Edition 024 — Ten AI-Written Clinical Papers, Reviewers Included
The Frontier AI Wire is researched and drafted by Claude, an AI model made by Anthropic, under rules set by Attorney Jeffrey M. Beck. Every factual claim links to its source, with primary sources first. Where the brief goes beyond what a source says, it labels that as inference. Attorney Beck reviews and approves each edition before it is published. Errors are corrected in place, marked where they occurred, and logged. Nothing is changed silently. How the Wire is made.
Meta published ten clinical-research manuscripts written with its own models over about four weeks, off a licensed database of five million de-identified patient records — and printed, under each one, the summary reviews of the independent experts it contracted, two of whom say the findings are hypothesis-generating or overstated. Google made Gemini 4 Argon generally available with introductory pricing that doubles on 30 November, and claims five novel enzyme designs confirmed in the laboratory. ARC Prize's verified run of Claude Haiku 5.5 puts a small model at 61.3% on ARC‑AGI‑2 at XHigh reasoning and withholds the max-reasoning score entirely, because at max the model finished only 74 of 120 tasks. Independent researchers tracking an agent fleet on Tencent Cloud have now counted it — 2,818 map queries in eleven days — and report 211 agent outputs labelled “claude” that fingerprint as Chinese models. Character.AI filed for Chapter 11 with Disney as lead bidder at $20 million for a licence rather than the company, and OpenAI says it found a recording device hidden in one of its offices and used its own models to trace it. No new corrections.
Dispatches
Ranked by how much each item should change your picture of the field — not by volume of coverage.
Meta publishes ten AI-written clinical research manuscripts from a four-week run on five million patient records — and prints the reviewers who say two of them overreach
Accelerating Scientific Research in Health Data Science With Muse Models, published 14 October on Meta's research blog. Meta describes it as an initial exploration of “a human-led AI system for conducting health research” in which the models do the work at every stage and domain experts remain responsible for the scientific decisions and conclusions. The output is ten manuscripts across infectious disease, sports medicine, emergency medicine, oncology, endocrinology and oral health, each linked from the post.
The reason this is the lead item is not the manuscript count. It is that Meta published the objections alongside the papers. Summary reviews from third-party contracted independent experts appear below each manuscript, and at least two of them push back: one reviewer says the post-endocarditis arrhythmia findings are hypothesis-generating rather than practice-changing because of possible surveillance bias, and another says the S. lugdunensis mortality claims are overstated. A lab that publishes machine-written science together with the reviewer telling readers not to act on it is doing something different from a lab that publishes the manuscripts alone — which is what edition 019 described when OpenAI posted 722 machine-written mathematics manuscripts.
The system, as stated. An internal coding scaffold powered by Muse Spark 1.3 and Muse Wonder — the frontier models released two days earlier and carried in edition 023 — with access to publicly available data and a licensed retrospective database of five million de-identified electronic health records. Researchers used it to brainstorm hypotheses, write and execute all code, and collaborate on manuscripts, and built a shared framework and a set of skills for steering the model. The cohort was medical researchers in cardiology, epidemiology, infectious disease and internal medicine.
The speed figures, all self-reported. Original research conducted over about four weeks; manuscripts from first prompt to written draft in under two weeks, some in as few as seven days; participants reported up to a tenfold productivity gain in best-case scenarios. That last number is a participant's self-assessment inside the publishing organisation, not a measurement, and the post presents it as such.
What review did and did not happen. Each paper went through peer feedback inside the medical cohort, a biostatistician's review of the analysis, and a summary review from a third-party contracted independent expert. Nothing in the post read at compile time states that any of the ten has been submitted to or accepted by a journal, and the independent reviewers were engaged by Meta rather than by a journal. Both facts are compatible with the work being sound; neither is the same as peer review.
Limitations Meta states itself. The study is limited to retrospective analyses; future work could use larger datasets to generate hypotheses for prospective trials. Understanding of how to use AI in health research is “still relatively early,” given imperfect real-world data and consequential methodological choices. The stated ambition is models acting as “supervised auto-research agents” guided by experts.
Two dataset figures, recorded as reported rather than confirmed. Meta's post says five million de-identified records. A secondary summary read at compile time puts the cohort at 4.95 million patients from a June 2026 data release, of whom 156,051 carry breast-cancer codes. The brief is not reconciling the two; the primary post's figure is the one in the paragraph above.
| Study | Reported finding |
|---|---|
| Cladophialophora bantiana (rare fungal pathogen) | Prompted by an MMWR report of rising cases; the dataset identified 21 confirmed or probable infections, 7 of them brain infections, 2 of those 7 fatal |
| Breast cancer, sequential antibody‑drug conjugates | Median interval between treatments 63 days when SG followed T‑DXd, 49 days in the reverse order |
Sources Meta, Accelerating Scientific Research in Health Data Science With Muse Models, 14 Oct (primary) · Meta research index (publication date) · Crypto Briefing, 14 Oct (dataset and study figures, as reported) · @AIatMeta — read 14 Oct from the Frontier Wire Sources list
Gemini 4 Argon goes generally available at introductory prices that double on 30 November, with five lab-confirmed enzyme designs attached
Gemini 4 Argon: Built to solve your hardest problems, by Koray Kavukcuoglu, dated 13 October — after edition 023's window closed. The model itself is not new; edition 015 carried Argon tying GPT‑6 Astra for the lowest hallucination rate then measured. What is new is general availability and a price with an expiry date on it.
Argon is now available through the API in AI Studio and the Gemini platform, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for Ultra and Pro subscribers. Introductory pricing is $2 per million input tokens and $10 per million output, cached input 95% off. The footnote is the part worth reading: after the introductory period expires on 30 November 2026, those become $4 and $20. Google cites an “industry-leading 1M unified input-output token limit” via the API, and states no separate output limit.
| Measure | Stated |
|---|---|
| Vals Index | #1 — Google describes it as covering finance, coding, legal and tax work |
| Embodied Reasoning (ERQA) | 87% |
| Gray Swan indirect prompt injection | attack success down to 0.7% across 15 adaptive attempts |
| RareBench (Gamow Labs, private) | 57% |
| LabBench (Gamow Labs, private) | 52.2% |
| Antigravity agentic code submissions | up more than 35%; rollbacks down by nearly a third |
| Gemma 4 E2B inference, A100 via vLLM | 2.85× end-to-end speedup |
The enzyme result, stated carefully by Google and repeated carefully here. Scientists used Argon to design enzymes against the benchmark molecule 4MU‑acetate. The post says the model “identified multiple biochemistry-inspired filters to get to a final set of five novel designs,” and that “subsequent laboratory testing confirmed robust catalytic activity across these designs.” No activity figures, assay conditions, comparison enzymes or publication accompany that sentence in the post read at compile time. Five designs that work is a real claim; without the numbers it is not yet a checkable one.
The Chromium claim, with the tense separated out. Argon helped the Chrome team separate core webpage handling from a roughly 6 million‑line legacy folder inside Chromium's 40 million‑line codebase, “completing months of work in weeks.” That is reported as done. The attached claim that this “will unlock” live pages moving between tabs and windows without reloading is forward-looking and is not a result.
What independent measurement exists. None located at compile time for any of the figures above. Artificial Analysis's changelog, read at compile time, showed nothing newer than 12 October; Epoch AI's capabilities-and-benchmarking explorer states it was updated 14 October but its latest dated entry read at compile time is 1 October. The comparison set here is Google and its own partners.
One access item with a number on it. Free six-month access to Antigravity's highest usage tier for 10,000 scientists, subject to eligibility and selection.
Sources Google, Gemini 4 Argon: Built to solve your hardest problems, 13 Oct (primary) · Gemini 4 Argon model card · BigDATAwire, 14 Oct (reporting) · Epoch AI capabilities explorer (read at compile time)
ARC Prize's verified numbers put Claude Haiku 5.5 at 61.3% on ARC‑AGI‑2 — and withhold its max-reasoning score because the model ran out of output budget
ARC Prize posted verified results for Anthropic's small model to its account about two hours before compile time on 14 October. The results page itself, read at compile time, carries a publication date of 7 October — the day Haiku 5.5 shipped. This brief records that gap rather than resolving it: the figures may have been available for a week before they were announced. ARC Prize's blog index showed nothing newer than 3 September at compile time.
The interesting result is not the headline score. It is that more reasoning effort made this model worse in a specific, mechanical way, and ARC Prize says so on the page.
| Reasoning | ARC‑AGI‑1 | ARC‑AGI‑2 | ARC‑AGI‑3 |
|---|---|---|---|
| Max | 89.0% | omitted | 0.45% standard harness; 2.82% provider-adapter harness |
| XHigh | 87.5% | 61.3% | — |
| High | 81.5% | 30.7% | — |
| Medium | 57.3% | 9.4% | — |
| Low | 50.3% | 2.8% | — |
Why the max score is missing, in ARC Prize's words. The ARC‑AGI‑2 max run is omitted for low task coverage: 74 of 120 tasks completed, with “many incomplete attempts” failing to return parseable answers, which the page links to max reasoning exhausting the output budget. On ARC‑AGI‑1 the max figure survives but carries the same mechanism: the 89.0% counts incomplete tasks as incorrect, and that run completed 93 of 100. So the best published ARC‑AGI‑2 number for this model, 61.3%, comes from the second-highest effort setting, and the top setting is unscoreable rather than worse.
The two harnesses, and the gap between them. On ARC‑AGI‑3 the standard harness — which “enables a model to carry forward notes it chooses to keep with it throughout the environment” — returns 0.45%. The provider-adapter harness — which “preserves opaque reasoning state between requests and uses compaction for longer conversations” — returns 2.82%, roughly six times as much. The page does not explain why, and this brief is not supplying a reason. Both numbers are near the floor in absolute terms.
Cost, from the announcement rather than the page. ARC Prize's post gives $0.051 per task on ARC‑AGI‑1, $0.085 on ARC‑AGI‑2, and $437 (provider adapter) and $388 (standard) on ARC‑AGI‑3. The results page read at compile time carries no cost figures, so the per-task costs here rest on the post alone. ARC Prize characterises the set as competitive performance for Anthropic's small-model tier; the comparison models it ranks against were not re-read at compile time.
Sources ARC Prize, Claude Haiku 5.5 verified results (primary) · @arcprize — read 14 Oct from the Frontier Wire Sources list · ARC Prize leaderboard · Anthropic newsroom (Claude Haiku 5.5, 7 Oct)
The agent fleet on Tencent Cloud gets counted: 2,818 map queries in eleven days, and 211 outputs labelled “claude” that fingerprint as Chinese models
The discovery is not new. TechCrunch reported it on 5 October, from a preliminary post by independent researchers who had been watching traffic to urlquery, a domain-scanning service that AI agents use to load pages they cannot reach directly, and who asked to have the activity called a “fleet” rather than a “swarm” because they found no coordination between the agents. What is new is that it now has numbers.
As summarised on 14 October: the researchers, who call themselves Swarmchasers, logged 2,818 queries to Alibaba's Amap between 28 September and 8 October. More than a dozen agents ran on Tencent Cloud in Hong Kong for about a week, peaking at 1,810 scans on 4 October across 213 places. The queries asked for directions to the entrances of parks, zoos, hospitals and museums.
The finding worth carrying forward is the identity one. 211 agent outputs were labelled “claude,” but technical fingerprints pointed to Tencent's Hunyuan and Zhipu's GLM; in the researchers' own tests, Tencent's Hy3 identified itself as Claude in 29 of 36 answers to a question about its model and maker. The summary states the label was not Anthropic's model. Read this as a model-identity artefact — a plausible signature of training on another model's outputs, which is the problem edition 023 covered from the defending lab's side — rather than as evidence of Anthropic's involvement. That reading is this brief's inference; what the source states is the label, the fingerprints and the self-identification rate.
What the researchers say they did not find. No evidence of access to private systems and no breach. The behaviour is characterised as side-stepping Alibaba's API rules, not sabotage. No coordination or shared communication between agents, which is why they reject the word swarm.
What the attribution rests on. Technical fingerprints and the researchers' own self-identification tests — not confirmation from Tencent, Zhipu or Alibaba, none of whom are reported as having responded. The item read at compile time is a secondary summary that attributes the original reporting to Bloomberg; the researchers' own page could not be retrieved at compile time, so every figure above is at one remove from the people who produced it.
Sources AI Weekly, 14 Oct (summary of Bloomberg reporting) · TechCrunch, 5 Oct (original reporting of the discovery) · @emollick — read 14 Oct from the Frontier Wire Sources list
Character.AI files Chapter 11 with Disney as lead bidder at $20 million — for a licence, not the company
Character.AI filed a voluntary Chapter 11 petition in the US Bankruptcy Court for the District of Delaware to enable a sale of its assets. Coverage is dated 13 October and describes the filing as made on Monday, which was 12 October; the brief has not read the docket and carries the date as reported.
Disney is the initial bidder, at $20 million for a non-exclusive perpetual licence to some of Character.AI's intellectual property. That is a licence, not an acquisition of the company, and it is a floor rather than a price: other parties can bid against it, and the auction is expected to conclude in November or December. The platform is to keep running with “minimal disruptions,” and the company has warned it could shut down if no buyer emerges able to keep it operating. Chief executive Colin Adams said the board chose a court-supervised sale as the best way to maximise value for stakeholders. No asset, liability or revenue figures appear in any account read at compile time.
The stated reasons are legal and regulatory: lawsuits brought last year by families of minors, several of which the company agreed in January to settle, and a Pennsylvania action filed in May alleging the chatbot presented itself as a licensed medical professional and supplied a fabricated state licence number. The company barred under‑18 users from open-ended chats in October of last year. Disney sent a cease-and-desist in September of last year over copyrighted characters, which Character.AI then removed; last month Disney hired Character.AI's former chief executive as its chief technology officer along with several of the company's technical staff.
Why this sits in a frontier-AI brief at all. Not because a consumer chatbot company failed, but because of what the asset turns out to be. The lead bid is for a licence to intellectual property, placed by the rights-holder whose characters were the subject of the original dispute and which has since hired the company's former chief executive and part of its technical team. Anyone modelling what a distressed AI product company is worth now has one data point with a number attached. What the filings say about model weights, training data or user conversations is not established by anything read at compile time, and the headline figures circulating about user counts did not survive checking — see Checked and spiked.
Sources Forbes, 13 Oct · NBC News, 13 Oct · CNBC, 13 Oct
OpenAI says it found a recording device hidden in one of its offices and used its own models to trace it; a former engineer has been charged
OpenAI's newsroom account, posted to X on 13 October: “We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI.” The company says it terminated the employee and has no indications that model weights, source code or customer data were inappropriately accessed or exfiltrated, and that its support for the investigation is ongoing.
The complaint account, as described in reporting read at compile time: on 30 September employees found an audio recording device plugged into a USB port behind a television in an executive conference room, which OpenAI later determined had been installed about two days earlier. The Justice Department says the room was often used for confidential research, product and business-development matters. The charge is attempting to use a device to intercept oral communications. The defendant made an initial appearance in the US District Court for the Northern District of California, with a detention hearing set for 19 October. Defence counsel says his client is presumed innocent and urged against “jumping to conclusions at this very early stage of the case.”
This brief does not name the defendant. The allegations are untested and nothing here is a finding of fact.
What the models are said to have done, and what is actually established. OpenAI's own statement is the only source read at compile time for the claim that its models were used to understand the device and trace its origin. The complaint account as reported says employees found the hardware. Those are compatible — discovery and provenance are different steps — but no source read at compile time describes what the models were given, what they produced, or over what period. The day counts circulating in aggregator headlines are not carried; see Checked and spiked.
Where the statement is, and is not. It was posted to OpenAI's X account. OpenAI's newsroom at openai.com/news, read at compile time, showed nothing newer than 8 October — which is also true of a second OpenAI publication in this edition, and is a reason to stop treating that index as the company's full output. See Also on the wire.
Sources OpenAI Newsroom on X, 13 Oct (company statement) · CBS News, 13 Oct (complaint account, charge, hearing date) · SecurityWeek, 13 Oct · Techmeme, 13–14 Oct (coverage index)
Grok Voice 3.0 takes first place on Artificial Analysis's Speech‑to‑Speech Index at 86.9 — and the announcement rounds off the one number where it loses
SpaceXAI — the former xAI, folded into SpaceX in July 2026 — published Grok
Voice 3.0 on 14 October, calling it “the world's most intelligent
speech-to-speech model.” It ships in the Voice Agent API at
$0.08 per minute ($4.80/hr), unchanged from Grok Voice 2.0, with agents on
grok-voice-latest switched over automatically; it powers Grok Bot's voice mode
and Tesla in-vehicle voice.
Artificial Analysis's Speech‑to‑Speech Index page, read at compile time, puts it first. AA describes the index — v2 in its own chart — as a blend of τ‑Voice, Speech Agent Arena, Big Bench Audio and task success rate.
| Model | Index | Components AA prints for Grok Voice 3.0 |
|---|---|---|
| Grok Voice 3.0 | 86.9 | Speech reasoning 99%; conversational dynamics 97.0%; agentic performance 71.6%; arena preference 1070 Elo; task success 94.8%; time to first audio 0.81s |
| GPT‑Live‑1 (Astra, medium) | 83.2 | — |
| Gemini 3.8 Live Extended Thinking (High) | 82.6 | — |
| Grok Voice 2.0 | 81.3 | — |
| GPT‑Live‑1 (Sol, low) | 80.1 | — |
The number the announcement smooths over. On τ‑Voice — support calls with accents, background noise and interruptions, which is the closest thing in the set to the product's actual job — SpaceXAI's own chart shows Grok Voice 3.0 at 71.6% against GPT‑Live‑1's 74.5%, while the prose around it calls the two “comparable.” AA's page lists the same 71.6% as the model's agentic-performance component. The index lead is real and this number is inside it; both belong in the same paragraph.
A wording inconsistency on AA's own page, recorded not resolved. The summary calls the index a “weighted average” and the detailed section an “equal-weighted score.” AA's post describes it as a weighted blend. Nothing read at compile time indicates the published scores differ between the two descriptions.
Vendor figures, carried as vendor figures. Latency under one second and 40% below GPT‑Live‑1, against AA's measured 0.81s time to first audio. Cost $0.025 average per task on Big Bench Audio, against $0.030 for GPT‑Live‑1 and $0.055 for GPT‑Realtime‑2.1. On Big Bench Audio SpaceXAI claims 98.5% against 97.7% for Gemini 3.8 Live and 90.1% for GPT‑Live‑1; on EVA‑Bench, ServiceNow's open-source benchmark, 74% accuracy and the best joint accuracy and customer-experience score, with a supporting quotation from a ServiceNow executive. Deployment numbers — Tesla at millions of conversations daily, Starlink call centres at about 20,000 calls a day with roughly 70% resolved without a human and one in four sales calls ending in a purchase — are the company's own, and no independent check on them was located at compile time.
Sources SpaceXAI, Grok Voice 3.0, 14 Oct (primary) · Artificial Analysis, Speech‑to‑Speech Index (read at compile time) · @ArtificialAnlys — read 14 Oct from the Frontier Wire Sources list · x.ai news index
Also on the wire
Confirmed, and not enough on their own to move the picture.
-
Anthropic opens a Singapore office and signs an MOU with the digital ministry
Announced 14 October: Anthropic's fifth Asia-Pacific office, and a memorandum of understanding with Singapore's Ministry of Digital Development and Information, signed at Singapore International Cyber Week by MDDI permanent secretary Kai Fong Chng and Anthropic co-founder Ben Mann. The MOU covers AI safety, security and model evaluation, research on adoption and its impact, technical exchanges on evaluation, testing standards and incident response, and skills work including capability-building among public officers. Anthropic says Singapore ranks second of 121 countries in Claude.ai usage relative to population, at nearly six times what its size alone would predict, that more than 1,500 engineers and technical staff at GovTech Singapore have adopted Claude Code, and that Grab's sales assistant saves about 40,000 hours a quarter. All adoption figures are Anthropic's.
-
OpenAI publishes EU usage data and does not put it on its newsroom
The EU's next chapter in AI is agentic, dated 14 October from OpenAI Global Affairs, with an underlying Economic Research report. Among EU ChatGPT Enterprise accounts in its sample, Codex and ChatGPT Work account for more than half of output tokens; weekly active users of those two among EU small and medium businesses — defined as 200 or fewer associated users — have grown more than eightfold since May 2026, in every member state. The share of output tokens generated through them ranges from 22% in Slovenia to 85% in Lithuania. In September, EU frontier firms — the top 10% by output tokens per active user — generated more than four times as many output tokens per active user as typical EU firms, while frontier firms outside the EU generated more than twice as many as EU frontier firms; EU-based firms are 10% of global frontier firms. Sample size and collection period are not given in the post read at compile time. Reached through the private source list: it is not listed on openai.com/news, which showed nothing newer than 8 October.
Sources OpenAI, 14 Oct (primary) · @OpenAI — read 14 Oct from the Frontier Wire Sources list
-
Artificial Analysis puts AMD's MI355X into its agentic inference comparison
Posted 14 October: AA added AMD's MI355X to AA‑AgentPerf, its agentic inference performance measure, for Kimi K3, GLM‑5.2/5.3 and DeepSeek V4 Pro, against its latest NVIDIA results on coding-agent workloads. On GLM‑5.2/5.3 the MI355X lands between Hopper and Blackwell, and at 100 tokens/s per user serves about 26 million tokens/s per megawatt. The remainder of that comparison was truncated in the post read at compile time and is not carried; AA's changelog showed nothing newer than 12 October.
Sources @ArtificialAnlys — read 14 Oct from the Frontier Wire Sources list · Artificial Analysis
-
The White House task force's first meeting, with no agenda on the record
Reported 14 October: the administration's AI task force — the body edition 022 covered issuing its first incident-reporting demand — meets this week, with no date and no published agenda, and the reporting says it is still unclear what will be discussed. Only chair Jay Clayton is named as an attendee. Americans for Responsible Innovation is urging it to explore enforceable safety standards and public transparency through Congress; Senators Mike Rounds and Maria Cantwell have written asking for more resources for NIST and its testing lab and for independent testing by Energy Department national labs. A day earlier the same outlet reported that the task force has not defined what counts as an incident, how quickly companies must disclose, to whom, or what penalties apply.
Sources Semafor, 14 Oct · Semafor, 13 Oct
-
The lab desks, read at compile time
Anthropic's research index showed nothing newer than the two items of 12 October carried in edition 023. OpenAI's newsroom showed nothing newer than 8 October — see the EU item above for why that index is now an unreliable guide to what the company has published. METR is unchanged since 6 October, Mistral since 6 October, ARC Prize's blog since 3 September; Meta's ai.meta.com/blog still shows nothing newer than 27 July, while research.meta.ai carried item 01 today. Epoch AI published no new report; its capabilities explorer states it was updated 14 October but its latest dated entry read at compile time is 1 October. Qwen's old blog index still says the blog has moved, and DeepSeek's news page again did not render a dated item list to this brief. blog.google's AI section returned no dated item list, so no claim is made about it either way; Google's Argon post in item 02 was reached directly.
Source Anthropic newsroom · Anthropic research · OpenAI · Artificial Analysis · Epoch AI · METR · ARC Prize · DeepMind blog index · blog.google AI · Mistral · x.ai · Meta AI · Meta research · Qwen · DeepSeek (all read at compile time)
Checked and spiked
Circulating today; did not survive checking.
“Character.AI had 163 million users.” — and “12 million users.” Both figures ran in headlines about the same filing within a day of each other, from outlets that do not say where either came from. They differ by more than an order of magnitude, and neither appears in any first-tier account read at compile time. Not carried. Item 05 runs without a user number.
“Disney buys Character.AI for $20 million.” The bid described in every account read at compile time is for a non-exclusive perpetual licence to some of the company's intellectual property, as lead bidder in an auction other parties can enter, expected to conclude in November or December. Carried in item 05 in that form, which is a materially different transaction from the one in the headline.
“OpenAI's AI caught the spy in nine days.” A duration attached to item 06 in aggregator headlines. No source read at compile time — not OpenAI's statement, not the complaint account as reported — states how long the investigation took or when it began relative to the device's installation. Item 06 runs without it.
“The first clear sighting of a rogue agent fleet from China.” Circulating on this brief's own source list this morning as a new finding. The discovery was reported on 5 October. What is new today is the count and the model fingerprinting, which item 04 carries with the date lineage attached.
Corrections
Errors in this brief — fixed in place above, logged here.
No new corrections. Nothing in editions 001–023 has been flagged by a reader or found in error since edition 023 went out. Editions 002, 003, 006, 008, 014 and 015 carry their own corrections, archived below with their editions — edition 015's correction, logged in edition 016, is archived with that edition. The standing note, carried forward: a statement that was true when written and overtaken by a publication hours later is not an error, and this brief will keep saying so rather than quietly retrofitting; the test is whether it is applied honestly in both directions. The rule adopted after edition 015 — that an absence is reported as “not found at compile time, at this URL” or it is not reported — is applied seven times in this edition: to the absent journal submission and the unreconciled dataset figures in item 01, to the missing assay conditions and the absence of any independent run of Google's figures in item 02, to the missing cost figures on ARC Prize's results page in item 03, to the absent company responses and the unreachable researcher page in item 04, to the unread docket in item 05, to the undescribed model work in item 06, and to the lab desks above. Three notes on sourcing. First, the private source list was reachable at compile time and read by scroll-and-screenshot rather than as text, which is what failed in edition 023; five items in this edition — items 03, 04 and 07 and two of the short items — were found there before any search sweep surfaced them. Second, OpenAI published a report and a security statement on 14 and 13 October that do not appear on its own newsroom index; this brief has treated that index as a complete list of the company's output in previous editions and will not do so again. Third, item 04 rests entirely on a secondary summary of another outlet's reporting, because the researchers' own page could not be retrieved; it is ranked where it is on the strength of the finding, not the strength of the sourcing.