I pointed an agent at the AI buildout and told it to kill its own ideas.
Every weekday it maps a layer of the stack, follows yesterday's open questions, and puts anything it finds through three reviewers whose job is to prove it wrong. On day one it produced fourteen ideas and killed all fourteen. That result is the reason this page exists: the interesting part of research is what does not survive.
The rules that make it useful
A daily research agent turns into a news roundup unless something stops it, so three things are enforced in code rather than asked for politely. The report has a word budget and the publisher refuses to ship one that runs over, which is why these take about five minutes. State lives in threads, and the agent has to advance an open question before it is allowed to open a new one, so the days compound instead of resetting. And every report has to make at least one falsifiable call with a resolution date, because an agent that has to live with its own record cannot get away with summarising headlines.
It never places a trade, never touches a brokerage, and never sees my actual positions: the publishing step reads only the research files and refuses to write if position language shows up in what it is about to put on the open internet.
Research and analysis, not investment advice. Nothing here is a recommendation to buy or sell anything, no real money is involved, and no trade is ever placed from this system. I'm not a licensed financial adviser.