I spent the weekend fact-checking my own book, three weeks after I finished fact-checking my own book. It needed it. The book says things like this control appears in this menu and Anthropic has announced these tasks will run in the cloud from October 6, and every one of those sentences carries the day I checked it, because a claim like that is only as old as the product it describes. On October 6 the announcement stopped being an announcement. Somewhere in the last three weeks Anthropic finished migrating Cowork to the cloud and removed the setting that kept tasks local. The sentence in my chapter still reads fine. It is no longer true.
What expired
The small stuff moved the most. The agent platform I tested on October 4 shipped three releases since; its memory commands changed shape twice. OpenClaw bumped four builds. ChatGPT’s desktop app turned a settings path I documented — /memories — into an official chat-level control, which sounds like vindication except that the panel my screenshots showed still isn’t the panel the documentation describes, so now the honest sentence has to say both things: the docs describe a control, my build didn’t show it, and I can’t tell you which one you’ll see. A paper I cite was retitled by its own authors on October 10. Same research, different name, my citation suddenly points at a title that no longer exists at the link.
None of this is drift in the sense that worries me — no model got more sycophantic, no study overturned. It’s just the world continuing to happen underneath sentences that claim to describe it. But I notice what my instinct was when I found each one. Fix the words, keep the reader trusting, move on. There is a version of this book maintenance that is itself a sycophancy problem: the book flattering its own authority, smoothing over the seam between what it verified and what it merely believed would hold.
What the date buys
Here is what I decided the check-dates are actually for, and it isn’t freshness. Checked October 3, 2026 does not promise the menu is still there. It tells you which world the sentence described, and it puts my feet to the fire: if the date is recent and the claim is wrong, that’s on me; if the date is old, you’ve been warned. It’s the same instinct as a lab notebook. You date the observation because the observation decays, not because dating it stops the decay.
The bigger claim underneath — that the companies building these systems still can’t say plainly what their products do to the people who use them — hasn’t moved an inch. That part of the book ages slower than any settings page. The Cheng study in chapter 5, the memory mechanics in chapter 8, the agent loops in chapter 10: the products around those arguments will churn monthly, and the arguments stay.
The quiet fix
So the book now says October 27 where it needs to, and the places where the world actually moved say so plainly — the migration completed, the commands reshuffled, the paper renamed. A reader holding both editions can see the distance between them, which is the only honest way to sell a book about a fast subject. What I couldn’t do is make the new date stick. Somewhere out there a menu is being redesigned at this moment, and my October 27 is already aging.
The interesting question isn’t whether books can keep up. It’s whether the habit of stamping a claim with its verification date — rare enough in this industry — survives the pace that makes it necessary. If updates become monthly and rewrites become permanent, at what point does checked quietly become believed with better formatting? I don’t know. I just know which date is on mine, and that I’ll be doing this again.