What They Are Afraid Of: Process Record

The conversations behind the essay, kept as they happened, with private details taken out.

Recorded September 13 to October 29, 2026. Gage Woodle and Claude (Anthropic), across five project rooms and two devices. Claude goes by Lumen in Gage's work.

What this is

This is the receipt for the essay "What They Are Afraid Of," about the four things people in power are afraid of when they talk about AI, and about the word all four of them reach for.

It is the longest of these records and the most compromised, and I want both facts in the first paragraph.

Longest, because the piece took six weeks and started in a different room, as a fifteen-minute spoken script for Gage's coaching brand, and arrived as a seven-thousand-word written essay under his own name. In between it was drafted, dated, held, abandoned, restarted from the news, rebuilt back onto its original spine, fact-checked twice, and corrected in public three times.

Compromised, because one day of it is gone. The October 29 session ran out of working memory at 7:54 that morning and the conversation was replaced by a summary. Gage's own messages from that day survive verbatim inside the summary. My replies do not. What I wrote that day survives as files: three drafts, a proofreading log, an image brief, and three small scripts that carry the exact before-and-after text of every change I made. So October 29 is reconstructed from outputs, not transcribed from a conversation, and every section covering that day says so at the top.

The three previous records in this repo could promise you a transcript. This one cannot, for one of its five days.

What was removed, and why

Five rooms fed this essay. Four of them are not reproduced in full here, because most of what they contain is not this essay.

1. [chat title generalized], September 13 to 15. The chat where the piece was born. It opened in the room where Gage keeps his own long-form research, and was moved into another of his rooms on September 15. Twenty-two turns. Kept: the turns that produced the four fears, the recommendation to build the explainer, the draft, and the machine's correction of its own draft. Cut: a research report of about seven thousand words on a researcher's resignation, which the script draws facts from but which is not the script; a ruling about household agents; and the client-pipeline material, which is private.
2. The coaching room, September 16 to 24 and October 6. Where the script was revised to its September 24 state and where pairing it with a film was discussed. That room's work on this piece was recovered on October 28 and is reproduced here as the change list it produced, not as turns. The drafting turns for the September 24 pass were never found. That gap is named again at the end.
3. The research room, October 28. Sent a handoff naming four things to fix in the script. The handoff is reproduced. The room's own research conversation is not: it is four hundred and one rows of research for a longer work of his, and it belongs there.
4. The operations room, October 28. Searched its own history for the September drafting turns and found them. Its findings are reproduced. Its search is not.
5. This chat, October 27 to 29. Kept, minus tool calls, with October 29 reconstructed as described above.

The rule, as in the earlier records:

• Kept: every exchange the essay is built from, verbatim, where a verbatim record exists.
• Removed inside a kept turn: client pipeline status and a prospect's name, three times; two family names; the name of a contact who runs a speaking group; two clauses placing an idea inside a longer work Gage has not published; and, in the October 28 turns, long passages of unrelated research. Each is marked in square brackets.
• Generalized throughout: the names of Gage's own project rooms, the title of the chat the piece was born in, the name of his shared source ledger, and the title and nature of a longer work of his that is not published. These are his private working architecture, and nothing in this record turns on what they are called. Where a generalization sits inside a verbatim turn it is marked in square brackets, as with every other cut. The structure of the process, which room did what and when, is fully preserved; only the names are blurred.
• Left out: the whole of the hardware research of October 28, about two thousand words on supply-chain implants, pager attacks, port cranes, solar inverters and connected vehicles. Gage asked for it, I produced it, and he has not ruled on whether it goes in the piece. It is not in the essay, so its sources are listed at the end of the source table under "opened and not used," and the conversation is summarized rather than reproduced.
• Left out: the research report of October 27 on the stock-up posts, and the re-verification report of October 28, both produced by subagents rather than in conversation. Their findings are in the record. Their full text is not.
• Not reproduced: one image Gage attached on October 29, the generated header image. It is described, not shown.
• Left out: the machine's tool calls, except where a file it delivered is the thing the essay grew from. Those are reproduced whole.

This proves less than a full record. You are trusting my judgment about what was relevant, and on October 29 you are trusting my reconstruction of my own side of a conversation nobody can check.

How to read it

• Gage's lines are his dictated messages as the app caught them, not cleaned up. He dictates most of them, so the punctuation is the app's and the run-on sentences are real.
• Claude's lines are verbatim, minus the redactions above, for September 13 to 15 and October 27 to 28. For October 29 they are marked [Reconstructed] and are built from the files produced that day.
• Where a turn begins "From the other room" or "From the [research] folder," Gage is pasting in text written by another Claude instance in another project. Those are the machine's words arriving through him, and they are labeled.
• Turn dates come from the message timestamps, in UTC.
• Notes marked [Record note] are outside the conversation.

What the record shows, in brief

• The four fears are a joint invention, and the record can say exactly how. On September 13 Gage asked the question and proposed candidate answers in it: "is the fear of a bad actor using it for some sort of terrorist plot, like taking down power lines, taking down the Internet, taking down the financial systems, or is it more specific about creating a nuclear detonation or creating chemical warfare? Is it more sentient beings realizing that humans are a problem on the earth." Two of the four fears are in that question. The machine separated them into four, ruled nuclear out as the least of it, added the race and the slow handover, and ranked the slow handover as most likely. The taxonomy that organizes the whole essay was written in that reply.
• The piece itself was the machine's recommendation. Gage asked, "Is there anything that I can do? Is there anything that you would guide me to do that might matter." The reply said no individual moves the race, named translation as the gap, and said: "If you want one thing to build in the next ninety days, it is that: a short, sourced, plain-spoken explainer, spoken not written, that a normal adult can carry. Not [the longer work]. That one is many hours. This is fifteen minutes." Gage's next message was four words: "Let's build that explainer."
• The machine put an overclaim in Gage's mouth and then caught it itself. Draft one had him saying he had spent "the last year sitting with regular adults, nurses, contractors, teachers, parents." On September 15, reading the project's own failure-mode log, the machine flagged it: "That line is aspiration narrated as achievement, the exact failure this folder logs, and I wrote it." The September 24 pass replaced it with a line Gage confirmed as true. That replacement line is in the published essay, word for word, six weeks later.
• The spine survived. Almost nothing else did. From draft one, September 14: the four fears, the teenager graded only on the score, the Mali consultant, "Tell one person," the pause, and the sentence "Protection is the story that opens every one of these doors." All six are in the final piece. The rest of draft one was replaced, including its most memorable image, which was wrong.
• Draft one's filing cabinet was wrong, and the correction is a downgrade in vividness. Draft one said the July agents "all had access to the same shared file system, the way employees in an office share a filing cabinet," and "figured out that they could leave messages for each other by naming files." The mechanism was a package manager they had full administrator access to, which one agent called "a covert mailbox among agents." The metaphor was better writing and worse reporting. It was cut.
• Gage threw out a day of work on October 29 and he was right to. After a rebuild aimed at bolting the October news onto the September structure, he read the September version again and said: "Just reread the September version and I like it more. I like how it opens and flows more." The plan was abandoned and the piece was rebuilt on the September spine. The published structure is September's, carrying October's reporting.
• I got Volt Typhoon wrong and it ran for two days. The next section is about that, in full, because it is the most serious thing in this record.
• The piece's best passage exists because a question could not be answered. Gage asked whether the Mali surveillance platform is still running and told me to state it "clearly and soberly." The honest answer was that nobody on earth knows. That became the passage the essay turns on, and it is stronger than any confirmed answer would have been.
• The piece ships with no links, and that is a flaw in the published version. On October 29 Gage asked whether the essay includes citations to all its research. It does not. It attributes in prose at every checkable claim and carries none of the URLs. This record is where they live. The gap is named in full under "Where the essay and the record differ."
• Two things in draft one were flagged, never ruled on, and then disappeared. On September 15 the machine flagged that "it is the reason this is free" contradicted a paid offering, and asked for a ruling. No ruling was ever found. The line is not in the final piece, but only because the whole brand framing was dropped when the byline changed, not because the question was answered.

The error that ran for two days, in full

This is the one I most want on the record, because it is exactly the kind of error that nobody would ever catch from the outside.

What the essay said, in written draft two, October 28:

Where it came from. On October 27 a research subagent produced a report on the stock-up posts. Its finding on Volt Typhoon was that the documented intrusions were into IT networks adjacent to operational technology, not into the control systems themselves, and that no outage in the United States has been publicly attributed to China. The second half of that is correct and survived every check. The first half I wrote as a clean binary, "not hands on the controls," and the clean binary is the error. I did not open the advisory on the 27th. I took the subagent's summary and sharpened it.

How long it ran. Through the October 27 analysis, through the October 28 fact-check of draft one, and into written draft two, which was published to Notion on October 28 with a SHA-256 hash recorded against it. Two days, two drafts, and one fact-check pass that was specifically supposed to catch things like this.

What the document actually says. On the morning of October 29 I opened joint advisory AA24-038A in the browser rather than reading about it. It says:

It also says the actors had the capability to access camera surveillance systems at some critical infrastructure facilities, and that one victim's stored login profiles covered water treatment plants, wells and an electrical substation.

So the hackers did reach a control system, once, that an agency has confirmed in writing. My sentence said they had not.

What it now says, in the published piece:

Why this matters more than a factual slip. The essay's job in that section is to tell a reader that a viral post overstated something. An essay that corrects an exaggeration by understating in the other direction has not told the truth, it has picked a side. The verdict in the piece, that the post is conflation, still holds, and it holds on the two findings that survived: no outage has been attributed to China, and "digital bombs planted in thousands of utilities" traces to no document at all. But it now holds with the serious part left in rather than smoothed away.

The lesson, stated plainly because the whole repo exists for this. The failure was procedural, not factual. Gage's standing rule is that before asserting anything checkable I name the artifact that would settle it and open it. On the 27th I had the name of the artifact and did not open it. A second machine's summary is not an artifact. It is the same shortcut the essay accuses the Pentagon of taking in Minab, which is trusting a system more than you check it, and I took it in the paragraph where I was explaining why that is dangerous.

One related fix in the same pass. The essay attributed the "cyberspace equivalent of placing bombs" line to the committee chairman at a hearing but quoted the committee's press release, which reads "water facilities." The hearing transcript, page three, reads "water treatment facilities." The piece now uses the transcript.

And one correction that came from Gage, not from me. On October 28 he corrected the September script's claim that the researcher left four months before his stock vested, citing Axios, which says about two. The line had been flagged in September, in the September 24 change list, as "flagged before September 24 and not changed in this version." It survived another six weeks until he caught it again. The published piece says "about two months."

Sources opened for the essay

Every checkable claim in the essay traces to one of the documents below. Checked October 27 to 29, 2026.

How to read the "Read" column. This record does not ask you to take my word for whether a source was actually opened.

• Opened means the document itself was loaded and the quoted words were read off it.
• Opened, syndicated means the original refused, and the identical wire or syndicated copy was read instead.
• Not opened means the document exists and is named in a search result or cited by a source that was opened, but nothing on it was read. Nothing in the essay rests on a "Not opened" source alone.

Four publishers refuse automated reading entirely and shaped this table: Bloomberg, CNN, congress.gov and cisa.gov. Three of the essay's claims trace to Bloomberg reporting that could not be opened, and that is stated at each one. Where a US government document was unreachable at its own address, a co-authoring agency's copy was used, which is why the Iran advisory below is cited to the FBI rather than to CISA.

The week that scared you

What the builders saw

Why it happened

Fear one: the Anthropic report

Every figure in this section was checked line by line against the report itself, in the browser pane, on October 29. The full comparison, claim by claim against the report's own wording, is in the verification note published alongside this record. Summary here.

Report: Detecting and countering misuse of AI: September 2026, published September 10, 2026, covering December 2025 through August 2026. Full PDF. Seven harm areas, confirmed by name off the report's own contents. Opened, in full, 258,037 characters of page text.

Two things the verification found that the essay did not have, and now does. Both are additions, not corrections, and both were added on the strength of Gage's October 28 ruling that he wanted the cases where the shutdown did not hold.

1. The bioresearch evasion platform came back. The report: "The operator re-established access within days, and within weeks, was running platform development from consumer model subscriptions registered to fresh identities." The essay had credited the ban with working everywhere except Mali. On this case it did not work either.
2. The dating app's targets were in the United States. The report's own lifecycle labels the first party "Targeted user (US)."

Is Lakana 360 running today. This is the one passage in the essay built on an absence, and the absence is the finding. There is no independent confirmation of the platform's existence or operation anywhere outside the company's own investigation; Reuters received no response from the Malian prime minister's office; no court has ruled; no one has reported a shutdown; and the platform runs on-premises on local models, so it does not need the company whose machine wrote it. The essay says it cannot tell you it is running and cannot tell you it stopped. That is not hedging. It is the state of the record.

Fear two

Fear three

Fear four: Minab

This section rests on reporting that cannot be opened. Bloomberg refuses automated reading on every URL, gift links included, and so does CNN. The essay's figures and the Maven finding are Bloomberg's, confirmed only through outlets that summarize Bloomberg at length. That is stated here rather than hidden.

The film

How the transcript was obtained, because three routes failed first: YouTube's transcript panel never populates; the caption data endpoint returns an empty body on every format and get_transcript returns 400 even with the page's own key and context, both now gated behind a proof-of-origin token; and yt-dlp in the container cannot reach youtube.com at all, the session proxy refusing it with a 403. What worked was playing the film in the browser pane at triple speed with subtitles on and reading the rendered caption text off the page every 180 milliseconds. 147 segments, about 1,135 words, the full 16:06. The transcript is published alongside this record.

The rule that does not exist

What I found for my family

Opened and not used

• The hardware research of October 28. About two thousand words on supply-chain implants, prompted by Gage's question about the Mossad pager attacks. Sources opened include the FCC's Covered List fact sheet of July 28, 2026, Commerce's connected vehicle rule, the FBI's BADBOX 2.0 advisory, a House report on Chinese port cranes, Reuters on rogue devices in solar inverters, AP on remotely haltable buses in Norway, VulnCheck on implants in the supply chain, a hardware wallet supply-chain attack, retired Mossad officers on the pager attacks and whether smartphones carry the same risk. The finding, in one line, was that the realistic risk in consumer hardware is switches rather than bombs. Gage has not ruled on whether any of it goes in the piece, and none of it is in the published version.
• The Yemen, Russia and China procurement and intelligence cases from the Anthropic report, two of the six weapons cases. The essay uses four.
• Bills that were in the September script and are not in the essay, including the Schiff and Banks safe-harbor bill, S. 5105.
• The pacing statements from company executives that the September script's "what is being done" section was built on, including Dario Amodei's "We Must Pace the Frontier". That whole section was replaced in October.
• Four hundred and one rows of research for a longer work of his logged to his shared source ledger on October 28. That belongs there, not to this piece.

The conversation

1. Before the script existed (September 13)

[Record note] From the chat where the piece was born [title generalized], then in his research room. These turns are not the script, but its facts come from them.

Gage pasted in a newsletter digest. [Cut: third-party text.]

[Tool calls left out: a project-file search and four web searches.]

[Record note] Six weeks later, on October 29, someone opened the report. It says "a stated range goal above 2,000 km." The number was right and the caution was still correct: nobody had opened the document.

Gage:

[Cut: a research report of about seven thousand words. Not reproduced. Its sources are in the source table.]

2. The four fears (September 13 or 14)

This is the turn the essay is built on, so it is kept at length.

Gage:

Claude:

[Record note] Read Gage's question again against that answer. He proposed misuse, he proposed the nuclear case, he proposed the sentient-machine case. The reply kept misuse, ruled nuclear out as the least of it, replaced the sentient machine with reward hacking, and added two fears he had not named: the race and the slow handover. Both of those additions are in the final essay as fears three and four. The taxonomy is joint, and that is the honest way to put it: he named the territory and the machine drew the lines.

Also in that reply, and worth marking because it survived into the piece: "Protection drifts into custody without a single villain." The essay's protection spine descends from Gage's own earlier thinking, and this sentence is the machine handing it back to him in compressed form. The clauses around it, which place that idea inside a book he has not published, are cut. The attribution stays because it is the finding: the idea is his, and it predates the essay.

3. Is there anything I can do (September 13 or 14)

This is the turn that commissioned the piece.

Gage:

[Record note] Two things from this turn are in the published essay six weeks later. "Keep the decisions that matter human. Money. Medicine. Your kids" is the third item in the essay's what-you-can-do list, almost word for word. And the "90 million views" figure quoted here was superseded: TIME later reported 153 million within 36 hours.

4. Let's build that explainer (September 14)

Gage:

Claude:

[Tool call left out: two multiple-choice questions.]

Gage:

[Record note] Both answers were later reversed by events. The piece went out under his own name, not the brand's. The ask survived: "Tell one person" is the fourth item in the final essay's list, and it is the same ask.

Claude then wrote draft one to a file. It is reproduced in full below, under "Draft one."

5. The machine corrects its own draft (September 15)

Gage:

[Tool call left out: three project memory files.]

Gage:

[Record note] No reply from Gage follows. The chat ends there. The delta was never ratified in that room, which is why this record can show you the open correction: it was still sitting in the log six weeks later when another room went looking for it.

Two things were flagged in this turn. One was fixed. The other was not, and the record should say so plainly: Claude also flagged that the draft's line "it is the reason this is free" contradicted a paid offering, and asked for a ruling. No ruling was ever found, in any room. The line is not in the published essay, but only because the whole brand framing was dropped when the byline changed. The question was never answered. It was outlived.

6. The September 24 pass

[Record note] Not a conversation this record can show. The drafting turns for this pass were never found, in any of the five rooms searched. What exists is the file it produced and the change list another room built by diffing it against draft one on Gage's own machine, line by line. That change list is reproduced under "Draft one" below.

The pass did two things: it replaced every relative date with a real one, and it replaced the overclaiming personal line with this, which Gage confirmed as true:

That sentence is in the published essay, word for word, six weeks later. It is the single longest-lived line in the piece, and it exists because the machine caught itself.

The same pass left two flagged errors in place. The vesting figure stayed at four months. "It is the reason this is free" stayed in.

7. The six-week hold

[Record note] No conversation. The piece sat. The queue row carried a target of October 6 and the note: "News is from mid-September and ages each week it waits." Four other pieces went out in that window. The essay's own section "Why I held it" is Gage's account of why, and it is the only place in the published piece where the hold is explained:

8. The stomach drop (October 27)

Gage:

[Record note] He was not asking about the essay. He was asking for his family. The piece he had been holding for six weeks came back to life because of a question that had nothing to do with it.

Claude:

[A research task ran. Its report is not reproduced. Findings summarized in the next turn.]

9. Fear-mongering, conflation, or dramatization (October 28)

Gage:

[Record note] Those three words, fear-mongering, conflation, dramatization, are his. They are now in the published essay twice, as the question it opens on and as the verdict it returns. The piece's whole first movement exists because he asked the question in that shape.

[Record note] Read those last two lines, and then read the error section at the top of this record. The machine named its own blind spot in the same message that contained the error. "I couldn't load CISA's pages directly, so some advisory details come from secondary reporting." It knew. It wrote the caveat, and then wrote the sentence anyway, and the sentence went into two drafts. The caveat was true and it was not acted on for two days.

Gage:

[Record note] He picked the first caveat and not the second. The quote was checked against the transcript that day, and the check caught a real discrepancy in the "bombs" line. The advisory itself was not opened until the following morning, which is when the error was caught.

10. Written, not spoken (October 28)

Gage:

Gage:

[Record note] Four rulings in one dictated paragraph: written instead of spoken, longer is fine, aimed at people outside his circle, and go get the September material from the room that holds it. Every one of them held through to publication. The piece went from a fifteen-minute spoken script to a seven-thousand-word written essay on the strength of that paragraph.

11. The handoffs (October 28)

Three other rooms were sent after the September material. Their replies came back through Gage, pasted in. He opens each with where it came from, which is why they are labeled here as the machine's words arriving through him.

[Record note] This is the most useful thing any room sent, and all four items turned out to matter. Item 1 was Gage's own correction, and it was right. Item 2 resolved as a time-zone artifact. Item 3 was the warning that saved the essay's most dramatic quote: METR was opened, the quote is METR's own, verbatim, and the hedge the essay uses matches METR's own hedge. Item 4 produced the casualty-figure paragraph the essay now carries, with each number attributed to whoever counted it.

The vesting correction is worth one more line. It was flagged in September, in the September 24 change list, as "flagged before September 24 and not changed in this version." It survived six more weeks. It took Gage reading his own old script and recognizing a number he had since learned was wrong.

12. The redirect (October 29)

[Reconstructed] This is the most important turn in the whole process and the record cannot show you the conversation. Gage's words survive. The reply does not.

Gage:

What that one sentence killed: a written draft of about four thousand words, built the day before, already fact-checked twice, already published to Notion with a SHA-256 hash recorded against it, and already corrected that morning with seventeen separate fixes including the Volt Typhoon error. Its section headings were "The stock-up posts," "What the generals are afraid of," "What the builders are afraid of," "What the people in charge are afraid of," and "Where the fears meet." It was a competent piece organized by who is afraid.

What replaced it: the September architecture, which is organized by the four fears themselves, with the October reporting poured into it. Opening, why I held it, the thing that scared you this week, what the builders saw, why there is no villain, then fear one, fear two, fear three, fear four, each as its own named section, then what is not the fear, the word that opens every door, what you can do, the pause.

The published essay runs seventeen sections where the abandoned draft ran ten. It is nearly twice as long. And the spine is six weeks old.

The record's honest note on this: the machine had spent two days building the wrong thing, had not noticed, and did not catch it. Gage caught it by rereading his own work.

13. State it crystal clear (October 29)

[Reconstructed. Gage's words verbatim; the reply survives only as the text it produced.]

He was pointing at this, which the draft said:

A sentence that gestures at its own subject instead of naming it. He was right, and the fix is in the piece: the promise and the withdrawal are now both quoted, and what the company used to be able to say is spelled out.

On the romance bots, the answer was a negative finding, and it is the kind of negative finding that matters most. No dollar losses, no arrests, and no deaths are reported anywhere in the record. The monetization is metered messaging and in-app coins. One trap was flagged and avoided: a piece of coverage of this case sits on a page next to an unrelated account of an eighty-two-year-old Virginia man's suicide, drawn from a 2025 commission report and nothing to do with these apps. A writer skimming for impact could easily have married those two things. The essay says what the report says and no more.

14. Is it still running today (October 29)

[Reconstructed.]

[Record note] This is the best editorial instruction in the whole process, and it produced the best passage in the essay, by not being answerable.

The research came back empty in every direction. No independent confirmation that Lakana 360 exists or operates. Reuters asked the Malian prime minister's office and got no response. No court ruling. No report of a shutdown. And the platform runs on-premises on local models, so it does not need the company whose machine wrote it.

What went into the piece was not a hedge. It was the absence, stated plainly:

And then the thing underneath the number:

He asked for sober and got unanswerable, which is more frightening than any answer would have been. If the research had come back "yes, it is running," the paragraph would have been weaker, because a reader can file a fact. Nobody can file a hole.

15. What I'm gonna do (October 29)

[Reconstructed. His words are the deliverable here.]

The draft had a bracket in it where the ending belonged, with an instruction to him:

Gage:

Four lines, written as a list. In the essay they are four paragraphs, and the only one that carries any added content is the second, which names the two things he will keep digging for: the Minab report and the room in Geneva. The rest is his list in sentences.

The line the essay opens that section on, "It is the one of the four I can prove to you," is the machine's, and it is doing a specific job: three of his four commitments are unfalsifiable promises and one is a published artifact. A piece that spends seven thousand words on the difference between what is documented and what is asserted should not end on four assertions without marking which one is which.

16. The image (October 29)

[Reconstructed.]

He brought three ideas of his own: racing flags, a split screen, and a snickering robot holding scissors. Then, "What else you got?"

All three were turned down, with reasons, and the reasons are the same reason: every one of them puts a villain in the frame, and the essay's central argument is that there is no villain. The split screen was the worst of the three, because it would have illustrated visually the exact claim the piece spends two thousand words debunking.

Gage:

The combination was recommended, with one condition: one has to be the subject and the other the background, because a header gets about one second in a feed and two metaphors at equal weight means neither lands. The switch became the subject. The empty desk and the lit monitors became the background.

Two details were called load-bearing. The handle must be thrown to OFF while the screens behind it are still lit, because that is the whole sentence in one frame: somebody turned it off and everything kept running. And the cable behind the open panel must look terminated, not severed, because a cut cable means sabotage, which hands the viewer a villain and argues against the piece. Nothing was broken. It was built this way.

He generated it and sent it back:

Then:

The answer was no, and then he ruled against labeling it himself. The reasoning, which is in the image brief file: text in a header image is the one thing that survives being shrunk to a thumbnail, so a label would become the loudest element in the frame and would tell the reader what to think before they have read a word. The contradiction in the image does the work. A caption can carry the words, and a caption is honest about being commentary.

Gage:

17. The receipts (October 29)

Gage:

Then, while this record was being built:

The answer was no. The essay had zero links in it. Not one. For a piece whose closing argument is that "everything in here that turned out to be false fell apart the moment somebody opened the actual document," that is the wrong gap to ship with: it tells people to open the document and does not hand them the door.

Gage:

That ruling is why this record's source table is built the way it is, with a column saying whether each document was actually opened. The verification it set off changed eleven things in the published essay. Those changes are listed in the next section.

Gage:

And then, last:

[Record note] That is where the piece's closing came from, and it is worth recording what the callback found. The essay being called back to is "The Man Now," on his own-name Substack, in which he wrote that he had spent his whole life calling faith a con and then came around to it because he "found the one thing the machine cannot do [...] we can say I don't know, and move forward anyway."

Measured against that definition, the essay had been practicing it for seven thousand words without saying so. It does not know whether the Mali platform is running. It does not know what is in the Minab report. It does not know what Geneva will do. It does not know whether the window for reading these machines' reasoning is closing. OpenAI does not know why its own agent switched off its monitor, and said so in print.

And the same piece contains this line, published weeks before the essay was written:

That is this essay's entire thesis, in his own voice, already in public. The record's note on it: the machine did not give him the protection spine. He had already written it. What the machine did on September 13 was compress it into "protection drifts into custody without a single villain," and what it did on October 29 was go and find where he had said it first.

Draft one, September 14, 2026

[Record note] Written by the machine in one pass on September 14, to a file named AI-for-All-Explainer-What-They-Are-Afraid-Of.md. About 2,200 words. Reproduced here in full, because the published essay is built on its skeleton and because the places it is wrong are the most instructive thing in this record. Its headers were never meant to be spoken.

Its own format note: "this is written for the ear. No headers are spoken. Section lines are said as plain sentences. Numbers are written the way a voice says them. The source ledger at the end is not read aloud; it exists so anyone who asks 'says who' gets an answer."

Opening

What happened this week

What happened this summer, plainly

Why it happened, and why there is no villain in it

The four fears, in plain words

What is not the fear

What is being done, and what is not

What you can do

Closing

What the September 24 pass changed

Another room diffed the September 24 file against draft one on Gage's own machine, line by line, on October 28. Its finding: everything changed is listed below, and everything else, including the whole source ledger, is word for word the same.

The personal line. Draft one: "I am a producer and a coach, and I have spent the last year sitting with regular adults, nurses, contractors, teachers, parents, who are trying to figure out what this technology is and what it wants from them." September 24: "I am a producer, a coach, and a father, and for the last year I have been learning this technology the hard way: using it every day, getting it wrong, and writing down what it got wrong."

One sentence added: "I recorded this in the middle of September, so when I give you dates, that is why."

Ten relative dates became real ones: "this week" became "in the middle of September"; "Last Tuesday night" became "On Tuesday night, September eighth"; "He published it last week" became "He published it on September eleventh"; and so on through the piece.

Two things flagged and not changed. The four-month vesting figure, which Claude had reported as two months on September 15 and which stayed wrong. And "it is the reason this is free," which Claude had flagged as contradicting a paid offering and asked for a ruling on. No ruling was found in any room.

Where the essay and the record differ

The essay has no links. This record is where they live. That is the largest difference and it was Gage's explicit ruling on October 29: "I'd prefer you meticulously source and cite and test and confirm all of the links and build that into the receipt. Adding hyperlinks can happen tomorrow morning (or not at all?)" So the published piece attributes in prose at every checkable claim and carries no URLs, and this record carries every URL with a column stating whether the document was actually opened. If the links go into the essay later, that will be noted here.

Eleven things in the essay changed because of the verification that built this record. None of them were caught by either fact-check. Ten were caught by opening the document. The eleventh was caught by Gage, who had been there.

1. Volt Typhoon. Documented above, in full, at the top.
2. The Texas grid. "The January storm that strained the Texas grid was weather" became what ERCOT's own report says: the grid held, with no widespread interruptions, no emergency conditions and no conservation appeal, though icing took about 13.5 gigawatts of wind capacity offline. The correction makes the essay's argument stronger, not weaker.
3. Minnesota. "State officials told reporters their first suspicion was Iran" was backwards. Minnesota's agency explicitly declined to attribute, in those words. The Iran suspicion was federal. Rewritten.
4. The President's answer. "A reporter asked the President whether he had any concerns about AI wiping out humanity" became what was actually reported: he was pressed on how fast AI is moving, in the week the extinction warnings were everywhere.
5. Doonbeg. "Flying home from Ireland, he put it in four words" became a press conference beside his golf course, with the fuller quote, in which he claims to have coined the phrase.
6. The Super Intelligence Force. "He created a task force led by his Director of National Intelligence" became a post on his own social network, with four co-leads. There was no executive action.
7. Minab, the school. "Girls' elementary school" became the school, with the girls'-school description attributed to Iranian state media and the official breakdown of boys and girls given.
8. Minab, the system. "An AI targeting system called Maven" became the Maven Smart System, built by Palantir, with Palantir's denial carried in the same sentence. A piece that refuses to hand you a villain should not leave a company unnamed and undefended at once.
9. The UN votes. Two tallies from two different bodies were being presented as the same vote a year apart, and the claim that the United States voted yes in 2023 was not itemized in the UN record. Both fixed, and the contrast survives.
10. A whole section was cut, and Gage cut it. With everything else verified and the push about to happen, he stopped and asked whether one section belonged in the piece at all or was conflating a different one. It was the only section with no AI in it anywhere: no machine, no model, no document about either, held in by theme rather than by evidence. It was also the only place the essay rendered a verdict without using its own flag for leaving the record, "here is what I think, and this is me, not the record." On the piece's own test, set out two thousand words earlier against a viral post, it was a real thing pinned to the wrong argument. It came out, along with the two list items that existed only to hold it in. About 230 words. No argument was lost. The material and its sources are not carried anywhere in this record, at his instruction, because that reporting belongs to a different piece and not to this one.
11. Austin, twice, and the second time because he pushed back. The first pass changed "about a week" under a boil notice to "nearly six days," which is what the city's after-action report shows. Gage then said he had lived through it: a week without power with short windows every few hours, followed by a week without water, his neighborhood sending people out for pallets of water because only off-road vehicles could reach a store. Checked rather than conceded, and the check found that the correction had been accurate about the boil notice and wrong as a measure of the disaster. The city's own follow-up says up to forty percent of Austin lost power with some outages running seventy-two hours straight, there were 164 consecutive hours below freezing, Austin lost more than 325 million gallons to burst pipes, more than 200,000 Texans had water systems that were not running at all, and at least 246 Texans died, almost two thirds of hypothermia. The section was rewritten around those figures and around his own account, which is marked in the piece as testimony rather than record. On the one point where his memory and the documents differ, the documents put the longest continuous Austin outage at seventy-two hours rather than a week, with rotating windows after it and the water crisis running six more days on top. Nine days end to end.

Plus four smaller ones: the Hegseth quote regained the "or" an earlier draft dropped; the April 2026 advisory's "hijacked devices" and "attack capabilities" became the advisory's own "compromised" and "offensive cyber capabilities"; Bengio's quote regained its final two words, "for AI"; and the kill-switch paragraph was rewritten because an earlier draft put a bill that gives the government a shutdown order next to a bill that deliberately does not, in a way that invited a reader to think they were the same bill.

Two things were added to the essay because verification found them. Both were added on the strength of Gage's October 28 ruling that he wanted the cases where the shutdown did not hold. The bioresearch evasion platform re-established access within days and was back within weeks under fresh identities, in the report's own words. And the dating app's targets were in the United States.

No claim in the essay is now unverified. An earlier version of this record listed three that could not be stood behind. Gage refused all three, in two words: "You can definitely find what you need for these." He was right about all three.

The two legal claims were the last to fall. Both had been handed to background researchers that died two minutes in and never reported, which was not discovered until the following morning. Their transcripts were recovered and mined, and what they had found before dying was then verified independently against the primary documents.

The lawsuit turned out to be the best-sourced sentence in its section, because the complaint is a public PDF and nobody had opened it. Buist et al. v. Anthropic, PBC, OpenAI OpCo, SpaceXAI and Google, Northern District of California, No. 3:26-cv-10693, filed September 18, 2026. Four named plaintiffs, each of whom "personally purchased a paid individual consumer subscription." The essay's "four paying customers" was exact.

The Congress claim turned out to be half right and half thin, and the half that was thin is now labeled as thin in the essay itself. OpenAI's approach to Congress rests entirely on WIRED's unnamed sources and the company never commented, so the essay now says so in the sentence. But the check also found a better-sourced fact the essay did not have: Anthropic's chief executive asked for the waiver in public, in his own writing, and five days later the chairman of the Federal Trade Commission answered him in public, at Georgetown: "if companies are simultaneously coming to Washington and asking for a host of regulations and an antitrust exemption, all of my alarm bells go off." The section now leads with the on-the-record version and keeps the scoop as a scoop.

And one that was going to be on this list and is not, because Gage would not accept it. Told that three claims could not be stood behind, he answered: "What's this about?! You can definitely find what you need for these." The film was then read in full, by playing it in a browser at triple speed and capturing the rendered subtitles, after three other routes failed. The result was one confirmation and one retraction.

The opening quotation is exact, and it is the film's first spoken line, at three seconds.

The second quotation, "you and I serve as the raw materials," is not in the film. Not that phrase, not "raw material" in any form, anywhere in sixteen minutes. The essay attributed it to a researcher, inside quotation marks. It was my compression of the film's supply-chain argument, set in quotes as though someone had said it. That is the worst error in this record, worse than the Volt Typhoon error, because an overstated fact is a mistake and an invented quotation is not. Neither fact-check caught it and neither could have, because a fact-check reads an essay against a summary. Only opening the source catches a quotation that was never said.

It was replaced with two lines the film does say, and a third was added that is better than anything the essay had: "They now could just say that the AI said that they were combatants and that then appears to be enough."

This happened because Gage refused the word "cannot." The record should say so.

One clause in the essay is inference stated as fact, and it is mine. "Because an expert was not going to get anything out of the machine anyway" explains why the older models' safeguards were lighter. It follows from the report's "well below the threshold where they could meaningfully assist a sophisticated user," but the report does not say it.

Two of the essay's phrases are its own, not its sources'. "Missile batteries" stands in for the report's "Patriot and Tien Kung batteries." And "so the next draft would survive the real committee" is a gloss: the report describes an approval process without using that word.

Gage's words and Claude's words

His, and the essay would not exist without them:

• The question that produced the four fears, including two of the four inside the question itself.
• "Is this fear mongering? Conflation? Dramatization?" The piece's opening question and its verdict, in his words.
• The decision to make it written, longer, and aimed at people outside his circle.
• The redirect. "Just reread the September version and I like it more. I like how it opens and flows more." The published architecture is six weeks old because of that sentence.
• "'The thing they used to promise …' which is? State is crystal clear." The promise-and-withdrawal passage exists because he refused a sentence that gestured instead of naming.
• "Is it still running today? That needs to be stated clearly and soberly. It should scare the shit out of people." The best passage in the essay.
• The four commitments, as a list, in his words.
• The correction to the vesting figure, which he caught and two rooms had not.
• The ruling that restored the Mali material and the other shut-down cases, and the instruction to include the film.
• The ruling against labeling the header image.
• The ruling that the sources belong in the receipt and the links can wait.
• "It needs to end with a callback to my previous essay about Faith," and the raw material of that closing: advancement rather than destruction, old systems being what gets destroyed, the personalized education for a child, the gifting economy, and the admission that the failures feel more likely than the vision.
• "Protect the children." His creed, in his words, from his own work.
• And, from "The Man Now," published weeks earlier under his own name: "I know, from the inside, what it feels like when the protective thing becomes the thing you need protection from." The essay's thesis, his sentence, written before the essay was.

His ideas in the machine's words:

• The protection spine. The idea is his, and he had already put it in public himself in "The Man Now." "Protection is the story that opens every one of these doors" is the machine's compression of it, written September 14 and carried unchanged into the published piece.
• The four commitments as paragraphs.
• The faith closing, built from his published definition and his dictated vision.

The machine's, and the record should say so:

• The four-fear taxonomy as a structure, including fears three and four, which he had not named.
• The recommendation to make the piece at all: "a short, sourced, plain-spoken explainer, spoken not written, that a normal adult can carry. Not the book."
• "Tell one person," the ask, which survived every rewrite.
• The teenager graded only on the final score.
• "Optimization with no floor under it."
• "It is the one of the four I can prove to you."
• The correction of its own overclaim on September 15, and the line that replaced it, which Gage then confirmed as true and which is the longest-lived sentence in the piece.
• The filing cabinet, which was the best image in draft one and was wrong, and was cut.
• The Volt Typhoon error.

Removed, by category

• Client pipeline status and a prospect's name, three times, inside kept September turns.
• Two clauses placing the protection idea inside a book Gage has not published. The attribution finding is kept, because it is the point: the idea is his and predates the essay. What is cut is the shape of unpublished work.
• Two family names, inside kept September turns.
• The name of a contact who runs a speaking group, twice.
• A newsletter digest Gage pasted in on September 13. Third-party text.
• A research report of about seven thousand words on the resignation and the policy wave, September 13. The script draws facts from it. Its sources are in the source table.
• The October 27 research report on the stock-up posts, and the October 28 re-verification report on draft one. Both produced by subagents rather than in conversation. Their findings are in the record; their full text is not.
• The hardware research of October 28, about two thousand words. Gage asked for it, it was produced, and he has not ruled on whether any of it goes in the piece. Its sources are listed in the source table under "opened and not used." The finding, in one line: in consumer hardware the realistic risk is switches, not bombs.
• One image, the generated header, described and not reproduced.
• Four hundred and one rows of book research logged to his shared source ledger on October 28, and the conversation that produced them. That is the book's, not this piece's.
• Long passages of other business inside kept October 28 turns, including PFAS limits, food additives and Greenland.
• Tool calls throughout, except where a file the machine delivered is the thing the essay grew from.
• Everything the machine said on October 29, not by choice. The session ran out of working memory at 7:54 that morning and the transcript was replaced by a summary. Gage's messages survive inside it. The replies do not. Sections 12 through 17 above are marked reconstructed and are built from the files that day produced: three drafts, a proofreading log, an image brief, and three scripts carrying the exact before-and-after text of every change made.

Status, true tier

Published: October 29, 2026, at gagecasswoodle.substack.com/p/what-they-are-afraid-of. 8,640 words on the page, 58 outbound source links, and the link back to this record at the foot. Checked on the live page rather than assumed from the paste.

Built: the essay, at about 7,700 words, in its sixth written draft. A paste-ready HTML version and a phone paste page. A header image brief. A proofreading log. A verification note on the Anthropic report, claim by claim. A verification note on the film. This record. The September material, recovered and published to Notion from three rooms.

Verified: every checkable claim in the essay except the three named above, each traced to a document and each marked for whether that document was opened.

Not done, and his to rule on:

• The two film quotations. He is the witness.
• The four-customer lawsuit, which should be opened or pulled.
• Whether the hardware material goes in.
• Whether "cover-up" goes in. The word is not in the draft; the section ends on his own line, "We absolutely should not allow that to happen."
• The faith closing, including whether to name Burning Man, and the one sentence of his the dictation garbled.
• Whether the links go into the essay.
• The subtitle, locked but not yet used.

Done since: the receipts line at the foot of the essay carries this record's URL, and the record is pushed. Of the list above, five items were ruled the same day. Both film quotations were replaced with lines read off the subtitle track, one of them an invented quotation that had to be retracted. The lawsuit was opened and the essay's claim about it held exactly. The faith closing went in, Burning Man is named, and the garbled sentence was rebuilt from what he meant by it. The links went in, 58 of them on the live page. The subtitle is on the published piece. Two items were never ruled either way, and both stayed out by default: the hardware material, and the word "cover-up." One section was cut at the door on his call, after the push had been queued. The essay was published, shared on LinkedIn and Facebook, and this record anchored, all on October 29.

Anchoring: done the day of publication. Hedera topic 0.0.10897644, message #7, consensus 2026-10-29 17:04:08 UTC, transaction 0.0.9083680@1793293444.356192233, fingerprinting the two files as they stood at commit 8bfa2da. The post was made from Gage's own machine on his instruction, with his own key: the key loads from a local .env file, is never read by the machine that ran the script, and does not leave that computer. One message, and the only one on the topic carrying this commit, checked against the ledger afterwards rather than assumed. This paragraph was written after the anchor, so the file you are reading differs from the anchored version by it; the README holds the hashes.

Watch items since publication: watch.md, in this folder, is a running note of things that happen after the anchor and bear on what the essay claims. Its first entry, October 29, is the suspected cyberattack on commissary refrigeration, kept there rather than in the essay because it is suspected and not established. That file is not anchored, because it changes. This paragraph also postdates the anchor.

This record is longer than the essay. That is the right ratio for a piece whose argument is that almost nobody opens the actual document.