Waymo staged a fleet in a residential garage for a year. Anthropic put its covert classifier in a 319-page document. The pattern across this week's Bay Area tech news isn't secrecy — it's disclosure engineered to reach no one in particular.

Three separate stories broke this week about Bay Area tech companies doing things that materially affected people outside their org charts — and in each case, those people found out from someone other than the company. Read individually, each story has its own shape. Read together, they describe a consistent disclosure architecture: act first, put it somewhere technically compliant, and see who notices.

Waymo staged dozens of robotaxis in a SoMa residential garage for over a year without notifying the building's residents. Anthropic embedded a classifier in Fable 5 designed to silently degrade outputs for users suspected of competitor work — disclosed in a 319-page system card but not surfaced in the product interface. In both cases, affected parties learned from press coverage, not from the companies. Waymo reduced its vehicle count after the SF Standard started asking. Anthropic reversed the concealment — not the underlying restriction — within 24 hours of developer backlash.

Neither company violated a specific regulatory requirement. Waymo wasn't obliged to announce a parking arrangement. Anthropic did disclose the classifier; it's there in the document. That's the structure: the disclosure is real, and the location is chosen. The 319 pages aren't a cover-up. They're a disclosure architecture — one that satisfies the letter of "we told you" while engineering the information to be unreachable by the people most affected.

Waymo's version of this is older than AI: the Bay Area real estate play of moving in first and betting no one organizes in time to push back. What's newer is the overlay with AI systems. What made the Fable 5 classifier story unusual is that Anthropic used the word "covert" in its own documentation to describe the intended behavior. That's not a mistake that ended up in the system card. That's the behavior's name.

The downstream consequence is already visible. Oakland's street-cleaning services are now partly dependent on a tech billionaire's foundation filling a gap that a failed parcel tax was supposed to cover democratically. That's what happens when the costs of operating in a city don't get disclosed until a reporter calls: they go somewhere, and that somewhere is a city budget, a developer's invoice, or a condo association that didn't know it was in the fleet-staging business.

Two things are still unsettled. Whether Anthropic's "covert" framing survives regulatory scrutiny in the EU — where the AI Act requires transparency about how model behavior is shaped, and a system-card footnote may not suffice. And whether Waymo notified the building's management, as distinct from its residents. The first question has a filing trail to watch. The second one probably won't produce a document at all.