The paper
Policies, leases, loans, public contracts, and the law behind them
Read by lawyers and brokers, one document at a time.
Sentinelle + Vigil9
Ground truth for contracts.
Sentinelle will read the contract, on a legal engine that already runs. Vigil9 will read the ground. Together: one answer per asset, with the evidence.
The gap
The paper
Read by lawyers and brokers, one document at a time.
The ground
Read by engineers and risk analysts, one map at a time.
Today they meet after a loss, in a dispute, by hand.
We are building the join that brings them together before.
The use case · fictional · designed, not built
A town insures its school through a policy it awarded by public tender. The river rises. Here is what the system is designed to do, step by step.
The property policy the town awarded by public tender covers the school. If a natural disaster is recognized by order, the law sets the deductible and the time to declare.
Copernicus maps the flood extent; radar sees through cloud. The water overlaps the school's parcel. Depth and damage stay with the adjuster.
Same building, same peril, inside the policy period. Weak matches go to a person. The clause is tested; the time left to declare is shown.
An agent asks the town to inspect the school and prepares a draft declaration, clause and image attached. Once staff have seen the damage, the town or its broker sends it. Before any loss, the same join flags gaps at renewal.
The system, floor by floor
Colour tells the side: terracotta for the paper, sage for the ground, navy and gold for the join. The line tells the state: solid navy fill, runs today · solid outline, in build · dashed outline, designed.
Choose a floor. The gold pin is one school: the same building on every floor. The agents' ring circles all five floors; a person decides.
Official texts from Légifrance, Judilibre and EUR-Lex, kept current as new versions are published. Next: the insurance terms towns publish in their tenders, and the policies pilot partners will share.
Built on our legal engine, which is constrained to its corpus and required to cite the article behind each answer (98.4% of answers attached to an openable source on our CITC test set). What is in build: turning policy wordings into structured covers, exclusions, limits and deadlines, each tied to its clause.
The join will give each asset one identity: address, parcel, building ID, contract and parties. Every fact will carry two dates: when it was true, and when we knew it. The same peril on both sides. Rules will test the clause against the event, and weak matches will go to a person.
Will map each building and the hazards it faces, from public sources, and what satellites observe, such as the extent of a flood seen by radar through cloud. Floods first, then wildfire.
Copernicus Sentinel-1 radar and Sentinel-2 imagery, French risk maps, the land registry and building outlines. In Canada: the Open Database of Buildings and provincial cadastres.
At each satellite pass, each new version of a text and each deadline, agents will compare the two sides and raise any gap with its clause and its image. They will draft; a person decides.
A person decides
Where it bites first: insured property
French towns feel it first. When an insurer cancels, the town re-tenders, and the new wording must be checked by hand against every school, hall and depot. In Canada, where overland flood cover is sold as an add-on, each portfolio has to be re-read against the hazard at renewal.
1 in 5French local authorities answering a Senate consultation have had a contract cancelled by their insurer since January 2023 (Senate, 2024)
~850,000Canadian homes still without access to overland flood insurance (Insurance Bureau of Canada, July 2026)
1.5 millionCanadian households highly exposed to flooding (federal task force, 2022)
First buyers: French towns and other owners of many buildings, with their brokers. They pay first for a diagnostic of their own contracts, then for a continuous watch.
What already runs
beta testers on French and European law
came from other legal research tools
of answers traced to a source you can open, on 123 questions validated by legal professionals
The engine is constrained to answer from our corpus and required to cite the article behind each answer. The 98.4% measures source attribution, not accuracy. Benchmark run with CITC, our research partner. About 70% of Sentinelle's building blocks already run in this engine (our estimate).
What we measure next: how often a policy is matched to the right building, on the pilot. It will be the first figure we publish.
Next
A 30-minute live demo, in English: the legal engine on questions you choose from French or EU law (it answers in French), then the architecture of Sentinelle and Vigil9.

Mehdi Kheredine · Founder and CEO
mehdi.kheredine@legibus.ai