We study adherence: how AI systems preserve task intent, domain constraints, role boundaries, safety criteria and user context inside real workflows.
Haku Labs is an AI research and product lab focused on context architecture for applied AI. Our near-term focus is adherence: the ability of an AI system to remain within its task, domain constraints, role boundaries, safety criteria and user context across complex workflows.
We work from a practical observation. Many fluent AI products still fail when they lose context. A system may have the right information available, yet treat a durable constraint as ordinary conversational content, let transient instructions override domain logic, or blur the status of a role, goal or safety boundary.
Our research asks how context should be structured, staged, verified and protected in products that people actually use. We develop bounded-context architectures, role-specific reasoning methods, failure tracking, and verification loops that reduce degradation across multi-step tasks — and we test that work through focused products where the cost of losing context becomes visible quickly.
Short essays on the ideas behind our work: how alignment is grown rather than imposed, why cognitive structure has to be designed in, and what it means to treat context as a place rather than a payload.
Use the arrows to move through the notes.
Haku Labs is the research and product lab. Haku Labs Research is its research arm. Two applied products put the same question — how do we keep AI faithful to context? — under pressure from opposite ends of the risk spectrum.
Adherence, context architecture, bounded reasoning and applied safety in AI systems. The shared research question that both products exist to test.
Sales intelligence. Commercial reasoning under clear task boundaries — data qualification, account research, ICP pre-checks and role-specific analysis for commercial teams.
Reflective and therapeutic support. A higher-sensitivity context: preserving safety constraints, respecting role boundaries, and collaborating appropriately within clinical or quasi-clinical settings.
The relationship between Haku AI and Pemma is methodological and operationally separate. They are deliberately different applied domains, kept distinct in data, users and workflows, used to stress-test the same core problem: useful AI must preserve the conditions that make its output appropriate.
Most AI is trained on behaviour: the output an expert endorses, the action a worker performs. That is a powerful signal, and a partial one. We are not only asking how a machine should reason in a commercial space. We build cognitive models of how skilled people actually reason inside a domain, the considerations and conditions that generate the right action, and use those models to drive adherence. The behaviour is the surface; adherence depends on what produces it.
This is why the work is interdisciplinary rather than purely technical. To model how an expert behaves, you have to understand how they think: where their attention goes, how they decide under uncertainty, where the boundaries of safe practice sit. We bring together expertise from psychology, mental health, philosophy and technology to build the high-level cognitive models that adherence depends on.
How expert practitioners read a situation, weigh it, and decide well under uncertainty.
How to hold safety, boundaries and care in high-sensitivity contexts where the cost of drift is real.
How meaning, context and conditions structure reasoning, and where machine cognition departs from it.
How to turn those models into bounded, verifiable architectures that hold up inside real products.
We start from the reasoning, not the output, and express it as atomic behavioural patterns and loops: small, well-bounded units of perceiving, deciding, acting and checking, each carrying only the context it needs. Bounded units keep the conditions that should govern behaviour from being crowded out by ordinary content, and they compose into workflows that hold their adherence as complexity grows.
We study how AI systems can remain faithful to the conditions that define appropriate behaviour. We treat broad alignment as a larger, emergent challenge — and focus on the practical layer where product failures can be observed, measured and improved: context design, constraint preservation, role-specific reasoning, staged workflows and verification.
How should AI systems represent durable constraints so they remain distinct from ordinary conversational content?
How can multi-step workflows preserve task intent, domain logic and role boundaries as context changes over time?
What failure patterns appear when ordinary content competes with value-priors, safety criteria or operational constraints?
How can bounded-context orchestration reduce hallucination, context drift and reasoning degradation in applied products?
What verification loops are needed to identify adherence failures before they reach users or downstream systems?
We apply adherence research to data qualification, account research, ICP pre-checks, role-specific reasoning, copy generation, verification and scheduling. The goal is to help AI behave more like a bounded commercial operator: clear about the task, the buyer context, the evidence base and the limits of its role.
We apply adherence research to high-sensitivity reflective contexts, where AI must preserve safety constraints, avoid role confusion, support human judgement and collaborate within clinical or quasi-clinical boundaries. The goal is to understand how AI can aid reflection while maintaining appropriate limits.
A narrow framework for a recurring failure mode in large language model applications: durable constraints are often stored and processed in the same context substrate as transient task content. System instructions, role boundaries, safety criteria, domain rules and user context then compete with ordinary conversational material inside a finite attention window.
The central claim is that adherence requires systems to distinguish between content that should be reasoned about and contextual conditions that should govern how reasoning proceeds — separating, staging and verifying different kinds of context so governing conditions are less easily displaced by transient material.
Type A — content competes with other content.
Type B — content competes with a constraint, value-prior, role boundary or safety
condition. Type B is where adherence breaks down, and where bounded-context architecture helps most.
Define the adherence problem and the distinction between ordinary content and governing context.
Formalise content-vs-content and content-vs-constraint failures using practical product examples.
Show how staged workflows separate research, qualification, generation, verification and action to reduce degradation.
Evaluation methods for whether systems preserve task intent, role boundaries, safety criteria and domain constraints over time.
Scope, deliberately bounded: broad AI alignment remains outside the immediate claims. Context architecture is one safety layer within a wider strategy. Commercial and therapeutic systems remain separate in data, users and workflows.
We are opening conversations with sponsors, research partners and collaborators who care about making applied AI faithful to the conditions that make it appropriate. If that is you, we would like to talk.
