It reads the law.
Then it proves it.
Most legal AI looks for text that sounds like your question and writes a confident paragraph from it. Ask Canada Safety was built the other way around: start from the binding law itself, find the provision that actually governs, read it the way a safety professional would, cite it, and check the answer against the law it came from.
Jump to: the difference · the path a question takes · compare mode · how it is measured · your documents · what's indexed and how it's verified
A typical AI search, and where it breaks on law
The standard recipe, called retrieval-augmented generation, is simple: find text that resembles the question, paste it into a prompt, and let a language model write. For many subjects that is good enough. Safety law is not one of them, because the law's meaning lives in places the recipe never looks: in definitions, in cross-references, in tables, and in the exceptions three subsections down.
The path a question takes
Every question runs through the same seven layers. Two things sit around them: the verified corpus that every layer reads from, and the measurement that every change has to pass before it ships. Where you see 🔒 Proprietary, we describe what a layer does, but not exactly how.
The corpus: the law, and only the law 🔒 Proprietary
Every jurisdiction's OHS Act and every binding regulation made under it, checked against the official text, and nothing else. It is the only thing the system reads from. What is in it, how current each document is, and how each one was checked against the official text are set out on the Methodology page.
Scope: the right law for where you work
Search is confined to the jurisdiction you picked. A document-scope framework keeps mine, diving and blasting rules out of a general question. In Ontario and Quebec, which each have separate construction and industrial rulebooks, the workplace named in your question decides which one applies.
Search, then a reader
The closest-sounding section is not trusted on sight. Candidate sections are read in full, and the ones that actually contain the duty asked about rise to the top, even when another section merely sounds more like the question.
Follow the law's own links 🔒 Proprietary
This is where it reads the law the way a safety professional does. A rule rarely lives in one section: the duty is in one place, the meaning of its key word in another, the exception three subsections down. So the system doesn't stop at the closest match.
Laws lean on each other: "subject to section 2.1.1", "as defined in section 1". So when a section depends on a definition or another section, that section comes with it, and the Act and its regulations are read together.
Some everyday questions never resemble the section that governs them. "If I think a job will hurt me, can I refuse?" doesn't look like a definition of "unusual danger", but it is still answered from the section that governs it.
Tables come back as tables: the exact row your question names, with each value beside its label and unit.
A grounded answer
The model answers only from the sections supplied, takes a short reasoning step before it writes, cites the section for each duty, and says so when the text doesn't cover the question rather than filling the gap.
Checks that earned their place 🔒 Proprietary
After the answer is written, it is checked for the dangerous mistakes we have seen happen, such as a figure read from the wrong row of a table. Every check is itself tested against answers known to be correct, to prove it leaves good answers alone.
Compare mode: four provinces, four full answers
The compare view answers one question for all four Atlantic provinces at once.
Each province gets a full answer from its own law, and the four answers are laid side by side, so a rule from one province never slips into another's.
- Measured. A compare benchmark grades all four provinces' answers, each against its own key.
- The cost. A compare answer takes a couple of seconds longer, because four full answers are prepared. We think that is the right trade for safety law.
Why we built it
Finding the answer took longer than doing the work.
It began as a productivity problem in Atlantic Canada. A plain safety question, such as how high a worker can be before fall protection is required, meant opening four provinces' Acts and regulations, each organised differently. The answer was often split across a definition, a section and a table. You read it once, then did it all again for the next province.
Four provinces, four rulebooks
Newfoundland and Labrador, New Brunswick, Nova Scotia and Prince Edward Island each write their own Acts and regulations, and split them up differently.
Hours, not minutes
A simple question meant digging through long PDFs, definitions and cross-references, then doing it again for the next province.
No quick cross-reference
Crews and contractors work across the Atlantic provinces. Nobody could see the four rules side by side.
The goal: the right section of the law, for the right province, in seconds, and the four Atlantic provinces side by side. That became compare mode, and it is still the heart of the site.
How a change gets in
Every change goes around the same loop, every time: build or fix one layer, test it on the real system, audit it against the official law, and keep it only if it measures better.
Measured, not believed
Our benchmarks are four sets of questions, 584 in all, spread across all 13 jurisdictions and written to mirror how people ask. The later sets were written after the engine was built, on sections it had never been tested on, so it could not have been tuned to them. Each question has an answer key written from the statute itself, section by section, with its conditions and exceptions. Every answer is read by hand against that key, and the benchmark is re-run whenever the engine changes. A separate benchmark grades the compare view, all four provinces' answers each against its own key.
We use these results to decide what ships: a change that doesn't measure better doesn't stay. We'll publish the figures once they have been independently reviewed.
We check the checkers. When a benchmark answer disagreed with its key, we went back to the law. In the first benchmark, all eight disagreements turned out to be the key's mistake, not the answer's. An automated judge flagged 22 answers as fabricated, and 20 of those flags were false. That is why our results come from people reading the law, not from a model grading a model.
Beyond the benchmark, more than 10,000 test questions have been asked through the real system and checked, and every change runs against more than 400 automated tests before it goes live.
What it does not do yet
A professional system says where its edges are. What the corpus does not contain (other codes and standards, guidance documents, the law of federally regulated workplaces, the few documents still to come, and why exposure-limit tables are answered by pointer) is listed on the Methodology page. These are the edges of the engine itself, today.
- Some answers still leave out a condition from a long section, even with the right section in front of the model. That is the work now under way, and it's why every answer links to the full text.
- Most official sources are checked automatically every week; the few whose sites block automated reading are checked by hand. A new amendment is published in the provincial gazette first and can take weeks to reach the consolidated copy, so there can be a lag. The date each document is current to is on the Methodology page.
- Compare mode covers the four Atlantic provinces. Other jurisdictions are answered one at a time.
- Toolbox Talk is not part of the benchmark.
- It explains the law; it isn't a substitute for advice from a qualified professional, and it says so.
Your documents. Same engine.
The method is not specific to safety law. Anything with numbered clauses and one official version could be read the same way, with every answer cited to its clause: safety manuals, operating procedures, collective agreements, employee handbooks, site rules, codes and bylaws.
A working example is live: our pilot on Newfoundland and Labrador's Labour Standards Act and regulations, built the same way as this site. Its text is taken from the official consolidation and checked against it section by section, every answer cites its section, and each section links to the government's own copy. It is a pilot, and says so.
Organisations can licence the engine, hosted by us, or have us build a version around their own documents, starting with a pilot on their own questions.
A professional system, built for Canadian safety law
It is an engineered system: a verified corpus of binding law from every province and territory, retrieval that follows the law's own structure, an answer that cites and admits silence, checks proven against correct answers, a compare view that answers each province on its own, and a benchmark graded by hand against the statute book. Every layer was built, torn down and rebuilt on measurement, and kept only when it earned its place.
Ask Canada Safety is built to be a system a safety professional can hold to the standard they hold themselves, because it shows its section, its limits and its method.
Try it free at askcanadasafety.ca. Pick a province or territory and ask in plain language; use compare mode for the four Atlantic provinces side by side; or get a Toolbox Talk a supervisor can read to a crew, with the law that applies cited to the section. Every citation opens the section's exact text and the government's own copy.
Every answer links to the official text, and that text governs. Figures on this page are as of 3 October 2026.