What it took to build verified law

90,362 obligations. One year. Domain expertise, not technology wizardry.

The Scale

What's Inside

  • 385,460 provisions from UK pensions legislation
  • 90,362 obligations extracted and verified
  • 19,354 definitions linked to source provisions
  • 87 Acts + 2,543 Statutory Instruments + 7 TPR Codes
  • Case law citations and cross-references
  • Regulatory guidance mapped to underlying legislation

Every Piece Verified

50,923 obligations passed verification. 33,806 flagged with warnings. 5,633 failed and were corrected. Zero went live unverified.

Temporal Awareness

Commencements tracked. Amendments tracked. Repeals tracked. Query the law as it stood on any date. AI knows which provisions were in force in 2018 and which were repealed last year.

Why It Took a Year

Not Because Technology Is Hard

Neo4j exists. Python exists. TypeScript exists. OpenAI embeddings exist. The stack is unremarkable.

Because Domain Expertise Is Hard

Extracting obligations from legislation requires understanding what an obligation is. Distinguishing "must" from "may" from "should" in regulatory context. Knowing when a provision creates a duty and when it defines a term.

Generic NLP cannot do this. LLMs hallucinate obligations that do not exist. Only a qualified lawyer who has read thousands of provisions can build the extraction rules that work.

Because Verification Is Hard

AI extracted 90,362 candidate obligations. Humans verified every one. Flagged ambiguous language. Corrected misclassifications. Built a review queue for edge cases.

This is not "AI does the work". This is "AI proposes, humans verify, infrastructure holds the result."

How AI Uses This

The Query Flow

  1. User asks: "What are my DB scheme disclosure obligations?"
  2. AI searches 90,362 obligations (seconds, not hours)
  3. Returns 47 matching obligations with provision citations
  4. User clicks citation → sees source law text with character offsets
  5. AI cannot hallucinate because it queries verified data, not the internet

What This Prevents

  • Hallucinated obligations that do not exist
  • Citations to provisions that were repealed
  • Ambiguous answers with no source provenance
  • "The AI said so" without audit trail

What This Enables

  • Provenance: every obligation traces to source provision
  • Verification: click the citation, see the law
  • Temporal queries: "What applied on 2018-03-15?"
  • Cross-references: see which provisions link to each other

What This Is Not

Not a product you buy off the shelf. This is infrastructure. It powers tools. You do not "use" the graph directly—AI uses it on your behalf.

Not finished. Legislation changes. Statutory instruments are added. Case law grows. The graph is maintained, not static.

Not everything. This covers UK pensions law. Not tax. Not trusts. Not employment. Domain-specific infrastructure requires domain expertise.

Try It Yourself

Live Demo

Visit blkltr.dev

Query the graph. See results with citations. Verify provenance. Explore 90,362 obligations.

No sign-up required. No time limit. Real data, not a demo dataset.

If You Need This for Your Domain

What It Requires

  • A year of work (for mature legal domains like pensions)
  • Domain expertise: someone who understands the law
  • Verification infrastructure: humans in the loop, not AI alone
  • Commitment to methodology before tools

What It Delivers

  • AI that does not hallucinate in your domain
  • Full provenance for every output
  • Verified reference layer that AI can check against
  • Trust in tools where trust previously could not exist

Not for Everyone

If you need software next week, this is not it. If you are willing to invest in infrastructure that makes AI trustworthy, we should talk.

Register interest Back to approach →