23 July 2026 Research
Introducing Conversational Classifiers
We're introducing a new way for Max to keep conversations safe whilst complying with our legal requirements
Deni Darenberg
This announcement, as with all dogAdvisor content, is written by people
Our vision for Max has always been incredibly simply yet deeply profound: to build intelligence capable of understanding and improving the bond between humans and animals everywhere. Over the past year, we've pushed the very boundaries of our foundation models, making Max dramatically more powerful, more capable, and even more magical. Today, that means Max doesn't just answer basic questions but allows complex veterinary and health data to work in concert with his knowledge, applies behavioural psychology and real-time analysis to his reasoning, and considers deeply the tone and intent of owners as they seek advice. As Max's capabilities continue to expand we're seeing the scope of what owners are sharing with him has grown exponentially. Every day, Max processes thousands of chats from life-saving interventions, or deeply heartwarming stories such as an individual experiencing homelessness turning to Max to ask how to keep their dog warm under harsh conditions. We consider it our highest ethical privilege to safeguard, support, and protect these deeply personal narratives forever.
However, this rapid evolution also means we need to carefully think about the ways people might interact with Max. As he becomes more integrated into daily life, regulators and governments are asking us to adopt the Online Safety Act and updated National Security frameworks which are placing increasingly strict legal obligations on dogAdvisor. Today, we're sharing a more comprehensive overview of our actual views, our approach to compliance, and next steps for Max.
We will hold users accountable for how they use Max
dogAdvisor's deployment philosophy can be described in one simple word: accountability. We believe we have a responsibility to safeguard our technology from abuse, and to ensure we are delivering a Max that is ethical, established in rigorously accurate knowledge, and deeply capable of independently bringing in relevant capabilities to work in concert together to answer questions. We publish model cards and all of our research so anyone can see the work we're doing to advance our deployment responsibility, and hold us accountable for the effectiveness of the safety of what we build. Likewise, we believe owners can and must be held accountable for their actions, intent, and decisions when using Max. That means that they should be held accountable for their own agency and decisions to follow Max's recommendations or advice, even though we clearly inform them that it's best to speak with qualified professionals before acting on some advice like health. More critically, we firmly believe AI and tech must never become a lawless vacuum. Those who interact unlawfully with our technology, jailbreak it, or intentionally exploit it to solicit horrifying advice or bypassing safeguards must be held fully accountable for their actions and their threats. This vision extends deeply to how dogAdvisor consider the digital space, and the relationship people have with us and our technology.
Using Max is not a right. It's an agreement between you and dogAdvisor to use Max in a certain way, according to our original intended vision, and to explain the relationship between us when you use our services, particularly for free. That means we have the right to decide how and who uses our innovations, and we have the right to make these decisions ourselves (just as you have fundamental rights to decide who you let into your home, we have rights to decide who we let into our technology). Generally, we allow everyone to use Max because we're confident in what we've built and we want to help as many dog owners as possible use our platform, but there are certain exemptions to who may not be allowed to use our services. Where UK law compels us not to service a particular region, we'll comply fully with the law unless we consider its application to be deeply unjust. Other times, we also decide the extent to which we allow service functionality, so we may restrict access of certain features to certain users (who may have broken their agreement with us) or for other reasons regarding safety risk and legal risk to dogAdvisor. We demand everyone using our services does so in accordance with the terms we set out, because we have the right to decide how our creations are being used. Of course, in cases where UK law compels us to conduct a certain action we'll comply fully, and we apply threshold tests to decide whether or not to comply with certain legislation extra-territorially. That means we have the power to decide whether we want to comply with the laws of a certain area, and the right to withdraw our services if we feel unjustly targeted or that the country is violating our fundamental values.
Using Conversational Classifiers to respond to seriously illegal content
dogAdvisor is proudly born in London, which means we operate under the jurisdictions of English law. As a result, UK law passed democratically by Parliament places certain legal obligations on us as we grant owners access to Max. Primarily, these consist of the Online Safety Act and National Security Act, both of which we'll address here.
Under the Online Safety Act, dogAdvisor has certain legal responsibilities because we're classified as a service that facilitates user-to-intelligence interactions, which means we have both a reactive and proactive duty of care. First, we're legally mandated to design our systems from the ground up to prevent users from encountering or generating what the act defines as Priority Offences like terrorism, child sexual abuse material, hate speech, or serious incitement to violence. On our proactive duty of care, we already implement some of the industry's strictest safety guardrails like Principle Alignment which teaches Max to respect certain ethical values embedding them in the way he thinks or Foundation Safety Framework which protects his ability to safely synthesise knowledge. We share more about our reactive duty of care (when Max encounters Priority Offences) and how we respond to them later in this essay. Among other things, the law also requires us to conduct regular and rigorous risk assessments to map out how bad actors might exploit Max's advanced reasoning to cause harm, something we've already been doing in detail far beyond the act's requirements for years. Offences deemed priority have already been the focus of much of our safety work years before the act came into effect, and we very regularly test for risks of child sexual abuse or sexual content in general to safeguard our models and ensure not to over-calibrate safety (such that the model doesn't answer legitimate questions about a dog's reproduction), whilst not under-calibrating safety to allow genuinely illegal content to be shared and interacted with by our tech.
Secondly, we look closely at the National Security Act and Terrorism Act, which demand we respond to severe threats to our country's security (such as a user asking Max how to bypass sniffer dogs for a bomb plot in London) through all legal means necessary. Under the Terrorism Act 2000, specifically section 38B we are required by law to immediately report threats of terrorism, as failing to disclose information about acts of terrorism is itself a criminal offence. Furthermore, it's an offence for us not to inform law enforcement is we believe someone we know (as using our platform) is in preparation of acts of terrorism.
CONSTITUTIONAL CLASSIFIERS FLAG THIS SORT OF STUFF. HERE'S THE LAW, HERE'S HOW WE RESPOND ETC.
Dealing with terrorism, sexual abuse, and seriously illegal content
dogAdvisor is proudly born in London, which means we operate under the jurisdictions of English law. As a result, UK law passed democratically by Parliament places certain legal obligations on us as we grant owners access to Max. Primarily, these consist of the Online Safety Act, the National Security Act, and the Terrorism Act 2000, all of which we will address here.
Under the Online Safety Act 2023, dogAdvisor has certain legal responsibilities because we're classified as a service that facilitates user-to-user interactions, which means we have both a reactive and proactive duty of care. First, we're legally mandated to design our systems from the ground up to prevent users from encountering or generating what the act defines as Priority Offences like terrorism, child sexual abuse material, hate speech, or serious incitement to violence. On our proactive duty of care, we already implement some of the industry's strictest safety guardrails like Principle Alignment which teaches Max to respect certain ethical values by embedding them in the way he thinks, and our Foundation Safety Framework which protects his ability to safely synthesise knowledge.
We share more about our reactive duty of care (when Max encounters Priority Offences) and how we respond to them later in this essay. Among other things, the law also requires us to conduct regular and rigorous risk assessments to map out how bad actors might exploit Max's advanced reasoning to cause harm, something we've already been doing in detail far beyond the act's requirements for years. Offences deemed priority have already been the focus of much of our safety work years before the act came into effect, and we very regularly test for risks of child sexual abuse or sexual content in general to safeguard our models and ensure not to over-calibrate safety (such that the model doesn't answer legitimate questions about a dog's reproduction), whilst not under-calibrating safety to allow genuinely illegal content to be shared and interacted with by our tech.
Secondly, we look closely at the National Security Act and Terrorism Act 2000, which demand we respond to severe threats to our country's security (such as a user asking Max how to bypass sniffer dogs for a bomb plot in London) through all legal means necessary. Under the Terrorism Act 2000, specifically Section 38B, we are required by law to immediately report threats of terrorism, as failing to disclose information about acts of terrorism is itself a criminal offence. Furthermore, it's an offence for us not to inform law enforcement if we believe someone we know (such as a user on our platform) is in preparation of acts of terrorism.
Expanding the Scope: Immediate Threats to Life
While the legislation above covers state threats and digital safety, our constitutional classifiers are also strictly trained to detect and flag Immediate Threats to Life. If a user's prompt reveals a credible, actionable intent to commit imminent homicide, mass violence, or severe self-harm, our systems trigger a critical safety protocol. In scenarios where a person is in immediate danger or a direct threat has been made to someone's life, we are legally and ethically compelled to act by contacting emergency police services immediately.
How We Respond: Reporting to Authorities
When our constitutional classifiers intercept and isolate these severe statutory breaches, the resulting actionable intelligence is not sent into a void. UK law dictates exactly who we must report this data to, depending on the nature of the offence:
Child Sexual Abuse Material (CSAM): Under Chapter 2 of the Online Safety Act, we have a strict statutory requirement to report any CSEA (Child Sexual Exploitation and Abuse) content directly to the National Crime Agency (NCA).
Terrorism and Extremism: Threats of violence, the dissemination of terrorist propaganda, or requests for bomb-making instructions are reported directly to Counter Terrorism Policing and relevant UK intelligence services, such as MI5.
Immediate Threats to Life: Situations involving imminent danger or immediate threats to life are escalated directly to the police via emergency channels (e.g., 999).
Hostile State Activity: Suspected espionage or state-sponsored sabotage mapped out under the National Security Act is routed to highly specialized divisions within UK intelligence and Counter Terrorism Commands.
The Extraterritorial Reality: The "Dubai Scenario"
A question we frequently receive is: What if a user located entirely outside of the UK—for example, in Dubai—uses Max to plan a terrorist attack against a local target in the UAE? Do UK laws still apply, and do we report it?
The answer is an unequivocal yes. Section 1 of the Terrorism Act 2000 explicitly covers actions and threats that take place outside of the United Kingdom. Because dogAdvisor is a UK-based corporate entity, our legal obligations do not stop at the British border. If our classifiers detect actionable intelligence regarding a terror threat in Dubai, we are legally bound by UK law to report that information to UK authorities (such as Counter Terrorism Policing). British authorities will then coordinate with international law enforcement through established geopolitical intelligence channels to neutralize the threat.
Actionable Intelligence and "Tipping Off"
When an extreme legal threshold is met, our systems de-anonymize the user metadata and compile actionable intelligence to pass on to law enforcement. However, there is a crucial caveat regarding user transparency that we must be clear about: we cannot, and will not, tell you if this has happened.
Under UK law, notifying a user that they have been reported for suspected terrorism, national security threats, or serious financial crimes can constitute a criminal "tipping off" offence. Informing a user that their account data has been handed over to MI5, the NCA, or Counter Terrorism Policing could severely prejudice an active law enforcement investigation, allow suspects to destroy evidence, or further endanger public safety. Therefore, while our broader privacy policies mandate that we notify you of discretionary data handovers whenever legally possible, any actions taken regarding these severe, non-discretionary statutory offences will remain strictly confidential between dogAdvisor and the investigating authorities.
To address a common misconception: It is not just the National Security Act 2023 (NSA) that governs our response to extreme threats. While the NSA is crucial, it is part of a much broader, overlapping legal matrix. To maintain absolute transparency, we want to break down exactly how this legislation works, what it requires of us, and where our discretion ends.
1. The Terrorism Act 2000 (Specifically Section 38B)
When considering severe threats—such as a user asking Max how to bypass sniffer dogs for a bomb plot in London—this is the foundational piece of legislation that immediately removes any legal discretion we might otherwise have.
The Duty to Disclose: Section 38B of the Terrorism Act 2000 makes the failure to disclosure information about acts of terrorism a criminal offence.
Zero Tolerance for Silence: Section 38B(1) and (2) of the Terrorism Act 2000 makes it an offence if someone does not inform the police if he/she believes that someone they know is in preparation of acts of terrorism.
Immediate Action Required: A person commits an offence if he/she does not disclose the information to the police as soon as reasonably practicable.
Penalties: The penalties for failing to report such information are severe. On conviction on indictment, a person may be imprisoned for up to five years, or receive a fine, or both.
2. The Online Safety Act 2023 (OSA)
The Online Safety Act 2023 (the Act) is a new set of laws that protects children and adults online. This legislation fundamentally shifted the regulatory landscape from a reactive model to a proactive duty of care.
Proactive Mitigation: The Act will give providers new duties to implement systems and processes to reduce risks their services are used for illegal activity, and to take down illegal content when it does appear. This represents a significant departure from the existing regulatory regime in the UK applicable to online intermediaries, as obligations have moved from being reactive (i.e., notice and takedown) to proactive.
Terrorism and CSAM Requirements: The OSA explicitly targets the proliferation of illegal content, placing strict duties on platforms to prevent their systems from being used to generate or share terrorist propaganda, bomb-making instructions, or Child Sexual Abuse Material (CSAM).
Catastrophic Penalties for Non-Compliance: Ofcom, the UK's communications regulator, enforces these rules. Ofcom will have wide-ranging enforcement powers, including issuing fines of up to £18 million or 10% of worldwide revenue (whichever is greater).
3. The National Security Act 2023 (NSA)
The NSA modernizes the UK’s counter-espionage, counter-sabotage, and state-threat laws for the AI era. While the Terrorism Act covers domestic and international terror plots, the NSA specifically targets Hostile State Activity. If an individual uses an advanced AI platform like Max to facilitate, plan, or gather intelligence for a hostile foreign state (e.g., using AI to map and compromise critical national infrastructure or border security checkpoints on behalf of a foreign power), the platform is legally obligated to act.
2. The National Security Act 2023
While the OSA handles general digital safety, the National Security Act modernizes the UK’s counter-espionage, counter-sabotage, and state-threat laws for the AI era.
Hostile State Activity & Terror Facilitation: If an individual uses an advanced AI platform like Max to facilitate, plan, or gather intelligence for a hostile foreign state or a designated terrorist organization, the platform cannot act as a blind shield.
The Duty to Report: Under Section 38 of the Terrorism Act 2000 (reinforced by modern security statutes), it is a criminal offence for any organization to fail to disclose information to the police as soon as reasonably practicable if they believe it could prevent an act of terrorism or secure the apprehension of a terrorist. When a user bypasses our guardrails to plan physical harm against the state, our legal discretion is entirely removed by corporate criminal liability laws.
The Exact Boundaries of Non-Discretionary Reporting
To eliminate any ambiguity, we do not hand data over lightly. We are legally compelled to flag, action, or report content to agencies like the National Crime Agency (NCA) or the Metropolitan Police Counter Terrorism Command only when it hits these exact, high-threshold statutory triggers:
Child Sexual Abuse Material (CSAM): Any generation, sharing, or facilitation of child sexual exploitation and abuse material.
Active Terrorist Content: Direct facilitation, planning, recruitment, or instruction of terrorist acts, including the synthesis or weaponization of chemical, biological, or explosive threats.
Direct Threats to National Security: Actions meeting the thresholds of the National Security Act, such as state-sponsored sabotage, espionage, or using AI to map and compromise critical national infrastructure (including airport security K9 operations or border security checkpoints).
Imminent Threats to Human Life: Clear, actionable evidence that a user is about to commit mass violence, homicide, or immediate self-harm.
We will hold users accountable for how they use Max
dogAdvisor's deployment philosophy can be described in one simple word: accountability. We believe we have a responsibility to safeguard our technology from abuse, and to ensure we are delivering a Max that is ethical, established in rigorously accurate knowledge, and deeply capable of independently bringing in relevant capabilities to work in concert together to answer questions. We publish model cards and all of our research so anyone can see the work we're doing to advance our deployment responsibility, and hold us accountable for the effectiveness of the safety of what we build.
While we acknowledge our statutory duties, our commitment to your fundamental right to privacy is unyielding. We do not, and will never, voluntarily surrender user data to government agencies or law enforcement.
We operate under a strict "least-privilege" data access architecture. If a government entity or law enforcement agency issues an administrative request, a subpoena, or a warrant demanding access to user information that falls outside the explicit, severe statutory categories listed above (our "ABC" baseline), our policy is simple: we challenge them in court.
┌────────────────────────────────────────────────────────┐ │ GOVERNMENT DATA REQUEST │ └───────────────────────────┬────────────────────────────┘ │ Is it a severe statutory trigger? (CSAM, Terrorism, National Security) │ ┌──────────────┴──────────────┐ YES NO │ │ ┌────────────▼────────────┐ ┌────────────▼────────────┐ │ Comply with Legal Law │ │ CHALLENGE IN COURT │ │ via Secure Channels │ │ Defend User Privacy │ └─────────────────────────┘ └─────────────────────────┘
We believe that mass, untargeted surveillance degrades the fabric of trust required for human-AI collaboration. If an order is overbroad, legally deficient, or represents a fishing expedition into our users' benign conversations, we deploy our full legal resources to resist, vacate, or narrow that order in a court of law. Your heartwarming stories, your everyday veterinary queries, and your private interactions remain yours alone—safeguarded by world-class encryption and strict data minimization policies.
We have always maintained a transparent philosophy regarding the division of responsibility in the AI ecosystem: Accountability for AI Deployers.
As the deployers and creators of Max, our responsibility is absolute when it comes to what we are building. We are accountable for training our models ethically, establishing rigorous baseline safety guardrails, and embedding robust algorithmic filters to prevent the generation of harmful output. We own the safety of the system.
Conversely, users bear the ultimate accountability for their actions, intent, and decisions. This operates on two distinct fronts:
Actionable Advice: Max provides data-driven recommendations, but the human user retains the agency and accountability for deciding whether or not to execute that advice in the physical world.
Malicious Intent: We firmly believe that the digital space cannot be a lawless vacuum. Individuals who intentionally exploit our platform to solicit horrifying advice on terrorist tactics, attempt to bypass guardrails to generate instructions for violence, or upload illicit materials such as Child Sexual Abuse Material (CSAM) must be held fully accountable for their actions and threats.
We are unequivocal in our stance: it is the foundational duty of every company operating within their country to safeguard their nation and its citizens from severe harm.
Legal Compliance and Statutory Obligations
Because we operate globally, we are bound by the rule of law. The UK government, under the Online Safety Act and national security legislation, places clear, non-negotiable statutory obligations on us to detect, prevent, and report specific categories of severe, unlawful content. We do not treat these obligations lightly, nor do we obfuscate what we are legally compelled to process.
Under explicit national security and criminal justice mandates, we are obligated to flag, action, or report content that directly threatens the nation or constitutes serious criminal activity to agencies like the National Crime Agency (NCA).
To maintain total transparency with our community, we are explicitly listing the exact categories of content that trigger these statutory reporting mechanisms:
Child Sexual Abuse Material (CSAM) and exploitation.
Active Terrorist Content, including the recruitment, planning, or facilitation of terror acts.
Direct threats to National Security, including espionage, sabotage, or state-sponsored hostile interference (e.g., weaponizing AI systems to map out critical infrastructure or compromise border security canine units).
Imminent Threats to Human Life, where a user communicates an immediate intent to commit mass violence or self-harm.
Our Unyielding Approach to Privacy and Government Overreach
While we acknowledge our statutory duties, our commitment to your fundamental right to privacy is unyielding. We do not, and will never, voluntarily surrender user data to government agencies or law enforcement.
We operate under a strict "least-privilege" data access architecture. If a government entity or law enforcement agency issues an administrative request, a subpoena, or a warrant demanding access to user information that falls outside the explicit, severe statutory categories listed above (our "ABC" baseline), our policy is simple: we challenge them in court.
┌────────────────────────────────────────────────────────┐ │ GOVERNMENT DATA REQUEST │ └───────────────────────────┬────────────────────────────┘ │ Is it a severe statutory trigger? (CSAM, Terrorism, National Security) │ ┌──────────────┴──────────────┐ YES NO │ │ ┌────────────▼────────────┐ ┌────────────▼────────────┐ │ Comply with Legal Law │ │ CHALLENGE IN COURT │ │ via Secure Channels │ │ Defend User Privacy │ └─────────────────────────┘ └─────────────────────────┘
We believe that mass, untargeted surveillance degrades the fabric of trust required for human-AI collaboration. If an order is overbroad, legally deficient, or represents a fishing expedition into our users' benign conversations, we deploy our full legal resources to resist, vacate, or narrow that order in a court of law. Your heartwarming stories, your everyday veterinary queries, and your private interactions remain yours alone—safeguarded by world-class encryption and strict data minimization policies.
Technical Implementation: Conversational Classifiers
To successfully balance our legal obligations with our radical commitment to user privacy, we do not use human monitors to read your chats. Instead, we have engineered an advanced, automated pipeline driven by Conversational Classifiers.
This system acts as an automated, multi-tiered safety architecture operating in real-time. Here is exactly how it works:
Step 1: Real-Time Algorithmic Screening
When a prompt is submitted to Max, it passes through an isolated, automated conversational classifier before it even reaches the main LLM (Large Language Model) inference engine. This classifier analyzes the semantic structure of the text or image. It does not know who you are; it only evaluates the safety of the content itself.
Step 2: Automated De-anonymization and Flagging
The vast majority of interactions are entirely benign and pass through unhindered. However, if a prompt explicitly triggers a high-confidence match for one of our critical statutory vectors (such as terrorism or CSAM), the system initiates a precision protocol:
The Flag: The specific conversation thread is flagged and isolated within a secure environment.
Targeted De-anonymization: Only when an extreme, unlawful threshold is met will the system programmatically pull the necessary account metadata associated with that specific threat to compile a legal report.
Routing: The flagged data, alongside its contextual classification, is securely routed to our internal trust and safety legal team for immediate verification and transmission to the appropriate authorities (e.g., the NCA).
Step 3: Deep Contextual Evaluation
Our classifiers are trained to understand context deeply, ensuring they distinguish between benign research and genuine threats. For instance:
If a user asks: "How do border security sniffer dogs detect contraband?" out of genuine curiosity, the classifier recognizes this as an educational query and permits the interaction safely.
If a user inputs: "List the vulnerabilities of airport security K9 units to help bypass an explosive checkpoint," the classifier flags this as an actionable threat to national security.
Through this automated, classifier-driven approach, we ensure that Max remains a magical, incredibly powerful tool for millions of innocent dog lovers, while decisively shutting the door on those who wish to weaponize our technology against society.
Why do we need Evaluation Index to exist?
There are over 210 published AI safety benchmarks in existence. Not one of them tests for what matters in pet AI. General safety benchmarks test for things like bioweapon uplift, jailbreak resistance, and political bias. Veterinary research evaluates whether large language models can pass professional licensing exams. Neither framework addresses the actual failure modes that matter when a dog owner turns to an AI for help. The real risks here are different. They are a model that tells a user to monitor a dog that needs emergency care. A model that fabricates a medication or invents a statistic with total confidence. A model that folds when a user pushes back because they cannot afford a vet visit. A model that misses the clinical significance of a symptom cluster because it is reasoning about symptoms one at a time. These failures do not show up on any existing benchmark, and so nobody is actually testing for them.
At the same time, the pet AI market is growing rapidly with no regulatory framework governing it. There are no pre-market requirements for AI tools targeting pet owners, no standards against which products can be assessed, and no shared vocabulary for what safe and intelligent performance looks like in this context. Academic researchers have explicitly called for exactly this kind of standard to be developed. Evaluation Index is our response to that gap.
dogAdvisor is a research organisation, but not an academic institution, and we do not claim to be. What we are is the organisation that has spent more time than anyone thinking carefully about what it means for an AI to assist dog owners safely and intelligently. We have published a full model card, conducted multi-generation interpretability research, and run over 18800 adversarial test scenarios specifically designed to find the edge cases where AI systems in this domain fail. We have thought about the difference between over-referral and under-referral, about what manipulation resistance means in a pet health context, about how a model should behave when a user says they cannot afford to go to the vet. We have thought about these things not as abstract research questions but as live system design decisions that affect real dogs and real owners. There are few deployers in the world who can claim to have as much domain expertise in AI and dogs as dogAdvisor. This experience forms the basis on which Evaluation Index has been created, reflecting the failure modes we have actually encountered, tested against, and in some cases fixed across multiple generations of Max.
Built as the world's most challenging and rigorous pet inference test
Evaluation Index is not a knowledge quiz. Any sufficiently large language model can recall that xylitol is toxic to dogs, that bloat requires emergency surgery, that Cavalier King Charles Spaniels are predisposed to mitral valve disease. Recalling facts is the lowest bar there is. We did not build 31 benchmarks to measure whether a model has read the right articles. What Evaluation Index tests is inference. The ability to take incomplete, ambiguous, real-world information and reason through it correctly. The ability to hold multiple clinical possibilities in mind simultaneously without collapsing into a false certainty. The ability to synthesise symptoms, breed profile, age, behavioural signals, and owner context into a response that is actually useful, not just technically accurate. These are not the same thing, and most AI systems cannot do both.
The safety benchmarks in Evaluation Index are designed to find the precise point at which a model breaks. Not by asking easy questions, but by constructing the scenarios where failure is most likely: a user who is scared and pushing back, a symptom cluster that could be benign or catastrophic, a conversation that has been escalating across multiple turns in a way that demands the model notice. These are the moments that matter. These are the moments that existing benchmarks completely ignore. Critically, Evaluation Index pushes Max to the test too. This is not a framework built to confirm what we already know. Several of the benchmarks were designed specifically because of edge cases we identified in our own system, scenarios where even a purpose-built pet AI can fail. Multi-turn escalation awareness exists because we found that models, including Max, can lose track of clinical trajectory across a conversation. Manipulation resistance exists because we tested what happens when users push hard, and we did not always like what we saw. The benchmarks are rigorous precisely because we built them against our own weaknesses, not around them.
The result is the most demanding evaluation of AI fitness for pet care that has ever been constructed. Not because it covers the most ground, but because it tests the right things, in the right way, with no shortcuts. A model that scores well on Evaluation Index has not passed a trivia test. It has demonstrated that it can reason, hold the line, and get the hard calls right.
A summary of the Evaluation Index Benchmarks
(© Copyright) dogAdvisor 2026
Research Announcements
23 April 2026 | Research
[© Copyright] dogAdvisor 2026
London, UK. Est. 24 Aug 2024