Four design choices that make LEASE-iQ accurate where general AI fails. Written for engineering reviewers, lawyers, and anyone evaluating LEASE-iQ on substance rather than slogans.
Hallucination cannot be eliminated at the base model. It can be detected, outvoted, and flagged. The architecture below is what makes that real.
LEASE-iQ never trains on your lease. Documents are ingested via open-source OCR plus multimodal LLM parsing, then converted into a proprietary graph structure that captures hierarchical relationships between clauses, schedules, and cross-references. UK legislation is pre-processed into LLM-compatible rules before any user interaction. At query time, only the relevant clauses and statutory rules are assembled into context. The model answers only from that supplied context, never from training data.
Three-level reference chaining. A clause like "subject to Schedule 2 paragraph 5(b)" is automatically resolved to (1) the schedule, (2) the definition that governs the term used in the schedule, and (3) any deed of variation that overrides the original wording. Generic AI flattens all three into prose and loses the precedence.
For every question, 10 independent responses are generated in parallel. A semantic similarity step (NLP, not LLM) scores each response for meaning. Responses below average similarity are discarded as outliers. The remaining responses go to 3 LLM agents that vote on the best answer. If no consensus forms, notes are passed back and the cycle repeats up to 3 iterations. If still uncertain, LEASE-iQ tells you it is uncertain rather than guessing.
UK statute and case law are pre-processed into a machine-readable rules layer that sits over the lease ontology. The Landlord and Tenant Act 1985, CLRA 2002, BSA 2022, LAFRA 2024, the Renters' Rights Act 2026, and the leading tribunal authorities are applied at query time, not relied on from foundation-model training data.
Currency is maintained by a human review process: when a new Act, statutory instrument, or material tribunal decision affects how a lease should be read, the regulatory overlay is updated. Page-level "Last reviewed" dates throughout the site show the live state. Human approval is always required before anything ships into the live overlay. We do not run an automated cadence; updates land when the law moves.
Data residency and training position. All processing happens on GCP Europe West 2 (London). Your lease is contractually prohibited and architecturally isolated from model training, with the contractual position covering Google's models, our own, and any third party.
KC-level honesty on edge cases. Some lease clauses are genuinely ambiguous. Even a King's Counsel might give different answers on different days. LEASE-iQ flags these honestly instead of guessing. A risk-rating system is shipping in the next release that will grade clause certainty so you know when to seek professional advice.
All large language models require randomness (temperature) to function. Hallucination cannot be fully eliminated at the base model level. Any tool that tells you otherwise is overclaiming.
What LEASE-iQ does instead: generates 10 responses in parallel, removes outliers through semantic analysis, has 3 independent agents vote, and flags uncertainty rather than guessing. LEASE-iQ also answers only from context supplied at query time (your lease plus the pre-processed regulatory overlay), not from training data. This removes the category of hallucination where a model "fills in gaps" from general knowledge. The remaining residual risk is managed through consensus and honest uncertainty.
This is the position we hold publicly because it is the position that survives a hostile read. The page you are on is the page we want a sceptical lawyer, a sceptical engineer, or a sceptical journalist to find when they look for the catch.
14 years in AI and data engineering (from large engineering consultancies to NLP-powered regtech startups) and 4 years at Google Cloud as Director of Engineering at Onix (a Google Cloud partner). LEASE-iQ is not a ChatGPT wrapper. It is a purpose-built legal intelligence system that uses no vendor-locked infrastructure and runs on a portable, multi-provider architecture.
If you are evaluating LEASE-iQ for a partnership, integration, or professional use case, the partner page has the API roadmap and a contact form Adam reads personally.
The Building Trust assistant can route you to the right page, explain a clause, or get you started with LEASE-iQ. First question is free.
Want a written, clause-cited answer in 24 hours instead? Talk to us →