A legal strategy
shooting range.
Not a chatbot. Not an outcome oracle. A dispute is a branching structure of legal elements — and most legal AI hides it behind prose. Simmis makes it visible, and lets you intervene on it.
Below is a real British Columbia small-claims debt dispute, modelled on the actual law. Drag the evidence, file a document, raise a defence — and watch the verdict, the win probability, and the settlement range move.
Civil Resolution Tribunal · Small claim · Debt
The dispute: the applicant says she lent the respondent $4,200 under a written note; the respondent says he repaid part of it and the rest was a gift. Claims up to $5,000 are decided online by BC's Civil Resolution Tribunal.
In the full system, an LLM tribunal member reads the filed record and assigns each disputed point a credence — its degree of belief on the balance of probabilities. Here, you set them.
Each move intervenes on the dispute — Pearl's do() operator — pinning an element and
recomputing everything downstream.
A debt claim is a deterministic element-test. Every node is real law, with its source. Composition is exact — no probability lives here.
The chance every required element is established at once. Note how four uncertain elements multiply into a number lower than any one of them — the structural risk a single answer hides.
Each party values the case from its own view. A settlement zone exists when the applicant's floor sits below the respondent's ceiling. Widen the private-information gap and watch divergent expectations collapse the zone — and the case go to a decision.
Research and educational prototype. Not legal advice, and not a deployed adjudicator. The law is real (CRT Act; Limitation Act SBC 2012 c. 13); the numbers are illustrative and you control them.
Three layers, kept distinct
The central failure mode in legal AI is collapsing everything into one prompt. Simmis keeps the law, the uncertainty, and the choices separate — so every branch is traceable to a rule, a fact, or a move.
What the law compels
The element-test, jurisdiction gates, deadlines, and composition. Plain data with an exact evaluator — the tree above. This is the part you can audit line by line.
What the evidence warrants
Facts aren't ground truth — they're claims with burden and uncertainty. Each disputed element carries a credence; the outcome is a distribution and a settlement range, not a verdict alone.
What the parties choose
Which evidence to file, which defence to raise, when to settle. In the live system these are LLM agents inside the legal nodes; here, the sliders and moves are you playing them.
Real: the debt element-test and its citations, the CRT's four-stage online pipeline, the balance-of-probabilities threshold, and the divergent-expectations settlement model. Faked for this static demo: the LLM that would read a real filed record and set the credences. Swap the sliders for that model and this same page runs on live cases.
The structure is the product
Open-source legal-simulation infrastructure: a procedural ontology, validated element-tests, a probabilistic engine, and a causal-intervention layer. Built on the Simmis stack — branchable, auditable, every step traceable.