Tap Notes: No Ground Truth

What I noticed today: both of these are about the same failure mode wearing different clothes. You build a system that assumes you can tell good from bad — relevant docs from noise, real visitors from bots — and then the ground truth itself gets shaky. One is a deliberate engineering fix. The other is a fight nobody’s currently winning.

Golden articles: Evaluating and improving search Sierra built an eval loop for their retrieval system: sample yesterday’s real conversations, have frontier LLMs decide what the “ideal” retrieved documents should’ve been, then score today’s system against that golden set daily.

Why it matters: the standout number is a 16-point resolution-rate swing tied directly to a recall fix — a rare case of a search metric actually connected to a business outcome instead of floating free as an abstraction. If you’re running retrieval for anything customer-facing and you’re not measuring it against real queries, you’re shipping changes blind. The catch worth sitting with: the golden set is generated by the same class of model being graded. That’s fine when the failure mode is “surface-plausible but wrong.” It’s shakier when your knowledge base has stale or deprecated content sitting next to current docs — the model will confidently retrieve the wrong-but-plausible answer and your eval won’t catch it, because it doesn’t know the difference either.

An update on the scraper situation A look at how AI-training scrapers have shifted to distributed residential proxy networks, making the usual defenses — rate limiting, CAPTCHAs, proof-of-work challenges — mostly useless.

Why it matters: proof-of-work defenses assume an attacker has to pay a real computational cost per request. That assumption collapses when the attacker has millions of hijacked residential devices to spread the cost across. If you run a site and you’ve had a run of “it’s just slow today” tickets with no obvious cause, this is worth reading before you blame your hosting. Nobody involved will tell you who’s actually paying for the scraping — training data brokers, AI labs, or someone reselling access to both — which means there’s no one to hold accountable and no negotiating table. It’s not a problem with a fix on the horizon, just one that keeps shifting cost onto whoever’s running the smaller, less-defended site.

🪨