
Questions from your security or procurement team?
Ask us about your requirements. We’d rather answer them early.