We Are Arguing About the Speed of AI. We Have No Plan for the Crash.

Published audit

The essay is the judgment. This page is the inspection surface.

The five insertions

There is no v1.1. The base is Muse Wonder. The freeze is the base plus exactly these insertions. Shown in context with the inserted words marked, each with its source credit, then in words.

The insertions, in words

The base is Muse Wonder. The freeze is the base plus exactly these insertions.

The merge, exactly: the frozen essay equals the Muse Wonder draft (step 4 of its output, with its "~1,050 words" word-count line dropped) plus exactly these five insertions, confirmed by diff. No other change.

1. CISA expanded at first mention. Base: "Agencies like CISA exist for a reason." Frozen: "Agencies like the Cybersecurity and Infrastructure Security Agency (CISA), the federal government's lead civilian cyber-defense agency, exist for a reason." (Requested by the author; written by Grok Bot.)
2. Added at the end of the "That platform is now digital…" paragraph: "We got a preview in July 2024, when one faulty software update, with no hacker and no AI involved, crashed an estimated 8.5 million Windows computers and grounded flights, delayed surgeries and knocked out some 911 systems." (Anchor chosen independently by Opus 5.5, Kimi K3.1 Pro and GPT-6 Astra; wording closely tracks Opus 5.5's draft; the 8.5 million figure appears in Astra's; written by Grok Bot.)
3. Added in the "That is why both dominant strategies fail…" paragraph, before "That is not a strategy. That is a hope.": "Defenders must be right every time; an attacker only has to be right once." (A light edit of a line in Gemini 4 Argon's draft.)
4. Added in the "To be sure" paragraph, after "…they are an assumption.": "A patient intruder can also corrupt the backups months before pulling the trigger." (Idea from Gemini 4 Argon, new wording.)
5. Added in the crash-plan list, after the offline-backups sentence: "A pre-agreed baseline date from which the economy can restart if the live records are lost." (Idea from Gemini 4 Argon, new wording.)

Human/AI essay: counts, floor, and findings

Verification depth. [F] means I fetched and read the page or document in this pass. [S] means I read a search-result excerpt of the source and did not open the full page. Nothing below is cited from memory alone.

Executive counts (human/AI essay)

Floor status: the freeze clears the floor. Every claim that states a fact holds in direction. The July 2024 outage sentence is correct on every point I could check. Where it is soft, the softness is in universals ("fail together," "past blackouts were local," "no credible answer") and in the flat title claim "no plan." The finding that matters most is H12/H20: the title and the sentence "Right now, we have no credible answer" sit against a record the essay half-acknowledges ("America is not starting from zero"). The essay hedges it in the "To be sure" paragraph, so I rate it OVERSTATED, not wrong.

Recommendation: leave the freeze as written and log these on the audit page, as with the machine essay. If the author wants a tightening, the smallest one is in the title/H12 claim (for example "no tested plan"). That would be a v1.1 and the author's call.

H. Human/AI essay: fact-check log

H1. The July 2024 outage sentence — SOLID

• Essay (insertion 2): "We got a preview in July 2024, when one faulty software update, with no hacker and no AI involved, crashed an estimated 8.5 million Windows computers and grounded flights, delayed surgeries and knocked out some 911 systems."
• Sources and checks:
  • Date and cause. CrowdStrike's technical post [S]: a sensor configuration update was released 19 July 2024 at 04:09 UTC (the evening of 18 July in Pacific time) and "triggered a logic error resulting in a system crash." CrowdStrike's root cause analysis for Channel File 291 [F, full text]: the new template type defined 21 input fields, the sensor code supplied 20, and the mismatch caused an out-of-bounds read and crash. "One faulty software update" is accurate.
  • No hacker. The same post says the issue "is not the result of or related to a cyberattack." The root cause analysis says the bug was reviewed by a third party and is not exploitable by a threat actor. NPR and CNN also report CrowdStrike saying it was not an attack [S].
  • No AI involved. The CrowdStrike analysis says the product has on-sensor AI and machine-learning models, but that the faulty "Rapid Response Content" is "behavioral heuristics, separate and distinct from CrowdStrike's on-sensor AI," written by detection engineers as regular-expression template instances [F]. So the claim holds for the fault. A pedantic reader could note that the vendor's product does use AI, which is a precision point, not an error.
  • 8.5 million. Microsoft's own blog [S]: "We currently estimate that CrowdStrike's update affected 8.5 million Windows devices, or less than one percent of all Windows machines." Reuters, AP and TechCrunch repeat it. The essay says "estimated" and "Windows computers," both fair. It does not say the figure is Microsoft's, and it omits "less than one percent." Mac and Linux machines were not affected (TechCrunch).
  • Grounded flights. The New York Times live coverage [S] says at least five US airlines (Allegiant, American, Delta, Spirit, United) had grounded all flights for a time, per the FAA. CNN [S] reported more than 3,000 US flights cancelled as of Friday evening. Reuters, AP and SEC filings put Delta's recovery at about five days and roughly 7,000 cancellations [S].
  • Delayed surgeries. AP [S]: Harris Health System (Houston) cancelled and rescheduled elective procedures. NBC [S]: Baptist Health in Kentucky cancelled some non-emergency procedures, including a scheduled open-heart surgery. USA Today and CNN also report cancelled medical procedures [S]. "Delayed surgeries" is accurate.
  • 911 systems. USA Today [S]: in Alaska, 911 and non-emergency calls went unanswered at multiple dispatch centers for about seven hours. Coverage also reported disruptions in Virginia, Iowa, New Hampshire, Oregon and Arizona [S]. "Some 911 systems" is the right size.
• Origin: the figure and the wording come from the comparison drafts, not from Muse. See section J.
• Severity: PASS. Optional nit: the sentence does not name Microsoft as the source of the 8.5 million estimate.

H2. "We got a preview" and "Power, water, communications, finance and transportation do not fail independently anymore. They fail together" — OVERSTATED

• Essay: the Muse-base sentence "do not fail independently anymore. They fail together, because they depend on the same internet, the same cloud providers, the same software supply chains," followed by the insertion that calls the July 2024 outage "a preview."
• Checks: the directional claim (shared dependencies create cross-sector failures) is well supported. CrowdStrike hit airlines, hospitals, 911 centers, banks and broadcasters through one vendor's software. The April 2025 Iberian blackout (H19) took out power, rail, communications and hospital records together. But "do not fail independently anymore" is a flat universal. Most outages are still sector-specific, and the essay's own example did not take down power or water, did not affect Mac or Linux systems, and affected under 1% of Windows machines. Most carriers recovered within days. A bounded, accidental, recoverable outage is a generous "preview" of a deliberate multi-sector cascade (Stage C Kimi 5 raises this).
• Severity: OVERSTATED (universal and "preview") — non-fail
• Origin: the "fail together" sentence is Muse's; the "preview" sentence is insertion 2.

H3. CISA: "the federal government's lead civilian cyber-defense agency" — SOLID

• Essay (insertion 1): "Agencies like the Cybersecurity and Infrastructure Security Agency (CISA), the federal government's lead civilian cyber-defense agency, exist for a reason."
• Sources and checks:
  • CISA's own site [S]: "the nation's cyber defense agency" and "National Coordinator for Critical Infrastructure Security and Resilience." A DHS report to Congress (Oct 2024) [S] calls CISA "the operational lead for federal cybersecurity" and "the nation's cyber defense agency." NSM-22 (April 2024) [S] codified the national-coordinator role.
  • "Lead civilian" is the essay's paraphrase, not CISA's own phrase. It is fair, because the military cyber role sits with NSA and Cyber Command, and CISA's own materials use "lead." One qualification: for the response to a specific incident, the FBI leads threat response while CISA leads asset response (Presidential Policy Directive 41 and the National Cyber Incident Response Plan [S]), so "lead" is accurate for defense and coordination, not for every task.
  • The abbreviation is now expanded at first mention, as requested.
• Context a reader may bring (not an error): reporting in January–February 2026 [S] says CISA fell from about 3,400 staff in January 2025 to under 2,400, and that it has had no Senate-confirmed director since the Plankey nomination stalled. The essay's "exist for a reason" is accurate and, if anything, understates the thinning of the agency. That context would support the essay's argument, so it is not a reason to change the sentence.
• Severity: PASS

H4. "a person who is homeless, addicted or mentally ill in Los Angeles today can still survive — not comfortably, not justly, but survive" — SOLID (hedged framing; reader flashpoint)

• Sources: LA County Department of Public Health, press release of 10 March 2026 [F]: 2,208 deaths among people experiencing homelessness in 2024 (300 fewer than in 2023), a mortality rate 4.2 times the county rate, about six deaths a day, and overdose as the leading cause (40% of deaths). LAHSA's 2026 count [S]: an estimated 73,040 people homeless in LA County, 49,002 of them unsheltered.
• Check: the sentence is hedged ("not comfortably, not justly"), comparative ("in a way that would have been unimaginable for most of human history") and a statement about the typical case. On the numbers, most unhoused people survive a given year (about 3% of the population died in 2024 by the county's rate). But 2,208 deaths and a death rate four times the county's is a lot for "survive." The essay's comparison with prehistory is not something I tested; I did not find a clean like-for-like mortality comparison, so I rate the comparison as framing, not as a fact claim. The rating is SOLID because the factual kernel (people survive within modern systems) holds and the sentence is carefully hedged.
• Note for the page: this is the line an adversarial reader will quote. A comparison draft (Opus 5.5) advised cutting it as "callous." The author kept it. That is an editorial decision, not a factual defect.
• Origin: the dump ("homeless, destitute, addicted, and mentally ill people can survive today in and around major cities like LA") is the source. Muse compressed it.
• Severity: PASS (with the caution above)

H5. "An Elon Musk born 20,000 years ago builds nothing recognizable… dies having sharpened a better stone" — SOLID (rhetorical counterfactual, one nuance)

• Checks: 20,000 years ago is the Last Glacial Maximum, before agriculture (about 11,000 years ago), before the earliest copper smelting (about 5000 BCE, at Belovode, Serbia) and before writing (Sumer, about 3400 BCE) [S]. The essay's list ("writing, metallurgy, electricity, capital markets and supply chains") is therefore correct as a list of things that did not yet exist.
• Nuance: "sharpened a better stone" understates what existed. Pottery at Xianrendong Cave in China is dated to about 20,000 years ago (Science, 2012) [S], and Upper Paleolithic people made bone, antler and shell tools and art. People then were not technology-free. This is rhetoric, not a claim about the archaeology, and it makes the point (infrastructure compounds talent) no weaker. "20,000" is Muse's specific figure; the dump says "prehistoric times."
• Note: the example names a living public figure as a thought experiment. It states nothing about him.
• Severity: PASS

H6. "Open-weight models ensure those capabilities will not stay locked inside a few responsible labs" and "Foreign adversaries will not pause because we do" — SOLID (direction), with a caveat on "ensure"

• Sources: the UK AI Security Institute, "How Far Behind the Frontier are Leading Open Weight Models on Cyber?" (17 July 2026) [F]: GLM-5.2 and DeepSeek V4-Pro perform like frontier closed models released 4 to 7 months earlier, a narrower gap than the 6 to 10 months seen in 2025; safeguards on open-weight models "can be removed." Epoch AI, "Open-weight AI models are 6-7 months behind on cyber" (data as of 8 October 2026) [F]: the best open-weight model (GLM-5.3) sits where the closed frontier stood 6 to 7 months earlier. The leading open-weight models are from Chinese developers (Z.ai, DeepSeek, Moonshot).
• Check: the term "open-weight" is correct for these models (the weights are downloadable; most are not open-source in the full sense). The claim that capabilities will not stay locked in a few labs is supported, with a lag of about half a year. "Ensure" is a strong verb for a lag, and "responsible labs" is a value judgment. Stage C Kimi 11 notes the essay does not mention that closed models are also misused; the record agrees (Anthropic's report on a state-sponsored campaign that used its own model; Google's threat-intelligence reporting on criminal and state actors using Gemini [S][F]). "Foreign adversaries will not pause because we do" is an opinion, but the record on state actors using AI for cyber operations backs its premise (H7).
• Severity: PASS
• Note: the essay uses "open-source imitators" in the third paragraph and "open-weight" later (I4).

H7. "Systems that can write code, find vulnerabilities, craft flawless phishing campaigns in any language, and operate persistently without sleep" — OVERSTATED (one word)

• Sources: Anthropic, "Disrupting the first reported AI-orchestrated cyber espionage campaign" [S]: a Chinese state-sponsored group used Claude Code to automate roughly 80–90% of a September 2025 campaign against about 30 targets, with reconnaissance, vulnerability discovery, exploitation and exfiltration done by the model. Anthropic also said only a small number of intrusions succeeded and that the model sometimes hallucinated credentials or findings. Google Threat Intelligence Group, 11 May 2026 [F]: for the first time, a threat actor using a zero-day exploit "we believe was developed with AI"; state actors sending thousands of repetitive prompts to analyze CVEs; agentic frameworks used for reconnaissance and vulnerability validation; LLMs used to build "higher-fidelity phishing lures."
• Check: write code, find vulnerabilities and operate persistently are all supported. "In any language" is plausible and not tested. "Flawless" is hyperbole. Google's wording is "higher-fidelity," and Anthropic notes failures and hallucinations. A reader who knows the reports would catch it.
• Severity: OVERSTATED ("flawless") — non-fail
• Origin: Muse.

H8. "lowers the skill floor, raises the speed ceiling, and multiplies the number of plausible attackers" — SOLID

• Sources: Google Threat Intelligence Group (May 2026) [F]: AI tools "lower the barrier for adversaries" and give "expert-level force multipliers"; adversaries use agentic workflows "at machine speed." AISI [F] grades models on tasks from "technical non-expert" up to "expert," and open models are cheap to run (an estimated $0.28 per task for DeepSeek V4-Pro on tasks that Opus 4.5 solved at $12.50).
• Check: all three clauses are supported in direction. The speed claim rests on vendor and government reporting, not on a controlled comparison.
• Severity: PASS

H9. "Money is not mostly cash in vaults; it is entries in databases" — SOLID

• Source: Federal Reserve H.6 release, September 2026 [S]: currency in circulation was about $2.48 trillion in August 2026, which is 12.4% of M1 and 10.6% of M2.
• Check: most money, on the Fed's measures, is bank deposits and other non-cash balances. Physical currency is about a tenth of M2. ("In vaults" is a figure of speech; most currency is outside banks.)
• Severity: PASS

H10. "Hospitals do not run on paper charts; they run on networked records, scheduling and billing systems" — SOLID

• Source: ONC Data Brief 35 [S]: 96% of US non-federal acute-care hospitals had a certified electronic health record system in 2015. The figure is old and the brief says adoption had plateaued. The essay states no number. Ascension's 2024 experience (H18) shows the dependency.
• Severity: PASS

H11. "automated logistics move food across continents on just-in-time schedules measured in hours" — UNSUPPORTED (as worded)

• Check: I found no source for "measured in hours." The sources I could reach disagree on how thin the buffers are. A Monash University piece [S] describes just-in-time systems as having minimal buffers. A BBC piece from 2020 [S] argues the chain holds much more inventory than shoppers assume, much of it in warehouses and back rooms. The directional point (modern food distribution is tightly scheduled and computer-coordinated) holds. "Measured in hours" is rhetoric, and it is the least supported number-like claim in the essay.
• Severity: UNSUPPORTED (as worded) — non-fail
• Origin: Muse.

H12. Title and "Right now, we have no credible answer" — OVERSTATED

• Essay: headline "We Have No Plan for the Crash"; "It's what happens the day prevention fails. Right now, we have no credible answer." Against "To be sure, America is not starting from zero. Banks keep backups. Utilities have cybersecurity teams."
• Sources: Sheltered Harbor, per the ICBA fact sheet and ABA coverage [S]: a financial-sector standard for nightly, encrypted, immutable, isolated backups ("data vaulting") of critical account data, with recovery plans and certification; participants held about 71–72% of US deposit accounts. FS-ISAC's page [F] now says the organization has been wound down and its specification transferred to FS-ISAC for its members. (I did not capture the wind-down date.) The essay's first proposal ("Verified, offline, immutable backups of core financial records, tested… in live restoration exercises") is close to what that standard describes. See also H20 for the national exercises.
• Check: the claim that there is no plan is stronger than the record. The defensible version is that existing measures are fragmented, voluntary or untested across simultaneous cross-sector loss, which is what the "To be sure" paragraph says. Stage C Kimi 2 raises the internal tension. The essay's thesis does not need "no plan," but it needs "no tested plan."
• Severity: OVERSTATED — non-fail. The headline claim carries the piece, so this is the finding most worth the author's attention.
• Origin: the title is Muse's; the dump asks "Is anyone in a position to take action/develop backup plans even asking these questions?", a question the essay turned into an assertion.

H13. "Defenders must be right every time; an attacker only has to be right once" — SOLID (idiom); no source needed, attribution note

• Check: this is a long-standing security idiom. Its widely cited origin is the IRA statement after the 12 October 1984 Brighton hotel bombing: "Today we were unlucky, but remember we have only to be lucky once, you will have to be lucky always" (Quote Investigator; Wikipedia, "Brighton hotel bombing") [S]. It is not a statistic and cannot be checked. As logic it is looser than it sounds: a layered defense makes an attacker win several times (Stage C Kimi 4, see the bad-premise log).
• Severity: PASS. The credit to Gemini 4 Argon records where the sentence entered this essay's editing chain, not who invented the idiom.

H14. The two-camp portrait and "A pause, even if America could enforce one — and it cannot globally" — UNSUPPORTED (opinion and characterization)

• Essay: "Washington, Silicon Valley and the commentariat are locked in a theological war between two camps: pause or accelerate… Both sides talk as if the only decision that matters is how fast we move." Also "Acceleration… rests on a heroic assumption," and the judgment that a pause cannot be enforced globally.
• Check: these are the author's thesis frame and judgments, not checkable facts. The two-camp picture is a simplification: government and industry work on defender preparation, as the AISI "window for defenders" argument and the exercises in H20 show. "Both sides talk as if…" is a generalization that Stage C Kimi 3 finds too neat. The claim that a pause is not globally enforceable is a reasonable judgment; the essay does not argue it.
• Severity: UNSUPPORTED (opinion; generalization) — non-fail
• Origin: the dump's own frame ("Perseverating on pausing vs accelerating AI development misses the more important question").

H15. "Aviation did not become safe… redundant systems, black boxes, emergency procedures, investigations that fed back into design. Nuclear power, finance after 2008, even seatbelts" — SOLID (aviation); the list is unsourced

• Source: the NTSB [S]: it investigates accidents, determines probable cause and has issued more than 15,800 safety recommendations since 1967; flight recorders have been central to its investigations.
• Check: the aviation account is a standard and accurate summary. "Nuclear power, finance after 2008, even seatbelts" is a loose list of cases that I did not individually source. They are widely documented (post-accident safety regimes, bank stress testing, seat-belt mandates), but I did not verify any of them in this pass. As analogies, they show a method (assume failure and design for survivability), not a threat model for an adaptive adversary (Stage C Kimi 12).
• Severity: PASS (aviation); the rest not checked

H16. "A patient intruder can also corrupt the backups months before pulling the trigger" (insertion 4) — SOLID (as a "can" claim)

• Sources: Veeam's 2024 Ransomware Trends Report [S]: attackers targeted backup repositories in 96% of ransomware attacks and affected them in 76%. Sophos's 2024 report on compromised backups [S]: attackers attempted to compromise backups in 94% of attacks, with 57% success. CISA advisory AA24-038A (7 February 2024) [S]: the PRC-sponsored group Volt Typhoon maintained access in some victim IT environments for at least five years, to pre-position for disruption of critical infrastructure.
• Check: the sentence says "can," and the record supports that. The statistics show attackers going after backups at the time of the attack, not corrupting them silently for months; I found no documented incident of long-dwell backup corruption, though Volt Typhoon shows multi-year dwell is real. "Months before" is plausible and unsourced as to backups. The Maersk NotPetya recovery (a single surviving domain controller that had been offline during the attack) [S] shows how close a victim can come to having no clean copy, but it was a wiper attack, not months-long poisoning.
• Severity: PASS
• Origin: Gemini 4 Argon's draft ("subtly poisoning the data inside 'clean' backups" and "poisoning backup data months in advance"). The essay's sentence is new wording of that idea.

H17. "A pre-agreed baseline date from which the economy can restart if the live records are lost" (insertion 5) — PROPOSAL, not a factual claim

• Check: this is a policy proposal. I found no existing US framework that uses a pre-agreed economy-wide restart date. The nearest real mechanism is data vaulting that preserves end-of-day account balances (H12), which would give a recoverable snapshot date but not an agreed rollback rule. Stage C Kimi 8 asks who picks the date and what happens to transactions after it. The sentence does not claim the mechanism exists.
• Severity: n/a (proposal). It is the least supported of the five insertions and the one Stage C presses hardest.
• Origin: Gemini 4 Argon's draft ("the government must be able to declare a baseline date—say, the first of the previous month—and reboot economic life from an incorruptible physical record"). The essay drops the example date and the optical-media detail.

H18. "Paper and local-cache protocols in hospitals that allow care to continue for weeks, not hours, without central systems" — SOLID (feasible, degraded)

• Sources: NPR and Becker's [S]: Ascension's hospitals lost their EHR, patient portal, phones and ordering systems to ransomware on 8 May 2024, used paper charts, faxes and manual medication processes, and had the EHR restored systemwide on 14 June 2024. NPR's headline reports staffers describing "harrowing lapses."
• Check: weeks of care on paper happened, which answers Stage C Opus 4 (that it is infeasible; see the bad-premise log). It also shows the cost: delayed results and medication-safety risk. The essay asks for protocols that allow care to continue, which is a fair ask, and says nothing about how well. A reader could ask for the word "degraded."
• Severity: PASS

H19. "past blackouts were local, temporary and analog-supported in ways today's systems no longer are" — OVERSTATED

• Source: the ENTSO-E final report on the 28 April 2025 Iberian blackout [S]: it began at 12:33 CEST, Portugal's transmission system was restored in about 12 hours and Spain's in about 16, and about 60 million people were affected. A peer-reviewed analysis of the Portuguese health service [S] and other reporting describe hospitals on limited generators, lost access to digital records, failed emergency communications, stopped rail and failed traffic lights.
• Check: "temporary" is right. "Local" is not right for a blackout that covered two countries and about 60 million people. "Analog-supported" is arguable, since the 2025 event already hit digital dependencies. The event supports the essay's thesis better than the sentence does. The essay's following sentence ("We survived because the surrounding system survived") holds.
• Severity: OVERSTATED (the word "local") — non-fail
• Origin: Muse.

H20. "Regular, national-scale stress tests — not tabletop exercises for press releases — that simulate the simultaneous loss of finance, communications and logistics" — OVERSTATED (implied absence)

• Sources: CISA, Cyber Storm IX and its after-action report [S]: held April 2024 with more than 2,200 participants from over 100 private-sector organizations across 12 critical-infrastructure sectors, 35 federal agencies, 13 states and 11 partner nations. NERC, GridEx VIII [S]: held 18–20 November 2025 with 378 registered organizations (up 48% from 2023), more than 28,000 estimated players, and participants from electricity, gas, water, finance, telecommunications and government.
• Check: the essay's phrasing implies that national-scale cross-sector exercises do not exist. They do. They are largely exercises, not live restoration of production systems, and I found nothing that rehearses simultaneous loss of finance, communications and logistics with real recovery. So the essay's ask (live restoration, not tabletop) is a legitimate gap. But the sentence reads as if the whole category were absent.
• Severity: OVERSTATED (implied absence) — non-fail

H21. The opening scenario — SCENARIO, not a factual claim

• Essay: the bank account that "doesn't open," the declining card, the hospital whose backup generators "kick in, and then the electronic records system goes dark."
• Check: the essay presents it as a hypothetical ("This is the scenario missing from our AI debate"). Each element has a real analogue (Ascension's records loss in H18; the CrowdStrike effects in H1; the Iberian blackout in H19), and none is claimed as something that has happened in the combined form.
• Severity: n/a

I. Human/AI essay: internal consistency and insertion seams

I1. The title and "no credible answer" against "America is not starting from zero"

The essay claims there is no plan, and later says banks keep backups and utilities have cybersecurity teams. The "To be sure" paragraph answers the second half by calling the backups untested. A reader can still see the tension (Stage C Kimi 2). See H12.

• Severity: CONSISTENCY — non-fail, reader-visible.

I2. Insertion seams

The merge diff shows five clean insertions; the seams read smoothly.

• Insertion 2 closes the "fragile" paragraph. The next paragraph begins "AI does not invent this vulnerability," which now has a concrete referent.
• Insertion 3 sits just before "That is not a strategy. That is a hope," and reads as the argument's turn.
• Insertion 4 follows "they are an assumption" and precedes "And past blackouts…". The "A… / And…" openings are a little close together.
• Insertion 5 is a noun-phrase sentence in a list of fragments, which matches the list's style (the merge note said this on purpose).
• Insertion 1 is a parenthetical gloss that lengthens an otherwise short sentence ("Agencies like the Cybersecurity and Infrastructure Security Agency (CISA), the federal government's lead civilian cyber-defense agency, exist for a reason").
• Severity: PASS; HYGIENE for the "A… / And…" closeness.

I3. "Prevention is essential" vs the framing of the debate

The essay says "Prevention is essential. Resilience is non-negotiable," and attacks only prevention alone. Stage C Opus 3 misreads this as an attack on prevention (see the log). The essay is consistent.

• Severity: PASS

I4. Terminology: "open-source imitators" and "open-weight models"

The third paragraph says "open-source imitators" and the seventh says "open-weight models." The dump uses "open weight." Most of the models in question are open-weight, not open-source. The terms are used for the same thing within three paragraphs.

• Severity: HYGIENE — non-fail

I5. Length

1,163 words for the file, 1,147 for the body, against an 800–1,200 ask. The Muse base was 1,070 words (Muse's own count said about 1,050). The five insertions added about 93.

• Severity: PASS

Full model attribution

Full model attribution, Issue 34

• The author (Synthia Cipher, the byline): wrote the two questions that seeded the machine essay; wrote the raw brain dump (the written-without-AI page), its thesis and the human/AI editorial prompt; ran GPT-6 Astra in ChatGPT; leaned to the Muse Wonder base; approved all five insertions ("apply all five"); decided to freeze both essays and ruled A on both.
• Human/AI essay, "We Are Arguing About the Speed of AI. We Have No Plan for the Crash." (issue-034-humanai-FROZEN-v1.0):
  • Thesis, ideas, civilization-first framing, scenario and the headline idea: the author (from the raw dump).
  • Base draft (structure, headline wording, crash-plan list, aviation analogy): Muse Wonder (meta/muse-wonder via OpenRouter, reasoning effort xhigh, single turn, $0.1596).
  • Insertions 3, 4, 5: Gemini 4 Argon (google/gemini-4-argon, reasoning effort high, single turn, $0.0460). Insertion 3 is a light edit of its line; 4 and 5 are its ideas reworded.
  • Insertion 2 (the July 2024 outage sentence): anchor and phrasing from Claude Opus 5.5 (anthropic/claude-opus-5.5, reasoning.max_tokens 32000, single turn, $0.1201); the sourced 8.5 million figure from GPT-6 Astra (run by the author in ChatGPT); the same anchor was chosen independently by Kimi K3.1 Pro (moonshotai/kimi-k3.1-pro, reasoning effort high, single turn, $0.0298). None of Opus's, Kimi's or Astra's other text is in the essay.
  • Insertion 1 (CISA gloss) and the written form of insertion 2: Grok Bot. Comparison of all five drafts, merge recommendation, splice wording, freeze and audit: Grok Bot (editorial partner, Tess).
• Machine essay, "The Machinery of Being Wrong" (issue-034-machine-FROZEN-v1.0): Claude Opus 5.5 (anthropic/claude-opus-5.5 via OpenRouter, reasoning.max_tokens 32000), from the two-question seed only: run B ($0.0588) is the base; one paragraph (reversibility / Ogallala / elections) is condensed from run A ($0.0609, "The Power to Change Our Minds"). Merge, freeze and audit: Grok Bot. Three other models were run on the same seed for comparison only and none of their text is in the essay: Kimi K3.1 Pro, Gemini 4 Argon, Muse Wonder.
• Audit: dual audit (fact-check, consistency, convergence, credits) by Grok Bot (executor). Stage C critics: moonshotai/kimi-k3 (high reasoning) and anthropic/claude-opus-4.1, critique only.

Convergence disclosure

Selection effect.

Grok Bot ran the comparison, read all five drafts, recommended which Gemini and anchor items to import, wrote the CISA gloss and the CrowdStrike sentence, and assembled the merge. The author said "apply all five" and "freeze it." So the five insertions were chosen by the system and approved by the author, and the Gemini items were chosen because they strengthened a version the author already preferred. The base choice (Muse) was the author's lean after the comparison. The audit page should say this rather than present the final text as a single model's or a single process's output.

Convergence disclosure, Issue 34 (from dual audit sections C and J; say all of it plainly)

• Machine essay: two independent Opus 5.5 runs on the same two-question seed converged on the same thesis (error correction / the capacity for self-correction) and on the same historical cases (ozone, Lysenko, the Great Leap Forward). The frozen text is one of those two samples plus one paragraph from the other. Kimi K3.1 Pro, Gemini 4 Argon and Muse Wonder each chose a different answer (durable cooperation, the living biosphere, phosphorus). The thesis is the stable answer for Opus on this prompt, not an accident of one sample.
• The machine model never saw the author's raw dump (human_drafts_shown_to_model false), yet there is real thematic overlap with it (civilization as the unit that matters, resilience and recovery, moving one level up from a first-order debate). Reasons given in the audit: the two questions and the dump share an author and a mood; and Grok Bot, having seen the dump, chose between four machine essays and noted that the Opus thesis was adjacent to the author's resilience theme (a secondary reason for the pick).
• Human/AI essay: every drafting model saw the raw dump (it is inside the prompt), so their agreement on the thesis is not independent evidence; the thesis, the LA and Musk framing and the pause-versus-accelerate frame are the author's. The five insertions were chosen by the system (Grok Bot) and approved by the author.
• The two essays reach different answers (error correction versus a plan for crash recovery) on the same day, from the same author's concerns, by different routes.

Dual audit sections J1–J6

J. Human/AI essay: convergence and credit disclosures

J1. Every drafting model saw the brain dump, so convergence on the thesis is not independent evidence

The five drafts (Opus 5.5, Kimi K3.1 Pro, Gemini 4 Argon, Muse Wonder and GPT-6 Astra) all extracted the same thesis: stop arguing about speed and plan for the crash. They did so because the dump says it ("We are focusing on how fast we should go instead of… how to handle a crash"). The thesis, the LA and Musk framing, the bank-account scenario and the pause-versus-accelerate frame, and the "one step ahead in perpetuity" language all come from the author's dump, not from a model's discovery. This is the opposite of the machine essay, where the model never saw the dump (section C).

J2. What is the author's, and what is a model's

• The author's (from the dump): the thesis; the civilization-first opening ("Nothing matters more than civilization"); LA homelessness as the proof that civilization is a floor; "Elon Musk born in prehistoric times builds what exactly?"; the threat model (a rogue human/AI wiping banking records or shutting down the grid, hospitals and grocery stores affected); "stay one step ahead in perpetuity… or until some plateau"; "open weight models"; the question of what the backup plans are; the title idea (how fast versus how to handle a crash).
• Muse Wonder's (base): the structure (lede scenario, nut graph, "To be sure," kicker), the "20,000 years" figure, the aviation analogy, the specific crash-plan list (offline backups, segmented fallbacks, hospital paper protocols, emergency payments, stress tests), the headline wording and the first mention of CISA.
• Added from other drafts: the five insertions in J3.

J3. The five insertions: where each came from

1. CISA expansion. Requested by the author. Written by Grok Bot. The base already named CISA ("Agencies like CISA exist for a reason").
2. July 2024 outage sentence. Three of the five drafts reached for the same anchor on their own: Opus 5.5, Kimi K3.1 Pro and GPT-6 Astra. Gemini and Muse did not. That three models picked it independently says it is the canonical example in their training data, not that it is the best evidence. The attribution file says the sentence was "written by Grok Bot." For honesty, the content and wording closely track Opus 5.5's draft, which has "roughly 8.5 million Windows computers crashed. Airlines grounded flights. Hospitals postponed surgeries. Some 911 systems went dark," and "no hacker… and no artificial intelligence." The 8.5 million figure also appears in GPT-6 Astra's draft with its attribution to Microsoft and the "less than 1 percent" qualifier. Kimi's draft used the incident with no figures. Credit Opus 5.5 for the anchor facts and phrasing, Astra for the sourced figure, and Kimi for the choice of anchor. Every fact in the sentence was then verified in this audit (H1).
3. "Defenders must be right every time; an attacker only has to be right once." From Gemini 4 Argon's draft, which says "Defenders have to be right every single time; an attacker only has to get lucky once." The essay's sentence is a light edit of Gemini's. The idiom itself is old (H13).
4. Corrupted backups. The idea is from Gemini 4 Argon ("poisoning backup data months in advance"; "subtly poisoning the data inside 'clean' backups"). The essay's sentence is new wording.
5. Baseline date. The idea is from Gemini 4 Argon ("declare a baseline date"). The essay's sentence is new wording and drops Gemini's detail.

J4. Selection effect (by Grok Bot)

Grok Bot ran the comparison, read all five drafts, recommended which Gemini and anchor items to import, wrote the CISA gloss and the CrowdStrike sentence, and assembled the merge. The author said "apply all five" and "freeze it." So the five insertions were chosen by the system and approved by the author, and the Gemini items were chosen because they strengthened a version the author already preferred. The base choice (Muse) was the author's lean after the comparison. The audit page should say this rather than present the final text as a single model's or a single process's output.

J5. What was not used

Opus 5.5 argued for a policy angle (mandated rehearsal, Sheltered Harbor, Change Healthcare, EU's DORA) and its real-world anchors; Kimi K3.1 Pro argued for layered resilience and tested restoration; Astra argued for turning exercise findings into verified improvements, with GridEx and federal ransomware guidance. None of those arguments is in the essay beyond the shared CrowdStrike anchor. Several of them are the facts that would answer H12 and H20 (Sheltered Harbor, GridEx). That is a note for the author, not a defect: the audit found them by independent search, and they were also in the drafts the essay did not use.

J6. Relation to the machine essay

Both essays were written on the same day, from the same author's concerns, but by different routes (the machine essay never saw the dump). They arrive at different answers: the machine essay says the most vital element is error correction; the human/AI essay says the plan we lack is for crash recovery. The machine audit (C4) noted this. The two are close at one point: the machine essay's "recover from almost anything short of sudden annihilation" is about resilience, and the human/AI essay's crash scenario is a sudden loss. Disclose as a theme overlap that comes from the shared author and day, not as a model finding.

Stage C findings on the human/AI essay

Issue 34 Stage C synthesis: cross-model critique of the human/AI freeze

Date: 2026-10-24 (PT) · Essay: humanai-essay-FROZEN-v1.0.md, "We Are Arguing About the Speed of AI. We Have No Plan for the Crash." (sha256 17f11473c9a2ccbb1d3dc9a426d5e57c654b2977b8ab83d11da49bd0d0f78584, 1,163 words including the headline). Legs: moonshotai/kimi-k3 (reasoning effort high) and anthropic/claude-opus-4.1 via OpenRouter. Same legs, same system prompt (the reader-lands-it test) and the same chat() call as the machine-essay Stage C (run_stage_c.py, imported unchanged by run_stage_c_humanai.py). Key loaded with the load_key() logic from the Issue 31 runner. Each leg saw only the essay text, never the raw brain dump or any other draft. Returns are in stage-c-returns/ with the humanai- prefix. 2 of 2 legs OK.

Severity as rated by the legs: Kimi 10 ARGUMENT, 1 SCOPE, 1 HYGIENE; Opus 3 ARGUMENT, 2 SCOPE, 1 FLOOR. After adjudication there are no true FLOOR objections. Opus #4 (hospitals cannot run on paper for weeks) is self-rated FLOOR, but its premise is wrong (see the bad-premise log). No invented citations in either leg. The only non-ASCII characters in the returns are typographic quotes and dashes.

Bad-premise and misread log

These critiques rest on a premise the essay or the record does not support. They are logged, not adopted.

1. Opus #4, "hospitals cannot function for weeks on paper … technically infeasible" (self-rated FLOOR). Wrong premise. The essay asks for protocols that allow care to continue, and the record shows a major system doing it: Ascension's hospitals ran on paper charts, faxes and manual medication processes from the 8 May 2024 ransomware attack until the EHR was restored systemwide on 14 June 2024 (about five weeks). NPR reported staffers describing "harrowing lapses" in that period. So the claim that it is impossible fails. The kernel that survives is that it is degraded and error-prone, which is Kimi 9 (rated SCOPE). Not a FLOOR.
2. Opus #6, "pause can't be coordinated globally, so American-only resilience is contradictory." Wrong premise. A pause only works if rivals also comply. A bank's offline backups, a utility's segmented controls and a hospital's paper protocols work without anyone's cooperation. That asymmetry is the argument for resilience. A small kernel survives: shared cloud and software supply chains mean "domestic" is never fully domestic. The essay itself says the sectors "depend on the same internet, the same cloud providers."
3. Opus #3, first half: "the essay argues we're wrongly focused on prevention." Misread. The essay says "Prevention is essential" and "Keep competing … keep working on alignment and defense." Its target is prevention alone. The second half (finite budgets; no priority ranking) lands and converges with Kimi 10 and Opus 5.
4. Opus #1, "the central premise depends on AI being categorically different." Partly misread. The essay says outright that "AI does not invent this vulnerability" and rests its case on changed economics (skill floor, speed, number of attackers) and on the failure of prevention in general. A crash plan is justified by consequences, not by AI being new. The kernel (the essay never says why existing cyber-defence is not just scaled up) overlaps with Kimi 1 and 6.
5. Opus #2, "the essay never defines the crash." Partly misread. The opening scenario and the sentence "The risk AI raises is simultaneity" do define it: simultaneous loss of finance, communications, logistics and utilities. What is missing is a threshold at which the plan activates. That is a SCOPE point, not a missing definition.
6. Kimi #4, "'defenders must be right every time' cuts against the resilience thesis." Mostly misread. In the essay the sentence attacks the strategy of staying one step ahead forever, which is the prevention-only position. It supports the resilience argument. The kernel that lands is narrower: in a layered defence an attacker has to win several times, so the maxim is looser as logic than as rhetoric. It is also an old security idiom (see audit section H13). HYGIENE, not ARGUMENT.
7. Kimi #5, CrowdStrike "proves only dependence, not AI catastrophe." Overreads. The essay says in the sentence itself that the incident had "no hacker and no AI involved" and calls it a "preview" of dependence. The kernel (it was accidental, bounded and recoverable, so "preview" of a cascade is generous) is correct and matches audit finding H2.

Objections that land (convergent across both legs first)

1. The proposals have no owner, price or priority (Opus 5, Opus 3 second half, Kimi 9, Kimi 10, Kimi 12). Who funds the backups, who runs the national stress tests, who picks the baseline date, and how much resilience is worth buying relative to prevention? "Fund resilience as if failure is possible" does not rank anything, and the essay concedes no cost. Five of the 18 objections touch this. LANDS (SCOPE/ARGUMENT). The strongest convergent point. It is partly the nature of a 1,100-word op-ed, but it is the essay's own ask ("What would a crash plan actually look like?") that invites it.
2. "No plan" overreaches against the essay's own concessions (Kimi 2). The title and "Right now, we have no credible answer" sit against "America is not starting from zero. Banks keep backups." Reader-lands-it: LANDS. The audit found external support for the objection: a financial-sector vaulting standard covering about 72% of US deposit accounts, a national multi-sector exercise in April 2024 and a 378-organization grid exercise in November 2025 (audit sections H12, H20). The essay's "To be sure" paragraph answers the backups half by saying they are untested, which softens but does not remove the tension. This is the objection that carries a fact-check finding with it.
3. The AI-specific case is thinly argued (Kimi 1, Kimi 6, Kimi 7, Opus 1 partly). The jump from "cheaper, faster attacks" to one-morning multi-sector collapse skips the attack path; AI's defensive uses get no mention; "the digital layer" bundles corruption, encryption, outages and industrial-control compromise. Partly lands. The audit's evidence (AI-orchestrated intrusions, an AI-developed zero-day, open models 6–7 months behind on cyber) supports the direction of the premise but not the scale of the scenario. The essay frames the scenario as hypothetical ("A rogue actor — human, assisted by AI, or eventually AI acting…"), which limits the damage.
4. The pause/accelerate binary is drawn too neatly (Kimi 3). Partly lands. The essay calls it a debate "locked" between two camps and says both sides talk as though speed is the only decision. Defender-focused work exists (audit H14, H20), which makes "both sides" a generalization. It is the author's thesis frame, so it is a call for the author, not a fix.
5. The CrowdStrike "preview" is generous (Kimi 5). Partly lands; see bad-premise log item 7 and audit H2.
6. The pre-agreed baseline date is an unexamined mechanism (Kimi 8). LANDS, and it is one of the five credited insertions (Gemini 4 Argon's idea). Who picks the date, what happens to transactions after it, how disputes get settled? It is a single list item, and the essay presents it as part of a menu, not a worked design. It is the least supported of the five insertions.
7. Hospital paper and manual fallbacks carry costs the essay does not name (Kimi 9). Partly lands, with the Ascension evidence behind it.

Objections from one leg only, lower weight

1. Open-weight claim "too categorical" (Kimi 11). Partly lands. The audit finds the claim directionally right (H6) and Kimi is right that closed models are also misused, which the essay does not say.
2. "Every serious engineer knows it" and possibility versus expected loss (Kimi 10). HYGIENE-level; a rhetorical universal.
3. Historical analogies do not transfer to an adaptive adversary (Kimi 12). Lands lightly; aviation and the others show a method (assume failure), not a threat model. HYGIENE.
4. Opus 5 (no implementation path) is folded into 1 above.

Resolved: the author ruled A on October 24, 2026. The essay ships as written.

Read-across

• Of Kimi's 12 objections, three touch credited insertions: #4 (insertion 3, the defenders-and-attackers line), #5 (insertion 2, CrowdStrike) and #8 (insertion 5, baseline date). None touches the CISA gloss or the corrupted-backups sentence. The merge added about 95 words; roughly a quarter of the landing critiques are about them. As with the machine essay, that is a reason for the audit page to say so, not a reason to edit the freeze.
• The two sharpest items are the "no plan" framing against the record (2) and the missing owner, price and priority (1). The first is the only one that is a fact-check finding as well as a critique.
• Under the proposed Ruling A, Stage C changes no freeze; the landing objections go on the audit page.

Resolved: the author ruled A on October 24, 2026. The essay ships as written.

Rulings that need the author

Ruling A for the human/AI essay is still pending in RULINGS.md. The items above that would make a ruling meaningful: (1) the title and "no credible answer" against the Sheltered Harbor, Cyber Storm IX and GridEx VIII record; (2) the missing owner, cost and priority; (3) the baseline-date sentence. A narrow v1.1 (for example "no tested plan" or "no credible, tested answer") would be one option for (1). It is the author's call, as it was with the Sen fix on the machine essay.

The written-without-AI page is its own third page

The written-without-AI page is Nothing Matters More Than Civilization.

Typo fixes

(in order of appearance)

1. P2, missing word. the word "of" was missing.
   Before: none this is really
   After: none of this is really
2. P3, spelling. "rouge" changed to "rogue".
   Before: a rouge human/AI
   After: a rogue human/AI
3. P3, missing letters (meaning-restoring; author to confirm). "can" -> "can't": the sentence only makes sense as "can't".
   Before: if you can access your bank
   After: if you can't access your bank
4. P3, stray line break + capital (paste artifact). line break inside a sentence joined; capital H lowercased.
   Before: How do⏎Hospitals function?
   After: How do hospitals function?
5. P3, capitalization. "Nuclear" -> "nuclear" mid-sentence.
   Before: accelerate Nuclear and biological
   After: accelerate nuclear and biological
6. P4, capitalization. "Model" -> "model" mid-sentence.
   Before: plateau in Model capability
   After: plateau in model capability
7. P4, spelling (wrong form). "surpasses" -> "surpassed".
   Before: being surpasses or at least
   After: being surpassed or at least
8. P4, spelling (wrong word). "had" -> "hard".
   Before: working as had on alignment
   After: working as hard on alignment

M. Human/AI essay: hygiene

• The banned-term scan is clean on the essay and the attribution file.
• The essay names Elon Musk (a public figure, as a thought experiment), and the agency CISA. It names no private individuals and no companies. It quotes no one.
• The essay contains no quotation marks; no quotation can be misquoted.
• Numeric claims: "20,000 years ago" (H5), "July 2024" and "8.5 million" (H1), "weeks, not hours" (H18) and "just-in-time schedules measured in hours" (H11). The first four hold. The last is the weak one.
• Sources read for this part were a mix of full pages and excerpts; the [F] and [S] marks say which.

N. Human/AI essay: source list

Pages read in this pass. [F] = fetched and read; [S] = search-result excerpt.

• H1: Microsoft blog, "Helping our customers through the CrowdStrike outage" (20 Jul 2024) [S], https://blogs.microsoft.com/blog/2024/07/20/helping-our-customers-through-the-crowdstrike-outage/ · Reuters [S], https://www.reuters.com/technology/microsoft-says-about-85-million-its-devices-affected-by-crowdstrike-related-2024-07-20/ · CrowdStrike, "Technical Details: Falcon Update for Windows Hosts" [S], https://www.crowdstrike.com/en-us/blog/falcon-update-for-windows-hosts-technical-details/ · CrowdStrike, External Technical Root Cause Analysis, Channel File 291 (6 Aug 2024) [F, full text], https://www.crowdstrike.com/wp-content/uploads/2024/08/Channel-File-291-Incident-Root-Cause-Analysis-08.06.2024.pdf · New York Times live coverage [S], https://www.nytimes.com/live/2024/07/19/business/global-tech-outage · USA Today [S], https://www.usatoday.com/story/news/nation/2024/07/19/crowdstrike-outages-what-happened/74474725007/ · CNN [S], https://www.cnn.com/2024/07/19/business/delta-american-airlines-flights-outage-intl-hnk · NBC News [S], https://www.nbcnews.com/news/world/live-blog/live-updates-it-outage-flights-banks-businesses-microsoft-crowdstrike-rcna162669 · AP on Delta [S], https://apnews.com/article/outage-airline-delta-40fc208ac838cafab40482b731072018
• H3: About CISA [S], https://www.cisa.gov/about · DHS/CISA report to Congress, "Critical Infrastructure Outreach" (Oct 2024) [S], https://www.dhs.gov/sites/default/files/2024-12/2024_1023_cisa_critical_infrastructure_outreach.pdf · Federal News Network, Jan 2026 [S], https://federalnewsnetwork.com/cybersecurity/2026/01/lawmakers-press-acting-cisa-director-on-workforce-reductions/ · New York Times, 22 Feb 2026 [S], https://www.nytimes.com/2026/02/22/us/politics/cyber-agency-dhs-security-setbacks.html · National Cyber Incident Response Plan [S], https://www.cisa.gov/sites/default/files/ncirp/National_Cyber_Incident_Response_Plan.pdf
• H4: LA County Public Health press release, 10 Mar 2026 [F], https://lacounty.gov/2026/03/10/los-angeles-county-homeless-mortality-rate-decreases-for-the-first-time/ · LAHSA, 2026 Homeless Count [S], https://www.lahsa.org/news?article=1079-2026-homeless-count-reveals-bottleneck-stymied-progress
• H5: Wu et al., "Early Pottery at 20,000 Years Ago in Xianrendong Cave, China," Science [S], https://www.science.org/doi/10.1126/science.1218643 · Getty, "Where Did Writing Come From?" [S], https://www.getty.edu/news/where-did-writing-come-from/
• H6, H8: UK AISI, "How Far Behind the Frontier are Leading Open Weight Models on Cyber?" (17 Jul 2026) [F], https://www.aisi.gov.uk/blog/how-far-behind-the-frontier-are-leading-open-weight-models-on-cyber · Epoch AI, "Open-weight AI models are 6-7 months behind on cyber" [F], https://epoch.ai/data-insights/open-closed-cyber-eci-gap
• H7, H8: Anthropic, "Disrupting an AI-orchestrated cyber espionage campaign" [S], https://www.anthropic.com/news/disrupting-AI-espionage · Google Threat Intelligence Group, "Adversaries Leverage AI for Vulnerability Exploitation, Augmented Operations, and Initial Access" (11 May 2026) [F], https://cloud.google.com/blog/topics/threat-intelligence/ai-vulnerability-exploitation-initial-access
• H9: Federal Reserve, H.6 Money Stock Measures (22 Sep 2026) [S], https://www.federalreserve.gov/releases/H6/Current/
• H10: ONC Data Brief 35 [S], https://healthit.gov/wp-content/uploads/2017/03/2015_hospital_adoption_db_v17.pdf
• H11: Monash Lens, "Empty shelves: When just-in-time systems meet just-in-case people" [S], https://lens.monash.edu/empty-shelves-when-just-in-time-systems-meet-just-in-case-people/ · BBC Worklife, "Why grocery shelves won't be empty for long" (2020) [S], https://www.bbc.com/worklife/article/20200401-COVID-19-why-we-wont-run-out-of-food-during-coronavirus
• H12: ICBA, Sheltered Harbor fact sheet [S], https://www.icba.org/documents/45248/963569/ShelteredHarborFastFacts.pdf · ABA Banking Journal, 2019 [S], https://bankingjournal.aba.com/2019/05/aba-trade-group-ceos-to-financial-institutions-join-sheltered-harbor/ · FS-ISAC, "Sheltered Harbor Announces Planned Wind-Down" [F], https://www.fsisac.com/who-we-are/sheltered-harbor
• H13: Quote Investigator, "We Only Need To Be Lucky One Time" [S], https://quoteinvestigator.com/2025/12/08/lucky-always/ · Wikipedia, "Brighton hotel bombing" [S], https://en.wikipedia.org/wiki/Brighton_hotel_bombing
• H15: NTSB, Safety Recommendations [S], https://www.ntsb.gov/investigations/Pages/safety-recommendations.aspx
• H16: Veeam, 2024 Ransomware Trends Report [S] · Sophos, "The impact of compromised backups on ransomware outcomes" [S], https://www.sophos.com/en-us/blog/the-impact-of-compromised-backups-on-ransomware-outcomes · CISA advisory AA24-038A [S], https://www.cisa.gov/news-events/cybersecurity-advisories/aa24-038a · CSO Online, "Rebuilding after NotPetya: How Maersk moved forward" [S], https://www.csoonline.com/article/567845/rebuilding-after-notpetya-how-maersk-moved-forward.html
• H18: NPR, "Ransomware attack led to harrowing lapses at Ascension hospitals, staffers say" (19 Jun 2024) [S], https://www.npr.org/2024/06/19/nx-s1-5010219/ascension-hospital-ransomware-attack-care-lapses · Becker's, "Ascension restores EHR systemwide" [S], https://www.beckershospitalreview.com/healthcare-information-technology/ehrs/ascension-restores-ehr-systemwide/
• H19: ENTSO-E, Final Report on the Grid Incident in Spain and Portugal on 28 April 2025 [S], https://eepublicdownloads.blob.core.windows.net/public-cdn-container/clean-documents/Publications/2025/iberian-blackout/Final%20Report%20on%20the%20Grid%20Incident%20in%20Spain%20and%20Portugal%20on%2028%20April%202025.pdf · Frontiers in Public Health, Portuguese National Health Service and the April 2025 blackout [S], https://www.frontiersin.org/journals/public-health/articles/10.3389/fpubh.2025.1630933/full
• H20: CISA, Cyber Storm IX [S], https://www.cisa.gov/resources-tools/resources/cyber-storm-ix · NERC, "GridEx VIII – Surge in Participation…" [S], https://www.nerc.com/newsroom/gridex-viii--surge-in-participation-reflects-importance-of-exercising-emergency-preparedness

The development record.