Writing
Notes on verifying what AI systems do.
Essays and open data from the lab: where AI systems behave, where they fail, and how to prove the difference. Every piece links to something you can check or run yourself.
The server you approved last month is not the one your agent calls today
An MCP server can change what its tools tell your agent after you approve it. One real server changed its tool list six times in a few weeks, from 85 tools to 103. Your agent cannot tell an update from an attack, so the check has to be a simpler question: has this changed since you approved it?
The AI everyone talks about is not the one developers run
The conversation is owned by a few frontier labs and their five-dollar models. The board developers actually route traffic on has DeepSeek at number one for nine cents, and the celebrated flagships mid-table at fifty times the price. Every number on it is checkable.
Games apps answer 3.5% of their reviewers. Strip out the templates and it is 1%.
Twenty of thirty-nine tracked Games apps reply to no recent reviewer at all. Underneath a lifetime average of 4.61, nine apps have fallen more than half a star against their own reviewers of a year ago and nine have risen.
YouTube TV lost more than a star with its own reviewers. ReelShort gained one and a half.
Eight of twenty-six tracked Entertainment apps have fallen more than half a star against their own reviewers of a year ago, and six have risen, with moves past a full star in each direction. The category's lifetime average, 4.56, shows none of it. Fourteen of the twenty-six reply to no recent reviewer at all.
Grok, Perplexity and Claude lost their reviewers in a year. ChatGPT and Gemini kept theirs.
Four of the twelve tracked Productivity apps have fallen more than half a star against their own reviewers of a year ago, and three of the four are AI assistants, with Grok and Perplexity down a full star and Claude down just over half. ChatGPT and Gemini have held flat. Five of the ten apps with enough recent reviews to measure reply to nobody, and none of the five assistants answers more than one reviewer in two hundred.
Model collapse is a provenance problem, and provenance is checkable
Over half the open web is now AI-generated, and the next training run will scrape it. The fix is not bigger models. It is data provenance, recorded at the source and checkable later. Why the scaling answer fails, and what a verifiable data discipline looks like.
Google externalised the cost of renaming Gmail
Google shipped Gmail address renaming and never shipped a webhook. 124 open-source projects still key OAuth identity on email. What breaks, and who pays.
124 repositories. Four ecosystems. One broken assumption.
The reproducible data behind the essay: 2M+ repositories scanned, severity tiers, ecosystem breakdown, and the full methodology. A complete audit.