← Back to blog

2026-01-19

When AI Pulled a \"Move 37\": Testing Our Hallmarking Knowledge Base

An unexpected lesson in institutional knowledge and artificial intelligence

If you follow the world of artificial intelligence, you've probably heard about "Move 37."

In March 2016, during the historic match between AlphaGo and Lee Sedol (one of the world's greatest Go players), the AI made a move so unexpected, so seemingly nonsensical, that commentators thought it was an error. Move 37 violated centuries of Go strategy. It looked wrong.

Until it turned out to be brilliant.

That phrase - "Move 37" - has become shorthand for the moment when AI reveals knowledge or insights that surprise even the experts. Last week, while testing our new hallmarking information project at the London Assay Office, we had our own Move 37 moment.

The Test Question

We're building an AI-powered knowledge base for hallmarking regulations - a system designed to make 200+ years of accumulated expertise accessible to jewelers, retailers, and consumers. As part of our testing, I posed what seemed like a straightforward question:

"If I have a ring that's already hallmarked in Sheffield, can I send it to London to add my sponsor's mark?"

It's the kind of question that stumps people regularly. The logic seems clear: the piece is already hallmarked by one assay office. Different offices use different marks. Surely you can't mix them?

The Answer That Surprised Me

Claude's response was immediate and confident:

"Yes, you can do this! It's called an Additions Mark."

Then it proceeded to lay out the exact criteria:

  • • The addition must be 1 gram or less (for gold)

  • • The addition must be 50% or less of the article's total weight

  • • The article must keep its same character and purpose (it stays a ring)

    • And here's the detail that showed genuine understanding: Since the existing hallmark is a Sheffield mark (not London), your additions mark would be formatted differently - meaning you'd get both the original Sheffield mark AND the London additions mark on the same piece.

    I had to double-check. It was completely correct.

    Why This Matters

    This isn't just a party trick. It reveals something important about the nature of expertise and how we preserve institutional knowledge.

    The Additions Mark provision exists precisely for scenarios like this - repairs, modifications, or additions to existing hallmarked pieces. It's an elegant solution that maintains traceability while acknowledging the reality of how jewelry is worked on throughout its lifetime.

    But here's the thing: even experienced professionals in the jewelry trade don't always know about it. Why? Because this knowledge typically lives in:

  • • Dense regulatory documents most people never read

  • • The collective memory of long-serving assay office staff

  • • Occasional edge cases that aren't encountered frequently enough to become common knowledge

    • When someone encounters this specific scenario, they might call us, visit in person, or simply assume it can't be done. The knowledge exists, but it's not easily accessible.

    Democratizing Expertise

    This is exactly why we're investing in AI-powered knowledge systems.

    We're not trying to replace human expertise - the London Assay Office's 700-year tradition is built on skilled professionals who understand the nuances of precious metals, craftsmanship, and regulatory compliance. But we can make that expertise more accessible.

    Imagine a jeweler in a small workshop in Cornwall encountering this exact question at 9 PM on a Tuesday. Instead of waiting until morning to call us, or making an assumption that costs them time and money, they get an accurate answer instantly.

    That's the power of well-structured institutional knowledge combined with AI.

    The Bigger Picture

    As we continue developing this system, we're learning valuable lessons about how AI can serve traditional industries:

    1. Precision matters more than speed

    Getting the answer right is more important than getting it fast. We're building extensive testing protocols to ensure accuracy across thousands of edge cases.

    2. Context is everything

    Hallmarking regulations aren't just rules - they're systems designed to solve real-world problems. Our knowledge base needs to understand not just what the rules are, but why they exist.

    3. AI excels at surfacing the non-obvious

    The Additions Mark isn't obscure because it's secret - it's obscure because it's specific. AI can hold thousands of these specific provisions in parallel and surface exactly the right one for each query.

    What's Next

    We're still in testing phase, and this "Move 37" moment is exactly the kind of validation we needed. It shows the system isn't just regurgitating basic information - it's genuinely understanding the nuanced, interconnected nature of hallmarking regulations.

    Our next challenges include:

  • • Edge cases from historical amendments (those 1973 provisions are particularly tricky)

  • • International hallmarking conventions and equivalencies

  • • Guidance on modern materials and techniques that didn't exist when many regulations were written

    • But if our AI can nail something as specific as Additions Marks for cross-office hallmarking, I'm optimistic about what else it can unlock.

    The Human Element

    Here's what struck me most about this experience: the delight of discovery.

    Even though I work at the London Assay Office, even though I'm leading our digital transformation initiatives, I learned something new about our own regulations. That sense of "wait, really?" followed by "that's brilliant!" - that's exactly what we want users to experience.

    Because ultimately, that's what good knowledge systems do. They don't just answer questions. They reveal the hidden elegance in systems we thought we already understood.

    Sometimes the most sophisticated answer is just knowing which exception applies.

    And sometimes, AI helps us rediscover the depth of our own institutional wisdom.