From Mergen AI's latest blog:Read more

Mergen AI
All articles

/2 min read/context graph, compliance, food, beverage, ai, fda

Evidence Trails of Compliance AI

We Mapped Every Canadian Food Regulation Into a Knowledge Graph

Tural Gulmammadov

We built the first comprehensive knowledge graph of Canadian food and beverage regulations.

Then, we mapped the whole text into 51,397 legal provisions, 131,069 requirements, and 132,202 machine-readable rules. All connected in a knowledge graph. All traceable to source.

Here’s what the data shows.

CFIA Runs the Show

Six authorities write Canada’s food rules. One writes 82% of them.

  1. CFIA | 81.7%

  2. Dept. of Justice | 9.1%

  3. Health Canada | 5.9%

If you sell food in Canada, CFIA is your primary regulator. Health Canada sets nutritional standards and health claims. Justice provides the legal framework underneath.

What Regulators Actually Focus On

Top five of the all requirement categories cover 91% of everything.

Half of all food regulation is about manufacturing process and paperwork.

  • Process (26%) — how food is made, handled, stored

  • Documentation (25%) — records, certificates, audit trails

  • Composition (17%) — ingredient limits, nutrient thresholds

  • Labelling (14%) — what goes on the package

  • Prohibition (9%) — what’s banned outright

9 Out of 10 Rules Are Non-Negotiable

Not all regulations carry the same weight. Some are mandatory. Some are recommendations. Some are just permissions.

  • Must / must not: 90.6% of all requirements

  • May: 5.0%

  • Should: 4.3%

Nine out of ten rules are hard obligations. Miss one, you’re non-compliant. This is why manual compliance doesn’t scale — there are simply too many mandatory requirements to track by hand.

Why We Built It as a Graph

Regulations form a web, not a spreadsheet. A single legal provision can create multiple requirements. Each requirement has a binding force. Each connects to a testable rule.

Legal provision → Requirement → Testable constraint

We structured it this way so we can do two things:

1. Constraint satisfaction — feed in a product spec, automatically check it against every applicable rule

2. Deontic modality analysis — distinguish between what’s mandatory, what’s recommended, and what’s merely permitted

The goal isn’t just “is this product compliant?” It’s why — with a full evidence trail from verdict back to the exact legal text. Audit-ready. Inspector-ready.

What’s Next

This was the overview. I will share more insights with you.

Appendix A

Click to download jupyter notebook HTML export for reading more.

Written by

Tural Gulmammadov

Founder and CEO, Mergen AI Eliminating the barrier to market entry for food and beverage manufacturers

Newsletter

New posts, straight to your inbox.

F&B compliance insights, curated from real compliance work. Free, and you can unsubscribe any time.