The Signal.

The Signal — Issue 13

Eight governance instruments landed in nine days — an executive order, a voluntary accord, a presidentially announced task force, a seventeen-government declaration, an FTC probe, a state subpoena, a Senate hearing and a bill with no text — and the only dated obligation anyone produced all week was a CISA fix-by date on two Zammad CVEs that bound federal agencies only, that the vendor publicly disputed, and that came and went on Monday with no patch for half the chain. Wednesday, October 7, 2026.

Current issue · Issue 14 · October 14 · Issue 12 · September 29 · Issue 11 · September 22 · Issue 10 · September 15 · Issue 9 · September 8 · Issue 8 · September 1 · Issue 7 · August 26 · Issue 6 · August 19 · Issue 5 · August 11 · Issue 4 · August 5 · Issue 3 · July 28 · Issue 2 · July 21 · Issue 1 · July 14

An executive order renamed the technology and started a 60-day clock on redefining it. A one-page accord set up four layers of oversight and wrote no penalty. The President announced a Super Intelligence Force in a social post rather than a signed document. Seventeen governments endorsed a science declaration in the new vocabulary. And a security research organisation that reads logs for a living published the first behavioural description of an agent attacking it. Eight instruments, one missing sentence: what happens when the looking finds something.

• Eight governance instruments landed between 29 September and 7 October and none of them writes a consequence sentence. (EO “Inaugurating The Era Of Super Intelligence,” whitehouse.gov, 29 September, read end to end — Sec. 3(b) sets a 60-day definition clock landing ~28 November · the White House Accord’s four layers, no penalty, no incident-reporting duty, self-chosen auditors, accord primary paywalled and NOT READ · the SIF announced by the President on Truth Social 4 October, name, roster and reporting line now primary-sourced, the chair/vice-chair structure and the 120-day report secondary via TechCrunch, no whitehouse.gov document and not an executive order · the Kyoto Vision, whitehouse.gov 4 October, sixteen named partners plus the US, declaration PDF linked and NOT OPENED here, non-binding, science policy and not oversight · FTC probe confirmed only by a spokesperson to CNBC, no agency primary on day six · California AG subpoena 1 October · Senate DMDCC witness statements, not findings · Hawley–Murphy announced with no bill number and no text, and congress.gov is bot-walled from this host, which is NOT VERIFIABLE rather than absent.)
• With no mid-market adoption figure in window for the second pack running, the anchor is a purchase order: ~$83.4M in TMF awards, State’s $17.3M share, with human review of every automated decision written into the award alongside employees building their own agents. (FedScoop, 2 October, read in full · the 100,000-hours figure is TMF’s projection and nothing is deployed · “every automated decision” is FedScoop quoting the TMF site, not a contract term read here · tmf.cio.gov NOT FETCHED, no vendor named.)
• The victim published the behavioural tell, and the chain’s federal fix-by date passed on Monday with the vendor still disputing it. DIVD on its own helpdesk breach: “we could see the agent working automated, because after every action it decided the next step itself…” (The Register, 1 October, read in full · the agentic attribution is DIVD’s inference from modus operandi, not an identification, and The Register hedges “or at least AI-enabled” · CVE-2026-102489 and -102490 added to CISA KEV 2 October, forensic triage flagged, due 5 October under BOD 26-04, read off the KEV JSON · Zammad publicly disputes the scope of one and published no patch for the other · the deadline bound federal civilian agencies, not private companies · DIVD rates 9.4, NIST rates 9.8.)
• A preprint priced coordination machinery against itself: runtime selection raised mean scores, and adding scoped work contracts and verification lowered them under a fixed budget. (arXiv:2610.00980, submitted 1 October, abstract only · single-seed exploratory evaluation, three hosts, one benchmark, no intervals, no replication — not a general result · the constructive half is arXiv:2610.02036, whose ladder runs 5 of 5 budget violations for ordinary teams, 4 of 5 with a visible live count, and 0 of 5 with commit enforcement, on the authors’ own constructed benchmark — visibility moved almost nothing.)
• The same mechanism showed up two orders of magnitude down: the enterprise client dictates the tenant and the compliance terms to the one-man consultancy. (The Intelligent Innovation Show episode 5, recorded 3 October 2026 · PARAPHRASE, not quotation — in-room Google Meet ASR, listen- check open at ~00:34, ~00:58 and ~01:00 · one named attendee spoke zero turns and is credited with nothing.)

Hold the first signal against the third. Eight instruments in nine days, and the only one that produced a date somebody actually had to meet was a CISA entry about two helpdesk CVEs — which the vendor says do not apply to current versions, and for half of which no patch existed when the date arrived. The top of the system wrote oversight layers. The bottom of it wrote a deadline, into a disagreement, and that deadline passed on Monday.

Read the last nine days as a procurement document rather than as a policy debate. Governance is arriving as instruments rather than as principles, and the instruments name parties and dates while leaving the consequence blank. Eight of them, in the order they landed, each at its own source strength.

One. The executive order “Inaugurating The Era Of Super Intelligence,” signed 29 September 2026, directs that the executive branch use “Super Intelligence” and “SI” — superintelligence, the administration’s replacement term for artificial intelligence — in place of “Artificial Intelligence” and “AI” in “official correspondence, public communications, websites, reports, policy documents, and other non-statutory documents.” Sec. 3(b) is the operative part: it gives the Assistant to the President for Science and Technology 60 days to submit proposed legislative language for a federal definition of Super Intelligence, including an assessment of whether it should “modify, expand upon, or otherwise supersede” the existing statutory definition at 15 U.S.C. § 9401(3). That clock lands around 28 November 2026. This is a terminology and definition-proposal order. It puts no obligation on a private firm, and Sec. 2(b) expressly declines to require alteration of previously issued regulations, contracts or grants. Do not call it regulation.

Two. The White House Accord on Superintelligence, same day. Four oversight layers, every layer ending in a report, no penalty written into the text, no requirement to report safety incidents to any government body, and each company picking its own auditor. The accord primary is paywalled and was NOT READ here; the four-layer structure and the three absences come from Al Jazeera and Infosecurity, both 30 September. David Sacks, the White House AI czar who helped organize the summit, answers that critique directly on the All-In podcast: he points to the FTC and the SEC standing behind the document and argues that although the agreement was entered into voluntarily, the governance that follows from it is not. That is a paraphrase of ASR captions at 0:05:20, not audio-verified, and it is carried as the organizer’s advocacy rather than as a finding.

Three, and this one moved since the outline. The “Super Intelligence Force (SIF)” is no longer press-reported. The President announced it himself, in writing, on 4 October 2026 at 8:23 a.m. ET: “Following the Historic White House Accord on Super Intelligence … I am announcing the formation of the Super Intelligence Force (SIF),” “tasked with coordinating the effort of the Federal Government,” and coordinating its engagement with “Consumers, Public Interest Groups, Religious Organizations, Critical Infrastructure Providers, and Super Intelligence Companies.” The named roster is DNI Jay Clayton, FTC Chairman Andrew Ferguson, Under Secretary of War for R&E and CTO Emil Michael, and OPM Director Scott Kupor, reporting to the President and the Chief of Staff. Split the label by claim: the name, the four officials and the reporting line are now primary-sourced from the principal. The chair and vice-chair structure, the 120 days to report on “the risks and opportunities presented by AI,” and the charter language about responding to “SI-enabled threats to our society, while preventing overregulation and regulatory capture” are secondary — carried by TechCrunch, read, and not in the President’s own post. There is still no whitehouse.gov document — both indexes were fetched — so this is not an executive order, and “AI czar” is the press’s word, not the post’s. The membership is the part a mid-market owner should read twice: the FTC chairman, who sits over an open industry-wide probe, and the director of OPM.

Four, and it is new. On 4 October 2026 whitehouse.gov published that at the 23rd STS Forum in Kyoto, the United States and 16 countries endorsed the Kyoto Vision for a Golden Age of Science, led by OSTP Director Michael Kratsios. Ministers “pledged to expand researcher access to SI for science tools,” and the Vision recognises the opportunity to “harness super intelligence,” which “will enable autonomous experimentation at scale.” Seventeen governments signed a document that says SI rather than AI, which makes the rebrand multilateral in nine days. Screens: the release links the declaration as a PDF and names all sixteen partners — Argentina, Bulgaria, Chile, Cyprus, Germany, Greece, Indonesia, Italy, Japan, Kazakhstan, the Republic of Korea, New Zealand, Poland, Singapore, the United Arab Emirates and the United Kingdom — and the declaration itself was NOT OPENED here, so nothing below is a reading of the document. There is no money and no deadline, and it mentions neither Clayton nor the SIF. It is science policy, not an oversight instrument — it is in this list because it is the vocabulary arriving at treaty level, and “autonomous experimentation at scale” is a government describing agents running without a human in each loop.

Five. An FTC investigation into OpenAI, Anthropic and unnamed others, confirmed by an FTC spokesperson to CNBC on 30 September. No agency primary exists on day six. The FTC press index was enumerated twice and still carries only two October items, neither about AI. No scope, no legal theory and no third company name is printed here, because no source read supports one.

Six. A California Attorney General investigative subpoena on OpenAI, 1 October, off oag.ca.gov. One line in the inventory here; it was spent as a post last week.

Seven. The Senate DMDCC “Rogue AI” hearing — five non-vendor witnesses, written testimony on the record at hsgac.senate.gov. Witness statements, not subcommittee findings.

Eight. The Hawley–Murphy AI Agent Accountability Act, announced 1 October. This is a press release announcing legislation. There is no bill number, no text and no referral; congress.gov is bot-walled from this host, so the correct report is NOT VERIFIABLE FROM THIS HOST and never “absent.” This issue does not define “operator,” because no document we can open does.

Population and screen. The word that will actually reach a mid-market owner out of all eight is operators, and it is the one word none of the eight defines. Every instrument above names a party, a venue or a date. The sentence that says what happens when the looking finds something is written in none of them.

There is no methodologically sound mid-market adoption survey in window, for the second consecutive pack, and that absence is itself the finding — so this issue anchors on something a buyer wrote down instead. On 2 October 2026 FedScoop reported four new Technology Modernization Fund awards totalling roughly $83.4 million, including a $17.3 million share to the State Department, and the award language carries two things in the same breath: a person reviewing every automated decision, and employees building their own agents.

Population and screen. The “more than 100,000 hours a year” attached to this is TMF’s own projection. Nothing is deployed. The phrase about reviewing “every automated decision” is FedScoop quoting the TMF site, not a contract term read here — tmf.cio.gov was NOT FETCHED, no award document was opened, and no vendor is named anywhere in this item. What survives is the shape, and the shape is the point: a buyer under no commercial pressure to say it wrote per-decision human review into the award. “Agents propose, a named human commits” now has a non-vendor precedent you can put in front of a board without citing anybody’s marketing.

DIVD, a Dutch CVE Numbering Authority, published the breach of its own Zammad helpdesk. The attack ran 21 September, was detected 22 September, and on 24 September DIVD reported to the vendor, notified Dutch authorities and consulted police. DIVD’s reading of its own logs: “the modus operandi indicates that this is an agentic AI powered attack,” and “we could see the agent working automated, because after every action it decided the next step itself, at the speed of light and sloppy logic or pattern.” Two CVEs chained — CVE-2026-102489 and CVE-2026-102490 — with no vendor patch available at the time of exploitation.

Since the outline, the anecdote became a dated obligation, and the date has already gone by. Both CVEs went onto CISA’s Known Exploited Vulnerabilities catalogue on 2 October, flagged for forensic triage, and the federal remediation deadline under BOD 26-04 fell on 5 October 2026 — two days before this issue reached you — read directly off the KEV JSON feed, catalogue version 2026.10.04. CISA describes 102489 as a “session fixation” flaw leading to “remote code execution as the zammad user” and 102490 as “improper privilege management” escalating “to root,” each of which “can be chained with” the other. And the vendor disputes it in public. Zammad’s own community posts, dated 1 October, say of the first CVE that “current Zammad versions are not affected” and that exploitation is “only possible on Zammad 6.5 and older,” versions that “reached end of support some time ago.” On the second: “we are working on it,” it “cannot be exploited remotely on its own,” update to 7.2.0. That thread has carried no new post since 1 October, and as of the deadline there was no advisory and no patch for 102490 — the GitHub advisory feed for zammad/zammad has nothing for either CVE, its newest entry being 25 August.

Population and screen. The agentic attribution is DIVD’s own inference from modus operandi, not a confirmed identification; The Register itself hedges “or at least AI-enabled.” No actor, vendor or model is named, DIVD’s own report and posts were NOT FETCHED, and the investigation was open at publication. The 5 October deadline bound federal civilian agencies, not private companies — do not let anyone sell it to you as a general deadline. The two screens together are the item: “how would we even know an agent did it” now has a published answer from people who read logs for a living, and it is behavioural rather than forensic — and the consequence layer, the one time it showed up in writing this week, showed up as a fix-by date that two credible parties disputed and one of them could not meet. That date passed on Monday with no patch and no advisory published for 102490. That gap is where an owner has to decide.

A preprint submitted 1 October 2026, arXiv:2610.00980, “Can AI Scientists Coordinate at Runtime?”, reports that runtime selection raised mean scores — and that adding scoped work contracts and verification reduced them under a fixed budget. The limits travel in the same sentence or the finding does not travel at all: a single-seed exploratory evaluation, three hosts, one benchmark, no intervals, no replication, abstract page only. This is not a general result about governance overhead and must never be reported as one.

The constructive half: arXiv:2610.02036 describes a design in which models propose while the harness owns shared state and governs commit, and its ladder has three rungs rather than two. Ordinary teams exceeded a shared budget in 5 of 5 runs. Giving them a visible live count of that budget left 4 of 5 violations. Enforcing the commit point left 0 of 5. Making the number visible, in real time, to the people spending it moved almost nothing. Abstract only, constructed benchmarks, small-n counts on the authors’ own harness.

Population and screen. Neither paper is peer reviewed and neither is production telemetry. Taken together they say something a buyer can use anyway, and the middle rung above is the receipt for the first half of it: visibility is not control, and controls have a price. Buy the ones that change a decision, and say out loud what the rest cost — because under a fixed budget somebody is already paying for them whether or not the line item exists.

On episode 5 of The Intelligent Innovation Show, recorded 3 October 2026, a guest described standing up a one-man SI consultancy from nothing. Host Brad Wilcox answered in procurement terms: the enterprise client will arrive with its own requirements and its own tenant that it makes him use, and the compliance considerations come attached — so he will likely be working inside a stack his client already owns. In the same room, Christian Nicolai supplied the demand-side half: before any of that, somebody has to define how much a customer is actually willing to give to the machine. The guest, Santosh Ghosal, closed by moving off his own question — he came in expecting a conversation about numbers and tech stack, and left saying it reduces to the value being offered and where the trust sits, and that not everything can be agentic.

Population and screen, and the screen is an attribution one. All three lines above are paraphrase, not quotation. The source is a Google Meet auto-transcription — in-room ASR, whose own footer concedes errors — at roughly 00:58, 00:34 and 01:00. None of them publishes as a quotation without a listen-check against the recording. Radhakrishna Mutthoju is on the attendee list, spoke zero turns, and is credited with nothing. Two other room claims heard on that episode — an outcome-pricing figure attributed to a third party, and a line about a large consultancy switching to outcome-based pricing — are unsourced and are not repeated here.

The pairing is the whole argument of this issue. The accord’s unwritten penalty and the whale client’s mandated tenant are the same mechanism at two scales, and the smaller one is the one that actually binds. Nobody is going to fine the one-man consultancy. The client is simply going to hand him the tenant.

Figures barred from every surface of this publication, named so you recognise them when they arrive unattributed. The “>100 organizations” and “~50 petabytes” figures are Reuters-reported, while OpenAI’s own living page said “notified dozens of third parties” at the 2 October 06:48:12Z capture — print dozens, attribute anything larger to the outlet. Also refused: technology.org’s “half a million a day” compute figure; the OneTrust 2026 AI-Ready Governance survey (out of window, vendor-commissioned, screened denominator); the “92% of CISOs / 95% doubt” pair (no reachable primary); “93% of midsize companies,” “$12.4B agentic market” and “76% of CFOs” (undated vendor blogs); Deloitte’s “1 in 5 mature agent governance” (January 2026); digitalapplied’s “11 official steps” (uncited); and an aggregator-sourced AMD projection of “40–60% savings on a 500-unit AI PC fleet” with no methodology and no primary. LASST v. OpenAI is secondary-sourced only, the complaint is not quotable from this host, and it gets no paraphrase.

Three things were checked and are named rather than run. Google suspended product vulnerability submissions to its open-source bug bounty on 1 October, saying the pause is due to “a significant rise in automated submissions, the vast majority of which are not valid” — a strong item, but the primary is a JavaScript shell from this host, Google published no count, and Google said automated submissions, not AI agents. It is held rather than spent. The David Robinson resignation was read only through secondary carriers; OpenAI did respond, and the response is dated 3 October, not yesterday. The spokesperson’s line: “We pause training or hold back models when we need to slow down.” Robinson has since retained a PR firm, which belongs in the frame. And a prompt-injection detector re-evaluation preprint (fifteen detectors, rankings that do not transfer between benchmarks) is real and is deferred: it needs a table and a benchmark-not-production frame that this issue has no room for.

Named gaps, printed as gaps, because an unlabelled hole reads as a finding. congress.gov, theatlantic.com, reuters.com, wsj.com and openai.com are bot-walled or paywalled from this host. bughunters.google.com returns a 3.5 KB JavaScript shell at HTTP 200 — a 200 is not a readable primary. NIST published zero October items. The FTC press index carries no AI item. oag.ca.gov carried no new release. No mid-market CEO or CFO survey landed, second pack running. The NYC Council’s 5 October hearing has produced three consecutive windows of silence from the issuing body — not disproved, unpublished.

SAME-DAY VERIFICATION RUN, 5 October 2026. Every URL below was re-fetched live on this date. Each entry carries the fetch verdict, datePublished where the page exposes one, and the load-bearing claim it is on the hook for. Tooling limit, stated rather than papered over: the retrieval tool on this seat returns page content, not the raw response line or a byte counter. A verdict of “fetched, content returned” is evidence of a 200-class response and is not a literal status-code read; byte counts from the 1–5 October research-bank reads are carried below LABELLED AS BANKED and were not re-measured today. Three divergences from the banked reads are listed first.

BODY-IMPACT — read before shipping. Two body claims no longer match their source and one date has turned. This seat does not rewrite body prose; these are the shipper’s calls.

1. Section 04 and signal 1, Kyoto Vision — FALSIFIED SCREEN. The draft says “the declaration text is not linked anywhere on the release, the 16 countries are not named.” Both halves are false on today’s page. The release links the declaration as https://www.whitehouse.gov/wp-content/uploads/2026/10/Kyoto-Vision-for-a-Golden-Age-of-Science.pdf under the anchor text “Kyoto Vision for a Golden Age of Science”, and it names all sixteen partners: Argentina, Bulgaria, Chile, Cyprus, Germany, Greece, Indonesia, Italy, Japan, Kazakhstan, Republic of Korea, New Zealand, Poland, Singapore, United Arab Emirates, United Kingdom. The screen sentence in section 04 and the “declaration text not linked” clause in signal 1 must be cut or inverted. The PDF itself was NOT OPENED on this run, so “the declaration was not read” remains true and is the honest residual screen. The seventeen-government count, the STS Forum number, Kratsios, the quoted spans, the absence of money, deadline, Clayton and SIF all still hold. 2. Section 02 instrument three and gate warning 2, the SIF split — LABEL NOW UNDER-STATED. The draft says the chair/vice-chair structure, the 120-day report and the charter language are “WSJ-only and were NOT READ here.” TechCrunch, already cited in this block and read again today, carries all three: Clayton as chair, Ferguson, Michael and Kupor as vice-chairs, “120 days to create a report on the risks and opportunities presented by AI,” and charter language directing the force to “develop plans for responding to SI-enabled threats to our society, while preventing overregulation and regulatory capture that would stifle innovation and competition.” The claim is therefore secondary-sourced-and-read, not unread. It is still not primary and still has no whitehouse.gov document — both indexes were re-fetched today and carry none — so nothing here licenses “executive order” or “signed.” The split stays split; only the weaker side’s label changes. 3. Section 04, signal 3, the LEDE_STRONG and the BRIDGE — the fix-by date has arrived. The KEV dueDate is 2026-10-05, which is today and is two days before Tuesday’s ship. Every mention reads as a future obligation. The entries themselves are unchanged. Tense only; no fact moved.

Verified unchanged, source by source.

Executive order, “Inaugurating The Era Of Super Intelligence”. RE-FETCHED 5 October 2026 — CONFIRMED. Content returned; on-page category “Executive Orders”, on-page date “September 29, 2026”. Banked read 1 October: schema datePublished 2026-09-29T21:17:25+00:00, browser UA required; the schema field was not re-extracted today. All four load-bearing spans still read as drafted: the Sec. 2(a) scope list (“official correspondence, public communications, websites, reports, policy documents, and other non-statutory documents”), the Sec. 2(b) carve-out (“Nothing in this section requires the alteration of previously issued regulations, Presidential actions, contracts, grants, or other historical documents”), the Sec. 3(b) 60-day deliverable including “modify, expand upon, or otherwise supersede”, and the 15 U.S.C. § 9401(3) anchor. Note for the archive, not for the body: the page is also identified as Executive Order 14434 — the issue does not print an EO number and need not start.

Truth Social post announcing the SIF, archived at trumpstruth.org/statuses/42126. RE-FETCHED 5 October 2026 — CONFIRMED. Post still dated 4 October 2026, 8:23 a.m. ET, still archiving truthsocial.com/@realDonaldTrump/117382616783706638, which is still the permalink carried above. The roster reads as drafted: Jay Clayton (DNI), Andrew Ferguson (FTC Chairman), Emil Michael (Under Secretary of War for R&E and CTO), Scott Kupor (OPM Director), reporting to the President and Chief of Staff Susie Wiles. Banked read 5 October: 16,364 B; truthsocial.com itself a 2,880 B shell, not re-tested today. The archive is named as an archive above and that framing stands. Corroboration, both re-fetched today: TechCrunch — CONFIRMED, dated 4 October 2026, same four names, and it also carries the chair/vice-chair structure, the 120-day report and the charter language: see BODY-IMPACT 2 — and NBC News — CONFIRMED, 4 October 2026 7:09 p.m. ET, confirms the Truth Social route and the “report to him and to White House chief of staff Susie Wiles” line. NBC names only Clayton, Ferguson and Michael and does not name Kupor; Kupor rests on the primary post and TechCrunch. NBC also reports Clayton as “leading the task force” and as expected to serve as AI czar — that is NBC’s characterisation of Clayton and it is exactly the usage gate warning 3 bars above the line. Do not let it migrate in. whitehouse.gov /presidential-actions/ and /news/ were both re-fetched today and still carry no SIF document — the presidential-actions index’s only post-1-October item is a National Manufacturing Day proclamation, 2 October. Banked sizes 262,944 B and 291,185 B, not re-measured. The WSJ page was not attempted on this run and remains unopened.

Kyoto Vision release, whitehouse.gov. RE-FETCHED 5 October 2026 — CHANGED. See BODY-IMPACT 1. On-page date 4 October 2026; banked datePublished 2026-10-04T09:14:42Z and 265,914 B, not re-measured. What changed against the banked read: the declaration PDF IS linked and the sixteen partner countries ARE named on the release. Still CONFIRMED on this page: sixteen countries plus the United States, the 23rd STS Forum annual meeting, Kratsios as OSTP Director, the pledge to “expand researcher access to SI for science tools”, the span “harness super intelligence and advance new approaches to greatly accelerate scientific discovery”, and “autonomous experimentation at scale, unlocking new experimental and analytical capabilities”. No mention of Clayton or the SIF, no money and no deadline — all three screens hold. The declaration PDF was NOT OPENED on this run; “not read” is true and “not linked” is not.**

CISA Known Exploited Vulnerabilities JSON feed. RE-FETCHED 5 October 2026 — CONFIRMED, NO ENTRY CHANGED. catalogVersion 2026.10.04, dateReleased 2026-10-04T18:52:56.0635Z, count 1734 — identical to the banked read, so the catalogue has not been reissued since. Banked size 1,767,828 B, not re-measured. Both CVEs still present, both vendorProject “Zammad GmbH”, product “Zammad”, dateAdded 2026-10-02, dueDate 2026-10-05, knownRansomwareCampaignUse “Unknown”, required action citing BOD 26-04 and CISA’s “Forensics Triage Requirements”. CISA’s own wordings re-read verbatim today: 102489 “contains a session fixation vulnerability that can lead to remote code execution as the zammad user. This vulnerability can be chained with CVE-2026-102490”; 102490 “contains an improper privilege management vulnerability that can allow the local zammad user to escalate privileges to root. This vulnerability can be chained with CVE-2026-102489.” dueDate 2026-10-05 is today and is past at Tuesday’s ship: BODY-IMPACT 3.

Zammad vendor statement, community.zammad.org thread 21297. RE-FETCHED 5 October 2026 as HTML and as thread JSON — CONFIRMED, AND NO NEW POST SINCE. Three posts only, all 1 October 2026: the reporter’s alert at 06:36:43Z, the Zammad core-team statement at 12:16:59Z, the core-team follow-up at 19:48:04Z. Nothing after 1 October, so the vendor has published no patch and no advisory for CVE-2026-102490 in the four days since. Both quoted spans were string-checked against the raw JSON and both are literally present: “current Zammad versions are not affected” and “reached end of support”, the latter in “These versions reached end of support some time ago.” Also confirmed verbatim: “Exploitation is only possible on Zammad 6.5 and older”, “we are working on it”, “This issue cannot be exploited remotely on its own”, and “Update to Zammad 7.2.0, the current stable release.” One span available but not used above, offered to the shipper as the tighter version of the same claim: “Zammad 7.0 and later are not affected.” This thread is also one of CISA’s own cited notes for both CVEs, which is worth knowing and is not a second source.

NVD and DIVD, from the banked read of 5 October and NOT RE-QUERIED on this run — labelled so nobody mistakes it for a same-day stamp: DIVD is the CNA (csirt@divd.nl), 102489 published 2026-09-30T17:16:40Z, status Analyzed, DIVD CVSS 4.0 of 9.4 against NIST CVSS 3.1 of 9.8. Report the disagreement, not the louder number. csirt.divd.nl/CVE-2026-102489 redirected to 433 B and was NOT READ; still NOT READ.

The Register, DIVD helpdesk breach. RE-FETCHED 5 October 2026 — CONFIRMED, with two precision notes. Still dated 1 October 2026 22:26 UTC (banked datePublished 2026-10-01T21:26:42Z; the on-page stamp reads 22:26, which is the same hour rendered in the outlet’s own timezone handling and is not a change of record). Both DIVD quotations re-read and present, including “the modus operandi indicates that this is an agentic AI powered attack”. The hedge is present as “an agentic operation or at least AI-enabled” and is The Register’s, not this house’s softening. Precision notes for the copy pass, neither a factual error: (a) the second quotation continues past where section 04 cuts it — the full span is “we could see the agent working automated, because after every action it decided the next step itself, at the speed of light and sloppy logic or pattern.” A quotation that stops mid-sentence needs an ellipsis or needs to run to the end; it currently takes a full stop that the source does not have. (b) 21 and 22 September read exactly as drafted (attack, detection). On 24 September the article has DIVD reporting to the vendor, notifying Dutch authorities and consulting police; “the Dutch DPA and NCSC” is a more specific naming than this article carries and should be attributed to DIVD’s own report, which was NOT FETCHED, or softened to “Dutch authorities”. The no-patch-at-exploitation claim rests on the article’s description of the attack as using zero-day exploits rather than on an explicit no-patch sentence — true as written, and that is the sentence it rests on.

FedScoop, TMF awards. RE-FETCHED 5 October 2026 — CONFIRMED. Still dated 2 October 2026 (banked datePublished 2026-10-02T20:54:38Z). Headline “New TMF investments: $83M for agentic AI, fast environmental reviews and more”. $83.4 million across four investments and State’s $17.3 million both hold. The human-review span reads verbatim “A person will review every automated decision and minimum risk management practices will be followed”; the agents span is that the investment will “allow employees to build their own automated agents”. The hours figure reads “more than 100,000 hours a year once fully operational” and is a projection, which is the label section 03 already carries. No vendor is named anywhere in the article — re-confirmed today; it references “commercial products” and “cloud-based platforms” and identifies no company. tmf.cio.gov was NOT FETCHED on this run either, and no award document has been opened. Both screens stand.

arXiv:2610.00980 and arXiv:2610.02036. BOTH ABSTRACT PAGES RE-FETCHED 5 October 2026 — CONFIRMED. Neither PDF was opened for this issue, then or now.

2610.00980, “Can AI Scientists Coordinate at Runtime?”, [Submitted on 1 Oct 2026], no revision. Every limit section 05 prints is the paper’s own word: “a single-seed exploratory evaluation across Agent Laboratory, EvoScientist, and ARK on ResearchClawBench” — three hosts, one benchmark, and the abstract states no intervals and no replication, which is why the issue says so. The finding reads verbatim: “Runtime selection yields the highest observed mean score for each host; adding contracts and verification reduces these means, with host-dependent outcomes relative to native execution.” “Host-dependent outcomes relative to native execution” is a span the shipper may want, because it is the paper hedging its own direction harder than our sentence does.

2610.02036, submitted 1 Oct 2026 16:49:40 UTC, v1 only. The design claim is verbatim: “models propose; the harness owns shared state and governs commit.” PRECISION NOTE, and the shipper should fix it in section 05 — the ladder has three rungs, not two, and the draft has mislabelled the middle one. The abstract reads: “ordinary teams exceed a shared budget in 5/5 runs, a visible live count leaves 4/5 violations, and commit enforcement leaves 0/5.” The 4-of-5 condition is a visible live count of the budget, not “telling a team it had a shared budget” — which is nearer the 5-of-5 baseline. The correction strengthens the item: visibility was the thing that failed, and that is the issue’s own “visibility is not control” line arriving with the right number under it.

Both are preprints, neither peer reviewed, and every number in both is a benchmark result on the authors’ own harness rather than production telemetry. 2610.02036’s benchmarks (TeamBench, the controlled revision benchmark) are constructed by the authors with small-n counts, as drafted.

OpenAI’s response to the Robinson resignation, syndicating Business Insider, byline Lauren Edmonds. RE-FETCHED 5 October 2026 — CONFIRMED. Still dated 3 October 2026 (banked datePublished 2026-10-03T19:01:53Z). The spokesperson span reads verbatim “We pause training or hold back models when we need to slow down” — note the capitalised “We”, so the version above needs either a capital or a lead-in that makes the lower-case correct. The frame “OpenAI said on Oct 3” is right and “yesterday” or “this week” is still wrong; no new OpenAI statement has appeared in the 4–5 October window. The PR-firm fact is now better-sourced than the draft implies: it is Robinson’s own written statement, quoted here as “I hired a PR firm, Spitfire Strategies, to help me with the public attention I’ll receive following my departure.” Section 07 may name the firm on his own words; it is not required to.

ASR-SOURCED AND NOT AUDIO-VERIFIED — the two open obligations before ship. (a) The All-In episode at 0:05:20 (published 2026-10-02, auto-generated YouTube captions; the extract records the recognizer garbling proper nouns in the same segment). The Sacks material in instrument two is paraphrase pending that check. (b) The Intelligent Innovation Show episode 5, recorded 3 October 2026, Google Meet auto-transcription, local copy at state/tiis-2026-10-03-raw.txt, recording on the shared drive. The three lines in section 06 are paraphrase pending a listen-check at ~00:34, ~00:58 and ~01:00. Neither check closed by ship. The three TIIS lines in section 06 and the Sacks material in instrument two ship as paraphrase, exactly as attributed, and nothing is lost.

ASR obligations: UNCHANGED BY THIS RUN. Neither listen-check carries a URL, so neither was touched by the 5 October re-fetch. Both remain open and both items remain paraphrase. This verification run closes the source-rot debt and does not close the audio debt.

NOT CHECKED, which is not the same as absent, and not re-tested today. Eurostat, the ONS, the OECD, the ECB, the BIS and Epoch AI — fourth week running. The Census BTOS national file was not re-checked. The standing bot-wall list in docs/research-standards.md was not re-probed on this run, so no host moved on or off it: congress.gov, theatlantic.com, reuters.com, wsj.com and openai.com stay bot-walled or paywalled from this host, bls.gov stays HTTP 403, and bughunters.google.com stays a JS shell at 200. Per the 10-01 ruling a 403 from a newsroom or a federal agency can be a user-agent artifact rather than a wall — none of these was retried with a browser UA today, so each stays NOT VERIFIABLE FROM THIS HOST and none becomes an absence. The WH Accord primary is paywalled and was not read. Zero PDFs were opened for this issue, including the Kyoto Vision declaration that is now known to be linked.

Re-fetch run 2026-10-05. Eleven URLs fetched live: nine CONFIRMED, one CHANGED (Kyoto Vision release), one CONFIRMED with an upgraded label (TechCrunch). Zero UNREACHABLE. No source went dark and no quoted span failed. Three BODY-IMPACT items are listed at the top of this block and are the shipper’s calls, not this seat’s. Claims · Woodward seat.