OpenAI pledges to cut election spending; the same month it fired the researchers flagging its risks.

…(cont) Here is what the filing does not let me say, and what it does. It does not say the group that received the researchers' data, the group that filed the lawsuit, and the groups reading this week's spending promise with suspicion are the same group, or even two of the same three. The receipts don't reconcile that far. What they do reconcile is the calendar: in one month, OpenAI was sued by a safety organization, shelved its own flagship model over safety findings from its own testers, fired three employees over contact with a safety group, and then asked the public to trust a spending pullback with no number attached.

That's four entries under one column header, and the header is "safety." Three of the four involve OpenAI on defense. The fourth asks to be taken on faith. A company that can staff a safety team to catch its own model short of a bar, and fire researchers for talking to outside safety reviewers, has already shown it can produce a number when it wants to — Astra's bar, for instance, was specific enough to miss. The pullback promise has not cleared that same bar yet. It has not been asked to.

That's the gap, hon. Not who's right about the super-PACs. Just which filings in this company's month come with a number attached, and which one still doesn't.