The Science of Legal Precision
Legal AI is only as good as its data and its constraints. Unlike generic LLMs that prioritize fluency, TruLexa prioritizes verifiability. Here is how we ensure every output is defensible in a Canadian court.
1. Canada-First Training Architecture
Most "legal AI" tools are fine-tuned on US case law (PACER data) with a thin layer of Canadian context. This results in "hallucinations" where American legal concepts (like "deposition" vs "examination for discovery") bleed into Canadian drafts.
We took a different approach. Our retrieval pipeline indexes:
- Supreme Court of Canada decisions dating back to 1876.
- Provincial Court of Appeal rulings for all 10 provinces.
- Quebec Civil Code doctrine and commentary (bilingual).
- Federal Court jurisprudence.
2. The Citation Verification Protocol
We do not trust the LLM to remember case citations. LLMs are probabilistic token predictors, not databases. Instead, we use a Retrieve-Then-Verify workflow:
- Semantic Search: We locate potential relevant cases in our vector database.
- Citation Check: The model drafts a response.
- Real-Time Cross-Reference: Before showing you the answer, our system queries the CanLII API to verify that the case name, year, principle, and AI-generated claims match the CanLII citations.
- Hallucination Blocking: If a citation cannot be verified, the system either flags it as "Low Confidence" or removes it entirely.
3. Province-Specific Procedural Rules
We maintain a structured database of court rules (e.g., Ontario's Rules of Civil Procedure, BC's Supreme Court Civil Rules). When you ask for a "Notice of Civil Claim," the system checks your jurisdiction and strictly adheres to the formatting requirements of that specific court registry.