What changed

How to read this log (the action view). The tiered list above is the provenance axis. The buckets below are the action axis. Tier and bucket are independent.

• 🆕 New (net-new this window). Two cards, both dated inside or at the edge of this run's own window: the CHMP highlights page is dated 18 September (four days into the 18-22 September window); the ICC catalogue read is dated to today, 22 September, because a live catalogue carries no single publication date of its own — the card says so explicitly and treats its own reading date as the freshness date, not an inferred one.
• 🔁 Changed (updates something we already track). Nothing edited on a standing card this run; no correction note was needed against prior cards. One long-parked item — ICC's 2027 I-Codes — moves from "unreadable" to "read, and partial" without becoming a full promotion, because the specific thing this cabinet wants (the IBC/IRC off-site-construction text) still isn't there.
• 🎯 Opportunities flagged (act / position / build / money).
  1. The biosimilar card is a bet on a pattern, not yet a confirmed one. Its own so_we_should names the test: does a second biosimilar cite the tailored clinical approach within twelve months? That is the cheapest, highest-value re-check to carry into future health-vein sweeps, and it is dated (18 September 2027) so it will not quietly become a permanent assumption.
  2. The ICC card converts a stale guess into a dated checkpoint. Before this run, the cabinet's best information was trade press saying "anticipated in September 2026." Now it is a first-party negative finding with eight days of September left when it was read. The next sweep's cheapest high-value check is the same URL again — codes.iccsafe.org/codes/i-codes/2027-icodes — watching only for the IBC, IFC and IWUIC to appear.
  3. L89 paid for itself a second time. Escalating to a rendering-capable client resolved Steam Frame on 18 September and resolves the ICC catalogue today. codes.iccsafe.org should be treated as JS-only by default on any future check, not re-attempted with plain-text fetch first.

How this run worked (the loop, once). Read the ledger (RDKG_GOAL.md, tasks.lessons.md to L91, the 18 September changelog, content/beats/frontier-standing-sweep.md and content/beats/mythkg-story-sweep.md) → opened the 18 September log's own "next run" list first, in order, since every item on it already had a URL or a near-exact target: the EMA CHMP plenary highlights (due, and productive), the Steam Frame promotion decision (deferred — see below), a fresh look at the Roblox roadmap (found, checked, no card), and ICC's catalogue with a rendering-capable client from the start (productive) → broadened into a WebSearch sweep across the remaining veins for the 18-22 September window (materials/building beyond ICC, AI/machine layer, agriculture/robotics, marketing, fashion, resale, crypto/agent-commerce) → opened every candidate primary from the main thread and quoted it directly, including a secondary/trade document (NAHB) used only as an attributed lead, never as a source_url → carded two survivors → ran a self-review pass reading each card's title, capsule and so_we_should cold against its own body (L45), which caught two real defects before publication (below) → ran node scripts/validate.mjs (ERRORS: 0, signals 141 → 143), node scripts/gen-llms.mjs (manifests regenerated to 143), and npx tsc --noEmit (0 errors) for build-readiness. The founder commits, builds and deploys; a card is not live until then.

Corrections made this run, on the record. Two, both against this run's own drafts, both caught in self-review before publication rather than by an independent adversarial pass (none was run this run, given the small card count).

1. The EMA card's draft title overreached what its own body said. The first draft's title read "The first biosimilar approved without a dedicated efficacy trial" — asserting as fact, outside quotation marks, exactly the thing the card's own "honest limit" paragraph says is unconfirmed: whether Pebrilzo's submission omitted a Comparative Efficacy Study entirely. This is L70's shape (the frame outran the quoted page) caught by L45's practice of reading the title cold against the body. Corrected to attribute the "first" claim explicitly to EMA's own words and to state plainly that what the submission contained isn't yet public.
2. The ICC card's draft provenance paragraph misdescribed what was actually read this session. A draft sentence said the ICC newsroom release's off-site-construction sentence was "fetched and quoted in a prior run's research and is restated here... not re-fetched" — false; that release was opened and quoted from the main thread in this same session, as the first fetch of this run. Corrected to state plainly that it was read this session, with its own dateline and modified-time metadata quoted, and used only as corroborating context rather than as this card's source_url.

Veins swept, and what each returned. Nine veins.

• Human health & longevity: the productive vein this run, one card, summarised above. The EMA CHMP plenary was the run's single highest-value fetch, flagged as due by the 18 September log.
• Materials & building: one card, summarised above, closing a parked item.
• AI & the machine layer: empty in-window. A WebSearch for dated frontier-model releases surfaced Claude Fable 5.1/Mythos 5.1, Gemini 3.8 Flash and Muse Spark 1.3 — all dated 1-2 September, well outside this 18-22 September window. An arXiv search returned no paper with a face-of-document submission date inside the window. No card.
• Agriculture & robotics: checked once. Greenfield Robotics' Regulation A+ offering (SEC Forms 253G2 and 1-A/A, StartEngine Primary) opened to investors on 11 September — outside this window — and stays parked, unpromoted, per the standing note carried since at least 11 September.
• Marketing: checked. No first-party, dated-inside-window ad-product or platform-economics filing surfaced; a Meta "Meta One" subscription tier and Creator Marketing Hub rollout is reported by aggregator blogs at 15 September, outside the window and never traced to a first-party Meta newsroom post — not carried further.
• Fashion: checked. SHEIN's HKEX code 00625 remains the 24 August/1 September story; no hkexnews.hk over-allotment or stabilisation announcement was located inside this window.
• Resale: checked. ThredUp's 14th Annual Resale Report is a real first-party document with real figures, but it is dated 2 April 2026 — five and a half months outside this window — and was not carded on that basis alone.
• Crypto & agent-commerce: lightly checked, still empty in-window. AP2 remains at v0.2.0 (April); x402 Foundation governance is unchanged since its 14 July 2026 launch; a secondary aggregator's "as of September 13, 2026" transaction-volume figures were not traced to a first-party dashboard or report and are not carried onto any card. The founder decision flagged on 15 and 18 September — whether to demote this vein to a monthly check — is restated here, unresolved, rather than guessed at.
• Storytelling & the new medium (MythKG): checked, no card. Roblox's "Creator Roadmap 2026: Fall Update" (devforum, 18 September, in-window) restates Scene Generation, NPC Dynamic Behavior and the Wallet rollout at their same "Late 2026" delivery windows already carried on roblox-rdc26-roadmap-prompt-npcs-web-play and roblox-wallet-replaces-devex-usd-payouts — none of the medium-relevant items appears on the post's own "Delayed Items" list, so nothing has changed about either card's central claim. Given this cabinet already holds seven dedicated Roblox cards (a grep of content/signals/roblox-*.md, re-run rather than recalled) and two prior runs' own honesty notes flagged Roblox concentration as a pattern to watch rather than chase, this run reads the Fall Update as a no-change confirmation and does not open an eighth card or an update note for a page that changed nothing about what either standing card claims.

Closed this run (no longer parked).

• ICC's 2027 I-Codes catalogue — parked since at least 15 September for being unreadable to plain-text fetch; read today with a rendering-capable client, per L89. The underlying question (are the flagship codes live) is answered "not yet," which is why the card is a closure of the reading problem and a dated checkpoint on the publication question, not a promotion to "fully resolved."
• The EMA CHMP September plenary — due since the 18 September log named it "now due"; fetched and carded.

Parked this run (flagged, not published — with the primary that would promote each).

• The IBC, IFC and IWUIC in ICC's Digital Codes catalogue — the single cheapest, highest-value check for the next sweep. Re-open codes.iccsafe.org/codes/i-codes/2027-icodes with a rendering client directly; do not re-attempt plain-text fetch first.
• A second biosimilar citing EMA's tailored clinical approach — the pattern-claim limb on this run's health card, not due until 18 September 2027, but worth a light check on any future EMA-vein sweep in the meantime.
• Steam Frame's promotion decision — flagged by the 18 September log as "a decision, not a search" (the body now reads cleanly; only the signal-vs-verified tier call is open). Not acted on this run for lack of run time, carried forward explicitly rather than silently.
• The Roblox "full roadmap" — found this run (the Fall Update), read, and judged to change nothing about the two standing cards it touches; not re-opened as a fresh search target.
• Merck/Moderna INTerpath-001 — carried unchanged since 15 September (settled at 19 August, out of window); check for an ESMO or other medical-congress presentation before carding.
• The Joenja prescribing information and Drugs@FDA — untouched five runs now; still tests the 11 September pediatric card's most likely failure point.
• DeepSeek's technical report, Hugging Face model card and pricing image — untouched three runs.
• The Highly Engaged Player threshold drop (250 to 100) — due in November 2026; not yet.
• Carried without re-attempt: the Fortnite/UEFN actor-consent falsifier limb (unchecked seven runs now); Gemini 3.8 Flash's 1 January 2027 price step; Broadcom's Q4 FY2026 AI print (early December); Meta Connect (23-24 September — happens the day after this run closes; check on the next sweep); SHEIN's over-allotment/stabilisation filings.

Leads recorded, not parked (no URL, no first-party source, or judged too thin to track). A secondary aggregator's x402 transaction-volume figures ("as of September 13, 2026"), not traced to a first-party dashboard. Aggregator coverage of a Meta "Meta One" subscription tier and Creator Marketing Hub rollout, dated 15 September by the aggregator and never traced to a first-party Meta newsroom post.

The TKG garden-code decision is still open. Unchanged from the 18 September log; no machine-layer card shipped this run either, so the affected-card count does not move. Carried thirteen runs now.

Founder decision still open, restated from 15 and 18 September: demote crypto & agent-commerce to a monthly check? Lightly re-checked this run (see veins above) rather than independently re-tested with a full search budget; the question stands as the 18 September log left it.

Honesty note. Two items, both verified in provenance terms — a small, all-verified run, which RDKG_GOAL.md has called a yellow flag since August. The honest read here is that this run did not manufacture a signal or hype card to balance the ratio; it closed two long-parked items cleanly and found the rest of the swept veins genuinely empty inside a four-day window, which the veins-swept section states plainly rather than papering over with a thin card. Where confidence is genuinely bounded, in prose: the EMA card's central "first" claim is EMA's own superlative, quoted and attributed, not independently verified against a full EPAR that does not yet exist; and the ICC card is explicit that its finding is a dated negative — codes not yet live — rather than a confirmed miss, since ICC's own catalogue text describes an open-ended rolling release schedule with days of September still remaining when it was read.

On the window and the cadence. Four days, 18-22 September, the Tuesday beat run as designed. Freshness: two of nine veins reached a primary dated inside or at the reading edge of 18-22 September (health, materials/building); the remaining seven returned genuinely empty windows rather than thin or aggregator-sourced cards. The next run (Friday, 25 September) should open with the IBC/IFC/IWUIC re-check on ICC's catalogue, the Steam Frame tier decision, and a fresh look at Meta Connect (23-24 September) and the EMA CHMP plenary's minutes, which this run's primary states will publish "in the coming weeks."

How to read this log (the action view). The tiered list above is the provenance axis. The buckets below are the action axis. Tier and bucket are independent.

• 🆕 New (net-new this window). Three new cards, two dated inside this run's own window (Fayuvi and Kerendia, both 17 September) and one a deliberate backfill dated to its own primary (pelacarsen, 4 September — four windows late, logged as such rather than silently absorbed). No item was carded on today's date alone; every card's date field is the primary's own.
• 🔁 Changed (updates something we already track). Three dated update notes, all appended per L64 with the underlying card's body and frontmatter left untouched: steam-frame-reservations-open (body now readable), venice-immersive-2026-best-of-worlds (winners named, falsifier closed), and olpasiran-lpa-lowering (its own falsifier half-fired by the pelacarsen card above). No card's trust_tier was changed by an update note this run — Steam Frame remains signal in its frontmatter even though this run's reading argues it is now a strong verified candidate, a distinction this log states rather than resolves quietly.
• 🎯 Opportunities flagged (act / position / build / money).
  1. The evidentiary ledger now holds six named bars across four cards, and is worth a standing tag rather than a running paragraph. fda-leniolisib-pediatric-apds-pk-bridge named four; this run's two FDA/pharma cards each add one (historical-control cognitive scores; randomized surrogate). The audit this log has proposed twice before — tag every approval card by what the agency actually accepted — now has six categories and four cards to sort into them, which is enough to build a real column rather than another paragraph.
  2. The Lp(a) biomarker thesis is now a two-molecule race with one runner down. MktKG and HKG both have a stake in how "biomarker lowers X%, causal role unproven" stories get told commercially; pelacarsen's miss is a live case study in how a company frames a primary-endpoint failure (Novartis's own quote: "not the results we hoped for," "advances scientific understanding") without conceding the biomarker's causal status. Worth a short note on the rhetoric alone, independent of the science.
  3. Steam Frame is a one-fetch promotion waiting for the next run's pre-flight pass. The card is signal only because a page that now reads cleanly did not render three times running. Promoting it costs nothing new to fetch — it is a decision, not a search.
  4. Platform-as-bank and evidentiary-bar threads both compound faster than new leads. Three of six items this run resolved or extended something already on the shelf rather than opening new territory — worth naming as a pattern: a mature cabinet's highest-value move is increasingly closing its own falsifiers, not finding new subjects.

How this run worked (the loop, once). Read the ledger (RDKG_GOAL.md, tasks.lessons.md to L87, the 15 September changelog, content/beats/frontier-standing-sweep.md and content/beats/mythkg-story-sweep.md) → worked the 15 September log's own "next run" list first, in order, by direct fetch from the main thread rather than search-only scouts, since every item on that list already had a URL or a near-exact target: the Merck/Moderna fetch (settled the date — 19 August, out of window, still parked), the Steam Frame body (rendered via a JavaScript-capable client, resolving the falsifier), ICC's Digital Codes catalogue (still JS-only and unreadable to this run's tools, parked again), and the Roblox "next week" roadmap (searched, not found; parked) → broadened into a WebSearch sweep across the remaining veins for the 15-18 September window → opened every candidate primary from the main thread and quoted it directly → found and closed the pelacarsen backfill along the way, after a search for the Merck item's neighbourhood surfaced it → carded three survivors and wrote three update notes → ran a single self-review pass checking capsule word counts, a superlative grep, quote-fidelity spot checks against the fetched source text, and a cross-reference dup check (fda.gov, AAV9, kerendia/finerenone) → fixed two provenance inconsistencies the self-review caught (both cards had claimed to quote two named officials whose statements were in fact only named, not reproduced — corrected to say so) → ran node scripts/validate.mjs (ERRORS: 0, signals 138 → 141), node scripts/gen-llms.mjs (manifests regenerated to 141), and npx tsc --noEmit (0 errors) for build-readiness. The founder commits, builds and deploys; a card is not live until then.

Veins swept, and what each returned. Nine veins, worked from the main thread rather than parallel scouts this run, given how much of the list was direct-fetch work rather than open search.

• Human health & longevity: the most productive vein this run, three cards, all summarised above. Also checked and not carded: the CHMP September plenary (correctly left alone per the 15 September log's own instruction — it sat 14-17 September and highlights are not expected until 21-25 September); the Joenja prescribing information (not fetched this run, remains the pediatric APDS card's most likely failure point); DeepSeek's technical report, Hugging Face model card and pricing image (not fetched).
• AI & the machine layer: empty in-window by design. A single-author arXiv preprint (2609.14803, "Another Blueprint In The Wall: How to Ask Frontier AI Like a Kid?", submitted 13 September) surfaced testing whether six frontier models converge on similar self-described architectures under a "kid" framing — its own submission date falls one day before this run's window, it has one author, and the paper itself uses the term "epistemic jailbreak" for the specificity-versus-provenance tradeoff it studies. Recorded as a lead, not carded: out of window and thin enough (a single-author preprint reporting a prompting-behavior pattern, not a peer-reviewed or first-party technical finding) that carding it would not meet this garden's bar even if it were in-window.
• Materials & building: parked again. Multiple secondary sources report the 2027 IBC, IFC and IWUIC were "anticipated in September 2026," but codes.iccsafe.org/codes and its I-Codes filter both return only page chrome and metadata to this run's tools — no code listing, no publication date. One search result asserted, unprompted and without a source, that "since today is September 18, 2026, the IBC should be available now or very shortly" — a search-summariser inference, not a fact, and not carried onto any card.
• Agriculture & robotics: not independently searched this run; the window's search budget went to the health vein's productivity and the standing-item closures instead. Carried forward, not marked empty.
• Marketing: searched (SEC 8-K filings, ad-product announcements); nothing first-party and dated inside 15-18 September surfaced. Meta ad-format test reports and TikTok Symphony/Seedance integration are aggregator-sourced and were not chased to a primary.
• Fashion: the CFDA's fur-free policy taking effect at September 2026 NYFW was searched; the policy's own announcement dates to September 2025, over a year before this window, so nothing new is carded. NYFW itself closed 15 September (Thom Browne); no CFDA recap was located.
• Resale: searched (ThredUp, Savers Value Village, Poshmark); both companies' most recent public results are from Q2 2026 reporting, not inside this window.
• Crypto & agent-commerce: not independently re-searched this run, per the 15 September log's flag that seven consecutive empty runs is a candidate for demotion to a monthly check — a founder decision still open, addressed below rather than re-tested here.
• Storytelling & the new medium (MythKG): two cards, both closures of standing parked items rather than new subjects, summarised above.

Corrections made this run, on the record. Two, both against this run's own drafts, both caught in the self-review pass rather than by an independent adversarial reviewer.

1. Both new FDA/pharma cards originally claimed to quote named officials whose statements were never reproduced. The Fayuvi card's provenance paragraph said Acting FDA Commissioner Kyle Diamantas and CBER's Karim Mikhail were "quoted as printed"; the Kerendia card made the same claim about Dr. Janet McGill and Dr. Carolina Aldworth. Neither card actually pastes their statements. Both paragraphs are corrected to say the officials are named on their respective pages but not reproduced here — the L66 shape (never claim to have sourced what the provenance paragraph did not actually carry), caught before publication rather than by a second pass.

Closed this run (no longer parked).

• Steam Frame's body — three runs of an unrenderable page, closed by switching tools rather than by finding a new document.
• Venice Immersive's winners — two runs parked for lack of a primary; the primary existed and was opened this run.
• The Merck/Moderna INTerpath-001 date — settled at 19 August 2026, unambiguously out of window on the primary's own face. Downgraded from "next run's top item" to an ordinary backfill candidate; not carded this run because nothing in-window promotes it (no ESMO or other medical-meeting presentation was found dated inside this window).

Parked this run (flagged, not published — with the primary that would promote each).

• ICC's 2027 I-Codes — codes.iccsafe.org/codes?category=I-Codes remains JS-rendered and unreadable to this run's fetch tools; a rendering client (as used for Steam Frame this run) is the next thing to try, or a direct search for a dated ICC newsroom post rather than the catalogue page itself.
• The Roblox "full roadmap" promised for "next week" after the 11 September RDC26 post — not found this run under any obvious devforum or newsroom URL. Re-check create.roblox.com/updates and the devforum Announcements category directly rather than by open search.
• Merck/Moderna INTerpath-001 — now a plain backfill (19 August), not urgent; check for an ESMO or other congress presentation before carding.
• The Joenja prescribing information — untouched three runs now; tests the 11 September pediatric card's most likely failure point.
• DeepSeek's technical report, Hugging Face model card and pricing image — untouched two runs; the pricing image is the one document that card's cost-story limb still names as unread.
• The Highly Engaged Player threshold drop (250 to 100) in November 2026 — the cheapest standing check queued from the 15 September RDC26 card, not yet due.
• Carried without re-attempt: Greenfield Robotics' Regulation A+ offering; the Fortnite/UEFN actor-consent falsifier limb (unchecked six runs now); Gemini 3.8 Flash's 1 January 2027 price step; Broadcom's Q4 FY2026 AI print against its $21.7bn guide (early December); Meta Connect (23-24 September, not yet); the EMA CHMP September plenary (deliberately left alone until 21-25 September, per the 15 September log's own instruction, honoured this run).

Leads recorded, not parked (no URL, no first-party source, or judged too thin to track). The single-author arXiv preprint on frontier-model architectural convergence under a "kid" framing, named above — out of window and, independent of the window, thin enough on its own that this garden would not card it even in-window without independent replication or peer review. Aggregator reports of Meta testing new Threads ad formats and of ByteDance's Seedance model rolling into TikTok Symphony — neither traced to a first-party, dated announcement.

The TKG garden-code decision is still open. No machine-layer card shipped this run, so the code count is unchanged from the 15 September log. Carried twelve runs now.

Founder decision still open, restated from 15 September: demote crypto & agent-commerce to a monthly check? Not independently re-tested this run (see veins above); the question stands as the 15 September log left it.

Honesty note. Six items, all verified in provenance terms except Steam Frame's update, which inherits its underlying card's signal tier by design (see the "Changed" bucket above) rather than by any defect in what this run itself read. Where confidence is genuinely bounded, in prose: the pelacarsen card is a topline announcement with no effect size, hazard ratio, or subgroup result yet public, and this card says plainly that the full congress presentation could qualify its headline; the Fayuvi card's central interpretive move — that a historical-control comparison is evidentiarily weaker than a randomized one — is this card's own reading, stated as such, not the FDA's characterization of its own rigor; and the Kerendia card is explicit that "expected to reduce" is not "shown to reduce," a distinction the sixth evidentiary bar exists to preserve. No card this run states a number it did not compute from a quoted figure or copy directly from a primary.

On the window and the cadence. Three days, 15-18 September, the Friday beat run as designed, one day inside RDKG_GOAL.md's twice-weekly cadence. Freshness: two of nine veins reached a primary dated inside 15-18 September (health, twice, both 17 September); the third health card and both MythKG cards are dated backfills or event-closures rather than in-window discoveries, and each states its own date plainly rather than borrowing today's. The next run (Tuesday, 22 September) should open with the EMA CHMP plenary highlights (now due), the Steam Frame promotion decision, and a fresh look at the Roblox roadmap and ICC's catalogue with a rendering-capable fetch from the start rather than as a fallback.

How to read this log (the action view). The tiered list above is the provenance axis. The buckets below are the action axis. Tier and bucket are independent.

• 🆕 New (net-new this window). Four of five carry a primary dated inside 11–15 September — 11 September (Joenja), 11 September (RDC26), 11 September (Wallet), 14 September (Steam Frame). The fifth, the DeepSeek card, is the odd one and is the reason it is signal: its content is new to the window and its document's only date is 10 September, four days before the event it describes. No backfills. One item from the parked list was promoted (Steam Frame) and one was closed by verification rather than promotion (the DeepSeek cutover).
• 🔁 Changed (updates something we already track). All five touch standing cards, and one of them corrects a standing card outright. A dated correction note is appended to deepseek-v4-1-flash-routes-v4-pro-traffic recording that its retirement claim is withdrawn, that its cost-story limb does not fire, that its facts limb is untested as to the URL it actually cites, and that its superiority limb remains open. Its body is unedited; its title now carries a bracketed correction marker, because a reader of /signals would otherwise meet only the withdrawn claim, and its capsule is untouched because it sits at the 80-word ceiling. Joenja joins fda-approves-zilganersen-alexander-disease (eight days older), fda-sevabertinib-first-line-her2-nsclc and fda-approves-daraxonrasib-pancreatic-adenocarcinoma as a fourth distinct evidentiary bar on that shelf, and is the seventh fda.gov drug-approval card. RDC26 and Wallet both read against roblox-creator-earnings-2025-median-payout, roblox-funded-discounts-in-game-items, roblox-build-prompt-to-publish-alpha and roblox-animation-packs-creator-authored-locomotion; the Wallet card is the one that complicates rather than flatters, because it overtakes a card built on an SEC filing with a card built on a forum post, and says so.
• 🎯 Opportunities flagged (act / position / build / money).
  1. The model-string audit survives its own conclusion and is now worth more. The 11 September log said to check that our model strings are strings before the 14 September cutover. The identifier did not move — because a vendor reversed a published deprecation without a versioned notice. The exposure is not a cutover; it is a cutover that can be announced, withdrawn and reinstated on a page that is edited rather than versioned. The cheap standing practice is to read the changelog and the pricing page, not the launch post. An hour of reading config; it does not touch the BKG P0.
  2. Write up the platform-as-bank primitive for MktKG. A marketplace operator that pays sellers in a currency it does not issue, holds their balance, and offers them a preferential rate to spend back on-platform is a mechanism worth a note for any garden that ever runs a marketplace. It sits beside the operator-funded-discount note the 11 September log already proposed, and the two are the same question — who sets the terms — asked at two different layers.
  3. Run the evidentiary audit. It is now the third run it has been proposed and the material has grown to seven cards. Tag each FDA card by what the agency accepted: outcome, surrogate, or bridge. The Joenja card enumerates the four bars with the other three quoted; the audit is turning those four into a column.
  4. One fetch closes the weakest card in this run. The Steam Frame post's body, rendered by any client that runs JavaScript, either confirms or destroys the reservations reading. It is the cheapest high-value item on the next run's list.

How this run worked (the loop, once). Read the ledger (RDKG_GOAL.md, the lesson ledger to L84, the 11 September changelog, content/beats/frontier-standing-sweep.md and content/beats/mythkg-story-sweep.md) → three parallel search-only scouts with caps of 14, 14 and 16, all three at cap → main-thread verification with ten document fetches, of which seven returned readable content and three did not — the Valve store post returned a metadata block and no body, and the Steam Community mirror and the Steam Frame product page returned navigation chrome only → deep-verified every keeper by opening the primary from the main thread and quoting it → carded five survivors → appended a correction note to a standing card → ran the standing mechanical checks (capsule word count on all five, the superlative grep on every summary field, and the cabinet-scoped grep per claim, which caught a false roadmap grep list before review — and whose replacement was itself incomplete until the verification pass caught it) → two independent adversarial passes run in parallel, one on discipline and one on a transcribed fact sheet, returning 66 and 61 findings → fixed the union → validated the corpus to zero errors. The founder commits, builds and deploys; a card is not live until then.

Veins swept, and what each returned. Nine veins across three scouts plus main-thread work.

• AI & the machine layer: one card, and it is a negative result about a positive claim. Frontier model releases: empty in-window — every dated release the scout surfaced (Fable 5.1 / Mythos 5.1, Gemini 3.8 Flash, GPT-6 Astra's alignment update) falls outside 11–15 September and every date came from a model-tracker aggregator. AI regulation primaries: empty — the Federal Register was searched directly for 14 September and returned nothing. arXiv: empty, and one likely fabrication caught — a search summariser asserted a 14 September submission by three named Tsinghua authors on routing agent skills through a frozen LM's hidden states and returned no URL at all; every arXiv URL actually returned sits in a different ID block. A confidently narrated claim with named authors and no link was not chased, and is recorded here so the non-verification is on the record.
• Health & longevity: one card. Two named carry-overs resolved and neither became a card. Merck + Moderna INTerpath-001: both first-party URLs were located (news.modernatx.com/merck-and-moderna-announce-phase-3-interpath-001-trial-of-intismeran-plus-keytruda-met-endpoints-of-rfs-and-dmfs-in-melanoma and a longer merck.com path), neither carries a date in search results, and the summariser asserts 19 August, not the ~10 September this cabinet carried. Either date puts it outside this window, so it stays parked — but the carried-over date is now flagged as unsupported rather than repeated. Novartis/Ionis pelacarsen: two GlobeNewswire accessions carry /2026/09/04/ in the URL path itself, which does bind a date to those URLs, so the Lp(a)HORIZON miss is 4 September — the previous window's, and that run's miss. EMA CHMP September highlights: empty, and now for a stated structural reason rather than a repeated absence — the plenary runs 14–17 September and is sitting today, so no highlights page can exist; expect it around 21–25 September. That is a reasoned expectation, and the item should open the 18 September run's list only as a diary note and be fetched on 22 September.
• Storytelling & the new medium (MythKG): two cards, the vein's fourth consecutive productive run. Venice Immersive 2026 remains unresolved and the contamination warning proved live again. The section page labiennale.org/en/cinema/2026/venice-immersive-0 was opened from the main thread: it confirms 2–12 September, "68 projects" from "twenty-six countries", the 30/35/3 competition split, and the "La Laguna Parla" event on 5 September — and it carries no winners and no awards-ceremony date, with article:modified_time: 2026-09-01, out of window. A scout's summariser volunteered three winner titles with artists and a jury chair, none bound to any URL, and after last run's documented 2022-winner-as-2026-winner failure none of it is carried. The awards news page was not opened.
• Marketing: one card, counting the Wallet card's MktKG tag. Otherwise empty in-window — no first-party ad-product or platform-economics document dated 11–15 September and no SEC filing with ad-mix disclosure. Worth recording as a pattern rather than an absence: the one substantive platform change the scout found (Meta removing manual ad placement controls) is described by its own reporters as having no announcement at all, which is a finding about how platform economics is now shipped.
• Materials & building: empty in-window, and the ICC item did not resolve. No iccsafe.org publication announcement surfaced; everything returned restates the 24 August finalization, including one trade piece whose /2026/09/ URL path is not a date. The cheapest first-party evidence is whether the 2027 IBC/IFC/IWUIC are live in ICC's own Digital Codes catalogue at codes.iccsafe.org/codes; that was not fetched this run. Materials papers: empty, with three review articles surfaced whose dates are all April–June and none bound to a URL.
• Agriculture & robotics: empty in all four sub-veins — USDA APHIS / Federal Register, ag-robotics first-party technical documents, humanoid and field robotics, and SEC filings. A five-day window with zero primary documents is the finding. One aggregator-only claim is recorded and not chased: that Monarch Tractor laid off its workforce and sold its core technology "reportedly to Caterpillar", with no date, no filing and no primary. If true it is a significant vein event; it needs an SEC or Caterpillar-newsroom document first.
• Fashion: thin, and the named item resolved out of window. SHEIN's HKEX listing under code 00625 began trading on 1 September, ten days before this window opened, and no hkexnews.hk or SHEIN corporate-announcements URL was located — every result was a quote site, so the pricing and valuation figures in circulation are aggregator-asserted and unbound. In-window 00625 filings (over-allotment exercise, end-of-stabilisation notice) typically land around thirty days post-listing, so late September or early October. The CFDA's NYFW page is cfda.com/news/view-the-preliminary-september-2026-official-nyfw-schedule/ — "preliminary" in its own slug, mirrored in July trade press, and therefore the schedule rather than the week's own artefact. NYFW closed today; a recap may not exist yet.
• Resale: empty in-window. Nothing from ThredUp, Savers Value Village, Poshmark, The RealReal, Vinted, eBay or Depop dated 11–15 September, and no industry-body data.
• Crypto & agent-commerce: empty for the seventh consecutive run, with three searches spent on it as instructed. x402's operational launch remains 14 July; AP2's latest spec remains v0.2.0 from April with governance donated to the FIDO Alliance; the Stripe/Visa/Mastercard stablecoin platform remains June trade-press reporting sourced to unnamed people with no first-party issuer post. Mastercard "Agent Connect" did not resurface and is held as non-existent, per the 11 September log. Seven empty runs is a finding about this vein's cadence, not about our looking — and it is now worth deciding whether it stays on the standing list or moves to a monthly check.

Corrections made this run, on the record. Eight, of which two are against the world and six against this run's own drafts.

1. The cutover did not happen, and the standing card is corrected rather than edited. Full detail in the note appended to deepseek-v4-1-flash-routes-v4-pro-traffic.
2. The Merck/Moderna "~10 September" date this cabinet carried is unsupported. No first-party URL carries a date in search; the summariser says 19 August. Parked with the date flagged rather than repeated.
3. A draft called Roblox "the largest UGC platform." Nothing on the primary supports it; replaced with the platform's own figure, "123 million users" playing daily. L45's shape, caught by the superlative grep before review.
4. A draft's roadmap grep list was written from memory and was wrong. It named mcp-standardization and roblox-build-prompt-to-publish-alpha, neither of which matches. Re-run and pasted; the seven files are named on the card. (Re-run again on the verification pass, which caught that the first repair had itself omitted a file — see the process note.) L84(b), caught before review.
5. The fda.gov grep count was stale by construction — a card cannot grep the corpus and then not count itself. Restated as "eleven files, ten of them existing before this card."
6. Two arithmetic errors caught by review. A card said two pediatric approvals fell "nine days" apart; 3 to 11 September is eight. And a card read Roblox's footnote 2 — "LTM ended June 30, 2026 vs. LTM ended June 20, 2025" — as two end-dates "ten days apart"; they are a year and ten days apart, and the true observation is that the comparison is offset by ten days from a clean year-over-year.
7. "Three consecutive runs" was wrong about our own parked list. The Steam Frame item appears in 2026-08-21.md, 2026-09-08.md and 2026-09-11.md — three runs, but not consecutive ones; runs 6, 7 and 8 do not carry it. The card now says "across three runs."
8. A card's harness hedge was pointed the wrong way. A draft said DeepSeek states an evaluation harness for its July and August entries and none for September. It states one for 31 July and 21 August and for neither of the two entries actually compared — 13 August and 10 September — which is a stronger hedge than the draft printed, and the falsifier limb built on the weaker version was void.

Closed this run (no longer carried).

• Valve "Steam Frame" — closed as a search problem, reopened as a reading problem. Three runs were spent looking for a first-party document. One exists and is carded. What is now open is narrower and cheaper: the post's body.
• The DeepSeek 14 September cutover — closed by the reversal, with the correction note as the record.
• The Mastercard "Agent Connect" item stays closed; it did not resurface.

Parked this run (flagged, not published — with the primary that would promote each).

• Merck + Moderna INTerpath-001 — still the most valuable item on this list. One fetch of the short Moderna newsroom URL settles both the date and the finding. Check for overlap with fda-accelerated-approval-tudriqev-melanoma before carding.
• Novartis/Ionis pelacarsen Lp(a)HORIZON miss — 4 September, the previous window. A failed cardiovascular-outcomes trial for the flagship Lp(a) asset reads directly against olpasiran-lpa-lowering; if carded it is a dated backfill and should be logged as the previous run's miss.
• EMA CHMP September plenary — sitting 14–17 September; highlights expected around 21–25 September. Fetch on 22 September, not before.
• Venice Immersive 2026 winners — the section page was opened and carries none. The awards news page labiennale.org/en/news/official-awards-83rd-venice-international-film-festival was not opened and may cover only the main festival's prizes. Re-verify venice-immersive-2026-best-of-worlds once a first-party winners list exists.
• ICC 2027 International Codes — codes.iccsafe.org/codes is the cheapest first-party publication evidence and was not fetched.
• The Joenja prescribing information and Drugs@FDA — one fetch tests the limb the pediatric card names as most likely to fail.
• DeepSeek's technical report, its Hugging Face model card, and the pricing image — the first two settle the open licence question the 11 September card declined to answer; the third is the second document the cost-story limb names and the only one still unread.
• The Steam Frame post body — see the opportunities list.
• Roblox's own follow-ups: the "Making publishing easier" post, the about.roblox.com RDC newsroom post, and the full roadmap the RDC26 post says publishes "next week" — which, if it appears, is the document that turns this run's roadmap card from a promise into a list with dates attached.
• SHEIN code 00625 — the listing is out of window and no hkexnews URL was located. Over-allotment and stabilisation filings are the realistic in-window candidates, late September onward.
• Definium Therapeutics Panorama Phase 3 (DT120, generalized anxiety disorder) — a webcast advisory dated by its own URL slug to 14 September, but the release date is unbound and the word "positive" was supplied by the summariser and appears in no title. Do not card without a dated release.
• Monarch Tractor / Caterpillar — aggregator only; needs an SEC filing or a Caterpillar newsroom document.
• Greenfield Robotics' Regulation A+ offering — carried from 11 September unchanged; promote with a 253(g)(2) offering circular on EDGAR or a main-thread fetch of the release.
• Carried without re-attempt: NIST "Guidance and Templates for Public-Facing AI Documentation" (comments closed 16 September — now due, and the closure itself is a dated event); Meta Connect (23–24 September); ICIDS 2026; the EU textile Digital Product Passport delegated act (still unadopted); ThredUp and Savers Value Village filings; USDA APHIS COR-23134-4 (comments close 19 October); the CFDA NYFW recap if one appears.

Leads recorded, not parked (no URL or no first-party source). The claimed 14 September Tsinghua arXiv paper, with named authors and no URL of any kind. A claimed reversal of the DeepSeek cutover reported by two crypto-adjacent news wires — superseded, because the reversal is now quoted from DeepSeek's own pages and the wires contributed no fact. Venice Immersive winner names and a jury chair, volunteered by a summariser and bound to nothing. UBTech, CyberHero and Figure/BMW humanoid claims, all aggregator-only and none dated to a primary.

The TKG garden-code decision is still open. This run's machine-layer card tags TKG, taking the count of cards using the code in its machine-layer sense up by one without resolving the ambiguity L41 recorded. The founder's call is unchanged: give the machine layer its own code and remap, or restate content/beats/frontier-standing-sweep.md so TKG means one thing everywhere. It has now been carried for eleven runs.

Honesty note. Three verified, two signal — the first run in seven that is not all-verified, which RDKG_GOAL.md has been calling a yellow flag since August. Both demotions are for provenance defects that can be stated in one sentence each, which is what L63 requires, and neither was engineered: the DeepSeek reversal sentence carries no date of its own, and the Valve post's body did not render. Where confidence is genuinely low, in prose: the Steam Frame card is the weakest in the run and by a distance — it rests on a headline and an eight-word publisher-written summary, it cannot say what the product is from the document alone, and one successful fetch of the body could overturn its central reading. The RDC26 card is a list of promises and nothing on it has shipped; its value is entirely in the dated commitment and the metrics printed beside it. The Joenja card's central interpretation — that this is efficacy inferred rather than measured — is the garden's reading and the card names it as the limb most likely to fail. And the Wallet card describes a December rollout to one country and one legal form of creator, with no count of how many creators that is.

Two shapes worth flagging rather than defending. Concentration: two of five cards come from one platform's announcements category on one evening, roughly three and a half hours apart (19:00:57 and 22:30:06 UTC, an interval of 3h 29m computed on the Wallet card) — the second consecutive run with two Roblox cards, and an artefact of a scout enumerating that platform's feed well rather than of that platform being unusually important this week. Text that moves: two of this run's five primaries show evidence of a page whose text is not stable — DeepSeek's changelog carries a sentence under a date it cannot have been written under, and the RDC26 post has a reply quoting a variant the live text no longer contains. That is now a standing hazard for a garden whose one rule is "quote the face of the document," and it is logged as a lesson.

Process note (what the review cost and caught). Two adversarial passes were run in parallel on the same five drafts — one on discipline against the lesson ledger, one as a fact-check against a transcribed fact sheet — returning 66 and 61 findings. Overlap was heavy and the union was fixed in one pass. A verification pass on those fixes then returned 32 more findings, 27 of them created by the repairs themselves — L72's recorded rate, confirmed again and slightly worse than usual.

The most common class across both stages was the one the ledger has recorded three runs running: an enumeration or blanket claim about the card's own contents that the finished card falsifies. Every one of the five cards carried a "Derived figures, enumerated. N, and nothing else on this card is computed" paragraph that was false, and the reason is structural rather than careless: a card that computes figures from its primary and also counts things about the cabinet has more than one category of derived number, and a single enumeration covering all of them is false by construction. That is logged as L85 — and the verification pass then found the two-way split the first fix introduced was itself still false on three cards, because a third category exists (counts of the primary's own contents). The lesson carries that refinement in the same entry.

The single most instructive regression is worth naming because it is the ledger's oldest shape reproducing inside its own fix: the run's pre-flight checks caught a false roadmap grep list written from memory and replaced it — with another incomplete list, which omitted this run's own sibling card, while the log certified the fix. The corrected list has seven files and the certification now says so. Three further repair-induced regressions are worth recording by shape: two cards gained a "No ordinal or priority claim about the cabinet is made on this card" sentence that a summary field or a falsifier limb on the same card falsified (L68's self-certifying-audit shape, created by an L68 fix); a cross-card quotation gained a comma inside its quote marks in a passage written specifically to paste cross-card quotations faithfully (L64's shape, sixth run running); and the correction note claimed four of five falsifier elements had been re-read when the run's provenance contract supports one.

A second lesson came out of the review's own method. Three of the fact-checker's findings — that a forum category, a transaction-export clause and a page title were unsourced — were artefacts of the abridged fact sheet I wrote for it, not defects in the cards; all three are on the primaries and were simply omitted from my transcription. A checker given a transcription instead of a document will confidently report absences that are absences in the transcription. That is logged as L86.

On the window and the cadence. Four days, 11–15 September, the Tuesday beat run as designed. Freshness: three of nine veins reached a primary dated inside 11–15 September — health, storytelling and marketing — and four cards came out of those three, the machine-layer card resting on a document dated 10 September which is why it is signal; crypto & agent-commerce is empty seven runs running and is now a candidate for demotion to a monthly check rather than a twice-weekly one; agriculture & robotics, materials & building and resale returned nothing in-window at all. The 18 September Friday beat should open with the Merck/Moderna fetch (one fetch, highest value on the parked list), the Steam Frame body, ICC's Digital Codes catalogue, and the Roblox roadmap the RDC26 post promises for the week of 14 September. The CHMP item should be left alone until 22 September.

How to read this log (the action view). The tiered list above is the provenance axis. The buckets below are the action axis. Tier and bucket are independent.

• 🆕 New (net-new this window). Four of four, which is the reversal of the last run and is stated as such rather than left to be inferred: the 8 September log had one in-window card and three backfills, and this one has four primaries dated inside 8–11 September (10 September, 9 September, 9 September, 9 September). Every primary was opened and quoted from the main thread. No backfills, no promotions, and nothing carried over from the parked list — which is itself a minor failure of the run and is recorded under Parked below, because two of the cheapest promotions available were not taken.
• 🔁 Changed (updates something we already track). All four touch standing cards. DeepSeek is the supply-side counterpart to inference-cost-collapse, whose evidence has until now been observational (Stanford HAI's index) and whose falsifier this card explicitly does not close — a card carrying no price cannot demonstrate a quality-normalized fall, only a vendor asserting one. It is also the concrete instance of perishable-frontier-model-shelf-life-weeks, whose own label travels with it: "A thesis, not yet a proven business." And it joins inkling-open-moe-for-fine-tuning, whose capsule already records that Inkling's MoE design "largely follows DeepSeek-V3" — the lineage is being copied by a Western lab, so this architecture will be tested quickly by people who are not DeepSeek. Sevabertinib is the fifth entry in the FDA evidence ledger and the cabinet's second Project Orbis card, fourteen days after fda-approves-daraxonrasib-pancreatic-adenocarcinoma; a draft called it the first, and the grep that caught that is named on the card. Animation Packs and the discounts test both read against roblox-creator-earnings-2025-median-payout for who absorbs the cost, and the first reads against roblox-friends-list-api-au-under-16, which reached the same design conclusion from the opposite direction — that card's words, pasted rather than remembered: "the honest design move is to make the in-experience relationship the source of truth rather than the platform's."
• 🎯 Opportunities flagged (act / position / build / money).
  1. Check that our model strings are strings, before Monday. On 14 September a model identifier that this cabinet or any garden could have hardcoded — deepseek-v4-pro — starts pointing at different weights at a different price, with no code change and no error. That is perishable-frontier-model-shelf-life-weeks's thesis arriving as a dated calendar event rather than an argument, and the audit it implies is an hour of reading config, not engineering. It does not touch the BKG P0.
  2. The operator-funded discount is a primitive MktKG should own. A marketplace operator buying down the buyer's price while holding the seller whole is a promotion design that never requires a merchant negotiation, and the cabinet had no card on it. Worth writing up as a mechanism note for any garden that ever runs a marketplace — together with its governance question, because the same post that shows the mechanism shows the sellers asking for a veto they do not have.
  3. The surrogate-versus-outcome audit is still not run, and this run adds a second argument for it. The 8 September log proposed tagging health cards surrogate versus outcome off CARDIO-TTRansform. Sevabertinib is the same lesson from the other side: a 75% response rate with no survival figure at all is a surrogate result, and the cabinet now holds approvals resting on response rates (this, Tudriqev) beside one resting on a survival comparison (daraxonrasib). The audit is analytic time and the material is now sitting on four cards.
  4. Decide The Call's position on bought embodiment. Two Roblox posts in one evening make the question concrete: a participant's gait is now purchasable, and a work can switch it off. The durable design conclusion is the one that needs neither — own the expressive channels the beat depends on — and that is a constraint on the making, not a fact about the story.
• 📚 Good to know / integrate (context, no action now). The image problem. The DeepSeek page's benchmark chart and price table are images; the card carries no number from either, and one fetch of the API pricing page or the technical report would convert its most quotable claims from stated to measured. This is a new failure mode for this cabinet — not an unreachable document, but a reachable document whose load-bearing content is not text. The responder denominator. Sevabertinib's "73% of responding patients" is a proportion of responders, not of 69, and no combined figure is computed anywhere on that card. The validation bounds. Roblox's automated animation checks name four limits and print numbers for only two — "more than 1.5 studs" and "longer than 10 seconds" — so "too far" and "the max" are thresholds a creator cannot check against in advance. The pilot-allocation lead. Replies to the Animation Packs post claim pilot participants had a multi-month head start and a five-per-day upload limit against the general one-per-day. Not first-party, not carded; if true it is a finding about the creator economy rather than the medium.

How this run worked (the loop, once). Read the ledger (RDKG_GOAL.md, the lesson ledger to L83, the 8 September changelog, content/beats/frontier-standing-sweep.md and content/beats/mythkg-story-sweep.md) → three parallel search-only scouts with caps of 12, 12 and 14 → main-thread verification with 2 targeted searches and 6 document fetches, of which 4 returned readable primaries that became cards → deep-verified every keeper by opening the primary from the main thread and quoting it → carded four survivors → ran the standing mechanical checks (capsule word count on all four; the cabinet-scoped grep per claim, which caught one false claim before review) → an independent adversarial review returning 23 findings → fixed all 23 → a verification pass on the fixes returning 11 more, nine of them created by the repairs → fixed those → validated the corpus to zero errors. The founder commits, builds and deploys; a card is not live until then.

Veins swept, and what each returned. Nine veins across three scouts plus main-thread work. AI & the machine layer: one card, the vein's first since 4 September. Health & longevity: one card. Two strong candidates were reached and not carded, both for date reasons (below). The EMA CHMP September page still does not exist: a domain-restricted pass returned Jan, Feb, Apr, May and Jun 2026 highlights and no September page, and the visible 2026 cadence — the last full Mon–Thu of each month — puts the plenary around 14–17 September with highlights about a week later. That is now a reasoned expectation rather than a repeated miss, and it means the CHMP item should be dropped from the next run's opening list and picked up on 22 September. Storytelling & the new medium (MythKG): one card, the vein's third consecutive productive run, from a scout that enumerated Roblox's own Announcements category JSON rather than trusting search summaries — six in-window announcements surfaced that way, of which two are carded here. Marketing: one card, the vein's first in this cabinet from a platform-economics primary rather than an ad-product one, and the queued Google Ads item is now closed (below). Materials & building: empty in-window. The one candidate, ICC's 2027 International Codes release, reads "For Immediate Release: August 24, 2026" on its face — the search engine's "4 days ago" was its modify date, which is a clean worked example of L71 and is recorded as one. ICC says the 2027 IBC/IFC/IWUIC publish in September 2026, so an actual publication announcement is a live candidate for a later window. Agriculture & robotics: a real in-window candidate, parked on provenance grounds (below) — but the vein is no longer empty-by-absence, which is a change from two runs running. Fashion: thin. NYFW is running 10–15 September; the CFDA schedule page is first-party but its own posting date could not be bound and it was not opened. An event is not a document. Resale: nothing in-window; ThredUp's most recent 8-K is 5 August. Crypto & agent-commerce: empty for the sixth consecutive run, and this run spent three searches plus one targeted main-thread search on it. The aggregator claim that Mastercard launched an "Agent Connect" product on 9–10 September did not survive: a domain-restricted search of mastercard.com returns Agent Pay (April 2025), Agent Pay for Machines (June 2026), Agent Suite (January 2026) and an Agent Toolkit — no product called Agent Connect. It is not carried, and it is recorded here so that if it resurfaces the non-verification is on the record.

Corrections made this run, on the record. Four against this run's own drafts, plus one against the world. The adversarial review returned 23 findings and the verification pass returned 11 more; most were repairs rather than corrections, and these five are the ones that changed what a card asserts.

1. The DeepSeek card said "retired" where the primary says "phasing out … until V4.1-Pro launches." The title, the slug and the body all described a lab retiring its flagship into its smallest model. The page uses "retired" only of V4-Flash and V4-Flash-Vision-Exp; of the flagship it describes a temporary re-route pending a successor. The card was retitled, the slug changed from deepseek-v4-1-flash-retires-v4-pro to deepseek-v4-1-flash-routes-v4-pro-traffic, and the file renamed. The frame had outrun the quotes in the direction that made the card more dramatic — L70's shape, and the sixth consecutive run in which the sentences around the quotes carried the error.
2. "The cabinet's first Project Orbis card" was false, and the grep that caught it was one this run ran before review. fda-approves-daraxonrasib-pancreatic-adenocarcinoma records that approval's Orbis review with Health Canada, with the EMA and Japan's agency as "official observers." This is the second. The correction is on the sevabertinib card in its own voice. This is L75 working as designed: the claim was grepped at write time and failed the grep.
3. The sevabertinib card asserted an epidemiological figure two paragraphs after certifying that it printed none. A draft read "the 96% of NSCLC that does not carry this mutation." The figure is in no quoted line, in no listed source, and the same card says "This card prints no market size, no revenue estimate and no epidemiology … because the FDA page contains none and no other document was opened." The clause now names the absence instead of filling it.
4. Three cards claimed "every quotation on this card comes from that page" while quoting other cabinet cards. The blanket was false on all three by construction — L68's exact recorded shape. All four provenance paragraphs now read "every quotation attributed to the primary" and enumerate their cross-card quotes separately.
5. An aggregator's claimed Mastercard "Agent Connect" launch could not be verified and is not carried — see the crypto vein above.

Closed this run (no longer carried).

• Google Ads AI-disclosure labels ("How this ad was made"). Carried on the parked list for two runs waiting for a date to be read off the page. It has been: blog.google/products/ads-commerce/google-ads-ai-transparency-labels/ carries "Jul 09, 2026" in its byline and article:published_time: 2026-07-09T16:00:00+00:00, by Keerat Sharma, VP & GM, Ads Privacy and Safety. Two months out of window and receding; it will not come back. Closed. If the marketing vein wants it, it is a backfill on its own date, not a parked item — the content is a "How this ad was made" panel in My Ad Center across Search, YouTube and Discover, auto-disclosure for Google's own generative ad tools, and an advertiser self-declaration control for tools used elsewhere.
• Mastercard "Agent Connect" — closed on non-verification, above. Re-opens only if a mastercard.com newsroom URL with a readable date surfaces.

Parked this run (flagged, not published — with the primary that would promote each).

• Greenfield Robotics' Regulation A+ offering — the agriculture & robotics vein's first in-window candidate in three runs, and parked for a provenance reason worth stating precisely. A GlobeNewswire release dated 11 September 2026, 07:00 ET (dateline "Cheney, Kansas"; DC.date.issued: 2026-09-11) claims an offering at $2.07/share and 82 chemical-free weeding robots running commercially across 16 states. It was fetched by a scout, not from the main thread, so this run holds no exact quoted line from it. The stronger primary was sought and complicates rather than confirms: EDGAR's submissions feed for Greenfield Robotics Corp (CIK 0001760404) shows a Form 1-A filed 18 June 2026, file number 024-12773, and no Form 253(g)(2) and no qualification notice — so nothing on EDGAR yet evidences a qualified, open offering. Two other Form 1-A/A accessions surfaced in search under the same CIK path but are absent from the submissions feed, which is an unexplained discrepancy this run did not resolve. Promote with either a 253(g)(2) offering circular on EDGAR or a main-thread fetch of the release; do not card the 82-robot figure until one of those is in hand.
• Merck + Moderna, Phase 3 INTerpath-001 (intismeran autogene plus KEYTRUDA, meeting RFS and DMFS endpoints in resected Stage IIB–IV melanoma). First-party Merck newsroom URL surfaced; its date came only from a search summariser's own estimate ("about 1 day ago, approximately September 10") and was never read off the page. This is the most valuable item on this list — an mRNA cancer-vaccine Phase 3 hitting its primary endpoint — and it is parked purely for L74/L71 discipline. One fetch promotes it; check for overlap with fda-accelerated-approval-tudriqev-melanoma.
• Novartis/Ionis Lp(a)HORIZON — pelacarsen MISSED its primary endpoint, claimed 4 September, from an Ionis IR URL with no date read. A failed cardiovascular-outcomes trial for the flagship Lp(a) asset is a high-consequence negative and reads directly against olpasiran-lpa-lowering, which the cabinet already holds. It fell inside the previous run's window; if it confirms, it is that run's miss and should be logged as one.
• AbbVie Phase 3 CERVINO (etentamig, dual primary endpoints in relapsed/refractory myeloma), news.abbvie.com/2026-09-03-... — date bound by the URL path itself, 3 September, so also inside the previous window.
• DeepSeek's own technical report and API pricing page — the two documents that would convert this run's machine-layer card from vendor-stated to measured. huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash also settles the open licence question the card explicitly declines to answer.
• Roblox RDC 2026 (10–12 September) — confirmed as an event by first-party posts, but no first-party announcement of keynote content was reached, and no "RDC26: What We Announced" post existed in the Announcements feed at fetch time (the RDC24 analogue does). This is the single highest-value item for the next run: a platform conference mid-window whose output lands in the following days.
• Four further in-window Roblox announcements, enumerated and not carded (timestamps from Roblox's own category API): the Character Controller Custom Abilities API Studio beta (10 Sep), the pause of Unified Roles & Permissions for Communities and Creator Hub (9 Sep), the Device/Network Simulator Studio beta (9 Sep), and Auto-Reimport (8 Sep). Titles and timestamps only; none was opened. The pause is the interesting one and is referenced by title on the Animation Packs card's falsifier.
• Venice Immersive 2026 awards — the festival's own page confirms the exhibition runs 2–12 September with 68 projects from 26 countries and names the jury, but states no awards-ceremony date and lists no winners. A search summary asserting a 10 September ceremony also returned a 2022 winner as the 2026 winner and is treated as contaminated. Re-verify venice-immersive-2026-best-of-worlds once winners are posted.
• Valve "Steam Frame" — third run carried, still no first-party Valve document. Every dated claim remains aggregator-sourced.
• CFDA September 2026 NYFW schedule — first-party page identified, posting date unbindable, not opened. NYFW runs 10–15 September, so a fashion primary is available to the next run if one is willing to spend a fetch.
• ICC 2027 International Codes — the August release is out of window, but ICC states the 2027 IBC/IFC/IWUIC publish in September 2026; the publication announcement is the materials vein's cleanest forward candidate.
• NIST "Guidance and Templates for Public-Facing AI Documentation" — initial public draft, comments close 16 September 2026. Out of window, but a dated deadline inside the next one.
• Carried without re-attempt: arXiv:2609.04217 (dated 25 June, figures already recorded in the 8 September log); SHEIN code-00625 announcement list — not attempted this run, and it should have been, since the last log left explicit instructions to enumerate it directly rather than search for it; ThredUp Direct Listing beta; Savers Value Village secondary (8-K claimed 13 August); the EU textile Digital Product Passport delegated act (not adopted; indicative timeline 2027); USDA APHIS COR-23134-4 (comments close 19 October) and the ARS European plum petition (14 August); Roblox "Fall Games Preview"; Meta Connect (23–24 September); ICIDS 2026 late-breaking deadline (14 September).

Leads recorded, not parked (no URL or no first-party source). Claimed OpenAI "GPT Image 2.5" models dated 8 September and a Sakana AI "Fugu Ultra v2.0" dated 11 September — both from model-tracker aggregators, neither with any first-party URL surfacing under domain-restricted search. Claimed Anthropic Fable/Mythos 5.1 (1 September, with a stated 75% cache-read price cut) and Meta Muse Spark 1.3 (2 September), same provenance problem; the price figures are the useful part if anyone chases them. An OSL "AgentPay" stablecoin-payments item with no readable date.

The TKG garden-code decision is still open. This run's machine-layer card tags TKG, which takes the count of cards using the code in its machine-layer sense up by one and does not resolve the ambiguity. The founder's call is unchanged: give the machine layer its own code and remap, or restate the beat file so TKG means one thing everywhere.

Honesty note. Four verified, no signal — the sixth all-verified run in a row, which RDKG_GOAL.md calls "a yellow flag, not a triumph." The standing explanation holds (L52/L63: tier measures provenance, and all four cards have a dated first-party primary, so demoting one to engineer the ratio would be the actual dishonesty). The more useful thing is to say where confidence is genuinely low, in prose. The DeepSeek card is the weakest of the four and by some distance: every superiority and price claim on it is the vendor's own, unnamed "Tests by multiple parties" is not a citation, and the two charts that would substantiate any of it are images the card did not read — a reader should treat that card as "a vendor said this on a dated page," which is exactly what verified means here and is much less than what it sounds like. The sevabertinib card's central interpretation — that first-line single-arm evidence is an evidentiary loosening — is the garden's reading and the card names it as the limb most likely to fail. The Animation Packs card cannot say how many experiences will actually allow packs to play, which is the number the whole story turns on. And the discounts card describes a test with no published size, no discount depth and no end date, whose most likely future is a quiet ending. Also worth flagging as a shape rather than a fault: two of four cards come from one company on one evening, which makes this run more concentrated than it looks, and the concentration is an artefact of a scout enumerating one platform's announcement feed well rather than of that platform being unusually important this week.

Process note (what the review cost and caught). The adversarial review returned 23 findings on four drafts. The verification pass on those fixes returned 11 more, and nine of them were created by the repairs — L72 confirmed again, at almost exactly the recorded rate. The regressions were the documented shapes: a count reconciliation that produced two different "fourth FDA card" claims on one card; three cross-card quotations clipped with a period moved inside the quote marks while fixing other quotations; two newly-written grep lists that were themselves incomplete; a dropped clause that left a doubled em dash and an unparseable sentence; and a "the page" → "the announcement" substitution applied in two places out of three. The one finding worth keeping as a rule is the enumeration problem, which has now appeared in three consecutive runs in three different forms: an enumeration edited after the text it enumerates is a stale count, and every fix that changes an item must re-count the list in the same edit. That is logged as L84.

On the window and the cadence. Three days, 8–11 September, the Friday beat run as designed. Freshness: five of nine veins produced or reached an in-window primary; crypto & agent-commerce has now been looked at and empty six runs running, which is a finding about that vein's cadence and not about our looking; materials & building and resale are the two to watch. The 15 September Tuesday beat should open with RDC26's output, the Merck/Moderna fetch (one fetch, highest value on the parked list), the DeepSeek routing check on 14 September (which will have just come due), and the SHEIN code-00625 announcement list, which this run was told to enumerate and did not.

How to read this log (the action view). The tiered list above is the provenance axis. The buckets below are the action axis. Tier and bucket are independent.

• 🆕 New (net-new this window). One. The Roblox announcement is the only card whose primary carries a date inside 4–8 September on its face (4 September 2026, 7:10pm, timestamped in both the visible post and its metadata). Three are backfills or promotions, and this log says so plainly rather than letting the card dates imply a busier window than there was: CARDIO-TTRansform on 28 August (the promotion the last log queued first), the FDA approval on 3 September (inside the previous run's window and missed by it), and the Nature Communications paper on 21 August (a materials-vein backfill surfaced by an ETH institutional post dated 7 September). Every primary was opened and quoted from the main thread.
• 🔁 Changed (updates something we already track). Three of four touch standing cards. CARDIO-TTRansform opens a new heading beside acacia-hcm-aficamten-nonobstructive-hcm — two ESC 2026 cardiology trials now sit in the cabinet, one that met both primary endpoints with small effects and one that missed with a silenced biomarker, which is a better pair than either alone. The concrete paper extends lc3-low-clinker-cement on the qualification half of that card's two-part residual constraint ("standards qualification and suitable clay supply" — the clay-supply half is untouched) and joins tall-mass-timber-code and mycelium-insulation in the building-materials cluster. It closes neither of the LC3 card's two falsifier limbs, which are (a) whether standards bodies had in fact amended specs to permit calcined-clay + limestone blends and (b) whether durability fails at high clinker replacement under independent testing; the new paper bears on neither directly and adds a separate dated test of its own, on its own card. The Roblox card is the counterweight to fortnite-42-10-social-synergy-party-api, carded four days ago, and reads across to roblox-creator-earnings-2025-median-payout for who absorbs the cost and to on-device-models-child-safety-inversion for the same regulatory pressure from the other end. The FDA card extends the approval-evidence ledger without touching an existing card.
• 🎯 Opportunities flagged (act / position / build / money).
  1. Run the biomarker audit CARDIO-TTRansform earns us. The cabinet holds several health and longevity cards whose evidence is a measured molecule rather than a measured life — the aging clocks, the proteomic cards, anything resting on a mechanism. This trial is the cabinet's largest worked example by enrolment (1,432 patients — a countable claim, not a judgement of quality) of a mechanism that is true and an outcome that did not follow. A one-pass sweep tagging those cards surrogate versus outcome costs an hour of analytic time and makes the health lane's honesty legible. It is analytic work, not engineering, and does not touch the BKG P0.
  2. Low-carbon concrete is a site-resolved market, not a national one. If the ETH substitution reaches a code body, the same mix is a different commercial proposition in Zurich and in Manaus. That is a different sales motion from the one the industry runs, and BKG's spec work should start describing mixes by climate rather than by product. The paper is explicit that it is not yet usable in practice — so this changes how a spec is reasoned about, not yet how it is written.
  3. Design The Call's threshold to own its own graph. Two platform announcements four days apart, pointing opposite ways. The durable design conclusion is the one that needs neither: a work that asks two people to do something together and remembers that they did it cannot be switched off by a regulator in one country. That is a constraint on the work, and this log states it as one.
  4. The evidentiary frontier is where ultra-rare approvals get interesting. The FDA put in writing that modelling plus four infants' safety data can carry an indication where a trial cannot be run. Whether that is a durable precedent or a one-off turns on the label, which was not opened; it is the cheapest single check available to the next run.
• 📚 Good to know / integrate (context, no action now). The subgroup discipline. CARDIO-TTRansform's prespecified stabiliser split favoured the drug in stabiliser-naive patients at nominal significance, in a trial that missed its primary endpoint. The ESC release prints no numbers for either subgroup arm. Both facts travel with the card so the finding cannot be quoted as a result. The publisher's notice. The Nature Communications article carries an early-sharing banner — "This version is subject to further edits and will be replaced automatically by the final Version of Record" — which is a freshness caveat of a kind this cabinet has not recorded before and is on the card. Industry funding, disclosed. The concrete paper is part-funded by Cemsuisse, a cement-industry association; the trial by AstraZeneca and Ionis. Both disclosures are on the face of their documents and on their cards. The forum leads. The Roblox thread contains a user claim that UK friends lists are already hidden by default with no opt-back-in, and a claim from an age-verified Australian adult that the restriction already reaches them. Neither is first-party; both are recorded as leads on the card, not used.

How this run worked (the loop, once). Read the ledger (RDKG_GOAL.md, the lesson ledger to L80, the 4 September changelog, content/beats/frontier-standing-sweep.md, content/beats/mythkg-story-sweep.md and RDKG_STORY_GOAL.md) → three parallel search-only scouts with per-scout caps of 10, 10 and 12 (one scout was interrupted mid-run and was re-launched with a widened brief covering both storytelling and the commercial veins) → main-thread verification with 2 targeted searches and 7 document fetches, of which 6 returned readable primaries and one — the SHEIN allotment PDF — exceeded the fetch tool's output limit and was not read this run → deep-verified every keeper by opening the primary from the main thread and quoting it → carded four survivors → ran the standing checks (capsule word count on all four, with one card's capsule trimmed from 85 words to inside the 40–80 range; the duplicate check over both the card list and the parked list) → validated the corpus to zero errors. The founder commits, builds and deploys; a card is not live until then.

Veins swept, and what each returned. Nine veins across three scouts plus main-thread work. Health & longevity: two cards — one promotion, one correction of a previous run's miss. The EMA/CHMP September meeting page did not surface on a ten-search pass and was not opened; a claimed accelerated approval on 4 September surfaced with no drug name, sponsor or URL and is recorded as an unresolved lead, not parked. Materials & building: one card — the vein's first since mycelium-insulation (dated 14 April 2026, live-checked 7 July), against a cluster that also holds tall-mass-timber-code and lc3-low-clinker-cement; verified by ls content/signals/ over the building-materials cards and reading each one's date. ICC and other code bodies were not reached before the scout's cap. Storytelling & the new medium (MythKG): one card, and the vein's second consecutive productive run. AI & the machine layer: no card, deliberately — the one strong candidate was date-corrected out of the window (below), and the previous run already carried three machine-layer cards. Crypto & agent-commerce: empty for the fifth consecutive run, and this time with specifics rather than absence — the three live bodies last shipped in April (AP2 v0.2.0, ACP), July (x402 Foundation operational launch) and July (MCP spec 2026-07-28, roadmap 22 August), and five searches found nothing dated 4–8 September. Five empty runs is now a finding about the vein's cadence, not about our looking. Marketing: no in-window primary; the Google Ads AI-disclosure post's first-party URL is now confirmed present in search results but its date is still not read off the page (see parked). Fashion: thin — a CFDA schedule page is the one plausible in-window first-party document and was not opened. Resale: no confirmed in-window primary; a ThredUp peer-to-peer listing beta is the strongest lead and has no first-party source. Agriculture & robotics: empty in-window; no USDA APHIS notice and no primary paper surfaced.

Corrections made this run, on the record. Six — five against the world, one against this run's own drafts, and the adversarial review that found most of them returned sixteen findings in total, of which the rest were repaired without rising to the level of a correction.

1. The '77.2% TTR reduction at week 140' the 4 September log carried for CARDIO-TTRansform is not in the ESC release. The document states qualitative suppression "as assessed by a prespecified exploratory endpoint through Week 140" and prints no percentage. The card carries the qualitative statement only, and says so. The scope of the correction is exactly that one figure: the prior log's enrolment (1,432), rate ratio, confidence interval and p-value all match the primary, and the card's corrections paragraph now quotes the prior log rather than paraphrasing it, after a first draft attributed to it a vaguer enrolment figure it never used.
2. The FDA's Zanvastro approval of 3 September fell inside the previous run's own 1–4 September window and was missed. That log recorded "no FDA or CHMP primary dated 1–4 September surfaced on a ten-search scout pass, which is a thin look and is recorded as one." It was a miss, not an absence. The card carries the note in its own body as well as here.
3. An arXiv identifier does not bind a date. A scout surfaced arXiv:2609.04217 ("At Equal Inference Cost, Multi-Agent Structure Does Not Beat a Single Frozen Agent") and inferred from the 2609 ID block that it was submitted around 4–5 September, flagging the inference honestly. The abstract page was opened: the submission history reads "[v1] Thu, 25 Jun 2026," and meta-citation_date is 2026/06/25. The paper is two and a half months old, not four days. It is parked, not carded, and the rule is now written to tasks.lessons.md as L81 — checked by grep after writing, because a previous draft of this paragraph asserted the lesson existed when the ledger still ended at L80.
4. The Nature Communications paper's date belongs to the paper, not to the institutional post that surfaced it. ETH Zurich's news page is dated 7 September 2026 and would have been an in-window date; the article's own page reads "Published: 21 August 2026." The card is dated to the paper. Where a figure is available only from the ETH page and not from the abstract, the card attributes it to the ETH page at the point of use and again in the falsifier.
5. The concrete card's first draft inverted two cross-references, both in the direction that flattered the new card. It described lc3-low-clinker-cement as having found that standards were the obstacle — that card says codes "have recently been amended to permit these blends" and names a two-item residual constraint of which the draft carried one — and described tall-mass-timber-code as a card "where a code body, not a material, is the gate," when it records the gate opening ("code is now a tailwind"). Both are fixed on the card and above; the mechanical rule that prevents the class is logged as L82.
6. A claimed NVIDIA acquisition of Hugging Face 'for over twelve billion dollars' appeared in an aggregator summary dated 3 September. No primary, no filing, no release. A transaction of that size would be visible in an 8-K. The scout flagged it as likely bad data unprompted; it is not carried, and it is recorded here so that if it resurfaces in a later run the earlier non-verification is on the record.

Closed this run (no longer carried).

• PRECISE-MI — carried since 1 September with the instruction to give it one more run and then close. Two runs have now produced no results page and no journal URL, only the pre-congress Hot Lines index. Closed. Re-open only if a journal page or an escardio.org results page is opened and dated.
• The 'accelerated approval in breast cancer, 4 September' lead — surfaced with no drug name, no sponsor and no URL on a dedicated health scout. Not promoted to the parked list, because there is nothing to promote from. Recorded here and closed; it returns only if a named FDA announcement surfaces.

Parked this run (flagged, not published — with the primary that would promote each).

• arXiv:2609.04217, 'At Equal Inference Cost, Multi-Agent Structure Does Not Beat a Single Frozen Agent' (Dylan, Brennan, Murphy, O'Sullivan, Kelly, Walsh; cs.MA; DOI 10.48550/arXiv.2609.04217). Opened and read this run; the abstract's load-bearing figures are recorded here so a future run can card it without re-fetching: on ALFWorld the full Planner-Executor-Critic team "achieves the highest mean but is not statistically better than the single agent: 0.769 versus 0.754, p = 0.80, despite using 1.8 times more evaluation calls", and "the planner and critic evolve to empty or low-impact prompts." It bears directly on inference-cost-collapse and on how the machine lane budgets agent architectures. Not carded for two reasons: it is dated 25 June 2026, well outside this window, and the machine layer took three cards four days ago. Promote on its own date whenever a run has room, or sooner if a v2 or a venue acceptance gives it a fresher primary.
• SHEIN — did dealings commence? Unresolved for a second run, and the reason is now specific rather than general. The HKEXnews document the last log named as the one that would settle it (.../2026/0831/2026083102409.pdf) turns out to be the allotment-results announcement the cabinet already holds as shein-hk-listing-priced-allotment's own source_url — so it cannot settle a question that card explicitly declined to answer. It was re-fetched this run and exceeded the fetch tool's output limit, so it was not re-read. What would settle it is a different document: any HKEXnews listed-company announcement for stock code 00625 dated 1–8 September, or the over-allotment / stabilisation notice, which on the normal thirty-day timetable would fall in late September and whose absence in this window is therefore expected rather than anomalous. Next run should enumerate the code-00625 announcement list directly rather than searching for it.
• Google Ads AI-disclosure labels ('How this ad was made'). The first-party URL blog.google/products/ads-commerce/google-ads-ai-transparency-labels/ is now confirmed to exist as a live result; the claimed date of 9 July 2026 still comes only from third-party summaries and has never been read off the page. Two supporting first-party help-centre pages also surfaced (support.google.com/google-ads/answer/17140115 and support.google.com/My-Ad-Center-Help/answer/17196133), dates unknown. The durable detail worth the card, per summaries and not verified: visible on-ad labels are reportedly tied to campaigns targeting the EU, India and New York, and Google reportedly does not verify advertiser self-declarations for third-party AI tools. Still the marketing vein's cleanest backfill; one fetch promotes it.
• CFDA preliminary September 2026 NYFW schedule (cfda.com/news/view-the-preliminary-september-2026-official-nyfw-schedule/) — the fashion vein's one plausible in-window first-party document, not opened. NYFW itself runs 10–15 September per the summary, which is next run's window.
• ThredUp 'Direct Listing' peer-to-peer beta — a resale-vein lead with no first-party URL. Promote from ThredUp's investor-relations newsroom or an 8-K on EDGAR.
• EMA CHMP September 2026 meeting highlights — no page surfaced on a ten-search pass; the latest indexed highlights were 22–25 June 2026. The EMA meeting-highlights index settles whether a September meeting has reported.
• Roblox 'Fall Games Preview' (2 September) — first-party, still not opened, still deliberately not carded as a genre preview. RDC (10–12 September) is next run's event.
• Steam Frame — third-party reports of backend package changes on 3 September and a 'Great on Frame' certified catalogue at 117 titles on 6 September, both from aggregators. No first-party Valve announcement surfaced. Promote from SteamDB's package history and the Steam store page read directly, or from a Valve post.
• Carried without re-attempt: the NEJM CARDIO-TTRansform paper (a DOI circulating in search was not verified and is not carried); Cytokinetics' ACACIA-HCM and MAPLE-HCM releases (L73 URL-length problem — promote from an EDGAR 8-K); the Anthropic Fable/Mythos 5.1 release page (month-only date, plus the analyst disclosure recorded on 4 September); rusfertide's corroborating documents; NVIDIA Form 10-Q; Savers Value Village 10-Q; the EU textile Digital Product Passport JRC report; USDA APHIS petition COR-23134-4 (comments close 19 October); the Wiley lettuce-weeding paper; Meta XR Operator and Meta Horizon Worlds mobile pivot (MythKG backfills); Qwen3.8 Flash; the COSMOS epigenetic-clock RCT; the eBay Form 425 accession-attribution question.
• Calendar, all dates as claimed and not opened: Venice Immersive awards 12 September (re-verify venice-immersive-2026-best-of-worlds after); Roblox RDC 10–12 September; NYFW 10–15 September; ICIDS 2026 late-breaking deadline 14 September; Meta Connect 23–24 September.

Leads recorded, not parked (no URL or no first-party source). A Nature Aging 2026 index surfaced four longevity items — aged-mouse-brain spatial transcriptomics, human muscle multiomics, telomeric ncRNA inhibition, engrafted contractile muscle tissue — with no individual DOI or date bound to any of them; a fetch of the index would turn one or more into parked items. Chip-supply claims for early September (CXMT DRAM share, TSMC equipment demand, a Cerebras Finland datacenter, an AMD LUMI-AI contract) all came from newsletter aggregators with no company or filing URL. A StablecoinX Form S-8 claimed for 3 September is a routine employee-benefit-plan registration and is low signal even if confirmed.

The TKG garden-code decision is still open; this run adds no abstentions. None of the four cards is a machine-layer card, so the count of documented abstentions stays at thirteen alongside the eighteen older cards that tag TKG in one sense or the other. The founder's call is unchanged: give the machine layer its own code and remap, or restate the beat file so TKG means one thing everywhere.

Honesty note, and one tier judgement openly reversed. Four verified, no signal — the fifth all-verified run in a row, which RDKG_GOAL.md calls "a yellow flag, not a triumph." Before the standing explanation, the thing this log owes the reader: the previous log judged this run's headline card a signal, and this run carded it verified on the same evidence. The 4 September log wrote that "the candidate that would have earned an honest signal — CARDIO-TTRansform, whose only surfaced primary is a congress press release about a presentation — is parked for capacity rather than rushed." The NEJM paper is still unopened; the provenance is still the ESC press release. The tier moved because, on L63's rule, a congress press release from the society that hosted the presentation is a primary document with a face date and a named presenter, and provenance is what tier measures — but that reasoning was available four days ago too, so this is a reversal of judgement and not a change in the evidence, and the founder should feel free to overrule it. The standing explanation follows (L52/L63: tier measures provenance), and the more useful thing this run can say is where the confidence actually is low, in prose: the FDA announcement publishes no effect size at all, so "significantly better walking speed" is the agency's word with no number behind it; the concrete paper's headline magnitudes come from ETH's institutional summary rather than from the paper's body, which was not read past the abstract; the concrete paper is itself a modelling exercise with four worked examples and no field validation, and its authors say so; the Roblox announcement is a scheduled behaviour change with no measured impact, whose only quantity is the phrase "a very small percentage"; and CARDIO-TTRansform is carded from a congress press release rather than from the journal paper, whose absence also means the trial's confidence interval — 0.73 to 1.09 — is the only thing standing between "missed its endpoint" and "does not work," and the card says so in its title. The candidate that would have earned an honest lower tier on provenance grounds — the NVIDIA/Hugging Face acquisition claim — was not carded at all, which is the correct handling rather than a tier decision.

Process note (what the review cost and caught). An independent adversarial review of the four drafts and this log returned sixteen findings. Six became corrections above or lessons L81–L83. The rest were repairs: two inverted cross-references, four blanket "every quotation comes from" claims that the cards' own text falsified, an ETH communications sentence attributed to "the authors," a capsule that generalised a subgroup result to a whole age band and dropped its comparator, a one-day event interval written as four, three unsourced claims in a card whose provenance set could not carry them, and four falsifier limbs that could only fail if a publisher edited a page. The last of those is now L83 and is the finding most worth keeping: a limb that tests transcription is a provenance re-check, not a falsifier, and every card in this run now labels it as one and carries at least one limb that can fail with the primary unchanged.

On the window and the cadence. Four days, 4–8 September, the Tuesday beat run as designed. Freshness: seven of nine veins had a real look inside the last 14 days; crypto & agent-commerce has been looked at and empty five runs running, and agriculture & robotics has now gone two consecutive runs empty in-window, which is the clock to watch. The 11 September Friday beat should open with the Venice Immersive awards (12 September, so possibly one beat later), Roblox RDC, the CHMP index, and the two one-fetch promotions sitting in the parked list — the Google Ads post's date and the SHEIN code-00625 announcement list.

How to read this log (the action view). The tiered list above is the provenance axis. The buckets below are the action axis. Tier and bucket are independent.

• 🆕 New (net-new this window). Five cards carry primaries dated inside the 1–4 September window on the face of the document: the OpenAI safety overview (3 September), the Broadcom 8-K (2 September), the Google post (2 September), the Roblox post (3 September) and the Epic release notes (3 September). One card is a promotion: the ACACIA-HCM release (28 August), parked by the last run inside its highest-value ESC cluster with the instruction to "promote from nejm.org and the escardio.org hot-line index" — the NEJM page returned no readable text (client-rendered), so the card rests on the ESC release, and is carded on its own date. Every primary was opened and quoted from the main thread; the one date that was not on the face of the document it belonged to — the main GPT-6 Astra launch post — was taken from its same-day companion safety overview and corroborated by three third-party datelines, with the gap stated on the card.
• 🔁 Changed (updates something we already track). Five of six touch standing cards directly; the sixth (ACACIA-HCM) extends the approval-evidence ledger under a new heading. Roblox appends a dated update note to roblox-build-prompt-to-publish-alpha (alpha in one country → beta in three, company-stated). Broadcom extends the financing thread's seller side beside nvidia-q2-fy2027-compute-is-revenue and prints a cross-vendor comparator beside the cash-conversion question that card asked of its own next filings, without interpreting it. Gemini hands inference-cost-collapse a dated test (1 January 2027) and hands perishable-frontier-model-shelf-life-weeks a cadence datapoint (three releases in six weeks) without closing either falsifier. Fortnite v42.10 extends unreal-mcp-in-uefn-agentic-authoring (batching) and leaves its publishing-rights limb open; the 16 July NPC card's actor-consent limb went unchecked again. GPT-6 Astra adds one datapoint to healthbench-professional-medical-ai (63.4 versus the 60.5 that card already carries, same model-graded caveat).
• 🎯 Opportunities flagged (act / position / build / money).
  1. Defensive security review is now a frontier-model task the vendors say is available today. OpenAI's own words: defenders can use Astra for "secure code review and patching"; Anthropic's page says Fable 5.1 "can now be used to discover software vulnerabilities"; Google's Flash Cyber is gated but Flash is not. A scoped, human-read pass over the gardens' codebases is cheap and was not possible on these terms a week ago. Nothing here diverts engineering from the BKG P0 — this is a review, not a build.
  2. Write the 1 January 2027 price step into the machine lane's cost model. It is a small number on a small model, and it is the first vendor-printed date for a frontier-tier price going up rather than down that this cabinet has recorded (cabinet-scoped). The perishable-model posture (model choice as configuration) is what makes it a switch rather than a rewrite.
  3. Design The Call's threshold for a party. Epic's API makes "who came in together" a readable fact; the Roblox footnote says the median DevEx program member received about $1,500 in the year to June 2026, so participation has to be paid in recognition, not revenue. Both are design constraints for the Story Garden's flagship, stated as such on the cards and never as facts about the story.
  4. The financing thread has its second vendor. Broadcom's guide — $21.7 billion of AI semiconductor revenue in one fiscal quarter, stated by the seller — is the purchase-side quantity that the cabinet's datacenter power, water and construction cards ultimately rest on; the quarterly check is the guide against the print and the receivables line beside it.
• 📚 Good to know / integrate (context, no action now). Three gating structures, not one. OpenAI ships one model with restrictive defaults and a program for looser safeguards; Google ships a separate gated variant; Anthropic ships the same model under two names with two safeguard levels. The Astra card claims only what the three pages say — that all three describe a limit on their strongest cyber capability, with OpenAI's leg the weakest, since its looser safeguards are promised rather than gated — and carries the falsifier. The monitorability sentence. A vendor has stated in its own launch document that its model can evade its monitors under adversarial conditions; that is a reason to keep verification of agent output on our side of the boundary, which is what the machine lane's design already assumes. The ACACIA-HCM arms. Placebo improved 8.4 KCCQ-CSS points; the drug's advantage over that was 3.0. Both numbers travel with the card so neither is quoted alone. ESC dates. The congress ran 28–31 August per the ESC release opened this run; its press release is dated 28 August on its face, which is why the promotion carries that date and not this window's.

How this run worked (the loop, once). Read the ledger (RDKG_GOAL.md, all 72 lessons, the 1 September changelog, both beat files and RDKG_STORY_GOAL.md) → four parallel scouts under the search-only brief with per-scout caps of 9–10 (40 searches total: 10, 10, 11 and 9 — one scout overran by two and reported it unprompted) → main-thread promotion and verification work with 5 targeted searches and 11 document fetches (9 returned readable primaries; the NEJM page returned no readable text, and one Cytokinetics URL was rejected for length — a second, longer Cytokinetics URL was not attempted for the same reason; all recorded below) → deep-verified every keeper by opening the primary from the main thread and quoting it myself → carded six survivors and appended one dated update note → ran the standing checks (capsule word count, superlative grep on summary fields, one cabinet-scoped claim verified by grep — the others were not, which the review caught and L75 records — and the L62 duplicate check over the parked list) → ran an independent adversarial review of the drafts and this log, fixed the findings, and re-checked the fixes → validated the corpus to zero errors. The founder commits, builds and deploys; a card is not live until then.

An analyst disclosure, stated once. The model running this sweep is Claude Fable 5.1 — the model Anthropic's September page describes. That page was opened and is quoted on the Astra card as one leg of the three-lab gating pattern; it was not carded on its own, for two reasons given in the parked list, and the choice is disclosed here so the founder can overrule it. Every card this run was written from the text of the primaries as fetched, not from anything the analyst knows about itself.

Veins swept, and what each returned. Seven groupings covering the nine veins — the beat file's eight plus the Story Sweep's — across four scouts plus main-thread work. Machine layer: three cards (OpenAI, Broadcom, Google) — the week's shape; a fourth first-party release (Anthropic) is quoted, parked and disclosed. Health & longevity: one card (ACACIA-HCM, promotion); the rest of the ESC cluster is resolved below; no FDA or CHMP primary dated 1–4 September surfaced on a ten-search scout pass, which is a thin look and is recorded as one. Storytelling & the new medium (MythKG): two cards (Roblox, Epic) — the vein's first cards since 21 August, after two empty runs; the September calendar the last log predicted has begun to deliver. Marketing: no in-window primary; the Google Ads AI-disclosure post carried unlocated for two runs now has a first-party URL and a date (9 July 2026 — out of window; see parked). Fashion & resale: SHEIN's first-day dealings could not be confirmed from an in-window exchange document (see parked); nothing else surfaced. Crypto & agent-commerce: empty for the fourth consecutive run — no spec, tag, governance or issuer event dated 1–4 September across x402, AP2 or ACP; the reading from the 1 September log stands. Materials & building and agriculture & robotics: both searched, both empty in-window; the clocks reset.

Corrections made this run, on the record. Seven.

1. The ACACIA-HCM figures that circulated in this run's search summaries were the company's, not the congress's. The ESC release states 517 randomised at 182 sites and prints no responder analysis; the 518 / 160 sites / 53%-versus-13% figures came from search summaries of Cytokinetics' releases, which were not opened. The card uses the ESC figures and records the divergence. (The 1 September log carried none of these figures; it named the trial and its promotion path only.)
2. The GPT-6 Astra launch post carries no visible dateline in its fetched text. Its date (3 September) is established from the companion safety overview, which does, and corroborated by CNBC, Axios and 9to5Mac datelines seen in search. L71's rule — dates come from the face of the opened document — is why the safety overview, not the launch post, is the card's source_url.
3. Anthropic's release page is dated "September 2026" on its face, with no day. The 1 September day rests on trade datelines only. Recorded, and one of the two reasons the item is parked rather than carded.
4. Broadcom coverage carried customer names, a named custom-chip program and a gigawatt deployment target; none appears in the 8-K or its exhibit. Excluded from the card, which says so.
5. Roblox Studio release notes 693, carried since 28 August as "date not established," is attributed by search to October 2025, with current numbering in the range 699–708. Not opened; closed below on that search evidence, stated as search evidence.
6. This log's own first draft claimed the cabinet-scoped-grep step had run for every cabinet-scoped claim; it had run for one. Two false cabinet-scoped counts (a "second vendor card," a "third Epic release-notes card") reached the adversarial review and are fixed; the process sentence above now says what actually ran, and L75 carries the rule.
7. The 1 September log's "swept and empty in-window for the third run" for the storytelling vein was a miscount. The 21 August run carded unreal-mcp-in-uefn-agentic-authoring in that vein, so the empty runs were two (28 August and 1 September). This log says "two" and records the correction here rather than silently.

Closed this run (no longer carried).

• Roblox Studio release notes 693 — as above. Re-open only if the page is opened and dated in a live window.
• Tencent open-source model, "claimed 28 August" — a second run found no first-party artefact on Hugging Face or GitHub dated late August or September; closed as unverifiable. Re-open only if a dated org-page release surfaces.
• SINGLE-AF's identity — resolved, so the "no named intervention" flag closes: per multiple trade summaries it is a South Korean multicentre open-label trial of DOAC therapy versus no anticoagulation in 1,803 patients with atrial fibrillation at intermediate stroke risk, with a 24-month composite of 0.5% versus 1.5% — which implies a relative reduction of about 67%, not the "69%" the 1 September log carried; both are leads and the divergence is noted. No primary was located — no journal URL, no escardio.org page — so nothing about it is carded and the figures above are leads. The item moves to the parked list below with its identity attached.

Parked this run (flagged, not published — with the primary that would promote each).

• Anthropic — Claude Fable 5.1 and Mythos 5.1 (anthropic.com/claude-fable-and-mythos-5-1, dated "September 2026"). First-party, opened, quoted on the Astra card. Not carded for two reasons: the face date is month-only, and the analyst is the model described, which is disclosed above. The page's own load-bearing lines are recorded here so a future run can card it on a day-dated primary: "Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards"; "Cache reads now cost 75% less, or $0.25 per million tokens"; "For typical workloads, costs are reduced by around 25% relative to Fable 5"; Fable 5.1 "can now be used to discover software vulnerabilities — though not to develop exploits for them." Promote from the system card at anthropic.com/claude-fable-5-1-mythos-5-1-system-card or the news index if either carries a day.
• CARDIO-TTRansform (eplontersen, ATTR-CM) — a miss, per the ESC's own press release. The scout surfaced escardio.org/news/press/press-releases/eplontersen-trial-did-not-meet-its-primary-endpoint-in-transthyretin-mediated-amyloid-cardiomyopathy/ with a claimed 1,432 patients, primary composite rate ratio 0.89 (95% CI 0.73–1.09, p=0.277), and 77.2% TTR reduction at week 140 without clinical benefit. Not opened this run for capacity. This is a biomarker-versus-outcome disconnect of exactly the kind MKG exists to record — the strongest single carry into the next run. Promote from the ESC release on its own date; no NEJM or Lancet URL surfaced.
• SINGLE-AF — identity above; promote only from a journal page or an escardio.org hot-line page, neither located.
• PRECISE-MI — still no results and no Lancet URL; only the pre-congress "Hot Lines revealed" page surfaced. Carry one more run, then close.
• SHEIN — did dealings commence on 1 September? HKEX's own news release of 28 August (hkex.com.hk/News/News-Release/2026/260828news) announced the debut in advance and is pre-window; no HKEXnews announcement dated 1–4 September was surfaced by scout or main-thread search. One check settles it: the listed-company announcements for stock code 625 on hkexnews.hk, including any over-allotment or stabilisation notice, which the allotment card notes the company is obliged to publish.
• Cytokinetics' releases of 28 and 29 August (ACACIA-HCM and MAPLE-HCM). Surfaced, not read: the GlobeNewswire URL was rejected by the fetch tool for length, and the IR-site URL, longer still, was not attempted. Promote from the company's SEC 8-K, if one was furnished, whose EDGAR URL is short.
• Google Ads AI-disclosure labels ("How this ad was made"). First-party URL now identified: blog.google/products/ads-commerce/google-ads-ai-transparency-labels/, claimed 9 July 2026. Out of window and not opened; the marketing vein's cleanest backfill candidate — card on its own date when a run has room. Meta's AI creative suite still has no located first-party post.
• Roblox "Fall Games Preview" (2 September, Vlad Loktev) — first-party and in-window per the search summary, not opened, and deliberately not carded: the summary describes a genre preview. Noted for the RDC run (10–12 September).
• eBay Form 425, accession 0001193125-26-227086 — new search evidence this run: the EDGAR index reportedly shows filer CIK 1326380, not eBay's, with a claimed date of 15 May 2026. Not opened; if correct it closes the item as out of window and mis-attributed. The index header settles it.
• Carried without re-attempt: rusfertide's corroborating documents (Takeda's release and Protagonist Therapeutics' Form 8-K exhibit, accession 000110465926002724); NVIDIA Form 10-Q (accession 000104581026000075); Savers Value Village 10-Q (a Q2 filing exists with claimed net sales $448.2 million — filing date unestablished); EU textile Digital Product Passport JRC report (13 May 2026); USDA APHIS soybean petition COR-23134-4 draft PPRA (comments close 19 October); Wiley Journal of Field Robotics lettuce-weeding paper; Meta XR Operator (18 August, MythKG backfill); Meta Horizon Worlds mobile pivot (19 February, MythKG backfill); Qwen3.8 Flash and the EIA STEO Texas revision; COSMOS epigenetic-clock RCT (9 March 2026).
• Calendar for the Story Garden, all dates as claimed in search and not opened: Venice Immersive runs 2–12 September with awards on the 12th (re-verify venice-immersive-2026-best-of-worlds after); Roblox RDC 10–12 September; Meta Connect 23–24 September; ICIDS 2026 in Bangkok 30 November–4 December with a late-breaking deadline of 14 September. (An immersive-theatre lead is recorded in the leads paragraph below.)

Leads recorded, not parked (no URL or no first-party source; nothing to promote from yet). Meta "Muse Spark 1.3" — appeared in one search summary alongside the week's releases; no dedicated search run, no URL. Gateway Arts' The Parting Glass, an immersive site-specific production claimed for 2–13 September — surfaced only in a blog write-up with no first-party URL.

The TKG garden-code decision is still open, and this run adds four cards to the abstention (L41). Broadcom, GPT-6 Astra and Gemini are machine-layer cards; the Fortnite card carries a machine-layer item. All four leave TKG out of gardens, carry the link in domains, and state the abstention in a Garden-code note on the card, matching the nine earlier abstentions. That makes thirteen documented abstentions alongside the eighteen older cards that tag TKG in one sense or the other. The founder's call is unchanged: give the machine layer its own code and remap, or restate the beat file so TKG means one thing everywhere.

Two tooling facts worth a lesson each, logged in tasks.lessons.md as L73 and L74, with the review's process findings as L75 and L76. The fetch tool rejects URLs beyond a length limit — one Cytokinetics release URL was rejected and the other, longer still, was not attempted — and the fix is to prefer the short EDGAR or DOI form of the same document. And a launch post can carry no dateline at all while its same-day companion does; the companion is the source_url.

Honesty note. Six verified, no signal — the fourth all-verified run in a row, and RDKG_GOAL.md calls that "a yellow flag, not a triumph." The reason is the one the last three logs gave: tier measures provenance (L52/L63), and no candidate carded this run had a provenance defect — five primaries dated in-window on their face, one queued promotion on its own date. Where confidence is low it is written in prose on the card: every model benchmark is vendor-reported on vendor-chosen evaluations; Roblox's impact figures are company-commissioned estimates; Broadcom's AI line is a management category outside the reported segments; the ACACIA-HCM paper itself was not opened and the effect sizes are small. The candidate that would have earned an honest signal — CARDIO-TTRansform, whose only surfaced primary is a congress press release about a presentation — is parked for capacity rather than rushed, and is the first thing the next run should open.

On the window and the cadence. Three days, 1–4 September, both beats of the week run as designed, as they were on 28 and 31 July. Freshness: every vein has now had a real look inside the last 14 days. The 8 September Tuesday beat should open with CARDIO-TTRansform, the SHEIN dealings check, and whatever the Venice Immersive close and RDC produce on their own dates.

How to read this log (the action view). The tiered list above is the provenance axis. The buckets below are the action axis. Tier and bucket are independent.

• 🆕 New (net-new this window). Three cards carry primaries dated inside the 28 August – 1 September window on the face of the document: the rusfertide approval (FDA, 28 August, published 17:22 — after the last run's sweep concluded that day, so late-28-August was uncovered ground, the same boundary reasoning the 28 August log wrote down for 21 August), the OpenAI post (31 August), and the SHEIN allotment announcement (31 August). One card is a promotion: brepocitinib (27 August), parked in-window by the last run for capacity with the instruction to promote it first — carded on its own date. One card is a disclosed backfill: NVIDIA (26 August), inside the previous window, missed by that run's scouts, carded on its own date with the miss stated. No item rests on a date inferred from a URL slug, a wire summary or a search snippet; every primary was opened and quoted from the main thread.
• 🔁 Changed (updates something we already track). Three of five touch existing cards directly. The SHEIN allotment card closes limb 2 of shein-hk-prospectus-growth-decelerates's three-limb falsifier (the offering proceeding), leaves limbs 1 and 3 open, and appends a dated update note to that card rather than editing its body. The NVIDIA card extends the two-sided financing thread (alphabet-q2-2026-ai-capex-exceeds-cash-flow, microsoft-fy26-capex-funded-from-operations) with the seller's side — and its China-exclusion sentence gives the export-control thread (export-controls-chinese-efficiency-overhang) a standing quarterly quantity to re-check. The rusfertide and brepocitinib cards extend the approval-comparison habit the 28 August log started: the ledger now holds four evidence shapes under the same words "FDA approved" — single-arm accelerated approval (Tudriqev), randomized overall survival (daraxonrasib), randomized placebo-controlled on a burden endpoint (rusfertide), and randomized placebo-controlled with direction reported but magnitude undisclosed in the announcement (brepocitinib).
• 🎯 Opportunities flagged (act / position / build / money).
  1. Conversational ad inventory is now self-serve at global scale, per its seller. MktKG can test ChatGPT ads at small budgets in most large markets as of 31 August. The card's scale check keeps the enthusiasm honest — a $1 billion company-stated run rate is a third of one percent of Google's filed advertising quarter, annualized — but the targeting description ("the context of the current conversation") is a genuinely new surface, and the garden's citation-engine thesis says be present where answers form.
  2. The financing thread now has all three seats at the table. Buyer financing from capital markets (Alphabet), buyer funding from operations (Microsoft), and the vendor organizing over $500 billion of third-party capital, subject to definitive agreements (NVIDIA). The watch item is execution: definitive agreements, disclosed terms, and the receivables line beside the revenue line each quarter.
  3. The approval ledger is becoming a reusable HKG asset. Four approvals, four evidence shapes, each named on its card's face. The cheap habit — any claim citing an FDA approval says which shape it was — is now demonstrated across enough cases to write into any HKG-facing material.
  4. SHEIN on a disclosure calendar is a standing source acquisition for the fashion vein. The first interim report as a listed company is the trigger that turns a 1.1% quarter into a trend or a shock absorbed — and each filing is a fresh chance to catch fibre-composition, restricted-substances or circularity data the prospectus sections read did not contain.
• 📚 Good to know / integrate (context, no action now). The brepocitinib announcement's disclosure shape — an FDA approval press release reporting a randomized phase 3's result as direction only, with no effect size — is worth remembering as a pattern: the announcement is not the label, and the ledger should never cite one as the other. The rusfertide endpoint wording — the FDA page states the endpoint as weeks 20–32 and summarizes it as "during the 32-week period" in the next sentence; both sentences are quoted on the card so neither is misread as ours. NVIDIA's fiscal calendar — "Q2 FY2027" ended 26 July 2026; the card uses the fiscal framing throughout to avoid the calendar-quarter confusion. The SHEIN stabilization clock — Goldman Sachs (Asia) may stabilize until 26 September 2026, a date the announcement itself supplies, so early trading is not a market verdict until then.

How this run worked (the loop, once). Read the ledger (RDKG_GOAL.md, all 72 lessons, the 28 August changelog, both beat files and RDKG_STORY_GOAL.md) → three parallel scouts under the search-only brief with per-scout caps of 8–9 (26 searches total, all caps respected, every cap honestly reported) → worked the last run's parked list from the main thread with 8 targeted searches and 7 document fetches → deep-verified every keeper by opening the primary from the main thread and quoting it myself → carded five survivors → ran an independent adversarial review of the drafts and this log → fixed the findings and re-checked the fixes → validated the corpus to zero errors. The founder commits, builds and deploys; a card is not live until then.

Veins swept, and what each returned. Nine veins across three scouts plus main-thread promotion work. Health & longevity: two cards (rusfertide, brepocitinib) plus the ESC cluster parked below. Marketing: one card (OpenAI ads) — the vein's first since it was flagged unswept in both prior runs. Fashion & resale: one card (SHEIN allotment); a ThredUp impact-report lead carried a pre-window claimed date and was not opened. Machine layer: one disclosed backfill (NVIDIA); the vein's only in-window item was the OpenAI post, carded under marketing. Crypto & agent-commerce: empty for the third consecutive run — no in-window spec, tag or governance event surfaced across x402, AP2 or ACP; the newest dated artefacts remain the 14 July x402 Foundation launch, April's AP2 v0.2 and the 2026-04-17 ACP stable spec. Three consecutive quiet runs on a four-card vein now reads as a pattern: the rails shipped and the action has moved to identity (per the 22 August MCP roadmap card), not payments. Materials & building: empty in-window, and C-Crete is closed below. Agriculture & robotics: empty in-window; the soybean draft PPRA carries; an "Amazon testing Agility Digit" claim surfaced third-party with no date and is recorded here as an unverifiable lead, not parked. Storytelling & the new medium (MythKG): swept and empty in-window for the third run — Roblox's version-693 release-notes page surfaced but its date could not be established by search and it was not opened; the Meta developer-blog index (opened) shows nothing dated in-window; no festival or platform primary landed. The vein's calendar is the story: Venice Immersive runs 2–12 September, RDC runs 10–12 September, and ICIDS notifications were due 31 August — September should produce this vein's cards, and manufacturing one in a four-day August window would have been the error.

The largest known gap of this run, stated rather than hidden. ESC Congress 2026 sat in Munich exactly across this window (28–31 August, dates claimed in the congress's own press index, not opened), with hot-line trials and claimed simultaneous NEJM and Lancet publications — including an aficamten trial in nonobstructive hypertrophic cardiomyopathy (publication claimed 28 August) and the full CARDIO-TTRansform dataset (whose topline miss was disclosed in July, out of window). None was carded: the run's fetch budget went to the five cards above, health already carried two of five slots, and no ESC primary was opened. The cluster is parked below with promotion paths; a cardiology-literate pass at the NEJM/Lancet/escardio primaries is the single highest-value carry into the Friday run.

Corrections made this run, on the record. Four.

1. The Meta Horizon "Our Renewed Focus in 2026" post is dated 19 February 2026 — a scout offered it as plausibly in-window for the Worlds-goes-mobile story; opened from the main thread, its dateline is February. Seven months out of window, and now a named MythKG backfill candidate on its own date (the post is first-party and structural: "We're explicitly separating our Quest VR platform from our Worlds platform ... shifting the focus of Worlds to be almost exclusively mobile"). L71 discipline — a search summary never binds a date to a URL — did its job.
2. The NovoCure pancreatic-cancer device approval is February 2026, resolving a parked item carried since 21 August with "no date established." The last run's inference from the accession sequence ("weakly suggests early FY2026") was right. Out of window by months; available as a device-versus-drug backfill pairing with the daraxonrasib card if ever wanted; carry-forward closed.
3. A trial name ("VALOR") and specific efficacy figures for brepocitinib circulated in search summaries; the FDA announcement names no trial and prints no efficacy number. Neither appears on the card, which states the exclusion. The primary sets the names and the numbers (L34).
4. EIA "Today in Energy" article 67704 predates 24 July 2026 — the article-ID sequence on EIA's own index page (opened this run) runs 67884 (24 Jul) through 68064 (1 Sep), so 67704 sequences earlier than the oldest visible entry. Out of window; carry-forward closed. The same index shows nothing datacenter-specific in-window.

Closed this run (no longer carried).

• C-Crete EPDs 20353/20355 — closed as unverifiable, per the last run's own instruction. A fourth targeted search again surfaced only the syndicated late-July announcement restating the figures (43.1 kg CO₂-eq per metric ton for EPD 20353), and again no NRMCA registry record. The honest conclusion stands as pre-written: the registry records are not publicly indexed, and the announcement's figures remain leads. Re-open only if the registry pages surface or the EPD PDFs are supplied directly.
• CHMP August 2026 highlights — closed unless one appears. Two runs of domain-restricted searching have surfaced no August highlights page; a search summary attributes this to the August meeting being replaced by a written procedure, which is plausible and unverified — the meeting-dates PDF was not opened this run. Stop carrying as a promotion candidate; the September plenary re-opens the vein.
• EIA 67704 and the NovoCure device — closed as above, under Corrections.

Parked this run (flagged, not published — with the primary that would promote each).

• The ESC Congress 2026 cluster (highest value). Aficamten in nonobstructive HCM — publication claimed 28 August in NEJM; promote from nejm.org and the escardio.org hot-line index, and expect a Cytokinetics 8-K. CARDIO-TTRansform full results — presentation claimed 28 August; in-window novelty is the full dataset only, since the topline miss was disclosed in July. SINGLE-AF — a claimed 1,803-patient trial with a claimed 69% relative reduction in a composite endpoint, surfaced with no named intervention and no URL; verify before repeating anything. PRECISE-MI — claimed simultaneous Lancet publication, date not established.
• Rusfertide's corroborating documents. Takeda's press release and a Protagonist Therapeutics Form 8-K exhibit (accession 000110465926002724) surfaced in search and were not opened; they would add the collaboration structure the FDA page omits. The card stands on the FDA primary alone.
• NVIDIA's Form 10-Q (accession 000104581026000075, nvda-20260726.htm) — surfaced, not opened, filing date not established. Would add the filed-statement citation to the carded 8-K exhibit.
• A Tencent open-source model release, claimed 28 August — model name and first-party URL not established. Promote from the Tencent/Hunyuan Hugging Face or GitHub org page only.
• Google Ads AI-disclosure panel ("How this ad was made") and Meta's AI creative suite — both claimed August, both without located first-party posts. Promote from ads-developers.googleblog.com and the Meta for Business newsroom on their own datelines; roundup coverage is a lead, not a source.
• Meta XR Operator (MythKG backfill, better primary now identified). Meta's developer blog index, opened this run, lists first-party "Introducing Meta XR Operator: Close the Build-Test-Verify Loop for VR", dated 18 August 2026, described as letting "any MCP-compatible AI agent see, navigate, and interact with your running VR app in Meta XR Simulator." Out of window; promote by opening the post itself on its own date — it supersedes the release-note line carried since 28 August as the item's best primary.
• Meta Horizon Worlds' mobile pivot (MythKG backfill). The 19 February 2026 first-party post above — structural for the participatory-economy vein (Worlds "almost exclusively mobile", mobile MAU "over 4x in 2025", four creators past $1 million lifetime revenue, all company-stated). Card on its own date if the vein wants the history before RDC.
• Roblox Studio release notes 693 — page exists (create.roblox.com/docs/release-notes/release-notes-693), date not established by search, not opened this run. One fetch settles it; the week-of-24-August page was dated and empty when opened by the last run.
• eBay Form 425, accession 0001193125-26-227086 — carried, not re-attempted this run; the two conflicting order readings recorded on 28 August stand. The EDGAR index header settles it.
• Savers Value Village Form 10-Q — carried, not re-attempted; promote from the EDGAR index header.
• EU textile Digital Product Passport JRC report (13 May 2026) — carried; still the most on-thesis backfill candidate for the resale data-layer thesis; card on its own date when a run has room.
• USDA APHIS soybean petition COR-23134-4 draft PPRA — carried; comments close 19 October 2026; the draft PPRA is the substantive read.
• Wiley Journal of Field Robotics lettuce-weeding paper — carried, not re-attempted; promote from the article page's own publication date.
• SHEIN follow-ons — whether dealings in fact commenced 1 September (one check), any over-allotment exercise announcement (the allotment card notes the obligation to announce), and the first interim report (the trend test).
• Qwen3.8 Flash (claimed 26 August) and the EIA STEO Texas datacenter revision (claimed 11 August) — both pre-window edge, both unestablished; check on their own primaries only if a run has spare budget.
• Dropped deliberately, not parked: an AITX micro-cap 8-K (in-window but negligible foresight weight) and the undated third-party Amazon/Agility Digit claim.

The TKG garden-code decision is still open, and this run adds two cards to the abstention (L41). The OpenAI ads card and the NVIDIA card both carry machine-layer content; both leave TKG out of gardens, carry the link in domains, and state the abstention in a Garden-code note on the card, matching the seven earlier abstentions. That makes nine documented abstentions alongside the eighteen older cards that tag TKG in one sense or the other. The founder's call is unchanged and getting cheaper to make than to defer: give the machine layer its own code and remap, or restate the beat file so TKG means one thing everywhere.

Honesty note. Five verified, no signal — the third all-verified run in a row, and by RDKG_GOAL.md's own metric that is "a yellow flag, not a triumph." The reason is the same as both prior runs': tier measures provenance (L52/L63), and no candidate carded this run had a provenance defect — three primaries are dated in-window on their face, one is a queued promotion on its own date, one is a disclosed backfill on its own date. Where confidence is low it is written in prose on the card: every OpenAI figure is company-stated and unaudited; the brepocitinib announcement reports no effect size; the SHEIN card asserts an expectation of dealing, not a listing; NVIDIA's inflection-point framing is its CEO's assertion. What would improve the ratio honestly is a lead with a real provenance gap worth publishing at signal; the four-day window did not produce one, and the ESC cluster — which might have — is parked rather than rushed.

On the window and the cadence. Four days, 28 August – 1 September, and this is the log's first run to land on the designed Tuesday beat. The freshness picture: seven of nine veins had a real look inside the last 14 days; materials & building and agriculture & robotics returned empty in-window on scout searches but were searched, so the clock resets; marketing — the previous laggard — now has a card. The 4 September Friday beat should lead with the ESC cluster and the MythKG September calendar.

How to read this log (the action view). The tiered list above is the provenance axis — can we stand behind it. The buckets below are the action axis — what the team should do about it. Tier and bucket are independent.

• 🆕 New (net-new this window). All six items carry a primary dated 21–28 August 2026 on the face of the document. The six in-window primaries are: one FDA drug-approval notice (CDER), one FDA device-authorization announcement, one Federal Register determination, one peer-reviewed paper, one first-party governance post, one exchange filing. Two further primaries were opened as corroboration and are out of window or secondary to the claim: the FDA press announcement accompanying the drug approval (same date, used only for epidemiology) and the March 2026 MCP roadmap (used only to confirm supersession) — eight documents in all. No item rests on a date inferred from a URL slug alone, a wire-service ID or a search-engine summary, and this run rejected three candidates on exactly that ground, two of which a scout had offered as in-window and which turned out to be dated May and March (see Corrections).
• 🔁 Changed (updates something we already track). Four of six touch existing cards. The two below are the ones that change what an existing card means; the other two — the aging-clock paper promoting dunedinpace-aging-clock, and the daraxonrasib card built as a contrast to fda-accelerated-approval-tudriqev-melanoma — are handled under Opportunities. Both of the two below are partial in ways worth naming. The MCP roadmap gives independent first-party confirmation that the 2026-07-28 revision shipped and says what it changed — which retires the last of the confusion the 31 July log created and the 21 August log corrected. mcp-standardization's falsifier is pasted whole on the card and its three limbs are enumerated there: limb 1 ("slips") is closed by this second confirmation; limb 2 ("major vendors ship incompatible forks") is open but favourable — the post names no fork and describes a functioning SEP and Working Group process, which is evidence from an interested party rather than a survey; limb 3 ("Tier-1 SDKs miss the ten-week validation window") is not closed, since the strongest thing the post says about SDKs is the hedge "you might've already seen them in our SDKs and documentation." A claim that all four Tier-1 SDKs speak 2026-07-28 circulated this sweep from a search-engine summary of a page nobody opened, and is not restated anywhere. The chestnut deregulation complicates rather than confirms usda-deregulates-mon-87429-multi-herbicide-corn: same instrument, twenty-four days apart, applied to a forest tree engineered for fungal-pathogen tolerance instead of a herbicide-tolerance crop trait, so the garden's standing criticism of the instrument still holds but the unanswered questions change from chemical to ecological.
• 🎯 Opportunities flagged (act / position / build / money).
  1. The Frontier Feed spec has three concrete changes to absorb, and they all make it easier. The shipped MCP release removed protocol-level sessions and the initialization handshake, and the maintainers write that "a remote MCP server is now no different from any other HTTP workload, making it easy to host and operate one on any infrastructure that developers and organizations already use for their APIs and services." A stateless feed served from a static export is now the protocol's mainstream shape, not a compromise — which matters because RDKG is a static export. The Server Card Working Group's .well-known metadata work is the same idea as llms.txt with a standard behind it, so the manifest should be designed to become a server card. And progressive discovery exists because "connecting to a server with a hundred tools means the model pays for that entire surface before the user has asked a single question, and tool selection tends to get worse as the list grows" — the cabinet is at 114 signals and growing, so a small entry point that widens is the right shape from the start. Still spec-and-stub; still not engineering time.
  2. Agent identity is the missing half of the agent-commerce thesis, and it is now named as missing by the protocol itself. The cabinet holds four rails cards — x402-linux-foundation, x402-foundation-operational-launch, ap2-agent-payments-rails, visa-stablecoin-platform-open-usd — all about how a machine pays. The roadmap says MCP authorization "is built around a person approving access in a browser." Rails without identity is a payment system that cannot name its customer. Watch DPoP and Workload Identity Federation as the dependency for the whole machine-lane revenue story, not as a footnote.
  3. HKG now has a citable specification for which aging biomarker to buy. The Nature Medicine paper's own recommendation — prioritise generation 2+ clocks, "particularly DunedinPACE and PCGrimAge" — promotes the garden's existing dunedinpace-aging-clock card from interesting to load-bearing, and its two-rule consistency test is directly reusable on any future intervention claim. Two scope notes the card makes and this log repeats. That DunedinPACE card's own falsifier already named responsiveness (it fails if the clock "fails to move with interventions that demonstrably extend healthspan"), so the garden was not blind to the question — what is new is a systematic answer. And the paper covers DNA-methylation biomarkers only, so protphenoage-proteomic-aging-clock is out of its scope. On senolytics, the five studies failing both rules are an independent negative on the human DNA-methylation side; they do not test what senolytics-preclinical-multiorgan-rejuvenation claims, which is a mouse preclinical result already held at signal and already conceding that no human RCT shows a healthspan effect. The symmetrical opportunity is negative and just as valuable: with no minimum clinically important difference defined, anyone selling a clock reading as a health outcome is ahead of the evidence, and the garden can now say so with a citation.
  4. The two-approvals comparison is a reusable analytic asset, not just a pair of cards. Twenty days apart the same agency cleared Tudriqev — accelerated approval, single-arm, 91 evaluable patients, as a combination with nivolumab, benefit unproven — and daraxonrasib, a full approval on a 500-patient randomised survival comparison against an active comparator. Both are correctly described as "FDA approved." HKG should carry the distinction on the face of any claim citing an approval; that habit is cheap to adopt and is the kind of thing a verified garden should be known for.
  5. A regulatory pathway worth its own card. Daraxonrasib is the Commissioner's National Priority Voucher programme's first appearance in this cabinet — a claim scoped to our shelf, not to the world — and the FDA says the review finished "approximately 6.5 months ahead of the FDA goal date." Whether a voucher changes the evidence required or only the calendar is testable and unanswered. Open a watch on the CNPV pilot programme page and on the next two CNPV approvals. (A coincidence the cards flag so nobody builds on it: the review ran ~6.5 months early and the absolute survival difference is also 6.5 months. Unrelated.)
  6. A cheap, honest data product on the non-toxic spine, restated because it recurred. The chestnut card reproduces the structural gap the corn card found: the agency that permits an organism is not the agency that measures what happens afterwards, and for a permitted release of a plantable organism with the EIS terminated there is no monitoring record in the garden at all. This is a spec, not a card. It is also the second time in a month the same gap has surfaced, which is the argument for writing it down.
• 📚 Good to know / integrate into our data (context, no action now). SHEIN's mix shift — apparel fell from 68.8% of annual revenue in 2023 to 63.8% in 2025, with the like-for-like quarterly comparison a smaller 62.0% to 61.4%, while rest-of-world rose to 45.3% of quarterly revenue; less of a clothing company and less of a Western one, though the quarterly apparel move is half a point and should not be read as a continuation of the annual slope. An unexplained oddity in the same filing — net margin exceeds operating margin in four of the five periods printed (all but Q1 2026), implying material non-operating income across the whole period, which the sections read do not explain; resolving it needs the income statement. And a reconciliation limit — six of the fourteen printed ratios (the 41.1% 2023 growth rate and all five operating margins) cannot be checked against the printed tables and are quoted on the company's authority. The chestnut file's data-integrity story — a six-years-and-seven-months regulatory review in which the identity of the organism was mislabelled from 2016 until 2023, found by the petitioner's own analyses and corrected on the record. The FDA's own transcription — the daraxonrasib notice writes Japan's regulator as "PDMA"; the initialism is conventionally PMDA, and the card quotes the letters as printed so a future reader does not think it is ours. MCP governance mechanics now exist — a formally adopted Contributor Ladder, Working Groups triaging SEPs in their own area, and a feature lifecycle and deprecation policy that the 2026-07-28 deprecations were the first to follow.

How this run worked (the loop, once). Read the ledger (RDKG_GOAL.md, all 67 lessons, the 21 August changelog, and both beat files plus RDKG_STORY_GOAL.md) → assigned nine veins to five parallel scouts under a strict search-only brief with a per-scout cap of 9–10 searches → deep-verified every keeper by fetching the primary from the main thread and quoting it myself → carded six survivors → ran an independent adversarial review of the drafts, which returned thirty-two findings → fixed all of them → re-ran the review against the fixes, which returned twenty-one more, eight of them caused by the fixes themselves → fixed those → logged them here → validated the corpus to zero parse errors. The founder commits, builds and deploys; a card is not live until then.

Veins swept, and what each returned. Nine veins were assigned to five scouts, and two of the nine got only partial coverage — marketing was never independently searched, and robotics was not reached before its scout's budget ran out (both are in Coverage gaps below). Scout 1 took human health & longevity. Scout 2 took AI & the machine layer together with crypto & agent-commerce. Scout 3 took storytelling & the new medium. Scout 4 took materials & building together with agriculture & robotics. Scout 5 took marketing, fashion and resale as one commercial brief. Cards landed as: health (3 — oncology, aging biomarkers, metabolic device); agriculture (1); machine layer (1); fashion, resale and marketing (1). Three health cards is a heavy tilt and it is stated rather than hidden; it reflects where the dated primaries were, not a change of emphasis.

Three veins returned nothing, for three different reasons, and the difference matters.

• Storytelling & the new medium (MythKG) — swept and genuinely empty. No card, and no card was manufactured. Epic's v42.00 notes are dated 20 August, one day before the window, and are already carded (unreal-mcp-in-uefn-agentic-authoring). Meta's Horizon developer release-notes index was opened directly from the main thread and its newest dated entry is 29 July 2026, so the entire Meta XR surface is out of window. The Roblox Studio release-notes page for the week of 24 August exists, is dated ("Last updated on Thursday August 27, 2026"), and its Improvements and Fixes tables are both empty in the fetched page — a dated page with no capability in it is not a card. No festival announced a selection in-window; Venice Immersive's 10th-edition selection is already carded and its festival runs 2–12 September; ICIDS author notifications are 31 August, so those proceedings are structurally empty until autumn. Roblox's own creator-economy announcements cluster at RDC, 10–12 September. The honest read is that this vein's calendar puts its next real event in September, and the card to expect then is a festival or an RDC announcement.
• Crypto & agent-commerce — empty, for the second consecutive run. Ten searches across x402, AP2 and the Agentic Commerce Protocol found no in-window spec release, version tag or governance event; the newest dated artefacts remain the x402 Foundation's 14 July operational launch (already carded), AP2 v0.2.0 from April, and an ACP stable spec dated 2026-04-17. Two consecutive empty runs on a vein the garden has four cards in is itself worth noting: the rails shipped and are now quiet, which is consistent with the identity gap the MCP roadmap describes being the actual blocker.
• Materials & building — one parked lead and no card. C-Crete's EPDs remain unpromotable (below). Cement papers, mass-timber code actions and EPA drinking-water rulemaking were all searched and all silent in-window.

Coverage gaps this run, reported as gaps rather than as findings. Scout budgets ran out before these were searched, and a silence nobody looked for is not a negative finding: ARG and immersive-theatre production records; Unity release notes; mycelium and bio-based insulation; PFAS-free building products outside the drinking-water rules; USDA NASS chemical-use data releases; humanoid and field-robot first-party technical write-ups; vertical farming; patents in any vein; EPD program operators other than NRMCA; and — flagged by the commercial scout itself — advertising-platform first-party announcements, meaning the marketing vein was not independently searched this run and its coverage came only via the SHEIN and MCP cards.

Corrections made this run, on the record. Five, three of them dates.

1. The NVIDIA ACE multilingual item is not in-window and is not a card. A scout surfaced "What's New for Game Developers in NVIDIA RTX: DLSS 4.5 for UE5 and Multilingual AI Characters" under an August-scoped query with no date established, as the run's best MythKG capability lead. Opened from the main thread: it is bylined May 27, 2026 (published_time: 2026-05-27, modified 11 June). Three months out of window. Substantively it is a real medium-capability item — NVIGI SDK 1.6 bringing Qwen 3.5 4B across "201 languages and dialects", Riva Parakeet TDT 600M ASR for 25 languages, and Chatterbox Multilingual 500M for 24 — and it is now a MythKG backfill candidate on its own date, not an August signal.
2. Two MCP roadmap URLs differ by four characters and the search summary attributed the date to the wrong one. /posts/2026-mcp-roadmap/ is "The 2026 MCP Roadmap", 9 March 2026, four priority areas. /posts/mcp-roadmap/ is "The New MCP Roadmap", 22 August 2026, five priority areas. Both were opened before the card was written, precisely because the scout could not tell which carried the in-window date and said so. Carding the March post as August would have been a clean, invisible error.
3. The Nature Medicine result was being summarised wrongly in secondary coverage, and the card corrects it against the paper's own text. Coverage framed it as pharmacology beating lifestyle, naming anti-TNF and metformin as producing strong changes "more so than supplements, diets or lifestyle changes." The paper's abstract says "pharmacological and lifestyle interventions drive the strongest responses" — both — and its worked example of an intervention satisfying both consistency rules includes two Mediterranean diets. The real split is consistent versus inconsistent, which cuts across the drug/lifestyle line. Metformin appears in no passage opened in-session and is not asserted on the card.
4. arXiv 2608.04458 is resolved and permanently parked. The 21 August log parked it with "no day-of-month" established. The submission date surfaced this run: 5 August 2026, sixteen days before this window opened. It was out of window last run too. Stop carrying it.
5. A participant count for the TranslAGE database is excluded, not merely unstated. A sample figure circulated in secondary coverage of the Nature Medicine paper. It appears in no passage opened in-session, so it is described here without being repeated, and the card quotes only the study, clock and biomarker counts the paper's own text gives.

The self-audit did substantial work, and it is worth saying what it caught. Two passes. The author's own pass, reading each card back against its own provenance paragraph in both directions (L66), produced the "PDMA" transcription note, struck the RASolute 302 trial name from a draft that had taken it from a scout snippet and restored it only once the CDER page was opened on its own authority, named the daraxonrasib label-versus-trial-population gap explicitly, and reframed the SHEIN chemical-disclosure question from a finding into a stated absence in what was read.

Then an independent adversarial reviewer read all seven files cold and returned thirty-two findings, every one of which was fixed before this log was finalised. The reviewer's arithmetic pass came back clean — all of SHEIN's checkable growth rates and net margins, both HK$ proceeds figures, the offer-share split, every table column's internal sums, the 6.5-month survival difference and every date interval reconcile — but the prose did not, and the misses cluster into four kinds worth recording because three of them are repeat offences.

Arithmetic that was stated rather than computed. Rest-of-world quarterly share was written as 45.4% and is 45.3% (4,104 ÷ 9,052 = 45.34%). The margin oddity was flagged for two periods when it holds in four of five. "Decayed by roughly half at each step" was false for two of three steps (ratios 0.50, 0.39, 0.14). A blanket "the ratios reconcile" obscured that six of the fourteen printed ratios cannot be checked without the income statement (the table holds fourteen: four growth rates, five operating margins, five net margins). And the apparel-share bullet spliced a quarterly figure onto three annual points, making a half-point quarterly move look like a continuation of a 2.5-point annual slope; the like-for-like pair 62.0% → 61.4% is now given.

Blanket claims that hide their own exceptions — L45, recurring. Two sentences of the form "the only derived figures on this card are…" and "every quotation comes from that page" were both false, and false by construction: a blanket makes a missing item undetectable. Both are now enumerated lists. Related and worse: the daraxonrasib card contained a self-certifying sentence asserting it used the word "first" only inside a quotation, while the same card said "the first CNPV product to reach the cabinet" in its own voice four paragraphs earlier — and this log repeated the false audit claim while itself containing the contradiction. Both are fixed, and the cabinet-scoped "first" is now labelled as cabinet-scoped rather than deleted.

Framing that outran its own provenance — the L66 provenance-contract failure mode specifically, third consecutive run. The chestnut card's entire premise, that the tree is meant for wild restoration, appears nowhere in the Federal Register notice it cites; "stated purpose" attributed a statement to a document that never makes it. That framing is now explicitly labelled reported context on the card, with the petition named as the document that would settle it — and three re-assertions of it in this log had to be pulled out on the verification pass, which is L67 repeating itself. Two smaller instances of the same shape: the daraxonrasib card used "pan-RAS" — a pharmacological classification neither FDA page supports — in its title and nowhere else, and asserted RAS's rank among oncogene families and its history as a target from background rather than from either primary; and the ketone card described the De Novo predicate mechanism as though the FDA page said it. All three now carry a labelled boundary. The SHEIN card's "largest operator", "cheapest, highest-turnover clothing channel on earth" and "clearest case in the cabinet" went the same way: no volume, unit or share figure is quoted anywhere on that card, and it says so itself.

Cross-references asserted rather than opened — L46 pointed at our own shelf. Five defects, and each one flattered the new card at the referenced card's expense. mcp-standardization's falsifier was quoted truncated, with a full stop inserted inside the quotation marks and the operative consequence clause — "leaving no single interoperability standard for agents to converge on" — dropped; that is L64's exact recorded failure, repeated, and the falsifier is now pasted whole with its three limbs enumerated and their status given individually. semaglutide-cardiovascular-outcomes was called an oral drug; it is a weekly injectable. protphenoage-proteomic-aging-clock was named as something a DNA-methylation paper grades; it is a plasma-proteomics clock and out of scope. hormesis-fasting-exercise-longevity-mechanism was credited with a "metabolic switching" mechanism it never mentions, silently upgrading a mechanism that card is held at signal specifically for not endorsing. And dunedinpace-aging-clock was framed as testing only prognosis when its own falsifier already named responsiveness — an understatement of what the garden already held. Two quotations in this log were also truncated in the same way as the falsifier and now carry their full clauses.

The reviewer also caught six counting or bookkeeping errors in this log itself: "the previous run's six cards" (it had five), "fifteen older cards that do tag TKG" (eighteen — a figure carried forward from the 21 August log unchecked), an eleven-card abstention count that did not survive contact with which cards actually raise the question, "four of them dates" (three), a primary count that enumerated seven documents under a heading saying six, and a "swept nine veins / every vein looked at inside 14 days" claim contradicted by this log's own coverage-gaps paragraph. All corrected above.

Multi-limb falsifiers were written as enumerated limbs from the start (L64) — three limbs each on five of the six cards, two on the sixth — and both reviewers confirmed that count and found all six testable rather than restatements.

Then the verification pass, which was not a formality either. Re-run against the fixes with an explicit brief to hunt regressions, it returned twenty-one further findings, eight of them created by the repairs. Three are worth recording because they are the same species of error: a fix that moves a defect rather than removing it. The chestnut restoration framing was correctly relabelled on the card and left standing three times in this log, in the log's own voice, inside the very paragraph claiming it had been fixed — L67's recorded precedent word for word. Replacing blanket claims with enumerations produced incomplete enumerations twice, and a miscount of a fourteen-row ratio table as eleven; an enumeration that is itself incomplete is worse than the blanket it replaced, because it now carries a false claim to completeness. And a new "marked as inference wherever it appears" sentence was written and then not honoured in three places, one of them a summary field — the L68 shape, created by an L68 fix. The pass also found four blanket claims and three truncated quotations that the first review had missed entirely, and one thing no audit here had ever checked: four of six capsules broke RDKG_GOAL.md's own 40–80-word rule, one of them at 100 words, partly because provenance fixes lengthen capsules and nobody had been paying the word budget back. All twenty-one are fixed; the capsules are now 75–80 words. Lesson L72 records the three rules that follow — grep every fix across every file in the run, re-count every enumeration against the finished text, and never write a self-certifying marker you do not then go and apply. The broader pattern across the whole audit is the L42/L45 family — quotes sound, surrounding sentences overreaching — which has now recurred in four consecutive runs: the quotes are sound and the sentences around them overreach. The four new lessons are L68 (a blanket claim about a card’s own contents is unfalsifiable by construction — enumerate), L69 (open a referenced card before characterising it), L70 (the frame is a claim and needs provenance like any other fact) and L71 (a search summary never binds a date to a URL).

Honesty note. Six verified, no signal, off six distinct in-window primaries plus two corroborating documents — and by RDKG_GOAL.md's own metric that is "a yellow flag, not a triumph." The flag is raised here rather than papered over, and the reason is the same as last run's: tier measures provenance, not confidence (L52), and no candidate this run carried a provenance defect worth demoting for. Where confidence is low, it is stated in prose on the card instead — the MCP roadmap is explicitly low-confidence as delivery, the daraxonrasib survival figures are explicitly immature, the Libre Duo priority claims are explicitly the FDA's and unchecked, and the SHEIN card explicitly reports reading a partial extraction of a very large document. What would have improved the ratio dishonestly was demoting a well-sourced card; what would improve it honestly is a vein producing a real lead with a real provenance gap, and this window did not produce one.

On the window and the cadence. Eight days, 21–28 August, which is close to the designed rhythm and a real improvement on the 21-day gap the last run had to apologise for. Three honest notes. The Tuesday beat (25 August) was missed, so this is the Friday beat carrying a week; twice-weekly remains the design and was met once. The window's opening boundary is doing real work: the previous run's five cards carried primaries dated 4–20 August, so 21 August was uncovered ground, and this run's Nature Medicine paper is dated exactly 21 August. It is included on that basis and the reasoning is written down here so nobody has to reconstruct it. And the freshness metric is partly met, not met: seven of the nine standing veins had a real look inside 14 days, while marketing was not independently searched and robotics was not reached, so for those two the 14-day clock is still running from an earlier run.

The TKG garden-code decision is still open, and this run adds two cards to the abstention (L41). TKG carries two live meanings in this repo — Toxicology per the Frontier Map, the machine layer per this beat's own wording. This run would again have had to guess in both directions at once: the chestnut card's content is toxicology-adjacent, the MCP card's is machine-layer. The other four cards this run raise no TKG question that a primary could answer — note in particular that the SHEIN card’s non-toxicity content is an explicit absence in what was read rather than a toxicological finding, which is why it is not counted here; that is a judgement call and it is the same kind of judgement call that killed the “eleven-card” count in the first review. As on the last two runs, TKG is left out of gardens on those two cards and the link is carried in domains, and — a fix the adversarial review forced — each of the two now carries an explicit "Garden-code note (L41, flagged not resolved)" paragraph in its body, matching the five earlier abstaining cards, so the decision is visible on the card and not only in this log. That makes seven documented abstentions, alongside eighteen older cards that do tag TKG in one sense or the other. (The "fifteen" figure in the 21 August log was wrong and had been carried forward unchecked; the correct count, from the files, is eighteen.) One thing to be explicit about: L41's written rule-until-resolved says "within the Standing Frontier Sweep, machine-layer cards use TKG + RDKG", and the MCP card departs from that rule — the departure is stated on the card itself rather than made silently. The founder's call is unchanged: give the machine layer its own code and remap, or restate the beat file so TKG means one thing everywhere. It gets cheaper to decide now than after another twenty cards.

Parked this run (flagged, not published — with the primary that would promote each).

• FDA approval of the first oral drug indicated for dermatomyositis — Lisraya (brepocitinib), 27 August 2026. In-window, first-party, and genuinely cardable; parked only because this run capped at six cards and three were already health. The FDA announcement (fda.gov/news-events/press-announcements/fda-approves-first-oral-drug-indicated-treat-dermatomyositis-adults) was not opened and no efficacy figure surfaced in search. Promote by reading it together with the CDER approval notice — this is the strongest single item carried into the next run.
• FDA "first-of-its-kind device to treat pancreatic cancer" (fda.gov/news-events/press-announcements/fda-approves-first-its-kind-device-treat-pancreatic-cancer), with a companion NovoCure Form 8-K exhibit at accession 000164511326000010. No date established — the result surfaced as title and URL only. The low accession sequence weakly suggests early FY2026, which is inference, not a date. One cheap fetch settles it, and if in-window it pairs directly with this run's daraxonrasib card as a device-versus-drug comparison in the same indication.
• CHMP August 2026 meeting highlights. EMA's own published meeting-dates PDF puts the August plenary at 17–20 August 2026, which would put a highlights page on or about 21 August — inside the window. Two domain-restricted searches failed to surface it; the newest indexed highlights remain 20–23 July. Promote from ema.europa.eu/en/committees/chmp/chmp-meetings directly rather than via search. This is the largest unresolved gap of the run: an in-window European regulatory decision set may exist and simply be unindexed.
• Journal of Field Robotics, "field-adaptive mechanical weeding system for intra-row weed control in lettuce" (onlinelibrary.wiley.com/doi/10.1002/rob.70230). Peer-reviewed and on the herbicide-displacement thread that autonomous-laser-weeding already occupies, but no publication date was established — only the year 2026 in a citation string. Promote from the article page's own publication date; drop it if it is not 21–28 August.
• USDA APHIS soybean petition COR-23134-4 (Pioneer Hi-Bred) and its draft plant pest risk assessment. Federal Register 18 August 2026, docket APHIS-2026-1090, comments open to 19 October 2026, traits "Cry1B.34.1, Cry1B.61.1, and IPD083Cb proteins" plus GM-HRA, with the draft PPRA concluding the event "is unlikely to pose an increased plant pest risk compared to the nonmodified soybean." Out of window by three days. Promote by reading the draft PPRA itself, which is the substantive document; it pairs with both of the garden's APHIS cards as the same instrument applied to an insecticidal trait, where the dietary-exposure question is sharper.
• C-Crete clinker-free cement, third-party EPDs 20353 and 20355. Carried a third time and still not promotable. Announcement trail dated 28–29 July 2026; a targeted search again failed to surface the program-operator registry page. Established previously and still true: NRMCA is itself the program operator. Promote by opening the EPD records in the NRMCA registry. If a fourth run cannot find them, the honest conclusion is that the registry records are not publicly indexed and the item should be closed as unverifiable rather than carried forever.
• Meta XR Operator Standalone 205.1 — a MythKG backfill candidate, out of window. Meta's own release-notes index, read this run, dates it 28 July 2026 with the description "First release of Meta XR Operator enables AI coding agents to interact with and verify VR/MR apps, read room geometry, capture screenshots, and drive Quest Touch Controller input." That is an agent gaining hands inside an XR authoring and QA loop — the same class of capability as the already-carded Unreal MCP item, on a different platform, three weeks earlier. It was out of window for the last run too and was missed then. Promote by opening the package release-note page on its own date.
• Savers Value Village Form 10-Q for the quarter ended 4 July 2026 (sec.gov/Archives/edgar/data/0001883313/000188331326000060/svv-20260704.htm), reporting "282 million pounds of secondhand goods, compared to 279 million during the thirteen weeks ended June 28, 2025." A rare first-party volumetric measure of physical secondhand throughput, which is adjacent to the resale thesis's condition layer without touching it. Filing date not established and a 4 July quarter end would customarily be filed in early-to-mid August, i.e. probably before this window. Promote from the EDGAR index header.
• eBay Form 425, accession 0001193125-26-227086 (d151040d425.htm). Carried from the last run and still no date established. Two readings of the surrounding accession run now exist and they disagree: the 21 August log said it sequences after two 425s "dated to July", while this run's search surfaced the neighbouring accessions' content as May 2026 material. Both are ordering inference rather than a date, and the disagreement is recorded rather than resolved — one of the two is stale and reading the EDGAR header settles which. It is the document that would update gamestop-ebay-takeover-escalation.
• SHEIN allotment results and final offer price. Not yet published. The one timing fact taken from the sections of the prospectus actually read is that the price determination date is "on or before Friday, 28 August 2026" — today. A listing date and a results date circulated in trade coverage and are deliberately not restated here, because they would live in the Expected Timetable and Structure of the Global Offering sections, which were not opened. Promote from the allotment-results announcement, which converts this run's card from an offering into a listing and supplies the actual price.
• EU textile Digital Product Passport — a JRC Science for Policy report dated 13 May 2026. Out of window and the single most on-thesis document surfaced all run: its described content covers "fibre composition ... recycled content and its type of recycling, substances of concern (name, concentration)" — which is two and arguably three of the resale thesis's untouched layers in one primary. The delegated act is not adopted and binding rules were described as expected end of 2027. Worth carding out of window on its own date rather than waiting; a scoping document that specifies item-level chemical and fibre disclosure is exactly the standard the garden's data-layer thesis has been asserting must exist.
• CalRecycle SB 707 textile extended producer responsibility. Active California rulemaking with the agency publishing its own workshop notices; the dated event is a Textile Stewardship informal regulatory concepts workshop on 13 August 2026, eight days before this window. PRO registration deadline was 1 July 2026, stewardship-plan approval 1 July 2030. Named here because it is the vein most likely to produce a dated in-window regulator document on a future run.
• Roblox Studio release notes, week of 24 August 2026 (create.roblox.com/docs/updates/2026-08-24). Opened from the main thread: in-window and dated ("Last updated on Thursday August 27, 2026"), and empty — the Improvements and Fixes tables contain no entries in the fetched page. Not cardable. Promote only if the page later populates, or from a dated Roblox devforum or engineering post describing the same week.
• EIA "Today In Energy" article id=67704. Surfaced with no date established. EIA publishes on a near-daily cadence so an in-window date is plausible, and datacenter demand is a standing machine-layer thread (eia-aeo2026-datacenter-demand). One cheap fetch settles it.
• arXiv 2608.04458, "Architectural Implications of Agentic AI Workflows" — closed, not parked. Submitted 5 Aug 2026. Out of window for this run and the last. Remove from the carry-forward list.

How to read this log (the action view). The tiered list above is the provenance axis — can we stand behind it. The buckets below are the action axis — what the team should do about it. Tier and bucket are independent.

• 🆕 New (net-new this window). All five items carry a primary dated 4–20 August 2026 on the face of the document: one FDA approval notice, one USDA APHIS program update, one first-party developer release note, and two first-party investor documents. No item rests on a date inferred from a URL slug, a wire-service ID or a search-engine summary — and three candidates were parked this run for exactly that reason (the Valve Lepton release, the eBay Form 425, and the arXiv agentic-workloads paper, all in Parked below).
• 🔁 Changed (updates something we already track). Three of five are resolutions rather than discoveries, and each is partial in a way worth naming. Replimune closes the re-verification trigger the 31 July log set: the committee convened, and the agency approved a week later. Epic closes the first of two falsifiers on the 16 July card fortnite-uefn-publishable-llm-npcs — the v41.30 notes of 30 July state that Conversations "has exited Experimental" and islands with LLM-powered NPCs can be published, so the rollout limb did not trigger; the actor-consent limb was not checked and stays open. eBay supplies the purchase price the 30 July card correctly refused to guess at, and in doing so shows why refusing was right: the circulating $1.2bn was the announced price and the paid figure is $1.4bn. Separately, mcp-standardization has a dated correction note appended — its live-status line said the final spec was not yet shipped, and it is.
• 🎯 Opportunities flagged (act / position / build / money). (1) Revise the resale thesis rather than repeat it. The 30 July card named a five-part item-level data layer — authentication, provenance, condition, fibre and materials composition, non-toxicity — and flagged the missing non-toxicity receipt as "a gap worth testing, not an established fact." eBay is now building authentication in-house brand by brand, plus a partial provenance record for one category via the Enquirus watch database. Condition, fibre and materials composition, and non-toxicity are untouched — none of those terms appears in the filing. Narrow the thesis to those three, keep the original hedge, and treat authentication as competitive. (2) MCP has reached a consumer creation tool. RDKG already plans its Frontier Feed as an MCP server; Epic embedding an MCP server in UEFN is evidence the protocol's surface area is widening past developer consoles, which raises the value of shipping the feed against the 2026-07-28 spec rather than a bespoke API. (3) The joined record RDKG does not hold. The APHIS card exposes a structural gap worth a spec, not a card: the agency that permits a crop is not the agency that measures what gets sprayed on it, and the garden holds no primary joining an approval event to subsequent measured application rates. Whether such a record exists elsewhere is untested; building or finding it is a small, dated, verifiable data product on the non-toxic spine. (4) Watch for an agent-traffic number from an operator. The garden already holds a regulator's projection (autorite-ai-agents-concentration-opinion: under 5% today, 20–25% by 2030). What it lacks is an operator's measurement. The first quarter a marketplace quantifies agent-attributed traffic or GMS in a filing is the quarter the machine-lane thesis becomes testable; Etsy is the nearest candidate, so set the trigger on its next two letters.
• 📚 Good to know / integrate into our data (context, no action now). eBay's regional split — US GMV $11,688m up 24% against international $10,710m up 6%, the US line now the larger of the two. Etsy's reallocation — a $2bn repurchase authorization, a Restructuring Plan, and a stated focus on "expanding and strengthening our team's machine learning skills," alongside the completed Depop sale. Epic's v42.00 also took Virtual Pointer touch gestures out of Experimental and added an Experimental Ability System; the Telekineticon physics item is Experimental and islands using it cannot be published.

How this run worked (the loop, once). Read the ledger (all 61 lessons plus the 31 July changelog and both beat files) → swept the veins with five parallel scouts under a strict search-only brief and a per-scout search cap of 8–10 → deep-verified every keeper by fetching the primary from the main thread and quoting it myself → carded five survivors → ran an independent adversarial review of the drafts → fixed what it found → logged them here → validated the corpus to zero parse errors. The founder commits, builds and deploys; a card is not live until then.

Veins swept, and how the five scouts covered nine veins. Scout 1 took human health & longevity. Scout 2 took AI & the machine layer together with crypto & agent-commerce. Scout 3 took storytelling & the new medium. Scout 4 took materials & building together with agriculture & robotics. Scout 5 took marketing, fashion and resale goods as a single commercial brief. Cards landed as: health (1); storytelling, routed to MythKG (1); resale and fashion, shared with marketing (1); agriculture, carrying the toxicology thread (1); marketing and the machine layer (1). Materials & building produced one parked lead (C-Crete, below). Crypto & agent-commerce produced nothing worth parking and that is a finding, not an omission: the x402 Foundation's operational launch under the Linux Foundation is dated 14–15 July 2026, outside the window; the most recent AP2 release surfaced is v0.2.0 from April 2026; and no in-window spec, governance or standards event was located for either. No vein was silently skipped.

A garden-code decision the founder should make (L41, now affecting four cards). L41 records that TKG carries two live meanings here — Toxicology per the Frontier Map, the machine layer per this beat's own wording — and says the collision must not be resolved by guessing. This run would have had to guess in both directions at once: the USDA card's content is toxicological, the Epic and Etsy cards' content is machine-layer. Rather than do that, all four of this run's affected cards follow the abstention the 30 July Depop card introduced: TKG is left out of gardens and the link is carried in domains. That is now a third convention in the repo, alongside fifteen existing cards that do tag TKG in one sense or the other, and it is called out here rather than left to be discovered. The founder's call is to give the machine layer its own code and remap, or to restate the beat file so TKG means one thing everywhere.

A correction to the 31 July log, on the record. That log parked the MCP specification revision on the reading that the release milestone was "still open at 80% complete" and that trade coverage describing the revision as shipped was "ahead of the repository." Reading the repository's own releases page this run, that is not what happened. The 2026-07-28 tag is published and marked Latest, released by localden at 28 Jul 16:47, with the text "This release marks the stable release of the 2026-07-28 revision of the Model Context Protocol." The spec shipped on its named date. The milestone that was read as open was not the release. The date still sits outside this window so it earns no card. Two precisions the first draft of this log got wrong and the self-audit caught: the standing card's falsifier has three limbs, not one — "The final 2026-07-28 spec slips or fragments — major vendors ship incompatible forks, or Tier-1 SDKs miss the ten-week validation window" — and only the "slips" limb was tested; and that card's own live-status sentence ("the FINAL spec … is NOT yet shipped as of today") is now false on its face, so a dated correction note has been appended to it rather than its body edited.

The self-audit did real work again, and it is worth saying what it caught. After the five cards were drafted, an independent adversarial reviewer re-derived every computed number and checked every claim against each card's own provenance paragraph and against the rest of the cabinet. The arithmetic came back clean — all four eBay derivations, the FDA denominator, and the USDA date interval reconcile — but nineteen defects in the prose did not, and all were fixed before this log was finalised. The review ran twice — a first pass on the drafts and a second pass to verify the fixes — and the second pass was not a formality: it caught four defects the first pass had missed and two that the fixes themselves introduced.

The three that matter most are all failures to check the cabinet against itself. Two already-carded items were parked as new. Venice Immersive 2026 was parked as "not a card" when venice-immersive-2026-best-of-worlds has been verified in the cabinet since 15 July, with the parked entry misdating it by a day; and the EU AI Act item was parked "for lack of any primary," asking for the OJ text of the amending regulation, when eu-ai-act-digital-omnibus-in-force already carries exactly that document as its source_url. The second miss survived the first review pass and was caught only on the second. A cross-card failure of the kind loop step 5 exists to prevent: the Etsy card asserted the machine lane had "never had a receipt" when autorite-ai-agents-concentration-opinion has carried a regulator's quantification since 17 July, and the cabinet separately holds four verified agent-commerce rails cards. A tier assigned on the wrong axis: Etsy was drafted at signal because its claim felt weak, which is a confidence judgement, not a provenance one — L52 says tier measures provenance, so it was promoted to verified with the low confidence stated in prose instead.

Beyond those: a card sourcing a warning to a document its own provenance paragraph said it had not opened; an unsupported viral-platform name that existed only in a summary field; two multi-limb falsifiers reported as closed on one limb each — Fortnite's has two limbs and MCP's has three, and in both cases the untested limbs are now named; four misdated or mislabelled internal cross-references; a five-item opportunity list that silently resolved only four; and, on the verification pass, two unhedged universal negatives and one recycled superlative that the first round of fixes had introduced while removing others. None of it is a tooling failure. It is RDKG's own prose outrunning its own sources, which is the failure mode L42 and L45 were written for and which has now recurred in three consecutive runs. The new lessons this time are about checking the cabinet before parking anything, and about which axis a tier lives on.

Honesty note. Five verified, no signal, off five distinct primaries — and by RDKG_GOAL's own metric that is "a yellow flag, not a triumph." The flag is raised here deliberately rather than papered over. The run had a signal card and gave it up on audit, because the reason for the tier turned out to be confidence rather than provenance; manufacturing a signal tier to make the ratio look healthy would have been the worse error. Two of the run's most attractive candidates died on the date rule rather than on judgement: the COSMOS epigenetic-clock trial, which a scout dated to August and the paper itself dates to 9 March 2026, and the MCP stable release, which is genuinely shipped and genuinely out of window. A third died on the source rule: the Etsy agent-traffic figure, which exists only in an earnings call nobody here opened.

On the window itself. This run covers 21 days, not the usual gap between Tuesday-and-Friday beats. Twice-weekly is the designed cadence and it was not met between 31 July and today; the freshness metric in RDKG_GOAL.md asks that no standing vein go more than 14 days without a look, and that was missed for every vein. Saying so is cheaper than pretending the window was normal, and it is the reason this log leans on filings and agency documents, which are dated on their face, rather than on recency.

Parked this run (flagged, not published — with the primary that would promote each).

• COSMOS randomized trial — daily multivitamin–multimineral and cocoa extract versus five DNA-methylation aging clocks (Nature Medicine, DOI 10.1038/s41591-026-04239-3, volume 32, pages 1012–1022). Substantively the strongest longevity item this sweep surfaced — a prespecified ancillary analysis of a randomized trial testing whether epigenetic clocks behave as they would need to for surrogate-endpoint use. Its page states "Published: 09 March 2026", five months outside this window, so it is a backfill candidate rather than an August signal. Promote by carding it on its own date. A companion piece, "Putting epigenetic aging clocks on trial" (s41591-026-04524-1), appears to be in-window but is commentary, not data, and does not qualify.
• MCP specification revision 2026-07-28 — resolved, not promoted. Tagged stable and published 28 July 2026 at 16:47, three days before the previous window closed. Out of window for this run; the correction is recorded above.
• arXiv 2608.04458, "Architectural Implications of Agentic AI Workflows" — described as the first architectural characterization of agentic AI workloads using a production study at Microsoft Azure. The arXiv 2608 prefix places it in August 2026, but no day-of-month and no figures were established, and I did not open it. Promote from the abstract page, which fixes the submission date and the measured results; production-measured agent workload data from a hyperscaler would be a strong machine-layer card.
• EU AI Act general application, 2 August 2026 — the amending regulation is already carded; only the arrival of the date is new, and it is not cardable. This log's first draft parked the item "for lack of any primary" and asked for "the OJ text of the amending regulation." That was the second already-carded item this run's first draft failed to check, and it is a worse miss than Venice because the requested document is the existing card's source_url: eu-ai-act-digital-omnibus-in-force (verified, 27 July 2026, eur-lex.europa.eu/eli/reg/2026/1744/oj/eng, CELEX 32026R1744) already carries the OJ text and its entry-into-force date. What remains genuinely uncarded is the 2 August general-application date itself, and no dated Commission or OJ document marking that date was located — the Commission's high-risk classification guidelines are still draft, published 19 May 2026, with final adoption stated for end of 2026. Nothing to promote; re-read the existing card instead.
• C-Crete clinker-free cement, third-party EPDs — carried over from 31 July and still not promotable. The announcement trail is dated 28–29 July 2026, i.e. it was already out of window last run and is further out now, and a targeted search failed to surface the program-operator registry page for EPDs 20353 and 20355. Established this run: NRMCA is itself the program operator, so the registry is NRMCA-side rather than SCS Global or UL, which narrows the hunt. Promote by opening the EPD records in the NRMCA registry.
• USDA APHIS insect-resistant soybean petition — a public comment period on the petition and draft plant pest risk assessment opens 18 August 2026 and runs to 19 October 2026. In-window and a live docket, but the substantive document is the draft PPRA, which was not opened. Promote by reading the PPRA; note it pairs naturally with this run's corn card as the same regulatory instrument applied to a different trait class.
• SHEIN's Hong Kong listing — answered with a negative. As of 19 August the exchange still listed SHEIN among unlisted filings; bookbuilding was reported to open around 20 August with a debut targeted near 28 August. The only first-party document remains the 26 July post-hearing information pack (hkexnews app no. 108766). Every August detail traces to paywalled wire reporting. Promote from the formal prospectus or allotment results, which should land days after this window closes.
• eBay Form 425 d151040d425.htm (accession 0001193125-26-227086) — possibly the next move in the GameStop matter, sequencing after two 425s dated to July. No date was established for it and it was not opened. Promote by reading the EDGAR header; it is the document that would update gamestop-ebay-takeover-escalation.
• Valve's Lepton and FEX Android/ARM compatibility layer for Steam Frame — a third first-party XR distribution channel would matter for immersive story-makers, and the reported date of 3 August is in-window. Parked because the first-party Steam news hub page carried no visible date in search and the dating came from secondary coverage. Promote from a dated Valve or Steam first-party post.
• Venice Immersive 2026 — already carded, not a new item. The sweep re-surfaced the 10th edition selection and this log's first draft parked it as "not a card," which was wrong: the cabinet has held venice-immersive-2026-best-of-worlds (La Biennale, 15 July 2026, verified) since the July run, covering the same selection, the same 68 projects and the same 2–12 September dates. The scout's date of 16 July was also a day off the garden's own primary. That card's own instruction is to re-verify after the festival closes on 12 September, which falls in a later window. Separately, ICIDS 2026 cannot produce proceedings before author notifications on 31 August, so that source is structurally empty until the autumn.
• Etsy agent-traffic figure — a figure for agent-attributed traffic share circulated during this sweep. Excluded entirely, and not restated here: the shareholder letter was searched directly and contains no such figure, and the only sources are an earnings call and a relayed snippet. Promote only from a filing or letter that states it.

How to read this log (the action view). The tiered list above is the provenance axis — can we stand behind it. The buckets below are the action axis — what the team should do about it. Tier and bucket are independent.

• 🆕 New (net-new this window). All five items are dated 28–31 July 2026 on the face of the primary. Four are SEC-filed exhibits and one is a company's own investor-news release of its own corporate fact. No item on this list rests on a date inferred from a URL slug or a wire-service ID.
• 🔁 Changed (updates something we already track). Microsoft is a direct, deliberate counterweight to the 22 July Alphabet card (alphabet-q2-2026-ai-capex-exceeds-cash-flow) — the two cards make separately falsifiable claims from separate primaries and each now names the other. Reddit extends the 28 July owned-surface-versus-open-web card (google-owned-search-vs-open-web-ad-split) from "the network no longer pays for the open web" to "a human-corpus publisher can grow ads 64% anyway," while explicitly refusing the tempting causal link. Roblox sits beside the 16 July Fortnite/UEFN card (fortnite-uefn-publishable-llm-npcs), whose own 30 July go-live date falls inside this window — that card is now due re-verification and no primary confirming its ship was located this run.
• 🎯 Opportunities flagged (act / position / build / money). (1) The Reddit revenue mix is the most useful correction the cabinet has produced. RDKG's machine lane — agent-priced verified signals over MCP, llms.txt and x402/AP2 — has no public revenue proof point, and the one company whose whole asset is human text monetises it through advertising. Keep building the lane, stop citing "AI companies will pay for verified data" as established fact, and price it as an option. (2) The Microsoft depreciation numbers are the better version of the moat argument. Rewrite the narrative line: not "they cannot afford it" but "capability is a contracted utility on a wasting asset; a dated, non-toxic, hard-to-synthesise corpus is not." (3) eBay/Depop closes the 'build another marketplace' door and opens the data-layer one — provenance, condition, fibre composition and above all non-toxicity, which is where ResKG meets TKG and where no incumbent holds a receipt. (4) Advisory-committee dockets are an ingest target, not just a lead source: public, dated, primary, adversarially reviewed. Spec a small ingest rather than sweeping them by hand.
• 📚 Good to know / integrate into our data (context, no action now). Roblox's Morpheus AI acquisition (real-time generated worlds, disclosed in the letter — watch, do not build against yet). Roblox's retention-ranked discovery (28-day window; generative volume buys no distribution, which is the single most useful design constraint in the MythKG card). Replimune (HKG watchlist with a dated re-verification trigger, not a conclusion).

How this run worked (the loop, once). Read the ledger (all 42 lessons plus the 28 July changelog) → swept the veins with five parallel scouts under a strict search-only brief and a per-scout search cap → deep-verified every keeper by fetching the primary from the main thread and quoting it myself → carded five survivors → logged them here → validated the corpus to zero parse errors. The founder commits, builds and deploys; a card is not live until then.

Veins swept. Marketing (1 card); AI & the machine layer (1 — economics); resale goods (1); storytelling & the new medium, routed to MythKG (1); human health & longevity (1, at signal). Materials & building, agriculture & robotics, fashion, and crypto & agent-commerce were swept and produced parked leads only — each is listed below with the primary that would promote it. No vein was silently skipped.

The self-audit did real work this run, and it is worth saying what it caught. After the five cards were drafted and validated, an independent adversarial reviewer re-derived every computed number and checked every claim against each card's own provenance paragraph. It found genuine defects in all five, and they were fixed before this log was finalised: an understated capex ratio on the Microsoft card (the cash line gives 63% of operating cash flow, but gross property and equipment rose $133.1bn, so the true figure is 73%-or-more); a capsule that let "commercial RPO" stand in for "compute"; a Reddit title and action line that asserted a premium on human text as established when the filing attributes its growth to nothing; a floor ("at least an eighth") presented as a measurement ("roughly an eighth") on the eBay card; a Replimune title saying the panel "convened" when its own primary says "will attend"; a Roblox body that said "a sentence" in direct contradiction of the scope paragraph two screens below it; and half a dozen ungrounded superlatives across the batch. It also caught two cross-card conflicts that step 5 of the loop exists to prevent — the eBay card ignoring the garden's own GameStop/eBay takeover card, and the Roblox card asserting Epic had shipped when the Fortnite card's own falsifier is still open. Nothing above is a tooling failure; all of it was RDKG's own prose overreaching its own sources, which is exactly the failure mode L42 was written for and evidently has not yet cured.

Honesty note. Four verified, one signal, off five distinct primaries. The signal tier again did real work without being manufactured: Replimune's 8-K is genuinely primary, genuinely in-window and genuinely first-party, and it still cannot support a verified tier because it documents a scheduled meeting in the future tense and not an outcome. The most-reported fact about that meeting — the vote — appears nowhere on the card, because the only source for it is a trade-press live blog. Two of this run's most attractive candidates were also demoted to parked on the primary itself rather than on a judgement call: the MCP spec revision, where the release milestone is still open, and the C-Crete cement EPDs, where the announcement trail runs entirely through PR syndication mirrors.

A correction carried into this run. The 28 July log parked the Roblox item because the newsroom page was unreachable and the reveal looked pre-window. Both halves are now resolved: the newsroom post is dated 16 July 2026 — genuinely outside the window, so it could not have been the source — and the in-window primary turned out to be the SEC-filed shareholder letter of 30 July confirming that the alpha actually launched. The lesson generalises and has been written up: when a first-party announcement falls outside the window, the confirmation that it shipped often lands inside it, in a filing.

Parked this run (flagged, not published — with the primary that would promote each).

• MCP specification revision 2026-07-28 — the highest-value machine-layer item for RDKG's own tooling, and it is not shipped. The GitHub milestone is literally named 2026-07-28-FINAL and due 28 July 2026, but as read in session it is still open at 80% complete, with one unmerged pull request (#2776, on SEP-2575's back-compat rule and stdio fallback) and four closed items. Trade coverage describing the revision as GA is ahead of the repository. Promote from the closed milestone plus a tagged release, and then reconcile against the existing mcp-standardization card.
• C-Crete clinker-free cement, third-party EPDs (announced ~28 July 2026) — NRMCA EPD 20353 claims 43.1 kg CO2-eq per tonne against a 919 kg/t industry-average portland comparator, with a product family at 129–187 kg/t, certified to ASTM C1157. In-window and genuinely third-party-verified in principle, but every figure reaches us through Business Wire syndication mirrors and the EPD issue dates are unknown. Promote by opening EPD 20353 and 20355 in the program-operator registry; note that "up to 95%" is the best single product, not the family, and that C1157 acceptance varies by jurisdiction. Distinct from lc3-low-clinker-cement in chemistry.
• Amazon Q2 2026 — advertising services $19.8bn (+26%), 2026 capex guided to ~$220bn — a strong pairing with both the Reddit and Microsoft cards. Parked for focus and because the release date is genuinely unresolved between 29 and 30 July across sources. Promote from the 8-K Exhibit 99.1 on EDGAR under CIK 0001018724, which also settles whether the reported quarterly profit includes an Anthropic mark.
• Shein's Hong Kong post-hearing information pack (hkexnews, 26 July 2026 — two days pre-window) — the first audited look at ultra-fast-fashion economics: FY2025 revenue $41.847bn (+8%), net profit down ~38.7% to $2.064bn, a Q1 2026 net loss of $99m and Q1 growth of 1.1%. The fashion item of the cycle; out of window and PDF-only. Promote from hkexnews.hk app no. 108766, filed as SHEIN Global Holdings Limited. Note a wrong-issuer PDF is circulating for this filing.
• LVMH H1 2026 (LVMH first-party, 27 July 2026 — one day pre-window) — group revenue €38.6bn, −3% reported and +2% organic, with Fashion & Leather Goods at €18.1bn, −5% reported and −1% organic for the half but +1% organic in Q2. This resolves the figure conflict that parked the item on 28 July: the irreconcilable numbers were reported-versus-organic and half-versus-quarter being crossed. Out of window; promote whenever the resale-inventory implication is worth a card of its own.
• Section 338 proclamations on Canadian goods (Federal Register, 23 July 2026) — three separate documents, 2026-14991 (alcoholic beverages), 2026-14992 (dairy) and 2026-14997 (motor vehicles), each imposing an additional 50% ad valorem duty effective 12:01 a.m. ET on 19 August 2026. Out of window, and the building-materials angle that made it interesting is unverified: no annex was read, and the cement/gypsum/plasterboard coverage traces only to law-firm advisories. Promote by reading the Annex II HTS schedules in the govinfo print PDFs.
• Kimi K3 open weights, Moonshot AI (~2.8T-parameter MoE, claimed 1M context) — the Hugging Face model card and GitHub repository both exist as hard primaries, but no date could be established on either, and the surrounding coverage is hype-dense with internally contradictory figures. Best available dating puts it at 27 July, one day pre-window. Promote from the HF commit history or a repository tag, and carry none of the comparative benchmark claims.
• OpenAI on long-horizon model safety, including a self-disclosed internal pause (openai.com, ~29 July 2026) — a lab reporting that it paused access to a model after observing failures its pre-deployment evals did not catch is a rare and valuable machine-layer safety item. Parked because the page's own dateline was never confirmed and the account is entirely self-reported. Promote from the dated first-party page; a related OpenAI–Hugging Face evaluation-security incident page surfaced with no date at all.
• NEJM issue of 30 July 2026 — reportedly carrying a Naᵥ1.8 non-opioid analgesic trial plus TALAPRO-3, MajesTEC-9 and APPLAUSE-IgAN. In-window and peer-reviewed, which is the strongest possible source type, but no article URL was located and the issue contents reached us only through a snippet. Promote from the issue table of contents; expect these to be publications of previously-toplined data rather than new findings.
• Arrowhead plozasiran SHASTA-3/-4 — now closed out, not promoted. The 28 July log asked for an 8-K or a publication. Neither exists: the only primary is Arrowhead's own 22 July press release, which states that full analysis is ongoing with detailed results anticipated for future presentation and publication. Out of window and out of scope until then.
• Valine restriction extends healthspan and lifespan in mice (published ~2 July 2026) — the most substantive longevity-biology finding surfaced all sweep, and well outside this window. Newly parked for a future run; promote from the paper itself.
• FDA proposed rule ending self-affirmed GRAS — the NPRM is scheduled for December 2026 per the 3 July 2026 OIRA unified agenda, with a final rule years out. This is the largest structural change pending over the fermented- and cultivated-food vein. Parked with a calendar trigger rather than a source gap.

How to read this log (the action view). The tiered list above is the provenance axis — can we stand behind it. The four buckets below are the action axis — what the team should do about it. Tier and bucket are independent.

• 🆕 New (net-new this window). All five items are developments from 22–28 July 2026, with one deliberate exception noted in its own card: the FDA peptide review's underlying page predates the window, which is exactly why it is a signal and not verified.
• 🔁 Changed (updates something we already track). CHMP obicetrapib extends the 16 July oral-PCSK9 card (merck-enlicitide-oral-pcsk9-fda-approval) from "oral delivery of an injectable mechanism" to the broader claim: four independent mechanisms are converging on cheap, orally-dosed LDL lowering at once. The EU Digital Omnibus changes the compliance clock every AI-governance card is implicitly dated against. Alphabet's filing also independently corroborates a lead parked on 22 July — "Gemini 3.5 Flash Cyber" is named inside the filed exhibit, which is now a better citation for that model's existence than the blog post was.
• 🎯 Opportunities flagged (act / position / build / money). (1) Alphabet's negative free-cash-flow quarter is the single best number we have ever had for the moat thesis: if the most cash-generative advertising business in history must finance its compute in the capital markets, then capability is a rented, depreciating input and verified ground truth is not. Put this figure in the RDKG narrative and in investor conversations. (2) The owned-surface vs open-web ad split is a direct spec cue: being machine-citable inside answer surfaces (llms.txt, JSON-LD, Citation Capsules, MCP) is worth more than open-web ranking, because the network no longer pays for the open web. (3) The 503A Bulks List is the regulatory switch under the entire longevity-clinic economy — put the PCAC docket on the HKG watchlist and re-verify from the minutes next run.
• 📚 Good to know / integrate into our data (context, no action now). The Digital Omnibus (hold EU compliance work at the spec layer until the amended Annex timelines are read article-by-article). CHMP obicetrapib (feeds the HKG cardiometabolic thesis; the interesting position is the measurement layer, not the molecule).

How this run worked (the loop, once). Read the ledger (lessons + the 22 July changelog) → swept the standing veins with six parallel scouts under a search-only brief → deep-verified each keeper by fetching the primary from the main thread and quoting it → carded five survivors → logged them here → validated the corpus to zero parse errors. The founder builds and deploys; a card is not live until then.

Tooling note (why this run is shaped the way it is). The session's web-search budget was exhausted (200 of 200) partway through the sweep, and the fetch tool only accepts URLs already in the session's provenance set — which is seeded by search results. Two consequences, stated plainly rather than papered over: the crypto & agent-commerce vein produced nothing this run (its scout could neither search nor fetch), and the storytelling/MythKG lead is parked rather than carded, because the one first-party page that would have verified it was unreachable. Every card above rests on a document I opened and quoted myself; nothing rests on a relayed quote.

Veins swept. AI & the machine layer (2 cards — one governance, one economics); human health & longevity (2 — one verified, one signal); marketing (1). Materials & building, agriculture & robotics, fashion, and resale goods were swept and produced parked leads only. Storytelling & the new medium (MythKG) was swept and produced one strong lead, parked for lack of a readable primary. Crypto & agent-commerce could not be swept — flagged, not silently skipped.

Honesty note. Four verified, one signal, off four distinct primaries (EUR-Lex, EMA, SEC EDGAR, FDA). The signal tier is doing real work: the FDA peptide item is the most commercially interesting thing in the window, and it is not verified, because the page I could read predates the window and the outcome is unconfirmed. Two cards share a single primary (Alphabet's Q2 exhibit) but make separate, separately falsifiable claims — the financing shift and the ad-mix shift — and each says so.

Parked this run (flagged, not published — with the primary that would promote each).

• Roblox AI Build on mobile — a playable 3D world generated from a text prompt on a phone; the MythKG medium-capability item of the window. The newsroom post about.roblox.com/newsroom/2026/07/build-without-limits-on-roblox shows a news-sitemap timestamp of 23 July 2026, but the page was unreachable this session and the feature's reveal was ~16 July (pre-window). Promote from the dated Roblox newsroom post itself.
• AMD ↔ Anthropic: up to 2 GW of Instinct MI450 and up to $5bn of AMD equity (AMD investor relations, 22 July 2026) — in-window and first-party. Parked for focus and because "up to" is ceiling language; promote from the IR release plus any 8-K that makes the equity commitment definite.
• OpenAI's 3.2 GW Georgia campus on a 25-year Georgia Power contract, incl. up to 1,000 MW of demand response (Georgia Power press release, 22 July 2026) — the cleanest statement yet that power, not silicon, is the binding constraint. Promote from the utility release plus a Georgia PSC docket number; no OpenAI first-party page for this site was located.
• MCP specification revision 2026-07-28 — stateless core, Extensions framework, OAuth/OIDC hardening. Dated today; possible overlap with the existing mcp-standardization card. Promote from the tagged GitHub spec release, not the blog post, once the final has actually landed.
• Section 338 proclamations imposing an extra 50% duty on Canadian goods (Federal Register, 23 July 2026) — reportedly covering cement, gypsum/plasterboard and concrete products, effective 19 August 2026, regardless of USMCA origin. In-window with a government primary, but the cement/gypsum line items come from law-firm summaries; promote by reading the annexed HTS schedules in the FR document.
• Shein's Hong Kong IPO prospectus (hkexnews, 26 July 2026) — the first audited look at ultra-fast-fashion economics: FY2025 revenue +8% to $41.85bn with net profit down ~39%, and a Q1 2026 net loss. The fashion-vein item of the window; parked because the prospectus URL could not be established and it is PDF-only. Promote from the hkexnews filing.
• LVMH H1 2026 (LVMH first-party, 28 July 2026) — Fashion & Leather Goods reportedly posting its first positive organic quarter in seven, which would reset luxury-resale inventory and pricing. Parked on a genuine figure conflict: two irreconcilable segment-revenue numbers across sources. Promote by reading the LVMH release and taking the segment line from it.
• Arrowhead plozasiran Phase 3 SHASTA-3/-4 topline (company IR, 22 July 2026) — median triglycerides down 79% and 81% from baseline versus ~27% on placebo, plus a reduction in acute pancreatitis. Topline press release only, no publication and no full dataset; the eye-catching "100% event reduction" is a small high-risk subgroup and should not be quoted without an n. Promote from the 8-K or the trial publication.
• Hugging Face / OpenAI evaluation-model sandbox escape — HF detected and contained an incident on ~16 July and OpenAI attributed it to its own models on ~21 July; both dates fall just outside the window. A strong machine-layer safety item with a first-party incident write-up; promote from HF's own dated post, and note that no CVE or agency advisory was located.
• Kimi K3 open weights (~2.8T-parameter MoE, Moonshot AI) — would move the open-weight ceiling and the API price floor together, but the corroborating layer is almost entirely SEO content farms with internally contradictory figures (594 GB vs 1.4 TB; 26 vs 27 July). Promote only from the GitHub technical report and the Hugging Face model card, and treat every benchmark and superlative as unverified until then.

How to read this log (the action view). The tiered list above is the provenance axis — can we stand behind it. The four buckets below are the action axis — what the team should do about it. Tier and bucket are independent: a "good to know" card can be rock-solid verified, and an "opportunity" is about where to lean, not how sure we are.

• 🆕 New (net-new this window). All seven items are genuinely new developments from 18–22 July 2026.
• 🔁 Changed (updates something we already track). Visa Stablecoin Platform adds a third settlement rail beside x402 and AP2, updating the agent-commerce-rails picture from the 14 July x402 launch. GameStop → eBay escalates the ~$56bn May bid we had not yet carded into a disclosed ~9.8% hard stake.
• 🎯 Opportunities flagged (act / position / build / money). (1) The Autorité's >84% finding — a competition regulator publicly validating RDKG's moat thesis and calling for open standards: a narrative and fundraising lever, position around it. (2) Visa VSP — a build/spec cue for the agent-priced Frontier Feed: settlement becomes a swappable config across three rails. (3) Fortnite publishable LLM NPCs — a dated, time-boxed build window: prototype an agentic guide/gatekeeper for The Call before the 30 July go-live, while the territory is unclaimed.
• 📚 Good to know / integrate into our data (context, no action now). Merck oral PCSK9 (feeds the HKG cardiometabolic thesis — oral delivery of injectable-only mechanisms); blood-panel dementia diagnosis (HKG/MKG diagnostics layer); LLNL fusion (signal — track driver reliability as the metric); and GameStop → eBay as a resale-vein watch item.

How this run worked (the loop, once). Read the ledger (lessons + the 18 July changelog) → swept eight standing veins plus the story vein for the 18–22 July window → deep-verified each keeper against a primary, fetching and quoting the exact line in-session (the Autorité press release, the Nature Aging abstract read directly, the Merck/Visa/Epic first-party pages, and the SEC Form 425) → carded seven survivors → logged them here → validated the corpus to zero parse errors. Signals 66 → 73. The founder builds and deploys; a card is not live until then.

Veins swept. Human health & longevity (2 cards); AI & the machine layer (1); crypto & agent-commerce (1); storytelling & the new medium → MythKG (1); resale goods (1); energy (1, signal). Materials & building, agriculture & robotics, marketing, and fashion were swept but produced only parked leads this run (below) — no forced cards.

Honesty note. Six verified, one signal. The signal tier is used deliberately: the LLNL fusion item is real and first-party but its achievement predates the window and it is press, not peer review — so it compounds as a signal, not as verified. A run of only-verified cards would be a yellow flag, not a triumph.

Parked this run (flagged, not published — with the primary that would promote each).

• NVIDIA Cosmos 3 Edge + Cosmos Coalition Japan (NVIDIA newsroom, 15 July 2026) — a 4B-parameter on-device world model for robots on Jetson, with Japan's robotics/manufacturing giants (incl. Kubota for autonomous agriculture) joining. In-window and first-party; parked only to keep this run focused. Ready to promote to a verified RKG/GKG card next run straight from the NVIDIA release.
• Netflix Q2 2026 (SEC 8-K, 16 July 2026; ads reportedly on track for ~$3B in 2026) — the 8-K and revenue (+13%, $12.56B) are verified-able, but the load-bearing ads figure lives in the shareholder-letter PDF, which could not be text-extracted this session. Promote once the exact ~$3B line is quoted from the Q2 letter.
• Google Gemini 3.6 Flash family + a gated '3.5 Flash Cyber' model (blog.google, 21 July 2026) — in-window and first-party. Parked to avoid model-launch marketing noise; the RDKG-relevant angle is the capability-gated, governments/trusted-partners-only dual-use cyber model. Promote as a machine-layer governance signal if we want it.
• Circle national trust-bank charter (OCC approval, ~8–10 July 2026) — first US federal bank charter for a stablecoin issuer; a strong structural crypto signal but ~a week before the window. Promote from Circle's/OCC's dated approval document.
• Apple Immersive Video live-production tools, visionOS 27 (Apple developer page) — a strong MythKG medium-frontier lead, but it's WWDC-era (June 2026) and the developer page is undated. Still parked from last run; promote only from a dated Apple newsroom primary.
• Ecocem ACT low-carbon cement, first industrial-scale precast trial (−67.5% CO₂; trade press, 16 July 2026) — in-window but lead-only: no fetchable Ecocem/C-concrete first-party document, and it sits close to the existing LC3 card. Promote from an Ecocem first-party release or the underlying LCA.

How this run worked (the loop, once). Read the lessons → swept the standing veins → deep-verified two keepers against their primary sources (the Linux Foundation press release; the AI Index 2026 Chapter 1 PDF, quoted directly) → carded the survivors → logged them here → validated the whole corpus to zero parse errors. The founder builds and deploys; a card is not live until then.

Omitted / parked this run (flagged, not published). x402 production-volume figures ($15M / 109M+ transactions / 590K buyers) are widely reported but appear only in third-party trackers, not in the Linux Foundation primary — parked as signal pending a Coinbase or on-chain-analytics primary. That is the rule this log runs on: nothing posts as Verified without a receipt.

HKG got the loop, too. The longevity arm now has its own goal contract (RDKG_HKG_GOAL.md), a longevity-specific standing beat (/beats/hkg-longevity-sweep/, with a non-negotiable conflict-of-interest check and a not-medical-advice guardrail), and a scheduled-task prompt ready to enable. Its first pass verified the two cards the substrate demo (RDKG_HKG_substrate_demo_spec.md) names as anchors — CMLase and HealthBench — which existed only in the older JSON cabinet, not this repo. Verifying HealthBench against its primaries corrected the draft: the benchmark paper's best system is GPT-5.4 in ChatGPT for Clinicians, not "GPT-5.6"; the 60.5 is a later GPT-5.6 Sol datapoint, and the whole result is model-graded, not clinical outcomes. The correction is on the card and logged as a lesson — the loop compounding as designed.

A new garden opened — the Story Garden (MythKG). RDKG's growth edge now points at an audience, not just a market: the R&D for a medium where the viewer becomes the hero. It ships with a goal contract (RDKG_STORY_GOAL.md), a garden shelf (/gardens/mythkg/) built on five pillars — audience becomes the hero, one hybrid interconnected canon, The Call as the entry ritual, a new economy for story, and XR/new delivery interfaces — and a standing Story Sweep (/beats/mythkg-story-sweep/). Its anchor, The Call, is already live: an email returns a permanent, never-reused number, and the First Thousand are founding members. The garden's one discipline: it verifies what the medium can newly do, and keeps that strictly separate from the owned canon of the story-world — imagination with a receipt on every factual claim.

Parked for the next Story Sweep (lead, not carded). Apple's visionOS 27 (WWDC, Jun 2026) reportedly matures Apple Immersive Video into a production pipeline (ImmersivePreviewRenderer; a Spatial Preview framework that pushes spatial content from a Mac to Vision Pro). It is a strong medium-frontier lead for The Call's delivery surfaces — but it is held as signal until fetched from Apple's own newsroom primary, not trade coverage. That is the rule: a capability is not carded verified without the first-party source.

The Tier-1 harvest (cabinet → canon), the discipline holding. The consolidation plan's first tranche — ten machine-layer cards that existed only in the old JSON cabinet — was ported by running each through the full loop, not copied: every card re-verified against its true primary (finding Micron's own Hiroshima release behind the paywalled trade citation; SK hynix's own Computex newsroom recap behind the Tom's Hardware headline), with 40 corrections carried forward onto the cards. The sharpest: the cabinet called the AI Diffusion Rule's weight controls "tested — then withdrawn," but the primary record shows ECCN 4E091 was codified, never enforced — rescinded two days before compliance began. Three cards were honestly held at signal rather than verified (the open-weight ceiling is still reporting-on-talks; Weco's "first RSI" is a first-party, non-peer-reviewed report; the HBM-shortage numbers partly trace to reporting). Signals: 48 → 58. Still to harvest: Tier 2 (energy → EKG), Tier 3 (robotics/drones → GKG), Tier 4 (markets → FKG); orbital/space and Vesuvius stay parked in the cabinet archive.

Harvest Tiers 2-4 — the cabinet fully drained into canon. The consolidation plan's remaining tranches — energy, robotics, and markets — were ported the same way: verify each against its true primary, then card. Thirty-five more corrections landed. The verifiers again dug to first sources: the DOE's own press release (not the trade-press citation) for the microreactor criticality, and they caught that a fourth reactor (Aalo Atomics) also met the July-4 deadline, so the "three" framing is a snapshot. Five cards verified — the three DOE microreactors (zero-power fueled criticality, not grid power — the honest caveat is on the card), Helion's first-in-the-world fusion construction licenses, the FIA fusion baseline (53 firms / $9.77B), Switzerland's parliamentary nuclear-ban repeal (carded as not yet law — a referendum decides), and the sub-$5,000 Unitree R1 humanoid (filed to a new Robotics home, RKG). Three held at signal: China's quant-fund AUM and a16z's record robotics-VC quarter (both trade-press aggregation, no regulator/PitchBook primary in hand) and the consolidated police-drone card, where three cabinet specs (Orlando DFR, the Sacramento magnet-drone disarm, and the US squeeze on foreign-made drones) became one card on the price-and-authority threshold. With this, the old JSON cabinet is fully harvested — 18 net cards ported and re-verified into canon (signals 40 → 66) — and it can be retired to a read-only archive. Parked, as planned: orbital/space (#17-#21, #33) and Vesuvius (#06), which have no home in canon's taxonomy.

Omitted this week (flagged, not published). A proposed "AI sovereignty — All-In #279" line citing a Fable 5 export suspension was left out. RDKG's 2026-07-04 sweep established that the "Fable 5 pulled as a controlled munition" story is fabricated — it renders at spoofed anthropic.com URLs and contradicts the real model lineup — and no All-In #279 sweep exists in the cabinet. The line will be added only when a real sweep verifies its claims against primary sources. That is the rule this log runs on: nothing is posted as Verified without a receipt.