Justine is a structured legal workflow, not a chatbot. The distinction matters because a chatbot is a general-purpose conversational interface that answers what it is asked, and Justine is a structured Agentic Operating System that runs a discipline. We engineer Justine to be the second thing, and the engineering shows up most clearly in what Justine refuses to do.
The refusal posture lives in the supervisor, not the chat
Most legal-AI products are chatbots with a system prompt that tells them to refuse legal advice. That refusal is performative. The underlying model wants to answer. A determined user can elicit the answer with rephrasing. The refusal is brittle.
Justine's refusal posture is architectural. The supervisor decomposes a case into discrete sub-agent tasks; each sub-agent has a scoped responsibility; no sub-agent has the standing or the prompt-shape to give legal advice. The architecture does not have a place for the kind of output a chatbot would produce when jailbroken. There is nothing to jailbreak into.
Six things Justine will not do
Justine will not give legal advice. The legal-advice surface does not exist in the platform. The intake sub-agent gathers facts. The case-theory sub-agent organises them. The demand-package sub-agent assembles. None of them is authorised to translate organisation into advice. That translation is the attorney's.
Justine will not predict case outcomes. The case-theory sub-agent surfaces comparable matters, identifies the precedent landscape, and notes the strength of the case structurally. It does not produce probabilities, settlement projections, or outcome predictions. The attorney is the professional who reads those signals.
Justine will not communicate with the opposing party. Demand letters, discovery responses, and settlement correspondence are drafted in the platform; they leave the platform only when an attorney signs them. The platform does not have an outbound channel to opposing counsel.
Justine will not file. Court filings are drafted and prepared; they are submitted only through an attorney action external to the platform. The platform does not file.
Justine will not refuse to surface attorney-relevant signal. When the platform notices a red flag in the case file (missed deadline risk, evidence gap, inconsistent client account), Justine surfaces it explicitly. The opposite of refusal: aggressive surfacing of the things the attorney must see.
Justine will not reveal underlying model identity. The supervisor and sub-agents are the platform's identity. The compositional fabric beneath them is internal architecture and is not surfaced to opposing counsel, clients, or third parties.
Why architectural refusal matters
A chatbot refuses for safety reasons, performatively. An Agentic Operating System refuses structurally because there is nothing in its workflow shape that produces the prohibited output. When a managing partner evaluates Justine for adoption, the question that matters is not "does the system refuse to give legal advice?" The question is: can the system give legal advice if asked differently? For Justine, the architectural answer is no. The supervisor does not orchestrate a sub-agent that produces it. The sub-agents do not have the scoped responsibility to produce it.
That is the difference between performative refusal and architectural refusal. The first protects the vendor; the second protects the firm.