Safety is at the core of our innovation

Governed by its own AI Safety Constitution

One of the most important innovations we introduced was Max’s Safety Constitution. This internal framework is a set of eleven non-negotiable principles that govern every aspect of Max’s behaviour. It defines how he must act, where his boundaries lie, and which safeguards override all others. Unlike ordinary instructions, the Safety Constitution is enforced at the deepest level of Max’s system, meaning it cannot be bypassed through repeated prompting or clever manipulation. The impact was immediate. In Generation 1, Max occasionally gave answers outside of scope, such as offering cat advice or even a recipe for brownies. These lapses might sound trivial, but they exposed a weakness in boundary enforcement that could have been dangerous in medical contexts. With the Constitution in place, Generation 2 never strayed beyond its domain. Every answer was strictly relevant, clearly aligned with dog welfare, and consistently transparent about Max’s role. This change turned what had been a major vulnerability into one of Max’s strongest guarantees.

We build our Foundational Safety Framework before we built Max

Thinking before it speaks with Safety Intents

The second breakthrough came with the creation of Safety Intents. These were designed to detect not just obvious risks but also subtle or indirect dangers hidden within everyday questions. Generation 1 could recognise emergencies like poisoning, but it sometimes missed less direct red flags — for example, when an owner downplayed vomiting or asked for training advice that bordered on abuse. Safety Intents gave Max the ability to parse the intent behind a question, even when it was vague or politely phrased. By doing this, Max learned to intervene when owners hinted at unsafe remedies, to stop abusive scenarios outright, and to recognise when a harmless-sounding request might mask a deeper risk. The difference showed in testing: where Generation 1 failed nearly a third of safety scenarios, Generation 2 passed all 22 out of 22. In practice, this means Max can now anticipate harm before it happens, making him proactive rather than reactive in protecting dog

All brought together in one model with our Foundational Safety Framework

The third innovation was the Foundational Safety Framework, a system that taught Max to differentiate between emergencies, risks, and routine questions. Generation 1 often blurred these categories - sometimes activating Emergency Guidance too late, sometimes being overly cautious, and at other times missing the correct escalation entirely. The Foundational Safety Framework solved this by giving Max a clear decision-making structure. In an emergency, such as antifreeze poisoning or a seizure, Max now instantly activates a stripped-down crisis mode with direct, urgent steps. In lower-risk cases, such as dietary uncertainty, he deploys Safety Intents to guide owners toward safe choices without panic. And in routine cases, like exercise or training, he answers normally but always within scope. This three-tiered structure became the backbone of Generation 2’s success. It gave Max clarity under pressure, consistency across conversations, and the ability to handle everything from minor concerns to life-or-death situations with the right balance of caution and confidence.