In 1989, Yann LeCun — often called one of the godfathers of modern AI — demonstrated something that would shape how we think about machine intelligence for decades. Working with the US Postal Service, he built a neural network that could read handwritten ZIP codes on envelopes. But the system had a design choice baked in that most people overlook: it didn't try to read every envelope. It only processed the ones it was confident about, and passed the rest to human operators.
That decision — to act selectively rather than universally — made the system powerful enough to process nearly 10% of all checks written in America within a decade. Not because it was perfect, but because it was trustworthy on the cases it touched.
There's a comfortable assumption that circulates in gold lending conversations: that a computer vision model can look at gold jewellery and reliably detect spurious collateral with 80% or better precision. In practice, this rarely holds up.
Gold is uniquely difficult to evaluate at scale. The same ornament looks different under different lighting. Items worn for decades develop surface characteristics no dataset fully captures. Presentation varies dramatically across branches, geographies, and even individual loan officers. The demo that impressed in a controlled environment rarely survives contact with production — where the data is messy, the lighting is inconsistent, and the edge cases are the rule, not the exception.
Chasing a headline precision number misses the more important question.
The question we stopped asking is "what precision can we achieve?" The question we started asking instead is: of all the spurious collateral sitting across a portfolio right now, how much can we actually surface?
This is a fundamentally different metric — recall over a defined set of alerts — and it turns out almost nobody in gold lending can answer it today. Not because the math is hard. Because the ecosystem to measure it has never existed. To measure spurious collateral recall at scale, you need historical repositories of flagged cases tagged and accessible, structured alert systems that surface anomalies consistently, branch-level labeling that builds ground truth over time, and production-scale feedback loops that let the system learn from what it gets right and wrong. None of these existed in the industry when we started. So we built them.
Today, Rootflo measures spurious collateral recall across 3,000+ live branches. Catching 40% of spurious collateral across our defined alert set isn't a limitation we're apologising for — it's a benchmark the industry didn't have before we created it.
That number will improve over time, as it always does when you have the infrastructure to measure and iterate. But the starting point — having a number at all — is itself meaningful. You can't improve what you can't measure. And for the first time in gold lending, this particular thing is being measured.
The postal system Yann LeCun helped build didn't win because it read every envelope. It won because it was the first system that could reliably read any envelope at scale — and because it knew, precisely, which envelopes those were.
Gold lending is at a similar inflection point. The infrastructure to detect, measure, and improve spurious collateral identification is being built right now. The question isn't whether the system can be perfect. The question is whether it can be trustworthy — and whether that trustworthiness can be measured.
We're reading the first envelopes.
