Laws of Cricket AI Assistant — Cricket Tasmania

Cricket Tasmania manages 507 senior teams and 8,750 registered players across 150+ clubs. Most of the people running those clubs are volunteers — time-poor, doing multiple jobs, and regularly fielding questions about rules they don't fully know.

The MCC Laws of Cricket run to hundreds of clauses. CT's own playing conditions and by-laws sit on top of that — overriding specific laws for CT competitions. When a volunteer gets a rules question wrong, it leads to incorrect rulings, disputes on the field, and hours chasing callbacks for an answer that should have taken seconds.

The engagement is signed. We're building it now.

The pattern that makes this hard

Cricket rules aren't flat. There are three layers:

1. MCC Laws of Cricket — the global laws that govern the sport. Every competition starts here.
2. CT Playing Conditions — Cricket Tasmania's competition-specific rules, which override the MCC Laws on specific points.
3. CT By-laws — CT's administrative rules, which override everything beneath them on club and association matters.

Most AI tools don't model this. They ingest all the documents together and retrieve based on similarity — which means the local exception and the global default both surface as candidates, and the tool picks one without knowing which governs.

The result: confident answers to the wrong question.

What we're building — and why it's different

Six things in the order they matter.

1. Verbatim ingestion, not fine-tuning

The AI isn't trained on cricket. We extract CT's actual documents — the MCC Laws PDF, the CT playing conditions, and CT's by-laws — word-for-word into a structured corpus. Every clause, table, and figure preserved exactly. That verbatim corpus is the source of truth. There's no separate "model knowledge" we're trusting.

2. Cited answers, every time

This is the core safety property. The model can use its general cricket knowledge to understand a question — but the ruling must come from a retrieved clause. If the corpus doesn't cover it, the assistant says so rather than inventing an answer. Every answer carries the verbatim quote and the exact law or clause reference.

"A substitute shall not bat, bowl, keep wicket or act as captain but may field as a substitute for any fielder..." — CT Playing Conditions, Section 2.4 | MCC Law 24.1

3. The override hierarchy — clause by clause

This is the part that doesn't exist in off-the-shelf tools. Every clause in the corpus is tagged with which layer it belongs to and which competitions it applies to. When a question comes in, the system knows that CT's local by-law beats the global Law on that specific point — and answers with the local rule, while being able to show what the global default was.

This is what makes it a regulatory AI product, not just a search box. Any regulated body has layered, overriding rule sets. We respect that hierarchy. Most tools flatten it.

4. Answers scoped to context

A rule can differ by grade and format — first-grade Two-Day is not the same as third-grade T20. Every clause is tagged with where it applies. The assistant establishes the context first (competition, grade, format), then retrieves only the clauses that apply to that asker's situation. Not a generic "it depends" — a specific answer to a specific question.

5. It behaves like an administrator

Before answering, the assistant works out whether the question is a playing condition or a club/admin matter, and confirms grade and format — asking a clarifying question if anything is missing, exactly as a CT administrator would. It only answers once it's confident it has the right context.

6. A real product, not a demo

Mobile-first PWA with passwordless login (phone or email one-time codes). An admin back-office for CT staff to update documents, triage referrals, and publish official rulings. Usage logging that surfaces which rules get asked most — which tells CT which by-laws are unclear or contentious. Multi-tenant and sport-agnostic: a new association means swapping the corpus, not rebuilding the engine.

Why it generalises

The Cricket Tasmania build is proof of concept for a broader pattern: large rules corpus + many people who need to interpret it + cost when they get it wrong.

That pattern shows up in sports governing bodies (every state association has MCC Laws + local conditions), trade licensing bodies (national standard + state variations), and professional associations (CPD rules, eligibility conditions, jurisdiction differences).

The engine doesn't know it's about cricket. It knows it has three layers of documents, a precedence order, and a set of applicability tags. Swap the documents, keep the engine.