Grounding is what keeps an AI honest: every answer anchored to something verifiable rather than a confident guess. In professional services, a confident wrong answer is a liability. So we ground our agents in three ways, in sources, in rules, and in the actions they are allowed to take.
General-purpose AI predicts the most likely next words. That makes it fluent, and it also makes it capable of stating something false with total confidence. For a law firm, that is not a quirk, it is a risk to clients, to compliance, and to reputation. Grounding closes that gap.
Every output traces back to real evidence: the conversation, the document, the matter record. The agent cites where a claim came from rather than inventing a plausible-sounding one. If the source is not there, the answer is not asserted.
retrieval + citations The rules that govern a task are written as formal statements in Lean 4, the same proof system used in mathematics. Compliance with a rule becomes something a machine can verify, not just something a model claims. This is what we build in LegalLean.
formal verification (Lean 4 / LegalLean) Learn more → Before one step hands off to the next, three properties are type-checked: the data fits, the capability is permitted, and the budget holds. The agent is bounded by construction, so its behaviour is auditable rather than emergent.
typed handoffs (The Stack) Learn more → Any vendor can claim their AI is accurate. Grounding is the part you can inspect: the citation you can follow, the rule you can read, the proof that type-checks. We keep the deep machinery, Lean 4 and the verified stack, one click away, so the people who want to see the proof can.