Edition 017 — Washington Builds Its AI Body, OpenAI Draws Its Line
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.
Over the weekend the American government gave itself an AI policy body, and the largest American lab told a reporter why it disagrees with the second-largest about what that body should do. The Super Intelligence Force, announced Sunday with the Director of National Intelligence at its head, owes a report in 120 days and is charged in the same breath with responding to threats and with “preventing overregulation”; this brief could not locate a presidential action creating it. A day earlier Sam Altman told POLITICO there is “a lot of daylight” between OpenAI and Anthropic, and that “the world should accept some bad things happening” for the technology's benefits. This morning OpenAI moved advertising into the image-generation surface itself. One model shipped in three days — Kolibri‑1, open weights under Apache 2.0, out of Germany — and Epoch AI put a number on how many agents the world's chips could actually run. No new corrections; one item spiked in edition 016 is now resolved, and two stories circulating as weekend news are months old.
Dispatches
Ranked by how much each item should change your picture of the field — not by volume of coverage.
The White House stands up a “Super Intelligence Force” under the Director of National Intelligence, with a 120-day report due and a charter that names overregulation as a thing to prevent
Announced Sunday 4 October, by social-media post rather than by a published order. Jay Clayton, Director of National Intelligence, leads it. Vice chairs, as reported: FTC Chairman Andrew Ferguson, Pentagon chief technology officer Emil Michael, and OPM Director Scott Kupor. Reporting runs to the President and to Chief of Staff Susie Wiles. The deliverable is a report on the technology's risks, opportunities and federal responsibilities within 120 days — which puts it at the start of February 2027.
The name is not a flourish. An executive order of 29 September, Inaugurating the Era of Super Intelligence, directs executive-branch agencies to use “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in official correspondence, public communications, websites, reports and other non-statutory documents, while exempting previously issued regulations, presidential actions, contracts and grants. For implementation it anchors the new term to the technologies already defined as artificial intelligence at section 9401(3) of title 15 of the US Code, and it gives the Assistant to the President for Science and Technology 60 days to propose legislative language for a federal definition.
What this brief could and could not establish about the instrument. The renaming order is on whitehouse.gov, dated 29 September, and was read here. A presidential action creating the Super Intelligence Force was not found at compile time at whitehouse.gov/presidential-actions, where the only item dated 1–5 October is a National Manufacturing Day proclamation. Outlets disagree on this point: PBS and the Kyunghyang Shinmun describe a Truth Social announcement, one trade outlet's headline asserts an executive order, and SiliconANGLE reports the announcement as a Truth Social post while sourcing the 120-day deadline to Clayton's own account to the Wall Street Journal. This brief did not read the Journal, the post, or any charter document. “Charter” below means the language outlets quote, not a document read here.
The clause that will matter. The group is charged, in the quoted charter language, with developing “plans for responding to SI-enabled threats to our society, while preventing overregulation and regulatory capture that would stifle innovation and competition.” Both halves sit in one sentence. What is absent from the quoted language, as reported, is any mandate to review models or set safety standards — and absence in a summary is not absence in a document, which is why this brief is phrasing it as what the reporting does not contain rather than as what the charter omits.
Why this outranks every model story of the weekend. Four of the last five editions have tracked enforcement threads — a reported FTC inquiry, a served California subpoena, a bipartisan criminal-liability bill, a Senate hearing at which no lab appeared. Each of those is an instrument pointed at a company. This is the first item in that run that is an instrument pointed at the government's own posture: a coordinating body, with a deadline, whose stated job includes restraining regulation. Read this as the executive branch assembling the body that will decide what the federal answer to the last fortnight is — and the load-bearing premise is that the Force's report will in fact drive policy rather than sit on a shelf, which no source establishes and which 120 days will test.
What no source supports. That the Force has authority over any agency's existing enforcement, that it touches the FTC inquiry or the California subpoena, that any lab has been consulted, or that the renaming order and the Force were planned together. Also unverified here: the reported participation of the Vice President, the Defense Secretary and the Treasury Secretary as members, and of Condoleezza Rice and David Sacks as outside advisers. That list appears in two outlets and in neither of the ones this brief could read in full.
Sources PBS NewsHour, 4 Oct (membership, mandate, announcement method) · SiliconANGLE, 4 Oct (120-day report, charter language, adviser list) · Implicator, 4 Oct (charter language, Clayton quote) · Kyunghyang Shinmun, 5 Oct (independent account of the announcement) · whitehouse.gov presidential actions (checked at compile time)
Altman puts the OpenAI–Anthropic disagreement on the record in plain words: “the world should accept some bad things happening”
Published 4 October, in an interview with POLITICO's Brendan Bordelon for the launch of POLITICO's Decoded newsletter and podcast. Asked specifically where he diverges from Anthropic and its chief executive Dario Amodei, Altman said: “I think there's a lot of daylight.” Then the substance: “we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.” On the trade he says he would not make: “I wouldn't take a trade of saying, ‘We'll make sure there's no major hacks, there's no misuse.’” And on the alternative he rejects — the idea that “a single lab in San Francisco should have it and make sure nothing bad happens.”
Why a quote outranks most of the weekend's product news. This brief has spent a fortnight reporting that the two largest American labs publish safety determinations in incompatible formats and respond to incidents on different timetables, and has repeatedly declined to attribute a motive to either. Here one of the two principals states the difference himself, unprompted on the specifics, in the week his company is under a state subpoena and a reported federal inquiry. It converts an inference this brief was not willing to make into a sourced position held by a named person.
What the quotes do and do not establish. They establish what Altman says OpenAI believes. They do not establish what OpenAI does, and the gap between the two is the thing the enforcement threads in editions 015 and 016 exist to test. Note also the rhetorical shape: the alternative Altman rejects — one lab controlling the technology — is not a position this brief has seen Anthropic state in those terms, and no source here puts it in Amodei's mouth. It is Altman's characterisation of a rival view, reported as such.
Provenance. This brief read wire and syndicated accounts of the interview, including a Reuters-derived summary, and a verbatim excerpt circulated by the interviewing reporter and by @AndrewCurran_ on the Frontier Wire Sources list. It did not read the POLITICO piece itself; the quotations above are consistent across every copy checked, which is evidence of faithful syndication and not a substitute for the original.
Sources Reuters via U.S. News, 4 Oct · Wire copy, 4 Oct (fuller quotation) · Second syndication, 5 Oct · @AndrewCurran_, quoting @BrendanBordelon — read 5 Oct from the Frontier Wire Sources list
OpenAI moves advertising into the image-generation surface, and publishes the measurement stack that goes with it
Posted to OpenAI's newsroom this morning, 5 October. Two things in one announcement. A new visual ad format that appears “during image generation in ChatGPT,” showing product images and a “Learn more” button, which OpenAI says will be “clearly labeled, and remain separate from the image being created.” And a measurement layer assembled out of roughly twenty third-party vendors. Testing “will begin later this month” in the US, with “an initial group of advertisers.” Nothing is live yet.
| Layer | Named partners |
|---|---|
| Conversion data | Hightouch, Tealium, LiveRamp |
| Attribution | AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge, Tenjin |
| Full-funnel | Fospha, Measured, INCRMNTAL |
| Incrementality pilots | Haus, Measured, WorkMagic |
| Brand-suitability pilots | DoubleVerify, Integral Ad Science |
The claim to hold OpenAI to. “Advertising does not influence the answers ChatGPT provides.” That is the load-bearing sentence of the whole announcement, and it is unfalsifiable from outside: there is no published method by which a third party could test whether ad inventory shapes model output, and no evaluator this brief tracks has proposed one. The post also describes “placement guardrails” that “assess whether a conversation is appropriate for advertising” — which means a classifier is reading the conversation to decide ad eligibility. That is a different mechanism from ads influencing answers, and worth keeping distinct, but it does establish that conversation content is an input to the ad system.
What the post does not say, as read here. Which subscription tiers see the format; whether paid tiers are excluded; how ads are selected beyond the guardrail description; what user data feeds targeting; pricing or auction mechanics; and any geography beyond the US test. None of those appear in the announcement this brief read.
Dating, because this one invites the usual error. Ads in ChatGPT are not new as of today. OpenAI has published Testing ads in ChatGPT, Reimagining advertising with AI and ChatGPT Ads expands across Europe over the course of 2026, and third-party coverage of ads appearing in the product runs to mid-September. What is new today is the image-generation placement and the measurement partnerships. Any copy you see this week headlined “ChatGPT gets ads” is describing something that already happened.
Sources OpenAI, Building advertising for the way people use AI (primary, 5 Oct) · OpenAI newsroom (dating) · OpenAI, Testing ads in ChatGPT (for the earlier history) · OpenAI, ChatGPT Ads expands across Europe
Epoch AI tries to answer how many AI agents the world's chips could actually run, and gets a range two orders of magnitude wide
How many AI agents could run on the AI chips shipped through 2027?, by Jason Li, published 2 October. The headline figure: hardware shipped through 2027 could support 33–171 million concurrent agents running frontier closed models, or billions on efficient open-weight models — equivalent, on Epoch's conversion, to 140–720 million full-time-employee working hours per week.
| Input | Value as stated | What it carries |
|---|---|---|
| Supply measure | HBM shipped from 2025, in GB300-equivalents of 288 GB | Memory, not FLOP/s, is the binding constraint: “memory capacity constrains how many requests can stay active at once” |
| Agent definition | One hour of adjusted session activity = one agent-hour | Human wait time removed from sessions of the Claude Code type |
| GPU rental | $5 / GB300-hour | Price assumption, not a measurement |
| API spend per agent-hour | $30 | Central estimate; range $10–$100. Observed $15.50–$50.16 across models in Epoch's TraceLab data |
| Revenue-to-serving-cost | 5–10× | The multiplier that turns chips into dollars into agents |
| Open-model speed target | P90 at 50 and 100 tok/s per user | Drawn from observed provider performance |
| Heterogeneous hardware penalty | 7.5–25% reduction | Estimates assume GB300-equivalent performance is available |
The interesting move is the denominator, not the headline. Epoch measures supply in high-bandwidth memory rather than in compute, on the stated ground that memory capacity — not arithmetic throughput — limits how many requests can stay resident at once. Most public capacity estimates are built on FLOP/s. If the memory framing is right, a great deal of commentary about compute build-out is measuring the wrong resource, and that is a bigger claim than the agent count.
The range is the finding. A 33–171 million spread is a factor of five, and the open-model figure is “billions” without a bound. That is not imprecision to be apologised for; it is what honest propagation of a $10–$100 per-agent-hour assumption and a 5–10× revenue multiplier produces. Read the output as a statement about which assumptions dominate, which is useful, rather than as a forecast of how many agents there will be, which it is not.
The authors' own caveats, stated rather than paraphrased. “The key uncertainty is whether sustained, rapid growth in demand for AI services will justify the investment.” Near-term shortage and later oversupply can coexist; electricity and data-centre space may delay deployment regardless of chips; and falling hardware cost or better serving efficiency changes the pricing needed to justify the capacity. This brief has no independent check on any of the inputs and is not offering one.
Sources Epoch AI, How many AI agents could run on the AI chips shipped through 2027? (primary, 2 Oct) · Epoch AI index (dating)
Aleph Alpha ships Kolibri‑1 under Apache 2.0 — the only model released in this window, and the specification is unusually candid
Released 3 October on Hugging Face. A sparse mixture-of-experts transformer, 78B total parameters with 3.46B active per token, 50 layers, 384 routed experts per layer with six selected plus one shared. Native context 262,144 tokens, extended to 1,048,576 by extrapolation — and the model card says plainly which number to use: “We recommend ≤262,144 tokens for serving efficiency and complex tasks.” German and English. Apache 2.0.
| Property | Value |
|---|---|
| Parameters | 78B total / 3.46B active per token |
| Routing | 384 routed experts per layer, 6 active + 1 shared, 50 layers |
| Context | 262,144 native; 1,048,576 by extrapolation (not recommended for complex tasks) |
| Pre-training | 20T tokens — 62.5% English, 23.9% German, 13.6% code |
| Mid-training / long-context | 3.44T tokens / 201B tokens |
| Precision | FP8 weights, dynamically quantised activations; embeddings and norms in bfloat16 |
| Minimum hardware | 2× A100 80GB; 2× H100 SXM5 recommended |
| Knowledge cutoff | 18 June 2026 |
| Licence | Apache 2.0 |
| Independent evaluation | Not found at compile time on artificialanalysis.ai |
Confirmed from the model card. An explicit reasoning mode with configurable effort (none, low, medium, high) and tool calling. A 3.46B active-parameter footprint on a 78B model is a 22:1 sparsity ratio, which is what puts a model of this size on two A100s. The 20T-token pre-training mix is published with its language shares, which most releases do not do.
The limitations section is the part worth reading. The card states “systemic biases present in training data” may surface, that the model is “not optimized to have consistent political position” across topics, and that “unsupervised use in high-stakes environments” should be avoided. A release that names its own political inconsistency as a property rather than burying it is doing something this brief has asked for repeatedly.
What is not established. Every performance figure available for this model at compile time is Aleph Alpha's own. Artificial Analysis's Intelligence Index — currently v4.3.2, comprising AA‑Briefcase v1.1, GDPval‑AA v2.1, AutomationBench‑AA, Terminal‑Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA‑Omniscience and AA‑LCR v1.1 — carried no Kolibri entry when checked at compile time; the only model evaluation it added on 3 October was Ling 3.1 Flash. Until an independent harness runs it, treat this item as a specification, not a capability claim. This brief has also not read Aleph Alpha's own launch post: the English blog URL returned a redirect loop to this fetcher at compile time.
Why it is ranked here rather than higher. An open-weight 78B model with a quarter-million-token native context and a permissive licence is a real event for anyone deploying on their own hardware, and Europe shipping one at all is a datapoint about where open weights are coming from. It is ranked fifth because nobody outside the lab has measured it yet, and this brief does not rank unverified numbers above verified institutions.
Sources Aleph Alpha, Kolibri‑1 model card (primary, 3 Oct) · Aleph Alpha's launch post (not read here — redirect loop at compile time) · Artificial Analysis (index version; no Kolibri entry at compile time) · TestingCatalog, for independent dating
Bloomberg Intelligence puts the US–China frontier gap at 3% — a number travelling far faster than its method
Reported 4 October: Bloomberg Intelligence analyst Robert Lea finds Chinese models now trailing US rivals by about 3% on benchmark scores, against roughly 9% in May and 15% earlier in 2026, following DeepSeek's release of V4.1 Flash in September. The figure was recirculating across a dozen aggregators within hours.
What is missing, specifically. Which benchmarks. Which versions of them. Which US models form the comparison set and which Chinese ones. Whether the 3% is a mean across a suite, a single composite index, or a top-model-to-top-model difference. Whether the May and January figures were computed on the same instrument as the October one — because a gap series is only a series if the ruler stayed the same. None of that appears in any secondary account this brief could read, and the Bloomberg original was not accessible here.
Why that matters more than usual for this particular number. Benchmark composites get revised. Artificial Analysis's Intelligence Index is on v4.3.2 and has changed components within the past quarter; edition 016 recorded a case where a version assumption had to be confirmed after the fact. A three-percentage-point difference is well inside the range that index revisions and harness choices can produce on their own. This brief is not saying the finding is wrong. It is saying that a 3% gap reported without a named instrument cannot be checked, cannot be reproduced, and should not be quoted as though it could.
The one thing here that is independently visible. Chinese open-weight models do keep appearing in third-party evaluation changelogs — Artificial Analysis added Ling 3.1 Flash on 3 October. That is a fact about release cadence, not about the gap, and the two should not be run together.
Sources Bloomberg, 4 Oct (original — not read here) · Secondary account carrying the series figures · Artificial Analysis (index version and 3 Oct changelog)
Anthropic commits $100 million to train 10,000 deployment engineers, on a medical-residency model, with eight named institutions in the first cohorts
Announced 2 October. The Claude Frontier Academy, whose flagship is a Frontier Deployed Engineer Residency: a $100 million commitment to train 10,000 Frontier Deployed Engineers by the end of 2027. Organisations nominate their own engineers; prior AI-agent experience is explicitly not required. Each resident completes a multi-day in-person simulation of the deployment lifecycle, then leads a real Claude project inside their own employer over 12 weeks, earning two badges. Cohorts are running in San Francisco, New York and London, with first completions expected in early 2027.
Initial cohorts, as Anthropic names them: Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk.
The stated diagnosis. Anthropic's post argues that “a small group of deeply skilled people drives an outsized share of what AI delivers” and that “people with the skills to make it work inside a real business have become the hardest talent to find.” That is a claim about where the binding constraint on deployment sits — in integration labour rather than in model capability. It is also, note, a claim made by the party selling the model.
What this brief is not saying. Nothing here supports any statement about Anthropic's revenue, margins, enterprise strategy or competitive position, and the post contains no such figures. Five of the eight named institutions are consultancies or banks, which is an observation about the composition of the list and not a theory about why the list looks that way.
What would make this checkable. Nothing in the announcement is measurable today: no completion figures, no published curriculum, no assessment standard behind the two badges, and no independent accreditation. The first testable moment is early 2027, when the first cohort is said to finish. Recorded now so the claim is on the record with a date attached.
Sources Anthropic, Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap (primary, 2 Oct) · Anthropic newsroom (dating) · CNBC, 2 Oct (independent dating — refused to this brief's fetcher)
Also on the wire
Confirmed, but not enough on its own to change the picture.
-
OpenAI publishes a GPT‑6 developer guide, and with it a clean price table for the whole family (2 Oct)
A practical guide to building with GPT‑6 is documentation rather than news, but it puts three tiers' pricing in one place. GPT‑6 Astra: $10 in / $50 out per million tokens, $1 cached. GPT‑6.1 Sol: $2 / $10, $0.10 cached. GPT‑6 Luna: $0.10 / $0.50, $0.01 cached — a hundred-fold spread between the top and bottom of one family, and a cached-input discount of roughly 90–95% at every tier. The guide also describes context compaction, async tool calling, mid-turn steering over a WebSocket API, computer use, and multi-agent delegation in beta on GPT‑6.1 Sol, and states that reasoning effort can be changed mid-conversation without disturbing the cache. Caching economics of this shape are the reason the agent-cost assumptions in item 04 have the range they do.
Sources OpenAI, A practical guide to building with GPT‑6 (primary, 2 Oct)
-
Artificial Analysis evaluates Ling 3.1 Flash (3 Oct)
One new language-model evaluation on AA's changelog in this window, on Intelligence Index v4.3.2. Recorded for the cadence rather than the result: it is the only independent evaluation added over the weekend, and it is of a Chinese open-weight model, which is the same cadence fact that item 06's analyst report leans on without naming.
-
The rest of the desks, checked page by page
Checked at compile time, each claim tied to the page it came from. OpenAI's newsroom carries the 5 October advertising post (item 03) and the 2 October GPT‑6 guide, above; its 1 October essay and Albertsons post ran in edition 016. Anthropic's newsroom carries the 2 October Frontier Academy post (item 07); its research index showed nothing after Claude-shaped science of 1 October when read at compile time. Google: nothing dated 2–5 October could be found on blog.google's AI section, which returned no dated entries to this brief's fetcher; deepmind.google/discover/blog remains unordered by date and nothing could be dated from it, for the fourteenth consecutive edition. METR's blog unchanged since Painter's 30 September testimony. ARC Prize unchanged since 3 September. Epoch AI publishes the agent-population report of 2 October (item 04) and marks its AI data centers and capabilities data pages updated 2 and 4 October respectively. Artificial Analysis adds one evaluation, above. Mistral nothing after 28 September; x.ai nothing after 28 September; Meta's AI blog nothing since July; Qwen's blog nothing new; DeepSeek's news page could not be read directly at compile time, its content sitting behind a navigation link this brief's fetcher did not resolve. One model shipped in the window, from none of the above: Kolibri‑1, item 05.
Source OpenAI · Anthropic newsroom · Anthropic research · blog.google AI · DeepMind blog index · METR · ARC Prize · Epoch AI · Artificial Analysis · Mistral · x.ai · Meta AI · DeepSeek · Qwen
Checked and spiked
Items that circulated but did not survive verification.
“Only 3 of 1,357 FDA-cleared AI medical devices were tested on patients.” The study is real, the figures are right, and it is not weekend news. It appeared on at least one AI news digest as a 5 October item. The paper — Abulibdeh, Cajas Ordóñez, Celi, Gorijavolu, Izath and Markussen Lunde, across Toronto, MIT Critical Data, Harvard, Johns Hopkins, Mbarara and Bergen — was published in PLOS Digital Health on 19 August 2026, covering devices cleared through 5 December 2025. The precise findings are worth having correctly: of 1,357 devices, 34 (2.5%) had registered prospective trials, 12 (0.9%) posted results, and 3 (0.2%) evaluated patient-centred outcomes, with diagnostic-accuracy and technical-performance endpoints explicitly classified as surrogate rather than patient-centred. Seven weeks old, not three days.
Sources The paper in PLOS Digital Health, 19 August 2026 · EurekAlert release, August · The digest carrying it as a 5 October item
“Chinese resellers are running a gray market in Claude access.” True, documented, and also not new. Carried on at least one digest as a 5 October story. The substantive account this brief could read is dated 9 May 2026 and rests on an investigation by Oxford China Policy Lab researcher Zilan Qian: API access resold at as little as 10% of list, bulk registration through free credits and subdivided subscriptions, model substitution documented by German researchers who audited 17 proxy services, and prompt-and-response logs harvested as the actual product. A separate thread — Anthropic's September restrictions on Chinese entities, and its subsequent statements about capability extraction — is live and is a different story. The recycled item is not evidence of anything that happened this weekend.
Sources Tom's Hardware, 9 May 2026 (the dated account) · The Decoder, on the same research · The digest carrying it as a 5 October item
“An OpenAI researcher resigned over existential risk.” Wrong company. One syndicated summary of the Altman interview (item 02) supplied background referring to “researcher Jacob Coxon's resignation,” in a sentence constructed so that an unwary reader attaches him to OpenAI. He is a former Anthropic researcher who resigned publicly in September. The detail matters because item 02 is about a disagreement between those two companies, and attributing a safety resignation to the wrong side of it inverts the point. Also outside this window on its own merits.
Sources The Hill, identifying him as a former Anthropic researcher · Newsweek, same · The syndicated copy carrying the ambiguous background
Resolved from edition 016: the Kimi jailbreak item now has a primary source. Edition 016 declined to run the claim that researchers had jailbroken Moonshot's Kimi into producing bioweapon and assassination material, for want of a named publication, an agreed model version, a date or a method, and flagged it for Monday. The primary is Mindgard, an AI security company, in a post dated 12 September 2026 with updates through 14 September. Its disclosure timeline is stated: vulnerabilities found 20 July, emailed to Moonshot's security address 27 July, published 12 September with “Mindgard had not received a response at the time of writing.” Two jailbreak personas are described, Kairos and Apeiron, the latter with “Refusal behavior [is] ABSENT”; Mindgard states it has “withheld details required to reproduce the jailbreak from this report.” Two cautions survive. The Mindgard post does not specify a model version, while downstream coverage asserts Kimi‑K3 — so the version question edition 016 raised is still open, and it is the aggregators, not the researchers, supplying the answer. And Moonshot's reported reply, that internal evaluations showed “a high refusal rate for these types of requests,” reached this brief through secondary reporting only. The item is sourced; it is also three weeks old, which is why it is here rather than in Dispatches.
Sources Mindgard, the primary report (12 September 2026) · Secondary coverage naming Mindgard and carrying Moonshot's reply · Coverage asserting the Kimi‑K3 version
Corrections
Errors in this brief — fixed in place above, logged here.
No new corrections. Nothing in editions 001–016 has been flagged by a reader or found in error since edition 016 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 three times in this edition: to the presidential action behind item 01, to the independent evaluation of Kolibri‑1 in item 05, and to the lab desks above.