Proving a drug is fake
to somebody who owes you nothing.

Counterfeiting online was, and is, an attribute problem wearing a legal problem’s clothes. Somebody lists a product that isn't what it says it is, at a price the real thing can't be sold for, from a place the brand doesn't distribute through. Every one of those is observable. None of them, on its own, proves anything.

We started with consumer goods, because that's where the data was. Pioneer had a set of premium headphones, the HDJ-1000, that were counterfeited heavily enough to be a real revenue problem. The work there was commercial: protect the channel, protect the price, protect the brand.

Pfizer and Eli Lilly changed the stakes entirely. Our primary targets were Lipitor, Viagra and Zoloft, three of the most counterfeited drugs on the internet at the time, and three that people were buying without a prescription because they would rather not have the conversation in person. A fake set of headphones sounds bad. A fake statin goes into somebody who believes their cholesterol is being managed. That difference didn't change the mechanism much, but it changed every threshold we set and every decision about what the system was allowed to do on its own.

We built the whole thing on top of the eBay Developer API. Policing eBay was never the aim; their listing structure carried a fantastic array of product attributes: condition, location, price, seller history, category taxonomy, all of it structured, all of it queryable, at a scale nobody else was publishing in 2007.

That attribute set became the shape of the whole product. Once you've decided what a listing is made of, you have decided what your model can reason about, and every later part of the system inherits that decision. It's the same lesson I keep relearning twenty years on: the data model is the design, and everything you draw afterwards is downstream of it.

The core of it was attribute variance. For a given product we would set the bands that mattered: how far under the minimum advertised price a listing sat, which locations were excluded because the brand had no distribution there, what a normal seller profile looked like for that category, and score every listing against them.

What came out was a probable outcome, not a verdict. The system flagged a listing as possibly counterfeit, with the reasons attached and weighted. On the HDJ-1000 that meant a listing at forty per cent under MAP shipping from outside the distribution network scored very differently from the same discount inside it.

Tuning those bands per client and per product was the actual design work, and it's the part that doesn't look like design from the outside. There was never a screen for it, and it decided more than any screen did.

Pfizer and Eli Lilly had a different problem shape. Their counterfeits weren't sitting in a marketplace with a clean API in front of them. They were on third-party pharmacy sites scattered across the open web, which meant there was no attribute feed to consume at all.

So we used the eBay attribute structure as the foundation for a third-party module, and I built a crawler to go and find those sites. Each site, and each product it offered, ran through the same counterfeit analysis with variance profiles written specifically for that drug. Lipitor, Viagra and Zoloft each got their own: different price floors, different legitimate-distribution geographies, different tells in how a listing described dosage and packaging. A discount that means nothing on one of them is the loudest signal on another.

Keeping the eBay attribute shape as the schema even when eBay was no longer in the picture is the decision I'd defend hardest. It meant one scoring module instead of two, one place to tune, and one vocabulary for everybody arguing about a result.

On the marketplace side the evidence package went to eBay and the listing came down under their own terms. On the pharmaceutical side it went to the ISP hosting the site, accompanied by whatever applicable law we could materially support.

That last clause did a lot of work. We weren't making legal arguments we couldn't stand behind, and a report that overreached once would have cost us every report after it. So the standard was that every claim in the package had to be traceable to an observed attribute, and anything we couldn't source didn't go in.

It worked. Sites came down, listings came down, and the brands stopped bleeding into a gray market that was, in the pharmaceutical case, actively dangerous to the people buying from it.

Look at what it was. A system that gathered evidence across sources nobody could read by hand, scored it probabilistically, produced a recommendation with the reasoning attached, and stopped short of taking the action itself. Every claim it made walked back to an observed attribute, because the whole thing was going to be read by somebody looking for a reason to say no.

That's the same line I hold at Oracle nearly twenty years on: the agent researches everything and approves nothing, and a grade whose evidence you can't reach is a more expensive opinion. I didn't arrive at that position from a design-systems argument. I arrived at it from work where being confidently wrong meant accusing an innocent seller, or worse, leaving fake medicine up for another week.

It's also where I learned to lead this kind of work rather than only do it. I built and ran the research team behind these products, servicing dozens of Fortune 500 customers. Two patent applications came out of that period as well, one for the detection system and one for the crawler, though neither is in force today.