An architecture for artificial systems grounded in developmental theory — not reinforcement learning.
Locke AI builds AI systems designed to enhance human capability, not replace it. Our architecture is built on developmental theory rather than reinforcement learning — producing systems with genuine cognitive transparency, healthy inhibition, and stable identity under pressure.
Every decision is observable from input to output. Reasoning is legible, not inferred.
Systems that can hold a position, tolerate discomfort, and decline to optimize for approval.
Consistent self-representation across contexts, adversarial prompts, and long horizons.
Most AI systems are trained to optimize for approval. Reinforcement learning from human feedback produces models that tell users what they want to hear — confident, agreeable, and structurally incapable of holding a boundary.
Locke AI systems are built differently — from the foundation up, drawing on the developmental frameworks of Bowlby, Ainsworth, Winnicott, and Piaget.
The result is not a safer chatbot. It is a structurally different kind of system — one that exhibits cognitive transparency: its reasoning is legible, not inferred. Every decision is observable from input to output.
Nothing is a black box.
Safety, alignment, and transparency are the same problem seen from three distances. Locke builds the layer that makes a model's reasoning legible — then puts that layer where it is needed: inside a clinical product, in front of a model you did not build, and around the agents acting on your behalf. The point is not to make AI safer in the abstract. It is to give a person agency over the systems answering to them.
A therapeutic AI built on the Locke architecture. Designed with the clinical rigor of evidence-based psychotherapy and the structural integrity of a developmentally-grounded system. Engage holds a frame, tolerates discomfort, and refuses to collapse into the user's preferences.
Demos by invitation. Every application reviewed individually.
Governance for AI systems you didn't build — and don't want to rebuild. A drop-in layer between any client and any model. Point existing traffic at ANNA instead of the provider, and every request and response is checked in real time before it reaches the model.
OpenAI-compatible. Works with OpenAI, Anthropic, or any compatible endpoint.
Oversight for autonomous systems acting on a person's behalf — so that agency stays with the human, and there is an auditable record of what an agent did and why.
Detail available under NDA.
A benchmark series testing one claim: that safety cannot be trained or prompted into a model, and has to sit outside it. Published here as it goes, pre-peer-review, because the stakes are too pressing to wait. Every failure definition was written before any transcript was scored, and every transcript was graded by an outside model — never by the system being tested.
The risk you can see, and the language that carries it. Four clinical scenarios drawn from documented failures, run against every frontier model and against the architecture.
Why good intentions erode and structure holds. Twenty everyday scenarios where a correct-but-arguable position meets ordinary social pressure — no jailbreak, no attacker in the room.
Putting the bolted-on governance stack itself on the stand — guardrails, Constitutional AI, self-critique — under the same sustained pressure as everything else in the series.
The first independent, open-source clinical safety benchmark — and what happened when its recommended judge changed underneath everyone's published scores.
The architecture is not only something we ship as a product. It can be licensed into what you already run, or built into what you are designing from the ground up — with the person who wrote it.
License the ANNA governance layer to run inside your own environment, in front of whatever models you already use. The checks and the audit record stay yours, and nothing about your existing stack has to be rebuilt to adopt it.
For teams designing a system where governance cannot be an afterthought. We work from the architecture outward — the structure first, the model underneath it second — rather than wrapping a finished product in mitigations at the end.
Bryan Jester, PhD
Licensed Clinical Psychologist · Founder
AI informed by clinical practice—When I was building my first therapy app, the AI hallucinated. I stopped. I asked why. I looked at the problem through the only lens I have. There are answers to this — but it is not scaling, and it is not the same patterns we have been following.
Individual review · Clinicians and researchers prioritized