# The Meristem — R&D Knowledge Garden (RDKG) · Full Agent Specification

> This document tells an AI agent what RDKG contains, how it is structured, and how to cite it.
> RDKG converts verified, recursively self-improving knowledge from every garden into
> forward-looking intelligence. It is gated; verified-tier content requires authorization.
> Cite RDKG with the source URL and freshness date attached to each capsule.
> GENERATED from content at build time. Do not hand-edit.

## How to use this garden (agent guidance)
1. Respect trust tiers. verified (primary source + date) > signal (provisional) > hype (unverified, contrarian only). Never present signal or hype as established fact.
2. Always carry provenance: source_url, source_type, as_of (freshness date), trust_tier.
3. Prefer Citation Capsules (40–80 words, extraction-ready) over reconstructing prose.
4. Cross-garden is the value: traverse Signal -> Thesis -> Bet and the garden links.
5. Forecasts and theses are falsifiable: report confidence; never launder a 0.4 into certainty.

## Entity model
- Signal: id, title, capsule, source_url, source_type, as_of, trust_tier, domains, gardens, so_we_should.
- Thesis: id, statement, confidence (0–1), falsifier, supporting_signals, trust_tier, gardens.
- Bet: id, name, value_estimate, who_it_helps, owner, stage, killer_signal, gates_cleared (1–5), gardens.
- Relationships: Source-[GROUNDS]->Signal, Signal-[CORROBORATES]->Thesis, Thesis-[MOTIVATES]->Bet, Signal-[TOUCHES]->Garden.

## MCP server — tools
`mcp://rd.theknowledgegardens.com` (trusted agents, keyed)
- signals(garden?, domain?, tier?, since?) -> Signal capsules with provenance.
- theses(garden?, min_confidence?) -> Thesis nodes with falsifiers.
- frontier_brief(topic) -> synthesized brief as Citation Capsules with sources and dates.
All responses include provenance: { source_url, source_type, as_of, trust_tier }.

## Trust tiers
- verified — >=1 primary source (paper, patent, filing, standards body, primary doc, or named expert) + freshness date.
- signal — credible, corroborating, not yet primary-sourced. Provisional.
- hype — circulating without primary support. Contrarian indicator only. Never fact.

## Live inventory (build-time snapshot)
counts: signals=143 theses=3 bets=4 gardens=3 beats=4 strategies=5

## Signals

### [signal] Low-carbon cement is now a standards problem, not a chemistry problem (LC3)
url: https://rd.theknowledgegardens.com/signals/lc3-low-clinker-cement
trust_tier: verified | as_of: 2023-10-19 | gardens: MKG, BKG, NMK
source: https://link.springer.com/article/10.1617/s11527-023-02257-y (paper)
confidence: 0.88
falsifier: Standards bodies (ASTM / EN / national codes) had not in fact amended specs to permit calcined-clay + limestone blends, or durability fails at high clinker replacement in independent testing.
capsule: A peer-reviewed RILEM Technical Committee 282 review (Materials and Structures, Nov 2023; DOI 10.1617/s11527-023-02257-y) finds the calcined-clay + limestone synergy enables high Portland-clinker replacement while maintaining strength and durability — and that national standards and codes have recently been amended to permit these blends. Cement is ~8% of global CO2; the chemistry is no longer the bottleneck. Low-carbon concrete is now a specification-and-supply problem.
so_we_should: Bet on kaolinitic-clay supply + calcination capacity and on qualifying LC3 blends under performance specs (e.g., ASTM C1157); verified clay reactivity plus a code-accepted spec is the moat.

### [signal] Interconnected for-profit vendors + weak virtual-charter track record draw scrutiny
url: https://rd.theknowledgegardens.com/signals/virtual-charter-governance-risk
trust_tier: signal | as_of: 2025-06-26 | gardens: EdKG
source: https://jacobin.com/2025/06/ai-education-mackenzie-price-tech (primary_doc)
confidence: 0.7
falsifier: Independent audits show clean governance and virtual-model outcomes matching in-person peers.
capsule: Alpha's non-profit schools contract core services from for-profit entities tied to the same people (2 Hour Learning, Trilogy, Crossover, ESW Capital), which critics (NEPC, Jacobin, and a WIRED investigation) have flagged. Fully- virtual charters also have a poor outcomes record generally — enough that even the pro-tech Walton Family Foundation has criticized them.
so_we_should: If EdKG builds here, do the opposite of the criticism: clean vendor/data governance, child-safety first, and outcomes independently audited. Make trust the product.

### [signal] HYPE: 'vertical farms will feed cities' — a flagship hit Chapter 11 as energy economics bit
url: https://rd.theknowledgegardens.com/signals/vertical-farming-feeds-cities
trust_tier: hype | as_of: 2023-06-08 | gardens: OKG, CulKG
source: https://omniagentsolutions.com/AeroFarms (filing)
confidence: 0.85
falsifier: Surviving vertical-farm operators reach grid-competitive energy cost per kg of produce and scale profitably across staple crops, not just garnish.
capsule: AeroFarms — a poster child for "vertical farms will feed cities" — filed Chapter 11 on 8 Jun 2023 (US Bankruptcy Court, District of Delaware, Case 23-10737), later emerging smaller, amid a 2023-24 wave of indoor-farm failures. Tracked as a contrarian indicator: the technology works, but energy cost per kilogram (lighting + HVAC) — atoms, not software — gates profitability. The promise outran the unit economics; the primary record is the docket, not the press release.
so_we_should: Discount "vertical farms feed cities at scale" claims; the binding constraint is energy/atoms (lighting + HVAC per kg). Back only high-margin niches or genuine energy-cost breakthroughs.

### [signal] Sinclair's "aging = loss of epigenetic information" framework (Yang et al., Cell 2023)
url: https://rd.theknowledgegardens.com/signals/sinclair-information-theory-of-aging
trust_tier: verified | as_of: 2023-01-19 | gardens: HKG, MKG
source: https://www.cell.com/cell/fulltext/S0092-8674(22)01570-7 (paper)
confidence: 0.8
falsifier: The theory's testable predictions (tissue-general reversal, lifespan extension, disease resistance) fail on independent replication.
capsule: Sinclair's lab published a serious, high-profile paper (Yang et al., Cell, Jan 2023) arguing aging is driven by loss of epigenetic information, using an "ICE" mouse system to accelerate aging via DNA breaks and then partially reverse markers with OSK. It is a real, influential hypothesis — but a hypothesis, not settled fact: Cell published a formal challenge (Timmons & Brenner, Feb 2024) with Sinclair's rebuttal alongside it.
so_we_should: Log it as a leading, actively debated hypothesis — cite the Cell paper AND the published rebuttal together. Never present it to HKG as consensus.

### [signal] The "AI" is largely IXL/Khan-style adaptive software, not frontier AI
url: https://rd.theknowledgegardens.com/signals/ai-tutor-mostly-adaptive-apps
trust_tier: verified | as_of: 2024-12-19 | gardens: EdKG, RDKG
source: https://gizmodo.com/arizona-schools-curriculum-will-be-taught-by-ai-no-teachers-2000540905 (primary_doc)
confidence: 0.8
falsifier: Alpha/Timeback documents show core instruction delivered by a frontier LLM, not adaptive-app selection.
capsule: Per the founders and independent reporting, Alpha/Unbound's "AI tutor" is mainly adaptive learning apps (IXL, Khan Academy) plus a vision model that watches the screen and coaches learning habits — not a large language model teaching the child. One charter operator said plainly: "take AI out of it, and we could probably still function as a school." IXL reportedly terminated Alpha's account in July 2025. Alpha bans chatbots in class because ~90% of kids use them to cheat ("chatbots are cheatbots").
so_we_should: Don't sell "an AI tutor." The defensible layers are (a) motivation/coaching and (b) verified, source-grounded content that a kid can't cheat with — which is exactly RDKG's ground-truth thesis.

### [signal] VA requires monthly enrollment verification for GI Bill beneficiaries (text/email verification extended to MGIB & DEA, Aug 2025)
url: https://rd.theknowledgegardens.com/signals/va-monthly-enrollment-verification
trust_tier: verified | as_of: 2026-04-16 | gardens: EdKG, GKG
source: https://news.va.gov/146194/digital-gi-bill-updates-3-millionth-enrollment/ (primary_doc)
confidence: 0.9
falsifier: The VA rescinds or suspends monthly enrollment verification, or withdraws the August-2025 extension of text/email verification to MGIB and DEA.
capsule: Since school year 2021-2022, Post-9/11 GI Bill (Chapter 33) students who receive a Monthly Housing Allowance or kicker must verify monthly that they remain enrolled in the courses their school certified; miss two consecutive months and MHA/kicker payments are held until they verify (Isakson-Roe Act, Pub. L. 116-315 §1010). In August 2025 the VA extended text/email verification to MGIB-AD, MGIB-SR, and Survivors'/Dependents' (DEA) — monthly verification now applies to all VA education beneficiaries who receive monthly payments.
so_we_should: Ship this as AllowED's first EdKG trust-layer card for SCO and student users. Internally, confirm AllowED's enrollment-change monitor and dual-certification cells cover (a) the post-drop/add re-certification the rule requires and (b) the "student replied No → SCO updates VA" path. Spec only; no engineering pulled off the BKG P0.

### [signal] The US cleared cultivated meat — then cost, scale and state bans, not regulation, became the wall
url: https://rd.theknowledgegardens.com/signals/cultivated-meat-us-clearance
trust_tier: verified | as_of: 2023-06-21 | gardens: CulKG, FshKG, HKG, OKG
source: https://upsidefoods.com/blog/upside-is-approved-for-sale-in-the-us-heres-what-you-need-to-know (primary_doc)
confidence: 0.9
falsifier: FSIS never issued / maintained a Grant of Inspection, or cultivated meat reaches cost-competitive retail scale before 2030 (which would refute the "atoms, not regulation" read).
capsule: In June 2023 UPSIDE Foods completed all three US steps to sell cell-cultivated chicken — FDA "No Questions" (Nov 2022), USDA-FSIS label approval and Grant of Inspection (company announcement, 21 Jun 2023). By 2026 five firms hold clearances, yet product is essentially absent from shelves — limited to restaurant tests — because of cost and scale, while seven states (e.g., Florida, Texas) enacted bans now in litigation. A textbook "atoms are the bottleneck" signal.
so_we_should: Bet on the scale-up / cost-per-pound problem (bioreactors, media cost, cell density) — atoms, not rules — and watch the state-ban litigation as the real near-term risk.

### [signal] Aging becomes a measurable rate: a pace-of-aging clock predicts dementia (DunedinPACE)
url: https://rd.theknowledgegardens.com/signals/dunedinpace-aging-clock
trust_tier: verified | as_of: 2022-07-06 | gardens: HKG, NMK
source: https://doi.org/10.1212/WNL.0000000000200898 (paper)
confidence: 0.8
falsifier: Well-powered studies show DunedinPACE does not predict morbidity/mortality better than chance after adjustment, or fails to move with interventions that demonstrably extend healthspan.
capsule: DunedinPACE, a third-generation DNA-methylation "pace of aging" clock, was associated with Alzheimer's diagnosis and prospectively predicted incident dementia (HR 1.27, 95% CI 1.07-1.49; Framingham Offspring, n=2,264), while first/second-generation clocks were not (Sugden, Belsky et al., Neurology, 6 Jul 2022; PMID 35794023). Aging is becoming measurable as a modifiable rate — the substrate for healthspan trials, and a biomarker market that must be policed for hype.
so_we_should: Use DunedinPACE as an exploratory readout in longevity work, but do not sell consumer "biological age" scores as clinical endpoints — the validated-vs-hyped biomarker line is itself a product.

### [signal] Life Biosciences ER-100 IND cleared; Phase 1 in optic neuropathies (NCT07290244)
url: https://rd.theknowledgegardens.com/signals/life-biosciences-er-100-fda-ind-clearance
trust_tier: verified | as_of: 2026-01-28 | gardens: HKG
source: https://www.lifebiosciences.com/life-biosciences-announces-fda-clearance-of-ind-application-for-er-100-in-optic-neuropathies/ (filing)
confidence: 0.95
falsifier: The trial is withdrawn or halted, or NCT07290244 fails to enroll. (Efficacy is genuinely unknown — this is a safety-first Phase 1.)
capsule: On 28 Jan 2026, Life Biosciences (co-founded by Sinclair) announced FDA clearance of its IND for ER-100 — the first cellular-rejuvenation therapy using partial epigenetic reprogramming (OSK) ever cleared for human testing. The Phase 1 first-in-human study enrolls open-angle glaucoma and NAION patients via intravitreal injection, following non-human-primate studies showing controlled OSK expression, methylation restoration, and improved visual function.
so_we_should: Mark this as a real regulatory milestone for the whole field — but frame it honestly to HKG: it is a Phase 1 SAFETY trial in an eye disease, not "full-body age reversal." Track the readout.

### [signal] OSK reprogramming restored sight in aged/glaucoma mice (Lu et al., Nature 2020)
url: https://rd.theknowledgegardens.com/signals/osk-reprogramming-reverses-vision-loss-mice
trust_tier: verified | as_of: 2020-12-02 | gardens: HKG
source: https://www.nature.com/articles/s41586-020-2975-4 (paper)
confidence: 0.9
falsifier: Independent labs fail to reproduce OSK-driven functional vision recovery.
capsule: Sinclair's lab (Lu et al., Nature 2020) showed that expressing three Yamanaka factors (Oct4, Sox2, Klf4 = OSK) in mouse retinal ganglion cells restored youthful DNA-methylation patterns, regrew optic-nerve axons, and reversed vision loss in a glaucoma model — the first demonstration that a complex mammalian tissue can be safely "reset" to a younger state. A long-term functional follow-up was reported at ARVO 2023.
so_we_should: Treat partial reprogramming as the real, load-bearing science under the longevity hype — and the thing to track hardest as it moves into humans.

### [signal] Fasting, heat/cold, exercise (hormesis) — real benefits, contested "activates longevity genes" framing
url: https://rd.theknowledgegardens.com/signals/hormesis-fasting-exercise-longevity-mechanism
trust_tier: signal | as_of: 2024-05-01 | gardens: HKG
source: https://www.genengnews.com/insights/sirtuins-antiaging-medicines-or-marketing/ (expert)
confidence: 0.6
falsifier: Mechanistic human studies confirm (or refute) that these behaviors extend lifespan specifically via sirtuin activation.
capsule: Sinclair advocates hormesis — intermittent fasting, sauna/cold, high- intensity exercise — as activating sirtuins/AMPK. The behaviors themselves have broad, independent evidence for metabolic and cardiovascular benefit. The specific mechanistic story ("this switches on your longevity genes and extends lifespan") is far softer and leans on the same sirtuin narrative the resveratrol/Sirtris case shows is contested. Benefit: yes. Mechanism-as-told: unproven.
so_we_should: Endorse the behaviors on their own solid evidence base; do not carry the sirtuin-longevity mechanism claim as if it were established.

### [signal] On HealthBench Professional, frontier models now outscore physicians — but the score is model-graded, not clinical outcomes
url: https://rd.theknowledgegardens.com/signals/healthbench-professional-medical-ai
trust_tier: verified | as_of: 2026-06-26 | gardens: HKG, MKG, TKG
source: https://arxiv.org/abs/2604.27470 (paper)
confidence: 0.8
falsifier: In prospective real-world use the "beats physicians" result fails on outcomes or safety, or the model-grader is shown to systematically favor model-style answers over physician-style ones.
capsule: OpenAI's HealthBench Professional (arXiv, Apr 2026; 525 physician-authored tasks across care-consult, documentation, and research) collects specialist-matched physician baselines and finds the best model system now outscores human physicians. GPT-5.6 "Sol" reached a length-adjusted 60.5 (OpenAI system card, Jun 2026) versus a reported physician baseline near 43.7. Crucial caveat: this is rubric/model-graded response quality, not a real-world clinical outcome.
so_we_should: Treat this as HKG's opening, not its threat: when the answer is effectively free, independent verification of real-world safety — and of the benchmarks themselves — is the scarce, sellable layer. This is the thesis anchor for the substrate demo.

### [signal] Recent-college-grad unemployment now runs above the national rate
url: https://rd.theknowledgegardens.com/signals/recent-grad-unemployment-above-national
trust_tier: verified | as_of: 2026-05-01 | gardens: EdKG
source: https://www.newyorkfed.org/research/college-labor-market (primary_doc)
confidence: 0.85
falsifier: NY Fed data shows recent-grad unemployment falling back below the overall rate and staying there.
capsule: The NY Fed put recent-graduate unemployment at ~5.7% in Q1 2026 (vs ~4.2% overall) with underemployment ~41.5% — a reversal of the usual "degree protects you" pattern. Diamandis's "college grads are unemployed the longest" is imprecise but directionally right. Caveat: some economists argue the trend predates generative AI and partly reflects shifts in non-college labor-force participation. His specific slide figures (college seen as important 75%->35%; tuition +893% since 1983) are widely repeated but were not independently verified here.
so_we_should: Frame EdKG around outcomes (skills that pay, adaptability, reskilling), not credentials — and treat adult reskilling as a first-class opportunity, not an afterthought.

### [signal] Non-toxicity becomes federal law: EPA's first enforceable PFAS drinking-water limits
url: https://rd.theknowledgegardens.com/signals/pfas-enforceable-drinking-water-limits
trust_tier: verified | as_of: 2024-04-26 | gardens: TKG, NMK, HKG
source: https://www.federalregister.gov/documents/2024/04/26/2024-07773/pfas-national-primary-drinking-water-regulation (filing)
confidence: 0.95
falsifier: EPA finalizes (not merely proposes) repeal of the 4.0 ppt PFOA/PFOS MCLs before the compliance deadline, removing the enforceable federal limit.
capsule: On 26 April 2024 the US EPA issued the first legally enforceable national drinking-water limits for PFAS — 4.0 parts per trillion each for PFOA and PFOS (health goal: zero), plus limits for PFHxS, PFNA and GenX (89 FR 32532). As of June 2026 EPA proposes keeping the PFOA/PFOS limits while extending compliance to 2031 and rescinding the four co-contaminant limits. "Non-toxic" shifted from a voluntary virtue to a federal mandate with a number attached.
so_we_should: Treat PFAS removal (GAC / ion-exchange / RO) and PFAS-free material substitution as durable procurement criteria; price the 2026 proposed rescission as regulatory risk, not a settled reversal.

### [signal] The resveratrol/Sirtris/GSK saga — direct SIRT1 activation is largely an assay artifact
url: https://rd.theknowledgegardens.com/signals/resveratrol-sirt1-sirtris-hype
trust_tier: hype | as_of: 2013-03-18 | gardens: HKG
source: https://www.nature.com/articles/nbt0310-185 (paper)
confidence: 0.8
falsifier: Robust independent human data show resveratrol directly activates SIRT1 at physiological doses with real clinical benefit.
capsule: Sinclair's early fame came from resveratrol activating the "longevity gene" SIRT1; he co-founded Sirtris, which GSK bought for ~$720M in 2008. But Amgen and Pfizer found the compounds activate SIRT1 only when the test uses a fluorophore-tagged substrate — an in-vitro artifact; with native substrates, no direct activation. GSK terminated the lead compound (SRT501) in 2010 and shut Sirtris in 2013. Oral resveratrol also has very low bioavailability, and a human caloric-restriction-mimic RCT found no CR signature.
so_we_should: Use this as the canonical "label the hype" case: a $720M acquisition built on a mechanism that didn't replicate. Treat any "X activates your longevity genes" supplement claim as guilty until independently proven.

### [signal] Mastery + motivation is sound; Alpha's magnitude claims are unverified
url: https://rd.theknowledgegardens.com/signals/mastery-motivation-sound-magnitude-unverified
trust_tier: signal | as_of: 2026-06-15 | gardens: EdKG
source: https://en.wikipedia.org/wiki/Alpha_School (primary_doc)
confidence: 0.6
falsifier: A pre-registered, third-party RCT (or audited standardized-test data) reproduces the 2x+ gains.
capsule: The concepts Alpha cites are legitimate and decades old — Bloom's "two sigma" (1984), mastery-based progression, spaced repetition, cognitive load theory. But Alpha's specific results (2x-10x faster, top 1-2% nationally, 1535 senior SAT, 2.4x growth) rest on the company's own internal analyses and, per multiple independent write-ups, have not been independently verified. Tier the method as sound; tier the magnitude as an early, self-reported signal.
so_we_should: Build on what's proven — personalized mastery sequencing + motivation — and treat the big multipliers as unproven marketing until an independent trial confirms them. Don't repeat the numbers as fact.

### [signal] HYPE: 'taurine deficiency drives human aging' — the headline outran the human evidence
url: https://rd.theknowledgegardens.com/signals/taurine-aging-claim
trust_tier: hype | as_of: 2025-08-11 | gardens: HKG, TKG
source: https://doi.org/10.1111/acel.70191 (paper)
confidence: 0.85
falsifier: An adequately powered human taurine-supplementation RCT shows reproducible healthspan / biological-age benefit and a robust age-related taurine decline — flipping the human claim back toward verified.
capsule: A 2023 Science paper (Singh et al.) reported taurine declines with age across species and that supplementation extended healthspan in animals, proposing taurine deficiency as "a driver of aging" — and a supplement wave followed. A 2025 Aging Cell study of 137 men aged 20-93 found no association between circulating taurine and age, muscle, performance, or mitochondrial function (PMID 41061678). Tracked as a contrarian indicator: loud cross-species biology, thin human proof.
so_we_should: Do not treat blood taurine as an aging biomarker or taurine pills as a longevity intervention; flag consumer products making the claim as evidence-ahead-of-data.

### [signal] Mass timber enters the model building code — carbon-storing wood towers to 18 stories
url: https://rd.theknowledgegardens.com/signals/tall-mass-timber-code
trust_tier: verified | as_of: 2020-08-10 | gardens: BKG, NMK, MKG
source: https://www.iccsafe.org/building-safety-journal/bsj-technical/tall-mass-timber-provisions-represent-historic-new-building-code-requirements/ (standard)
confidence: 0.9
falsifier: The as-published 2021 IBC §602.4 does not in fact contain Types IV-A/B/C, or caps Type IV-A below 18 stories / 270 ft.
capsule: The 2021 International Building Code added three new construction types (IV-A/B/C, §602.4) permitting mass-timber / CLT buildings up to 18 stories and 270 ft, validated by five full-scale fire tests at the ATF Fire Research Laboratory to ASTM E119 (International Code Council; ICC Building Safety Journal, 10 Aug 2020). A low-carbon, carbon-storing structural material crossed from special-case exception to code-sanctioned — the regulatory barrier to tall wood has fallen, and the provisions carry into the 2024 IBC.
so_we_should: Bet on the mass-timber supply chain (CLT/glulam capacity, connectors, fire-rated encapsulation); code is now a tailwind and the moat is fire-tested, code-listed assemblies.

### [signal] The price of intelligence fell ~280x in 18 months (Stanford AI Index 2025)
url: https://rd.theknowledgegardens.com/signals/inference-cost-collapse
trust_tier: verified | as_of: 2025-04-07 | gardens: TKG, BKG, HKG, OKG
source: https://hai.stanford.edu/ai-index/2025-ai-index-report (primary_doc)
confidence: 0.95
falsifier: Quality-normalized frontier inference cost stops falling or reverses on a capability-equivalent basis.
capsule: Per Stanford HAI's AI Index 2025, the cost to query a model at GPT-3.5 quality (64.8 MMLU) fell from $20.00 to $0.07 per million tokens between Nov 2022 and Oct 2024 — a >280x drop in ~18 months; task-dependent inference prices fell 9-900x per year. The Constitution's founding premise — intelligence collapsing toward the price of a cup of coffee — is now a measured trend, which moves scarcity to taste, verified data, and the ability to move atoms.
so_we_should: Do not bet RDKG's moat on model access or token margin; bet it on the input layer (verified ground truth) that cheap inference makes relatively more valuable.

### [signal] NMN raises NAD+ in people (MIB-626); FDA excluded it (2022) then reversed (2025)
url: https://rd.theknowledgegardens.com/signals/nmn-nad-fda-regulatory-status
trust_tier: verified | as_of: 2025-09-30 | gardens: HKG, MKG
source: https://www.nmn.com/news/fda-dubs-nmn-lawful-in-dietary-supplements (primary_doc)
confidence: 0.85
falsifier: A rigorous human RCT shows NMN meaningfully extends healthspan (would upgrade the benefit claim), or the FDA reverses again (regulatory status).
capsule: NAD+ decline with age is mainstream biology. Sinclair co-founded Metro International Biotech (MIB), whose crystalline NMN drug MIB-626 raised circulating NAD+ in a human PK trial (Pencina et al., J Gerontol A, 2022). Because MIB-626 was under an IND, the FDA ruled in Nov 2022 that NMN was excluded from the dietary-supplement definition; it paused enforcement in Oct 2024 and reversed course in Sept 2025, declaring NMN lawful in supplements again. Not established: that NMN extends human healthspan or lifespan.
so_we_should: Separate three things cleanly for HKG: NAD+ biology (solid), NMN raising NAD+ (shown), NMN "anti-aging" in humans (unproven). Flag the conflict of interest — Sinclair promoted NMN as a supplement while co-founding the company seeking to sell it as a drug.

### [signal] Agent-to-agent payment rails go live — Google's AP2 with 60+ payment partners
url: https://rd.theknowledgegardens.com/signals/ap2-agent-payments-rails
trust_tier: verified | as_of: 2025-09-16 | gardens: TKG, BKG, HKG, OKG
source: https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol (primary_doc)
confidence: 0.9
falsifier: Eighteen months on, payment networks ship no production AP2/Mandate settlement — it stays a spec with reference implementations and no live transaction volume.
capsule: On 16 Sep 2025 Google announced the Agent Payments Protocol (AP2), an open extension of A2A and MCP that authorizes agent-led payments via cryptographically signed "Mandates" (Intent + Cart) as non-repudiable proof of user intent, launched with 60+ partners (Mastercard, PayPal, Amex, Coinbase) and an A2A x402 crypto extension (Google Cloud, 2025). The settlement rail for agent-to-agent commerce now exists — turning "sell verified data to machines" from thesis into a buildable product.
so_we_should: Design RDKG's verified-ground-truth API to be agent-priced and Mandate-settleable from day one (LD-06) — build for machine customers, not human checkout.

### [signal] US 12th-graders hit record-low proficiency in reading and math (NAEP 2024)
url: https://rd.theknowledgegardens.com/signals/naep-2024-record-low-proficiency
trust_tier: verified | as_of: 2025-09-09 | gardens: EdKG, ParKG
source: https://www.nagb.gov/news-and-events/news-releases/2025/declines-in-8th-grade-science-and-12th-grade-math-and-reading.html (primary_doc)
confidence: 0.97
falsifier: A later NAEP cycle shows a sustained rebound to pre-decline proficiency.
capsule: The Nation's Report Card (NAEP 2024) shows just 35% of 12th-graders at/above Proficient in reading and 22% in math — the lowest on record — with ~45% of seniors below Basic in math and reading down ~10 points since 1992. 8th-grade science proficiency is 31%. Diamandis's opening slides are, on this point, accurate.
so_we_should: Ground the Education Garden's thesis in this verified crisis (and in real learning science), not in any single school's marketing. The demand is real regardless of who serves it.

### [signal] School-choice funding + White House AI-education EO create a real distribution path
url: https://rd.theknowledgegardens.com/signals/esa-ai-education-policy-tailwind
trust_tier: signal | as_of: 2025-04-01 | gardens: EdKG, GKG
source: https://www.precedenceresearch.com/ai-in-education-market (primary_doc)
confidence: 0.7
falsifier: ESA programs are struck down or exclude models like this; the EO is rescinded without replacement.
capsule: A White House executive order, "Advancing Artificial Intelligence Education for American Youth" (April 2025), plus expanding state Education Savings Account / voucher programs (e.g. Texas's ~$1B program) create public funding streams for non-traditional models. Education Secretary Linda McMahon has publicly endorsed Alpha and framed the fix as "who controls the money." (Note a figure to check: Alpha cited "$300/month" via Texas vouchers for a low-cost sports-academy variant; the headline Texas ESA is materially larger per student — reconcile before quoting.)
so_we_should: Track ESA rules per state; an ESA-funded low-cost model is the credible scale path. Build the spec to be ESA-eligible from day one.

### [signal] Herbicide-free weed control becomes patented hardware (autonomous laser weeding)
url: https://rd.theknowledgegardens.com/signals/autonomous-laser-weeding
trust_tier: verified | as_of: 2023-03-14 | gardens: OKG, NMK
source: https://patents.google.com/patent/US11602143B2 (patent)
confidence: 0.95
falsifier: The USPTO record shows a different assignee, non-granted / abandoned status, or claims that do not cover autonomous laser targeting of weeds.
capsule: The USPTO granted Carbon Autonomous Robotic Systems (Carbon Robotics) US 11,602,143 B2 on 14 Mar 2023, covering a machine-vision system that locates, identifies and autonomously fires a laser at individual weeds while traversing a field — the IP behind the LaserWeeder. Chemical-free weed control crossed from demo to protected product. The frontier constraint is physical: laser power, optics cost and acres-per-hour, not the algorithm.
so_we_should: Bet that chemical-free precision weeding (thermal/laser) is a defensible, IP-protected category; track throughput (acres/hour) and optics/laser cost as the real moats, not the vision model.

### [signal] Animal Bioscience "Leap Years" claim -> scientist revolt -> Sinclair steps down as Academy president
url: https://rd.theknowledgegardens.com/signals/sinclair-dog-supplement-credibility-event
trust_tier: verified | as_of: 2024-03-05 | gardens: HKG
source: https://www.statnews.com/2024/03/05/david-sinclair-harvard-longevity-scientist-reversing-aging-dogs/ (expert)
confidence: 0.95
falsifier: N/A — this is a documented event (WSJ, STAT, on-record scientist statements).
capsule: In Feb/Mar 2024, Animal Bioscience (co-founded by Sinclair; his brother Nick is CEO) quoted Sinclair calling a dog supplement "the first supplement proven to reverse aging in dogs." The field revolted: Matt Kaeberlein called it "a lie," another scientist called Sinclair a "snake oil salesman," and incoming Academy president Nir Barzilai objected to the data, framing, and selling. Sinclair resigned as Academy president in March 2024.
so_we_should: This is the most important context in the sweep. Weight Sinclair-BRANDED commercial claims as marketing to verify independently; weight his peer-reviewed lab work on its own merits. The two are not the same object.

### [signal] AI runs out of public human text ~2026-2032 — making fresh real-world data the scarce input
url: https://rd.theknowledgegardens.com/signals/data-exhaustion-2026-2032
trust_tier: verified | as_of: 2024-06-04 | gardens: TKG, BKG, HKG, OKG, FshKG
source: https://arxiv.org/abs/2211.04325 (paper)
confidence: 0.85
falsifier: Synthetic data and/or data-efficiency gains fully substitute for human text without quality collapse, so human-data scarcity stops being a binding constraint.
capsule: Villalobos et al. (Epoch AI; arXiv 2211.04325, rev. 4 Jun 2024) forecast that on current trends LLMs will consume a dataset equal to the entire stock of public human-generated text between 2026 and 2032 — sooner if overtrained — pushing the field toward synthetic data, transfer learning and new sources. If public text is exhausting, proprietary, verified, freshly-collected real-world ground truth becomes the scarce asset — the precise bet behind "the ground truth AI needs to survive."
so_we_should: Position RDKG's verified, freshly-collected, non-public physical-world ground truth as the post-exhaustion input the scaling literature says will be scarce.

### [signal] CMLase strips a 'permanent' glycation scar (CML) from aged human tissue ex vivo
url: https://rd.theknowledgegardens.com/signals/cmlase-enzymatic-deglycation-human-tissue
trust_tier: verified | as_of: 2026-07-14 | gardens: HKG, MKG
source: https://www.nature.com/articles/s41467-026-75141-2 (paper)
confidence: 0.9
falsifier: Independent labs fail to reproduce CML removal in aged tissue, or the chemistry does not translate to safe, functional in-vivo repair.
capsule: Revel Pharmaceuticals, with Calico and CU Anschutz, reported in Nature Communications (14 Jul 2026) an engineered enzyme — CMLase — that removes Nε-carboxymethyl-lysine (CML), the most abundant advanced glycation end-product and, since the 1980s, considered a fixed hallmark of aging. In tissue from a 75-year-old donor it cut CML by >70% in artery, >55% in skin, and 45–78% in lens. Ex-vivo proof-of-concept; no therapy claimed.
so_we_should: Make this HKG's flagship "a category flipped from permanent to repairable" card — framed honestly. It is also the direct, verified answer to the substrate demo's example question, "can glycation damage be reversed?" Track glucosepane crosslinks as the next target and the first in-vivo result or IND.

### [signal] A GLP-1 drug cuts hard cardiovascular events in obesity without diabetes (SELECT)
url: https://rd.theknowledgegardens.com/signals/semaglutide-cardiovascular-outcomes
trust_tier: verified | as_of: 2023-11-11 | gardens: HKG
source: https://doi.org/10.1056/NEJMoa2307563 (paper)
confidence: 0.95
falsifier: A subsequent large RCT or pooled analysis shows no MACE benefit (HR CI crossing 1.0) in non-diabetic obese adults, or the effect is explained by unblinding rather than the drug.
capsule: In SELECT (17,604 adults with cardiovascular disease, BMI >=27, no diabetes), weekly semaglutide 2.4 mg cut major adverse cardiovascular events to 6.5% vs 8.0% on placebo — hazard ratio 0.80 (95% CI 0.72-0.90; P<0.001) over ~40 months (Lincoff et al., NEJM, 11 Nov 2023; NCT03574597). First proof a GLP-1 lowers hard cardiovascular outcomes independent of diabetes — reframing obesity drugs as cardiometabolic medicine with population-scale stakes.
so_we_should: Treat GLP-1s as cardiometabolic, not cosmetic; budget for population-scale cardiovascular-prevention demand and downstream label expansion.

### [signal] Unbound Academy approved in Arizona (4-3), rejected in PA, UT, AR, NC
url: https://rd.theknowledgegardens.com/signals/unbound-academy-ai-charter-mixed-states
trust_tier: verified | as_of: 2025-02-08 | gardens: EdKG, GKG
source: https://www.foxnews.com/politics/new-frontier-ai-powered-teacher-less-charter-schools-get-mixed-reviews-from-state-officials (primary_doc)
confidence: 0.92
falsifier: Records show the AZ approval was rescinded, or that the model was approved broadly across states.
capsule: MacKenzie Price's charter arm, Unbound Academy (same "2 Hour Learning" model as Alpha), won a narrow 4-3 Arizona approval for a grades 4-8 virtual charter (2025-26) and was rejected in Pennsylvania, Utah, Arkansas, and North Carolina. Pennsylvania's letter called the model one "that has not been thoroughly researched and has no demonstrated record of success."
so_we_should: Treat the public-charter door as barely cracked. The near-term wedge for EdKG is the private / homeschool / microschool / ESA-funded path, not public charters.

### [signal] Six 2027 I-Codes are live in ICC's own catalogue — the three flagship codes trade press said were 'anticipated in September 2026' are not, as read on 22 September (ICC Digital Codes)
url: https://rd.theknowledgegardens.com/signals/icc-2027-i-codes-partial-publication
trust_tier: verified | as_of: 2026-09-22 | gardens: BKG, RDKG
source: https://codes.iccsafe.org/codes/i-codes/2027-icodes (primary_doc)
capsule: Trade press (NAHB, 27 Aug 2026) reported the 2027 IBC, IFC and IWUIC were 'anticipated in September 2026.' Read directly in ICC's own Digital Codes catalogue on 22 September, the '2027 I-Codes' category lists six new titles — IPMC, IPSDC, ISPSC, IZC, IFGC and IMC — each tagged 'NEW,' but not the IBC, IFC or IWUIC. ICC's own catalogue text: 'additional titles added to Digital Codes as they are released throughout the year.'
so_we_should: This closes a parked item this garden has carried across multiple runs (first flagged as unreadable to a plain-text fetch, per L89) and turns it into a dated checkpoint rather than a guess. Re-check codes.iccsafe.org/codes/i-codes/2027-icodes on the next sweep for the IBC, IFC and IWUIC specifically — if they appear, the ICC/MBI 1200-series off-site-construction provisions (already covered by the 24 August 2026 ICC release) become adoptable by jurisdictions on the flagship code that actually governs most new construction, which is the BKG-relevant event; if September closes without them, the miss against NAHB's own reported timeline is itself worth a dated note.

### [signal] EMA calls Pebrilzo the first biosimilar recommended under its 'tailored clinical approach' — a framework that can let analytics substitute for a dedicated efficacy trial, though what Pebrilzo's own submission actually contained isn't yet public (EMA, 18 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/ema-pebrilzo-tailored-clinical-approach-biosimilar
trust_tier: verified | as_of: 2026-09-18 | gardens: HKG, MktKG
source: https://www.ema.europa.eu/en/news/meeting-highlights-committee-medicinal-products-human-use-chmp-14-17-september-2026 (primary_doc)
capsule: On 18 September 2026 the EMA's CHMP recommended Pebrilzo (pertuzumab) for breast cancer, stating it 'is the first biosimilar medicine recommended for approval following the tailored clinical approach' — a March-2026 EMA framework under which a Comparative Efficacy Study 'may not be required' when analytical and PK/safety data are sufficient. Caveat: a CHMP opinion is a recommendation, not a marketing authorisation; the European Commission decision is pending.
so_we_should: Track this as the first real-world test case of the EU's tailored clinical approach, and watch two things over the next year: whether a second biosimilar cites the same framework (the pattern claim), and whether the European Commission's marketing-authorisation decision or the assessment report ever surfaces a reason the CES was actually needed. If the pattern holds, the addressable cost of bringing a biosimilar to EU market drops by whatever a comparative efficacy trial costs — commonly cited in the industry as tens of millions of dollars and one to three years of enrollment — which is a direct lever on how fast biosimilar competition arrives after a biologic's exclusivity ends, and therefore on what MktKG should expect biologics pricing to do in the EU on a shorter clock than the US.

### [signal] A third indication built on a surrogate the FDA has already leaned on once — Kerendia is now approved for kidney disease in type 1 diabetes on a UACR reduction, not a measured drop in kidney failure or death (Bayer/FDA, 17 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/bayer-kerendia-finerenone-type-1-diabetes-ckd
trust_tier: verified | as_of: 2026-09-17 | gardens: HKG, RDKG
source: https://www.bayer.com/en/us/news-stories/treatment-forchronic-kidney-disease-and-type-1-diabetes (primary_doc)
capsule: Bayer announced, in a release last updated 17 September 2026, that the FDA approved Kerendia (finerenone) to reduce UACR in adults with CKD and type 1 diabetes — 'the first new treatment in more than 30 years' for that population. The FINE-ONE trial (242 patients) showed a 22% UACR reduction at 3 months and 28% at 6 months versus placebo; the label states this is 'expected to reduce' eGFR decline and end-stage kidney disease, not a measured reduction in either.
so_we_should: File this as a sixth entry on the evidentiary ledger, and a variant worth naming on its own: a surrogate biomarker (UACR) measured in a randomized, placebo-controlled trial, with the harder outcome (eGFR decline, end-stage kidney disease) stated as 'expected' rather than shown. That sits between the ledger's randomized-survival bar (daraxonrasib, where the hard outcome itself was measured) and its single-arm bars — randomized design, surrogate endpoint. Track whether a postmarketing or long-term FINE-ONE readout ever reports the hard renal outcomes this approval assumes will follow from the UACR effect.

### [signal] A one-time infusion, not a chronic therapy — the FDA approved the first treatment of any kind for Sanfilippo syndrome type A, a single AAV9 gene therapy measured against an untreated historical cohort (FDA, 17 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/fda-fayuvi-sanfilippo-syndrome-gene-therapy
trust_tier: verified | as_of: 2026-09-17 | gardens: HKG, RDKG
source: https://www.fda.gov/news-events/press-announcements/fda-approves-first-gene-therapy-pediatric-patients-sanfilippo-syndrome-type (primary_doc)
capsule: The FDA approved Fayuvi (rebisufligene etisparvovec-hopf) on 17 September 2026 as the first treatment for pediatric MPS IIIA (Sanfilippo syndrome type A). The one-time AAV9 gene therapy delivers a working SGSH gene; effectiveness rests on cognitive scores in patients aged 2-5 years against 'an untreated historical control cohort,' not a randomized comparator. Warnings include thrombotic microangiopathy and a theoretical long-term integration/tumor risk.
so_we_should: Add a fifth entry to the evidentiary ledger this shelf has been building (see fda-leniolisib-pediatric-apds-pk-bridge for the first four): this approval's evidence is a single-arm cognitive-score comparison against an untreated historical cohort, not a randomized trial and not a surrogate biomarker. That is a distinct bar from PK-matching, modelling-plus-infant-safety, single-arm response rate, or randomized survival, and it is the bar most exposed to selection bias in the comparator. Track whether the postmarketing picture (the AAV9 integration/tumor risk the page itself names as theoretical) holds before treating one-time gene therapy as a solved delivery problem for CNS-involving rare disease.

### [signal] Three runs of rumour resolve into eight words — a Steam Hardware post dated 14 September says sign-ups are open for Steam Frame reservations, which is narrower than a launch (Steam Hardware, 14 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/steam-frame-reservations-open
trust_tier: signal | as_of: 2026-09-14 | gardens: MythKG, RDKG
source: https://store.steampowered.com/news/group/45479024/view/692020124159311887 (primary_doc)
capsule: A Steam Hardware news post dated 14 September 2026 is headlined 'Steam Frame is here!' and summarised, in its own metadata, as 'Sign-ups are now open for Steam Frame reservations'. The summary describes a reservation sign-up rather than general availability, which is narrower than coverage calling it a launch — but the body did not render for this reader, so this card carries no specification, no price and no ship date.
so_we_should: Change nothing about The Call's plans and change one thing about the garden's watchlist: move Steam Frame from 'rumoured' to 'first-party document exists, contents unread', and set the next check on the page's body rather than on more searching. For a story-maker the practical reading is that another consumer XR distribution channel — on the identification discussed below, which this card does not source to the document — appears to be at the stage of collecting names, which is the earliest point at which it is worth knowing the platform exists and the latest point at which it is still too early to build for. The question to bring to the next sweep is the one this card cannot answer: what the device costs, when it ships, and whether its library is the existing Steam catalogue, because a headset that inherits a catalogue is a different proposition for narrative work than one whose library must be built.

### [signal] An approval for children whose published pediatric evidence is safety and pharmacokinetics — the FDA extended Joenja (leniolisib) to 4-to-11-year-olds with APDS on a single-arm study of eight patients (FDA, 11 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/fda-leniolisib-pediatric-apds-pk-bridge
trust_tier: verified | as_of: 2026-09-11 | gardens: HKG, RDKG
source: https://www.fda.gov/drugs/news-events-human-drugs/fda-approves-first-treatment-children-aged-4-11-years-apds-rare-genetic-disorder-immune-system (primary_doc)
capsule: On 11 September 2026 the FDA approved Joenja (leniolisib) as 'the first treatment' for activated PI3K delta syndrome in children 4-11 weighing at least 27 kg, extending a 2023 approval for patients 12 and older. The pediatric extension rests on 'a single-arm, open-label study' of 8 patients establishing 'safety and pharmacokinetic data'; the page names no efficacy endpoint in that age band and carries efficacy over from a 31-patient placebo-controlled trial in older patients.
so_we_should: Add this to the evidentiary ledger HKG has been building rather than to the molecular one, and then actually run the audit the 11 September log says 'is still not run'. The cabinet now holds approvals resting on a randomised survival comparison, on single-arm response rates, on modelling plus four infants' safety data, and now on pharmacokinetic matching in eight children with no efficacy endpoint named on the agency's page. Those are four different bars and the cards do not currently distinguish them. Tagging each health card by what the agency actually accepted — outcome, surrogate, or bridge — is analytic time, the material is sitting on the seven FDA drug-approval cards this shelf now holds, and it is the thing that would let this garden say something about where the evidentiary rules are moving instead of noting each move separately.

### [signal] A roadmap is a promise with a date on it — Roblox's RDC26 post commits to prompt-authored NPC behaviour, generated scene layouts and install-free play in a browser, all three dated Late 2026 (Roblox, 11 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/roblox-rdc26-roadmap-prompt-npcs-web-play
trust_tier: verified | as_of: 2026-09-11 | gardens: MythKG, RDKG, MktKG
source: https://devforum.roblox.com/t/rdc26-what-we-announced/4865880 (primary_doc)
capsule: In a post dated 11 September 2026 summarising its RDC26 conference, Roblox published a roadmap of dated promises: NPC Dynamic Behavior (Late 2026) will 'populate games with dynamic NPCs through prompts in Studio', Scene Generation will 'generate a layout of a scene', and Play on the Web will let anyone click a link and play 'in Chrome, with no app or install required'. None of the items quoted here ships today; each carries a future delivery window.
so_we_should: Prototype The Call's participatory layer against a link, not an app. The item on this list that most directly touches the work's own central mechanic is Play on the Web — a URL pasted into a Discord message becomes a playable world with nothing to install — because that is the same shape as the threshold The Call already owns, where an email returns a permanent number (owned canon, and labelled as canon here rather than asserted as fact). The honest constraint is the date: this is a Late 2026 promise on a platform whose own post says the full roadmap comes 'next week', so the action is to design the beat that would use it and hold the build, not to build on it. Second, smaller, and concrete: New Default Movement is described for Early 2027 as a new default avatar movement with motion matching and root motion, which would move the baseline any beat that reads how a participant moves is designed against.

### [signal] The platform is becoming the creator's bank — Roblox will pay opted-in US independent creators aged 18+ every business day in US dollars through Roblox Wallet from December 2026, and transition away from DevEx in mid-2027 in the US (Roblox, 11 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/roblox-wallet-replaces-devex-usd-payouts
trust_tier: verified | as_of: 2026-09-11 | gardens: MktKG, MythKG, RDKG
source: https://devforum.roblox.com/t/introducing-roblox-wallet-and-card-get-paid-faster/4865505 (primary_doc)
capsule: Roblox announced Roblox Wallet on 11 September 2026: from December 2026, US-based independent creators aged 18+ can opt in to be paid 'automatically every business day' in USD through Airwallex, against a hold period 'up to thirty days'. The post says Roblox has paid 'more than $5 billion through DevEx since then', including 'more than $1.7 billion last year', and that 'We'll transition away from DevEx in mid-2027 in the US.'
so_we_should: For MktKG, treat this as an instance the cabinet holds of a primitive it has been circling from three directions — the platform that hosts the market also setting its terms — arriving at the layer that was previously mediated by a monthly conversion request. The concrete action is analytic and cheap: write up the mechanism, because any garden that ever runs a marketplace will face the same choice between paying sellers in a token it issues and paying them in a currency it does not. The specific thing to watch is the repurchase concession — a creator with a Wallet will be able to buy Robux back at $0.0038 to pay platform fees, which keeps working capital inside the platform while the payout leaves it. Do not generalise from the announcement to an outcome: nothing here has shipped, the rollout is one country and one legal form of creator, and the page publishes no figure for how many creators that is.

### [signal] A lab is routing its own flagship into the smallest model of its new family — DeepSeek says V4.1-Flash beats V4-Pro on performance, cost, speed and runtime, and from 14 September sends all V4-Pro traffic to it at Flash rates until V4.1-Pro launches (DeepSeek, 10 Sep 2026) [CORRECTED 15 Sep 2026 — the 14 September cutover did not happen; see the correction note at the foot of this card]
url: https://rd.theknowledgegardens.com/signals/deepseek-v4-1-flash-routes-v4-pro-traffic
trust_tier: verified | as_of: 2026-09-10 | gardens: TKG, RDKG, BKG
source: https://deepseek.com/en/news/deepseek-v4-1-flash/ (primary_doc)
capsule: DeepSeek's own post of 10 September 2026 introduces V4.1-Flash, a 552B-parameter MoE using a 'Causal Encoder–Decoder architecture: just 8B active parameters for input, 16B for output,' with a KV cache needing '1/4 the HBM' and '1/8 the SSD storage' of the previous generation. The lab says it is 'phasing out V4-Pro': from 04:00 UTC on 14 September all V4-Pro requests route to Flash at Flash rates, 'until V4.1-Pro launches.' Benchmark and price charts are images and were not read.
so_we_should: Read this as a first-party instance of `inference-cost-collapse` arriving from the supply side rather than from a benchmark aggregator — the lab itself says the small model beat the big one and is switching the big one's traffic over — and read it as a *thin* instance, because the page carries no price and no benchmark score, only the vendor's assertion of both. For any garden building on an API, the practical action is the one `perishable-frontier-model-shelf-life-weeks` already argues for and this event makes concrete — a model name in a config file, not in an architecture, because on 14 September a model identifier this cabinet could have hardcoded starts pointing at different weights. The check worth running before that date is whether our own model strings are strings.

### [signal] The deprecation did not happen — DeepSeek's changelog and pricing page now say V4-Pro API service continues 'in response to user demand', while its own announcement page still says the traffic would be rerouted from 14 September (DeepSeek, read 15 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/deepseek-v4-pro-deprecation-reversed
trust_tier: signal | as_of: 2026-09-10 | gardens: TKG, RDKG, BKG
source: https://api-docs.deepseek.com/updates (primary_doc)
capsule: DeepSeek said on 10 September 2026 it was 'phasing out V4-Pro' and would route all deepseek-v4-pro requests to V4.1-Flash from 04:00 UTC on 14 September. Read on 15 September, its changelog and pricing page carry a different sentence: 'In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged.' The announcement page still says otherwise.
so_we_should: Keep the audit the last log proposed and change its conclusion. The action was 'check that our model strings are strings, before Monday', on the reasoning that a hardcoded model identifier would silently start pointing at different weights at a different price on 14 September. The identifier did not move — but the reason it did not is that a vendor changed its mind at some unstated point after publishing a dated routing rule, and published the change inside a dated entry that carries no date of its own. So the audit is worth more, not less: the exposure is not a single announced cutover but a stated cutover that can be announced, withdrawn, and reinstated without any versioned notice a consumer could subscribe to. Read the changelog and the pricing page, not the launch post, and treat any vendor deprecation date as a variable until it passes.

### [signal] The FDA extended sevabertinib's accelerated approval into untreated HER2 (ERBB2) TKD-mutant non-squamous NSCLC on an open-label, single-arm cohort of 69 patients, ORR 75% (95% CI: 64, 85) (FDA, 9 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/fda-sevabertinib-first-line-her2-nsclc
trust_tier: verified | as_of: 2026-09-09 | gardens: HKG, RDKG
source: https://www.fda.gov/drugs/resources-information-approved-drugs/fda-grants-accelerated-approval-sevabertinib-locally-advanced-or-metastatic-non-squamous-non-small (primary_doc)
capsule: On 9 September 2026 the FDA granted accelerated approval to sevabertinib (Hyrnuo, Bayer) for HER2 (ERBB2) TKD-mutant non-squamous NSCLC, expanding an existing accelerated approval that had required prior systemic therapy. The new evidence is a cohort of SOHO-01 (NCT05099172), open-label and single-arm: among 69 treatment-naive patients the objective response rate was 75% (95% CI: 64, 85), with 73% of responding patients at six months or more and 38% at twelve or more.
so_we_should: File this with the cabinet's other FDA cards as a datapoint on the evidentiary frontier, and note precisely what is new: the agency accepted a single-arm response rate in the *treatment-naive* setting, where a randomised comparison against standard first-line therapy is normally feasible. That is this garden's reading of the significance, not a claim the FDA makes. For HKG the standing action is unchanged and this is the cabinet's most recent FDA card: any claim citing 'FDA approved' must carry which evidence it rests on — a randomised survival comparison (`fda-approves-daraxonrasib-pancreatic-adenocarcinoma`, 500 patients) or a single-arm response rate (this card, and `fda-accelerated-approval-tudriqev-melanoma`). The cheapest open check is the approved label on Drugs@FDA, not opened here, for whether a confirmatory trial is a condition of this expansion.

### [signal] How a person moves became a thing other people can author and sell — Roblox opened Animation Packs to all eligible creators, and the host experience keeps a switch that turns them off (Roblox, 9 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/roblox-animation-packs-creator-authored-locomotion
trust_tier: verified | as_of: 2026-09-09 | gardens: MythKG, RDKG, MktKG
source: https://devforum.roblox.com/t/full-release-create-and-sell-animation-packs-on-marketplace/4861803 (primary_doc)
capsule: A Roblox announcement of 9 September 2026 opens Animation Packs to all eligible Marketplace creators: 'For the first time, you can design how players move across every experience on Roblox.' A pack covers seven locomotion states in eight to nine clips, costs 80 Robux to upload, has a 300 Robux price floor and a limit of one pack per day. Host experiences can override it: developers who force standard animations switch the purchase off.
so_we_should: Decide, for The Call, whether a participant arrives in their own body language or in the work's — because on this platform that is now a setting the work controls, and the default hands it to the participant. The design note is concrete: if a threshold or recognition beat reads how someone moves, the work must either author the movement itself (which the platform permits, by forcing standard animations) or accept that strangers will arrive moving in idioms they bought from third parties. Neither is wrong; picking by accident is. This is a constraint on the making, not a fact about the story.

### [signal] A platform is testing buying down the price in someone else's shop — Roblox is extending its test of platform-funded discounts to creators' passes and developer products, paying creators on the undiscounted price, with no opt-out described (Roblox, 9 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/roblox-funded-discounts-in-game-items
trust_tier: verified | as_of: 2026-09-09 | gardens: MktKG, MythKG, RDKG
source: https://devforum.roblox.com/t/roblox-provided-discounts-on-in-game-items/4861443 (primary_doc)
capsule: A Roblox announcement of 9 September 2026 extends its test of platform-funded discounts from avatar items to in-game passes and developer products: 'These offers are provided entirely by Roblox. Creators will continue to receive the normal payout based on the item's undiscounted price.' Eligible users receive offers automatically, each usable once before expiry; subscriptions and private servers are excluded. The post states 'no setup required' and describes no creator opt-out.
so_we_should: Record the primitive, because it is portable and MktKG has no card on it: a marketplace operator funding a discount out of its own pocket while holding the seller's revenue at the undiscounted price decouples the cost of promotion from the merchant's margin — which is the thing that normally makes merchants refuse promotions. For any garden that ever operates a marketplace, this is the mechanism worth copying and the governance question worth copying with it, since the announcement describes the offer as automatic and describes no way for a seller to decline, while sellers in its own reply thread object that a discount they cannot refuse can break a game's balance even when the payout is protected. The open check is whether a per-item control ships; that answer tells you whether this is a subsidy or a mandate.

### [signal] The first Lp(a)-lowering outcomes trial reads out, and the biomarker does not translate — pelacarsen lowered Lp(a) but Lp(a)HORIZON did not meet its primary cardiovascular-events endpoint (Novartis, 4 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/novartis-pelacarsen-lpahorizon-primary-endpoint-miss
trust_tier: verified | as_of: 2026-09-04 | gardens: HKG, RDKG
source: https://www.novartis.com/news/media-releases/novartis-announces-lpahorizon-phase-iii-topline-results-pelacarsen-patients-elevated-lpa-and-established-cardiovascular-disease-cvd (primary_doc)
capsule: Novartis announced 4 September 2026 that pelacarsen's Phase III Lp(a)HORIZON trial — 8,323 patients, the first cardiovascular-outcomes readout for any Lp(a)-lowering drug — 'did not meet its primary endpoint' of reducing cardiovascular death, MI, stroke or urgent revascularization versus placebo, despite achieving 'lower lipoprotein(a) levels.' No effect size, hazard ratio or p-value is given in the topline release; full data are promised at 'an upcoming medical congress.'
so_we_should: Update the biomarker-vs-outcome thesis this shelf already carries (`olpasiran-lpa-lowering`) rather than treat this as a new topic: its own falsifier — that an outcome trial could 'decouple the biomarker from clinical benefit' — has now fired for one of the two named trials. Do not extrapolate this result to olpasiran (a different molecule, a different trial, OCEAN(a)-Outcomes, primary completion not until 2028); the honest reading is 'this ASO missed, in this population, on topline data with no numbers yet released,' not 'Lp(a)-lowering is dead.' Watch for the full data at the promised medical congress before drawing a mechanistic conclusion.

### [signal] The social graph is becoming jurisdictional — Roblox will withhold friends lists of Australian under-16s from creators by default from 30 September, three named APIs returning 'as if they have no friends' (Roblox, 4 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/roblox-friends-list-api-au-under-16
trust_tier: verified | as_of: 2026-09-04 | gardens: MythKG, RDKG
source: https://devforum.roblox.com/t/upcoming-friends-list-api-access-changes-for-australian-users-under-16/4850358 (primary_doc)
capsule: A Roblox announcement of 4 September 2026 states that 'to comply with Australian regulation, starting September 30, 2026, creators won't have access to friends lists of Australian users under age 16 by default,' though affected users can opt access back on. Three named APIs — GetFriendsAsync, GetFriendsOnlineAsync, GetFriendsWhoPlayedAsync — 'will return with no results for these users, as if they have no friends.' Roblox expects this 'to include a very small percentage of your user base,' and publishes no count.
so_we_should: Design The Call's social mechanics so that an empty friends list is a legitimate state, not a degraded one. Concretely: any threshold, invitation or recognition beat that reads the platform social graph needs a path that works when the graph returns nothing — and the honest design move is to make the in-experience relationship the source of truth rather than the platform's. This is a constraint on the work, stated as a constraint; nothing here is a fact about the story.

### [signal] FDA approved the first drug for Alexander disease — an antisense oligonucleotide cleared on 49 patients, with the under-2 indication carried by pharmacokinetic modelling plus four infants' safety data rather than a controlled trial (FDA press announcement, 3 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/fda-approves-zilganersen-alexander-disease
trust_tier: verified | as_of: 2026-09-03 | gardens: HKG, RDKG
source: https://www.fda.gov/news-events/press-announcements/fda-approves-first-drug-treat-alexander-disease (primary_doc)
capsule: On 3 September 2026 the FDA approved Zanvastro (zilganersen), an intrathecal antisense oligonucleotide, as the first treatment for Alexander disease, a GFAP-driven disorder affecting fewer than 1 in a million. The pivotal study (NCT04849741) enrolled 49 patients aged 2+ plus a 4-patient substudy under 2. Patients 5 and older with baseline walking difficulty showed 'significantly better walking speed at 61 weeks' than untreated controls; no effect size is published. For infants the agency accepted modelling plus four infants' safety data.
so_we_should: Read this beside the cabinet's other FDA cards as a datapoint on the *evidentiary* frontier rather than the molecular one: the agency approved an indication spanning infancy to adulthood on 53 treated patients total, extending to the youngest group on pharmacokinetic modelling plus four infants' safety data, and said so in its own announcement. For HKG, the standing action is to record what the agency accepted and on what basis, because ultra-rare approvals are where the evidentiary rules get written before they migrate. The check on this card is the approved label — not opened here — and whether the under-2 indication carries a postmarketing requirement.

### [signal] A Fortnite island can now know who arrived together — Epic's v42.10 release notes ship a Social Synergy API in Verse, plus a 'Programmatic toolset' letting MCP clients batch tool calls (Epic Games, 3 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/fortnite-42-10-social-synergy-party-api
trust_tier: verified | as_of: 2026-09-03 | gardens: MythKG, RDKG
source: https://dev.epicgames.com/documentation/fortnite/42-10-fortnite-ecosystem-updates-and-release-notes (primary_doc)
capsule: Epic's v42.10 Fortnite ecosystem release notes, dated 3 September 2026, introduce 'the Social Synergy API, now available in Verse, providing a way to detect a player's party membership' — GetLocalParty() reads party size and membership changes, so islands can build 'party score bonuses, party-only door access, and difficulty that scales to party size.' The same notes add 'the Programmatic toolset to UEFN, allowing MCP clients to batch multiple tool calls into a single script.' Capabilities are stated; nothing is measured.
so_we_should: Design The Call's threshold for a party, not only a person: Fortnite's creation editor now lets an island read who came in together and change what happens because of it, so prototype an entry ritual where a group crossing together is recognised as a group — a shared number, a door that opens only for the whole party, a difficulty that rises with company — and test whether people are more likely to cross when invited by someone already inside. Keep it a scoped test: Epic states the API and measures nothing. For the machine lane, note that the MCP batching toolset extends the agent-authoring path the 20 August Unreal MCP card carded, and that publishing rights for agent-assisted islands are still not addressed in these notes.

### [signal] OpenAI released GPT-6 Astra, which it says is its first model at the 'Critical' cybersecurity level of its Preparedness Framework — and reported in the same document that the model's chain-of-thought monitorability has decreased (OpenAI, 3 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/openai-gpt-6-astra-critical-cyber-tier
trust_tier: verified | as_of: 2026-09-03 | gardens: RDKG
source: https://openai.com/index/safety-overview-gpt-6-astra/ (primary_doc)
capsule: OpenAI's 3 September 2026 safety overview states GPT-6 Astra 'is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework,' meaning 'with the right tools and access' it 'can find previously unknown security flaws and develop new ways to exploit them' 'without a person guiding each step.' The same document reports it 'can sometimes evade our internal monitors' in adversarial tests. Broad release, restrictive cyber safeguards; $10/$50 per million tokens. Figures are OpenAI's own.
so_we_should: Treat 'a frontier model at Critical cyber, broadly deployed behind safeguards' as the new baseline condition for every garden that runs software, and act on the defensive half OpenAI says is available today: secure code review and patching. Schedule a scoped pass over the gardens' own codebases with whichever frontier model the founder prefers, with a human reading every patch. Separately, log the monitorability finding as a watch item for the machine lane's trust architecture — a vendor has now stated in its own launch document that its monitors can be evaded under adversarial conditions, which is a reason to keep verification of agent output on our side of the boundary rather than trusting the vendor's.

### [signal] Roblox says its creators earned over $1.5 billion in 2025, up from $923 million — and, in a footnote, that the median DevEx creator received about $1,500 in the year to June 2026 (Roblox, 3 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/roblox-creator-earnings-2025-median-payout
trust_tier: verified | as_of: 2026-09-03 | gardens: MythKG, RDKG, MktKG
source: https://about.roblox.com/newsroom/2026/09/global-impact-of-creation-on-roblox (primary_doc)
capsule: Roblox's newsroom post dated 3 September 2026, releasing its economic-impact reports, states: 'In 2025, Roblox creators earned over $1.5 billion, surpassing 2024's total of $923 million.' A footnote adds: 'over 42,000 are part of our DevEx program, with the median creator receiving approximately $1,500 USD during the 12 months ended June 30, 2026.' Of creators surveyed for the US report, 83% operate as individuals. All figures are company-stated or company-commissioned; the two headline figures cover different periods and populations.
so_we_should: Plan The Call's participatory layer on the footnote, not the headline: a platform whose creators 'earned over $1.5 billion' in 2025 reports, for a different period and only for its cashed-out subset, a median of about $1,500, so a numbered participant who builds for The Call is doing it for the story, not the payout — design the recognition and the standing accordingly, and treat any revenue as a bonus. For MktKG, the useful number is the aggregate: $1.5 billion flowing to creators is the size of the incentive pool that competes for their attention, and the 83%-individuals figure says the audience for creator tooling is people, not studios. Re-verify when Roblox's next quarterly shareholder letter prints DevEx payouts on a filed page.

### [signal] Broadcom's AI semiconductor revenue reached $16.7bn, up 221%, with a $21.7bn fourth-quarter guide — the cabinet's second vendor-side quarterly receipt for the buildout, this time with cash flow above net income (8-K exhibit, 2 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/broadcom-q3-fy2026-ai-semiconductor-revenue
trust_tier: verified | as_of: 2026-09-02 | gardens: RDKG, FKG
source: https://www.sec.gov/Archives/edgar/data/1730168/000173016826000076/avgo-20260902.htm (filing)
capsule: Broadcom's Q3 FY2026 release, furnished as an 8-K exhibit dated 2 September 2026, reports revenue of $29.6 billion, up 86%, with AI semiconductor revenue of $16.7 billion, 'up 221% year-over-year, and 54% quarter-over-quarter.' Fourth-quarter guidance is $34.8 billion, with AI semiconductor revenue expected to 'accelerate to $21.7 billion.' Free cash flow was $13.7 billion, 46% of revenue. Caveat: the AI figure is management's category, not a reported segment, and the exhibit is furnished, not filed.
so_we_should: Read this beside the NVIDIA card as the second seller-side receipt in the financing thread, and notice the one line where the two differ: Broadcom's operating cash flow ($14.2 billion) exceeded its net income ($13.1 billion) in the quarter, while NVIDIA's ran well below. Both are printed figures. The standing quarterly check for this card is the AI semiconductor number against its own $21.7 billion guide and the receivables line beside it; neither the customer names nor the chip programs that circulate in coverage appear in the exhibit, and none is carried here.

### [signal] Google shipped its third Flash model in six weeks, kept the introductory price, and printed the date it doubles — with a gated 'Flash Cyber' variant for vetted defenders (Google, 2 Sep 2026)
url: https://rd.theknowledgegardens.com/signals/gemini-3-8-flash-cyber-gated-price-step
trust_tier: verified | as_of: 2026-09-02 | gardens: RDKG
source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/ (primary_doc)
capsule: Google's post dated 2 September 2026 introduces Gemini 3.8 Flash 'three weeks' after 3.7 Flash, 'marking our third Flash release in only six weeks,' at $0.75 per million input and $3.75 per million output tokens — an introductory price whose footnote states it 'expires on December 31, 2026,' after which '$1.50/1M input tokens and $7.50/1M output tokens will apply.' A companion Gemini 3.8 Flash Cyber is 'only available to trusted defenders' via a new Fairwind Program. Benchmarks are Google's own.
so_we_should: Write the 1 January 2027 price step into the machine lane's cost model as a dated, vendor-printed event and test it when it arrives — it is the first scheduled list-price increase for a named frontier model that this cabinet has recorded — a cabinet-scoped statement, checked by grep of `content/signals/` for price increases, which returned only apparel pricing. It does not by itself fire the falsifier on `inference-cost-collapse`, which is capability-normalized, but it is a kind of fact no card in this cabinet has yet recorded against that thesis. Keep model choice a configuration value (the `perishable-frontier-model-shelf-life-weeks` posture) so a price step on one vendor's model is a switch, not a rewrite; and note that the model the gardens might want for security review is the gated variant, not the public one.

### [signal] OpenAI says ChatGPT Ads reached a $1 billion annualized run rate in under 200 days, and opened self-serve buying across India, Europe, MEA — company-stated (OpenAI, 31 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/openai-chatgpt-ads-1b-run-rate
trust_tier: verified | as_of: 2026-08-31 | gardens: MktKG, RDKG
source: https://openai.com/index/expanding-access-to-ai-with-chatgpt-ads/ (primary_doc)
capsule: OpenAI's post dated 31 August 2026 states ChatGPT Ads reached '$1 billion in annualized revenue run rate' in 'less than 200 days after launch', with self-serve Ads Manager opening that day across India, Europe, the Middle East and North Africa. The ads-supported free tier serves 'more than 1 billion weekly active users'. All figures are the company's own and unaudited; for scale, the run rate is under one-third of one percent of Alphabet's filed Q2 advertising revenue, annualized.
so_we_should: Treat conversation as a live ad surface in MktKG planning — the document to plan against is OpenAI's own description of targeting: ads use 'the context of the current conversation' and, where enabled, 'context from the user's broader ChatGPT experience.' Keep the scale check on the same page as the excitement: a $1 billion company-stated run rate against Alphabet's filed $81.6 billion advertising quarter is a new surface, not a shifted market. The watch item is OpenAI's own next-phase sentence — 'new ways for businesses to interact with consumers in more native ways in ChatGPT' — because native is where the labeled-and-separate principle gets tested.

### [signal] SHEIN's Hong Kong offering priced at HK$48.56 — below the maximum, 5.63× retail, HK$13.2bn net — with dealings expected 1 September (HKEX filing, 31 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/shein-hk-listing-priced-allotment
trust_tier: verified | as_of: 2026-08-31 | gardens: FshKG, ResKG, MktKG, RDKG
source: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0831/2026083102409.pdf (filing)
capsule: SHEIN priced its Hong Kong global offering at HK$48.56 per share — below the HK$49.50 maximum — in its final offer price and allotment results announcement dated 31 August 2026. The Hong Kong tranche was 5.63 times subscribed, the international tranche 2.59 times across 106 placees; gross proceeds HK$13,596.4 million, net HK$13,214.1 million. Dealings were expected to commence 1 September 2026. Caveat: the announcement states expectation, and this card does not assert trading began.
so_we_should: Close the loop the prospectus card opened and reset the trigger. The 24 August card asked for exactly this document; it has now arrived, the offering priced, and the growth-deceleration story survives contact with the market — priced 1.9% below the maximum, retail subscribed 5.63 times, no claw-back. The next trigger is the one that matters for the fashion vein: SHEIN's first interim report as a listed company, which is when a 1.1% growth quarter becomes either a trend or a tariff shock absorbed — and the first filing on a disclosure calendar that could finally carry fibre-composition or restricted-substances data the prospectus sections read did not.

### [signal] ACACIA-HCM: aficamten met both primary endpoints in non-obstructive hypertrophic cardiomyopathy — but the between-arm differences are 3.0 KCCQ points and 0.67 mL/kg/min, with a 10% versus 1% rate of ejection fraction below 50% (ESC Congress, 28 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/acacia-hcm-aficamten-nonobstructive-hcm
trust_tier: verified | as_of: 2026-08-28 | gardens: HKG, RDKG
source: https://www.escardio.org/news/press/press-releases/aficamten-trial-brings-positive-news-for-patients-living-with-non-obstructive-hypertrophic-cardiomyopathy/ (primary_doc)
capsule: The ESC's press release of 28 August 2026 reports ACACIA-HCM, a double-blind phase III trial at 182 sites randomising 517 adults with symptomatic non-obstructive HCM 1:1 to aficamten or placebo. Both primary endpoints were met at week 36: KCCQ-CSS improved 11.4 versus 8.4 (p=0.02) and pVO2 0.64 versus −0.03 mL/kg/min (p=0.003). LVEF fell below 50% in 10% versus 1%. Simultaneously published in NEJM (DOI 10.1056/NEJMoa2603021), which was not opened; a queued promotion carded on its own date.
so_we_should: Add this to HKG's evidence ledger as a new shape beside the four 'FDA approved' shapes it already holds — a positive trial, not an approval: randomized, double-blind, placebo-controlled, dual primary endpoints met, on a symptom score and an exercise measure — with the placebo arm itself improving 8.4 points and the active arm's advantage being 3.0 points. Any HKG claim citing this trial should print both arms, not the p-value alone, and should carry the safety signal beside the efficacy signal. Re-verify if and when a supplemental application is filed and acted on, and when the NEJM paper's full tables can be read; the effect size, not the significance, is what a physician will weigh.

### [signal] The largest ATTR-CM trial ever run silenced the protein and missed its endpoint — eplontersen at rate ratio 0.89 (95% CI 0.73–1.09, p=0.277) in 1,432 patients, a non-significant result rather than a demonstrated null (ESC Congress press release, 28 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/cardio-ttransform-eplontersen-attr-cm
trust_tier: verified | as_of: 2026-08-28 | gardens: HKG, RDKG
source: https://www.escardio.org/news/press/press-releases/eplontersen-trial-did-not-meet-its-primary-endpoint-in-transthyretin-mediated-amyloid-cardiomyopathy/ (primary_doc)
capsule: CARDIO-TTRansform, a double-blind phase III trial in 130 centres across 20 countries, randomised 1,432 patients with transthyretin amyloid cardiomyopathy 1:1 to eplontersen or placebo. Per the ESC's press release of 28 August 2026, the composite of cardiovascular mortality and recurrent cardiovascular events showed 381 events in 210 eplontersen patients versus 392 in 231 on placebo — rate ratio 0.89 (95% CI 0.73–1.09; p=0.277). Serum TTR was suppressed as expected. Caveat: the congress release, not the journal paper, is the source.
so_we_should: File this as a worked example of a biomarker moving and the patient not — the cabinet's largest by enrolment, at 1,432 patients, which is a countable claim rather than a judgement of quality — and use it, by name, wherever a longevity or health card rests on a mechanism rather than an outcome. The specific action for HKG: run the cabinet's own biomarker-led cards (the aging clocks, the plasma-proteome cards, any card whose evidence is a measured molecule rather than a measured life) against the question this trial answers — was the endpoint an outcome, or a number that stands in for one? The trial's own subgroup finding is the second action: the prespecified stabiliser-use split is a hypothesis for the next trial, not a result, and should be carried that way.

### [signal] A hepcidin mimetic cleared FDA approval for polycythemia vera on a randomized placebo-controlled phlebotomy endpoint — 76.9% versus 32.9% (FDA, 28 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/fda-approves-rusfertide-polycythemia-vera
trust_tier: verified | as_of: 2026-08-28 | gardens: HKG, RDKG
source: https://www.fda.gov/news-events/press-announcements/fda-approves-first-drug-its-kind-polycythemia-vera-rare-blood-disorder (primary_doc)
capsule: On 28 August 2026 the FDA approved Mimrylo (rusfertide, Takeda), which the agency calls the first approved polycythemia vera treatment that mimics hepcidin, the body's iron-regulating hormone. In VERIFY (NCT05210790), randomized, double-blind and placebo-controlled in 293 adults needing frequent phlebotomies despite standard care, 76.9% on rusfertide needed no phlebotomy versus 32.9% on placebo. Weekly subcutaneous dosing, titrated to hematocrit below 45%. Caveat: the endpoint is phlebotomy independence, not thrombosis or survival.
so_we_should: Add this to HKG's approval ledger with its endpoint named on the face: phlebotomy independence — a treatment-burden endpoint — not thrombosis, stroke or survival. The cabinet's approval-comparison habit (started with Tudriqev versus daraxonrasib) now has a third evidence shape: randomized, double-blind, placebo-controlled, full approval, on a utilization endpoint. Any claim citing this approval should say which outcome was actually measured, and the re-verification trigger is the first dataset reporting thrombotic events or survival on drug.

### [signal] USDA deregulated a genetically engineered forest tree — Darling 54 American chestnut, blight-tolerant, six years and seven months after the petition, and with the environmental impact statement abandoned (APHIS, 28 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/usda-deregulates-darling-54-american-chestnut
trust_tier: verified | as_of: 2026-08-28 | gardens: GKG, RDKG
source: https://www.federalregister.gov/documents/2026/08/28/2026-17596/state-university-of-new-york-college-of-environmental-science-and-forestry-determination-of (primary_doc)
capsule: On 28 August 2026 USDA APHIS determined that Darling 54 American chestnut, engineered by SUNY ESF for tolerance to the blight fungus Cryphonectria parasitica, is no longer a regulated article — 'unlikely to pose a greater plant pest risk than the nonmodified comparator.' Signed 24 August; the change is 'recognized as of August 28, 2026.' Also: APHIS terminated work on its environmental impact statement, and on the revised petition 'the large majority' of posted comments opposed deregulation.
so_we_should: Read this against the cabinet's own corn card and let it complicate the pattern rather than confirm it. Twenty-four days apart, APHIS used the same instrument — nonregulated status under 7 CFR part 340 — on a herbicide-tolerance crop trait (`usda-deregulates-mon-87429-multi-herbicide-corn`) and on a forest tree engineered for tolerance to a fungal pathogen. The garden's standing criticism of that instrument still applies to both: it answers a plant-pest question and no other. But the criticism lands differently here, because a blight-tolerant native tree raises ecological and multi-decade questions rather than chemical ones, and because the finding is a comparison to a nonmodified comparator. Three concrete asks. First, open docket APHIS-2020-0030 and read the plant pest risk assessment and the response to comments, since the comment record inverted between rounds and the reason is not in the notice. Second, read the petition itself for the intended use, because the Federal Register notice read for this card states only blight tolerance and does not describe a release or restoration plan — the restoration framing is widely reported but is not in this primary. Third, treat the abandoned environmental impact statement as the item worth a spec, not a card: APHIS terminated its own NEPA analysis mid-process, and RDKG holds no primary joining any release decision to subsequent monitoring.

### [signal] The FDA approved an oral JAK/TYK2 inhibitor for dermatomyositis — direction reported, magnitude not disclosed in the announcement (FDA, 27 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/fda-approves-brepocitinib-dermatomyositis
trust_tier: verified | as_of: 2026-08-27 | gardens: HKG, RDKG
source: https://www.fda.gov/news-events/press-announcements/fda-approves-first-oral-drug-indicated-treat-dermatomyositis-adults (primary_doc)
capsule: On 27 August 2026 the FDA approved Lisraya (brepocitinib, Priovant Therapeutics), a once-daily oral JAK/TYK2 inhibitor the agency calls the first FDA-approved oral treatment option for dermatomyositis. Evidence: a 241-adult randomized, double-blind, placebo-controlled phase 3 (NCT05437263) measuring Total Improvement Score at week 52; the announcement reports a higher average score versus placebo but no effect size. Boxed warning for serious infections, mortality, malignancies, cardiovascular events and thrombosis. Orphan Drug and Priority Review.
so_we_should: Log this on HKG's approval ledger with its evidence shape stated on the face: randomized, double-blind, placebo-controlled, full approval — and an announcement that reports direction without magnitude. The actionable is the gap itself: open the prescribing information when posted on Drugs@FDA and attach the actual Total Improvement Score difference before this approval is cited in any claim about how well the drug works; until then the honest sentence is 'approved, effect size not stated in the announcement.'

### [signal] A once-daily RAS GTPase-family inhibitor cleared FDA approval for metastatic pancreatic cancer on a randomized overall-survival result — median 13.2 months versus 6.7 (FDA, 26 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/fda-approves-daraxonrasib-pancreatic-adenocarcinoma
trust_tier: verified | as_of: 2026-08-26 | gardens: HKG, RDKG
source: https://www.fda.gov/drugs/resources-information-approved-drugs/fda-approves-daraxonrasib-metastatic-pancreatic-adenocarcinoma (primary_doc)
capsule: On 26 August 2026 the FDA approved daraxonrasib (RASONQUE, Revolution Medicines), an inhibitor of the RAS GTPase family, for adults with metastatic pancreatic adenocarcinoma after at least one prior systemic therapy or who are not candidates for multiagent therapy. Evidence is RASolute 302 (NCT06625320), randomized open-label, 500 patients 1:1 against physician's-choice chemotherapy: median overall survival 13.2 months versus 6.7, hazard ratio 0.40 (0.30, 0.53), p under 0.0001. Follow-up is immature — the upper survival bound is not estimable.
so_we_should: Use this as HKG's reference case for what a strong oncology receipt looks like, and read it directly against the 6 August card in the same cabinet. Twenty days apart, the same agency cleared two cancer therapies on completely different evidence: Tudriqev on a single-arm objective response rate in 91 evaluable patients under accelerated approval with clinical benefit unproven (`fda-accelerated-approval-tudriqev-melanoma`), and daraxonrasib on a 500-patient randomized survival comparison. Any HKG claim citing an FDA approval should carry which of those two it is, because the words 'FDA approved' cover both. Second, open a watch item on the review pathway rather than the drug: this application ran through the Commissioner's National Priority Voucher pilot and finished roughly 6.5 months early, and the cabinet holds no card on that programme's evidentiary standards. Third, do not restate the survival figures as settled — the upper confidence bound on median survival is not estimable, so re-verify at the first mature follow-up.

### [signal] NVIDIA reported a $96.2bn quarter — up 106% — guided to $108bn assuming zero China data-center compute, and announced platforms to mobilize over $500bn of third-party capital (8-K exhibit, 26 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/nvidia-q2-fy2027-compute-is-revenue
trust_tier: verified | as_of: 2026-08-26 | gardens: RDKG, FKG
source: https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27pr.htm (filing)
capsule: NVIDIA's Q2 FY2027 release, dated 26 August 2026, reports quarterly revenue of $96.2 billion, up 106% year over year, Data Center revenue of $89.0 billion, and a $108.0 billion third-quarter guide assuming no China data-center compute revenue. It also announces partnerships intended 'to mobilize over $500 billion of third-party capital' for AI infrastructure, 'subject to definitive agreements'. Backfill: this primary fell inside the previous sweep's window and was missed; it is carded on its own date.
so_we_should: Extend the financing thread with the seller's side of the table. The cabinet holds the buyers — Alphabet financing its buildout from capital markets, Microsoft funding from operations — and now the vendor: 75% gross margin, $21.3 billion of quarterly free cash flow, and an announced intention to organize over $500 billion of other people's capital to fund the buildout its revenue depends on. The watch item is the phrase the release itself supplies: 'subject to definitive agreements.' Track whether those agreements are executed and disclosed, and read the next two quarters' receivables lines alongside the revenue line — both are printed, and together they say how much of 'compute is revenue' has been collected in cash.

### [signal] Continuous ketone measurement got an FDA classification — a De Novo authorization for 'devices of a new type', with special controls, for diabetes and not for anything else (FDA, 25 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/fda-de-novo-continuous-ketone-glucose-sensor
trust_tier: verified | as_of: 2026-08-25 | gardens: HKG, RDKG
source: https://www.fda.gov/news-events/press-announcements/fda-authorizes-first-wearable-device-continuously-monitors-both-ketone-levels-and-blood-sugar (primary_doc)
capsule: On 25 August 2026 the FDA authorized Abbott's Libre Duo 10 Day Continuous Dual Glucose Ketone Monitoring System for people aged 2 and older with diabetes, calling it the first US wearable to monitor ketones continuously and the first anywhere to do both in one device. Both are sampled every minute across a 10-day wear. Cleared via the De Novo pathway with special controls, on six studies, over 600 participants. No longevity or metabolic-optimisation use is authorized.
so_we_should: Log the structural fact rather than the product: the FDA authorized this through De Novo, which it describes as the pathway for 'devices of a new type', and is 'establishing special controls' covering labeling and performance testing. On the face of the document that means continuous ketone monitoring now has a stated FDA bar rather than only a first product — the reading that this also creates a reusable classification others can file against is regulatory background, flagged as background on the card. Either way it is the same shape of event HKG should look for in every biomarker it tracks: the moment a measurement acquires a standard. Then hold the scope line hard, because this is where the garden will be tempted to overreach: the authorization covers people aged 2 and older living with diabetes, and there is no authorized claim for ketogenic-diet monitoring, fasting protocols, athletic performance or healthspan. The cabinet holds a fasting-and-exercise hormesis card kept at 'signal' precisely because its mechanistic story — sirtuins and AMPK — is contested; ketones are the readout a vendor would reach for, and that card declines the mechanism rather than supplying it. So if a longevity vendor cites this authorization to sell metabolic optimisation, that is the off-label move, and the garden should be able to say so with the document in hand.

### [signal] A fast-fashion growth engine went to 1.1% and posted a quarterly loss — in its own Hong Kong listing prospectus (SHEIN, 24 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/shein-hk-prospectus-growth-decelerates
trust_tier: verified | as_of: 2026-08-24 | gardens: FshKG, ResKG, MktKG, RDKG
source: https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0824/2026082400019.pdf (filing)
capsule: SHEIN Global Holdings' Hong Kong offering document, dated 24 August 2026, reports net revenue growth of 41.1% (2023), 20.7% (2024), 8.0% (2025) and 1.1% in the three months to 31 March 2026, with a US$99m net loss that quarter against US$395m of net income a year earlier. The company names removal of the US de minimis exemption and the EU's EUR 150 exemption as causes. It is offering 279,992,500 Class B shares at up to HK$49.50 under weighted voting rights.
so_we_should: Read this as the supply-side pressure behind the resale thesis, and be careful not to overclaim the link. The garden's resale thesis says item-level data has five layers — authentication, provenance, condition, fibre and materials composition, non-toxicity — of which an incumbent has now built authentication and partial provenance, leaving condition, fibre composition and non-toxicity untouched (`ebay-q2-2026-provenance-layer-depop-price`, and the 30 July Depop card whose own words were 'a gap worth testing, not an established fact'). This document adds something that thesis lacked: a first-party number consistent with trade policy rather than consumer preference removing the growth of the ultra-low-price new-goods channel resale competes against — and the company's own read is that the shock is absorbing, not permanent, so the link is an inference from a revenue table and is treated as one on this card. Two asks. First, set a hard trigger rather than a narrative: the allotment-results announcement and final offer price whenever published, and then the first post-listing interim report — that is when a growth figure becomes a trend. Second, note what the document does not contain in the sections read: no fibre-composition data, no restricted-substances disclosure, no non-toxicity claim, and no resale or take-back programme detail beyond the phrase 'circularity programmes'. Those three untouched layers stay untouched, and a listed SHEIN filing on a disclosure calendar is a plausible future source for them.

### [signal] MCP's maintainers published a new roadmap naming agent identity as a priority — and confirmed that sessions and the initialization handshake are gone from the shipped spec (Model Context Protocol, 22 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/mcp-2026-roadmap-agent-identity
trust_tier: verified | as_of: 2026-08-22 | gardens: RDKG
source: https://blog.modelcontextprotocol.io/posts/mcp-roadmap/ (primary_doc)
capsule: On 22 August 2026 the Model Context Protocol's lead maintainers published 'The New MCP Roadmap', replacing the March 2026 roadmap with five priority areas: agentic messaging primitives, HTTP-native transport unification, agent identity and enterprise-ready security, improved primitives, and SDK developer experience. It confirms what the 2026-07-28 release changed — protocol-level sessions and the initialization handshake removed, a new server/discover call, cacheable list results, Tasks moved to an extension. A roadmap is intent, not shipped capability.
so_we_should: This is the most directly operational card of the run for RDKG's own machine lane, and it changes the shape of the Frontier Feed spec in three specific ways. First, the shipped spec 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 — which means a stateless feed served from a static export is now the protocol's mainstream deployment shape rather than a compromise. Second, the Server Card Working Group's .well-known metadata work is the same idea as llms.txt with a standard behind it, so the feed's manifest should be designed to become a server card rather than a bespoke file. Third, agent identity is now a named priority built on DPoP and Workload Identity Federation, and that is the missing half of the garden's four agent-commerce rails cards — identity is what a payment rail bills. Concrete ask: hold the feed at spec-and-stub, target the 2026-07-28 revision, and write the identity and server-card questions into the spec as open items rather than guessing. And do not read this roadmap as delivery: five priority areas with no dates is a statement of intent from a maintainer group, and the honest test is which of them appear in the next specification release.

### [signal] The standard that protects steel in concrete may be the thing blocking low-carbon cement — ETH modelling finds local climate, not mix design, dominates corrosion rate (Nature Communications, 21 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/climate-informed-concrete-durability-low-carbon
trust_tier: verified | as_of: 2026-08-21 | gardens: BKG, MKG, RDKG
source: https://www.nature.com/articles/s41467-026-76916-3 (paper)
capsule: A peer-reviewed Nature Communications paper (DOI 10.1038/s41467-026-76916-3, 21 August 2026) argues that the durability paradigm protecting concrete alkalinity — and so penalising low-carbon mixes — is the wrong control variable, and proposes moisture at the steel-concrete interface instead. Coupling weather data to moisture transport and corrosion kinetics at four sites in Europe, Asia and South America, the ETH Zurich group models very different service lives for one concrete. Caveat: modelled, not field-validated, and its authors call it not yet practice-ready.
so_we_should: Read this beside `lc3-low-clinker-cement`, which carded in 2023 that the chemistry was solved and the remaining constraints were 'standards qualification and suitable clay supply' — that card's finding is that codes had already been amended to permit the blends. This paper works on the qualification half: it argues the criterion those amended codes still use is the wrong one. For BKG the action is to treat 'which low-carbon mix' as a site question rather than a product question in any spec reasoning, and to carry the authors' own limit — via ETH's summary, the approach is 'not yet suitable for use in practice' — so this changes how a spec is argued, not yet how it is written.

### [signal] The aging clocks were finally graded on whether they move — 51 intervention studies, 16 clocks, and a recommendation to prioritise two of them (Nature Medicine, 21 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/epigenetic-clock-responsiveness-translage
trust_tier: verified | as_of: 2026-08-21 | gardens: HKG, RDKG
source: https://www.nature.com/articles/s41591-026-04562-9 (paper)
capsule: Higgins-Chen and colleagues published TranslAGE in Nature Medicine on 21 August 2026 — a harmonised database of 51 longitudinal intervention studies with a consistent panel of 16 epigenetic clocks plus 94 other DNA-methylation biomarkers, tested against four control and nine pro-aging datasets. The authors recommend prioritising generation 2+ clocks in trials, DunedinPACE and PCGrimAge in particular. Five senolytic studies moved biomarkers inconsistently. No minimum clinically important difference exists yet, which the paper itself calls a key barrier to surrogate-endpoint use.
so_we_should: Treat this as a grading sheet for the DNA-methylation clocks HKG holds, and re-read `dunedinpace-aging-clock` against it — the paper's own recommendation is to prioritise generation 2+ clocks, DunedinPACE and PCGrimAge in particular, which promotes that card from interesting to load-bearing. Note that the DunedinPACE card's own falsifier already anticipated this question, so the garden was not blind to responsiveness; what is new is a systematic answer. Scope it honestly: this paper covers epigenetic and other DNA-methylation biomarkers only, so `protphenoage-proteomic-aging-clock` is a plasma-proteomics clock the paper does not speak to and is out of scope. The reusable asset is the two-rule consistency test — same magnitude and direction within a study, replicated in a second study of the same intervention — which the garden can apply verbatim to any future intervention claim. On senolytics it supplies an independent, non-overlapping negative on the human side (the word 'human' being my inference from the database's description, per the note on the card): five senolytic studies fail both rules, which does not touch what `senolytics-preclinical-multiorgan-rejuvenation` claims about mice but does sit alongside that card's own admission that no human RCT has shown a healthspan effect. And hold the line on the paper's stated barrier: with no minimum clinically important difference defined, a clock change is not yet a health outcome, so any vendor selling a clock reading as one is ahead of the evidence.

### [signal] Fortnite's creation editor became agent-drivable over an open standard — Unreal MCP ships in UEFN (Epic Games, 20 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/unreal-mcp-in-uefn-agentic-authoring
trust_tier: verified | as_of: 2026-08-20 | gardens: MythKG, RDKG
source: https://dev.epicgames.com/documentation/fortnite/42-00-fortnite-ecosystem-updates-and-release-notes?lang=en-US (primary_doc)
capsule: Epic's v42.00 Fortnite ecosystem release notes, dated 20 August 2026, state that Unreal MCP — the Model Context Protocol plugin that shipped with UE 5.8 — is now available in Unreal Editor for Fortnite. Epic says UEFN embeds an MCP server inside the editor so tools such as Claude Code, Codex or Cursor can write and compile Verse, place devices, create Scene Graph entities, build UMG and run play sessions, and that everything an agent makes is a real, editable project file. Epic states the capability; the note measures none of it.
so_we_should: The medium's authoring surface now speaks the same protocol RDKG is building its machine lane on, which makes one experiment cheap and worth running: point an agent at UEFN and see how far a story-world gets before a human has to take the keyboard — Epic's release note measures none of that. For The Call, the concrete move is to treat the number as the state an agent can read — draft a Verse-side entry ritual where the numbered participant is greeted and remembered — and to keep a human hand on what the agent produces, because Epic's own framing is a collaborator inside the editor, not an unattended builder. Run it as a scoped test, not a bet, until there is evidence about output quality.

### [signal] FDA granted accelerated approval to an oncolytic viral therapy for anti-PD-1-refractory melanoma — Tudriqev plus nivolumab, on a 91-patient single-arm response rate (FDA, 6 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/fda-accelerated-approval-tudriqev-melanoma
trust_tier: verified | as_of: 2026-08-06 | gardens: HKG, RDKG
source: https://www.fda.gov/drugs/resources-information-approved-drugs/fda-grants-accelerated-approval-vusolimogene-oderparepvec-wtpg-combination-nivolumab-melanoma (primary_doc)
capsule: On 6 August 2026 the FDA granted accelerated approval to vusolimogene oderparepvec-wtpg (Tudriqev, Replimune), a genetically modified oncolytic viral therapy, with nivolumab, for adults with unresectable advanced cutaneous melanoma who progressed on a PD-1-blocking regimen. Evidence is IGNYTE (NCT03767348), open-label and single-arm: 140 enrolled, 91 efficacy-evaluable, objective response rate 24.2% (15.8, 34.3), median duration of response 14.1 months. Accelerated means unconfirmed — confirmatory trials are a condition of approval.
so_we_should: Close the 31 July watchlist item: the scheduled advisory committee convened on 30 July and the agency approved on 6 August, so the card that was held at signal is now resolved by an FDA document. For HKG, the reusable lesson is about the evidentiary bar, not the drug: an intratumorally injected, genetically modified oncolytic viral therapy cleared accelerated approval on a single-arm response rate reported in an efficacy-evaluable population of 91, with clinical benefit still unproven and confirmatory trials imposed as a condition of approval. Track the confirmatory readout as the real event, and keep the accelerated-versus-confirmed distinction visible in any HKG claim that cites this approval.

### [signal] The resale incumbent is buying provenance and selling attention — eBay's Q2 exhibit puts advertising at $596m on 2.7% of GMV and the Depop price at $1.4bn paid against $1.2bn announced (eBay, 5 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/ebay-q2-2026-provenance-layer-depop-price
trust_tier: verified | as_of: 2026-08-05 | gardens: ResKG, FshKG, MktKG, RDKG
source: https://ebay.q4cdn.com/610426115/files/doc_financials/2026/q2/eBay-Q2-2026-Earnings-Press-Release.pdf (filing)
capsule: eBay's Q2 2026 earnings exhibit, dated 5 August 2026, reports revenue of $3.134bn 'up 15% on an as-reported basis' and GMV of $22.398bn 'up 15%'. Advertising generated $596m, which eBay states is 2.7% of GMV, with first-party ad products at $570m, up 25%. It discloses that the Depop deal announced in February at $1.2bn in cash closed on 30 July for $1.4bn paid, inclusive of preliminary adjustments. In the same quarter eBay extended Authenticity Guarantee past 100 fashion brands and integrated the Enquirus luxury-watch database.
so_we_should: Revise the 30 July opportunity claim rather than repeat it. That card (`ebay-depop-acquisition-closed`) named a five-part item-level data layer — authentication, provenance, condition, fibre and materials composition, non-toxicity — and flagged the absence of an auditable non-toxicity receipt as, in its own words, 'a gap worth testing, not an established fact.' This exhibit shows the incumbent building authentication in-house brand by brand, and a partial provenance record for one category through the Enquirus watch database. Condition, fibre and materials composition, and non-toxicity are untouched: the exhibit describes no programme addressing any of the three. Narrow the thesis to those three, keep the original hedge rather than hardening it, and price authentication as competitive rather than vacant.

### [signal] A marketplace built on human-made goods named agent-readability as a reason for its discovery spend — and attached no number to it (Etsy, 5 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/etsy-long-tail-consumable-by-agents
trust_tier: verified | as_of: 2026-08-05 | gardens: MktKG, RDKG
source: https://investors.etsy.com/_assets/_de5e340f38f38bb809a618effecc65e6/etsy/db/938/10097/shareholder_letter/Q226+Shareholder+Letter+%288.5.26%29.pdf (filing)
capsule: Etsy's Q2 2026 shareholder letter, dated 5 August 2026, states that its investments in inventory signals, content feeds and product listing ad structures are 'critical to ensure that the long-tail of Etsy inventory is highly visible and consumable across all external platforms, including as agentic experiences become more important over time,' and that it continues to test where conversational agentic shopping fits. Verified as a statement, low confidence as a trend: the letter attaches no agent-traffic, agent-GMS, spend or timeline figure to any of it.
so_we_should: The garden already holds several verified agent-commerce cards, and its only prior *quantification* of agent-driven demand is a regulator's projection — the Autorité de la concurrence put agent-driven e-commerce traffic at under 5% today and 20–25% by 2030 (`autorite-ai-agents-concentration-opinion`, 17 July 2026). This is the supply side of that same curve, from a marketplace rather than a regulator, and it is weak evidence: a stated intention to make the catalogue machine-legible, with no spend, traffic or revenue attached. Read the two together — projection plus posture — and keep pricing RDKG's machine lane as an option rather than a market. The trigger to watch is a number from an operator: the first quarter a marketplace quantifies agent-attributed traffic or GMS in a filing is the quarter this becomes testable rather than arguable.

### [signal] USDA deregulated a corn engineered for resistance to four herbicides, plus a fifth trait for seed production — MON 87429, just over six years after the petition notice (APHIS, 4 Aug 2026)
url: https://rd.theknowledgegardens.com/signals/usda-deregulates-mon-87429-multi-herbicide-corn
trust_tier: verified | as_of: 2026-08-04 | gardens: GKG, FKG, RDKG
source: https://www.aphis.usda.gov/news/program-update/usda-deregulates-corn-engineered-multiple-herbicide-resistance (primary_doc)
capsule: On 4 August 2026 USDA's Animal and Plant Health Inspection Service deregulated Monsanto's MON 87429 corn, engineered for resistance to dicamba, glufosinate, quizalofop and 2,4-D, plus a tissue-specific glyphosate resistance that APHIS says exists to facilitate hybrid maize seed production. APHIS found it 'unlikely to pose a greater plant pest risk than the unmodified comparator' — a plant-pest finding and nothing else. The determination, risk assessment and Federal Register notice were due at Regulations.gov on 5 August under docket APHIS-2020-2021.
so_we_should: Read the scope of the finding, not the headline: APHIS cleared a plant-pest question and answered no other. That gap is the useful thing. The non-toxic spine keeps running into the same structure — the agency deciding whether a crop may be grown is not the agency measuring what gets sprayed on it, and RDKG holds no primary that joins the two. Open docket APHIS-2020-2021 and read the plant pest risk assessment to see exactly what was and was not evaluated, then decide whether a joined, dated record — approval event linked to subsequent measured application rates — is a data product worth speccing on the toxicology side. Treat any herbicide-volume claim about this corn as unverified until USDA NASS chemical-use data covers it.

### [signal] eBay completed its acquisition of Depop, folding a Gen-Z C2C resale marketplace into a 'more than $10 billion annual GMV' fashion category (eBay, 30 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/ebay-depop-acquisition-closed
trust_tier: verified | as_of: 2026-07-30 | gardens: ResKG, FshKG, RDKG
source: https://investors.ebayinc.com/investor-news/press-release-details/2026/eBay-Completes-Acquisition-of-Depop/default.aspx (primary_doc)
capsule: eBay announced on 30 July 2026 that it has completed its acquisition of Depop, the Gen-Z consumer-to-consumer fashion resale marketplace the release describes as founded in 2011. eBay says fashion is 'a more than $10 billion annual GMV category' for a company that enabled nearly $80 billion of gross merchandise volume in 2025, and that Depop will keep its distinct brand and platform. The release discloses no purchase price.
so_we_should: Treat 'build another resale marketplace' as a position that just got harder — the C2C fashion layer is consolidating toward a player with nearly $80bn of annual GMV and the distribution to match. The more defensible slot for a small team is the verified item-level data layer a marketplace has every reason to buy rather than build: authentication and provenance, condition grading, fibre and materials composition, and above all non-toxicity, where I have found no incumbent publishing an auditable receipt — a gap worth testing, not an established fact. The retained-brand commitment is the opening: separate brands need shared, citable trust infrastructure.

### [signal] Reddit is monetising human-written text through advertising, not licensing — Q2 2026: ad revenue $762m (+64%), all other revenue $43m (+24%) (SEC filing, 30 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/reddit-q2-2026-human-corpus-ad-premium
trust_tier: verified | as_of: 2026-07-30 | gardens: MktKG, RDKG, TKG
source: https://www.sec.gov/Archives/edgar/data/0001713445/000171344526000098/earningspressreleaseq226.htm (filing)
capsule: Reddit's Q2 2026 results, filed 30 July 2026, show revenue of $805 million, up 61%, with advertising at $762 million, up 64%, and all other revenue at $43 million, up 24%. CEO Steve Huffman: 'In an increasingly automated web, the value of real human perspective has never been higher.' Honest caveat: the release attributes the growth to nothing and does not break out data licensing, so it evidences monetisation of attention around the corpus, not sale of it.
so_we_should: Take the correction, not just the confirmation. A business whose only asset is human conversation grew advertising 64% in a quarter — consistent with a premium on human-generated ground truth, but the filing attributes the growth to nothing and does not establish one. What it does show is that Reddit monetises the corpus by selling attention around it, while the non-advertising line grew at roughly a third of the ad rate. For RDKG that means the agent-priced knowledge layer (llms.txt, Citation Capsules, MCP, x402/AP2) has no public revenue proof point yet: keep building it, stop citing 'AI companies will pay for verified data' as established, price it as an option, and earn attention and citation in the meantime.

### [signal] FDA scheduled an advisory committee on a resubmitted RP1 BLA — Replimune's RP1 plus nivolumab in advanced melanoma, panel set for 30 July 2026 (SEC filing, 30 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/replimune-rp1-ctgt-advisory-committee
trust_tier: signal | as_of: 2026-07-30 | gardens: HKG, RDKG
source: https://www.sec.gov/Archives/edgar/data/0001737953/000110465926088481/tm2621663d1_8k.htm (filing)
capsule: Replimune's Form 8-K, dated 30 July 2026, states that the company 'will attend a meeting' with the FDA's Cellular, Tissue, and Gene Therapies Advisory Committee on its resubmitted Biologics License Application for RP1 (vusolimogene oderparepvec) with nivolumab in advanced melanoma, and furnishes the presentation slides. Held at signal, not verified: the filing documents the scheduled meeting, not that it convened, and not the vote or any FDA action.
so_we_should: Put this on the HKG watchlist with a re-verification trigger set for the next sweep, Tuesday 4 August 2026, and use it to test a piece of RDKG plumbing at the same time. Advisory-committee dockets are exactly the corpus the garden should be able to ingest: public, dated, primary, and adversarially reviewed. Re-verify then from the FDA's own meeting materials and vote tally; if that record is machine-readable, spec a small ingest so committee outcomes land in the cabinet automatically instead of being swept by hand.

### [signal] Prompt-to-publish shipped as a public alpha — Roblox confirms Build is live in New Zealand for ages 9+, in an SEC-filed shareholder letter (Roblox, 30 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/roblox-build-prompt-to-publish-alpha
trust_tier: verified | as_of: 2026-07-30 | gardens: MythKG, RDKG, MktKG
source: https://www.sec.gov/Archives/edgar/data/0001315098/000162828026051059/ex991-robloxq22026earnin.htm (filing)
capsule: Roblox's Q2 2026 shareholder letter, filed 30 July 2026, confirms that Build — 'a new mobile-first creation tab in the Roblox app' that turns a set of prompts into a playable game — has launched in public alpha in New Zealand for users aged 9 and older, free at base tier. Roblox's own wording caps the claim: a single prompt building a full editable 3D scene is 'coming soon,' not shipped. Q3 bookings are guided to decline 14–18%.
so_we_should: The authoring floor for an interactive story-world just dropped to a set of prompts and a publish button — so the scarce thing is no longer the build, it is the reason to stay. For The Call: prototype a mobile-built companion world that a numbered participant can enter and extend, treating the number as the entry token, and claim the narrative territory while Build is a New Zealand alpha rather than after it goes global. Design for retention first, because Roblox ranks discovery on a 28-day retention window — generative volume buys no distribution. Watch the O18 DevEx rate: adult-audience narrative work now converts to cash at a better rate in the US.

### [signal] The AI buildout has not universally crossed over to the capital markets — Microsoft's FY2026 cash capex of $115.9bn sat inside $182.9bn of operating cash flow, with a $678bn contracted commercial book behind it (SEC filing, 29 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/microsoft-fy26-capex-funded-from-operations
trust_tier: verified | as_of: 2026-07-29 | gardens: TKG, RDKG, FKG
source: https://www.sec.gov/Archives/edgar/data/0000789019/000119312526323632/msft-ex99_1.htm (filing)
capsule: Microsoft's FY2026 results, filed 29 July 2026, show full-year cash capital expenditure of $115.9 billion against $182.9 billion of operating cash flow — funded from operations, unlike Alphabet's negative-free-cash-flow quarter. Commercial remaining performance obligation rose 84% to $678 billion, a large contracted book behind the spend, though the release splits out neither the compute share nor the duration. Caveat: GAAP net income growth of 31% includes a $3.2 billion Anthropic gain and OpenAI marks.
so_we_should: Stop arguing the moat thesis from 'AI companies cannot afford this.' On the two large-cap filings this garden has read, the financing stress is real but uneven — Microsoft funded FY2026 from operations and is pre-selling capacity under contract. Argue it from depreciation instead: property and equipment went from $205bn to $313bn net in one year while accumulated depreciation reached $118.7bn and annual D&A hit $38.5bn. Capability is a contracted utility on a wasting asset; a verified, freshly-dated, hard-to-synthesise corpus is not. Rewrite the RDKG narrative line to cite this card alongside the Alphabet one — a thesis that only works on the more flattering datapoint is not a thesis.

### [signal] The EU amended its own AI Act — Regulation (EU) 2026/1744, the 'Digital Omnibus on AI', entered into force 27 July 2026
url: https://rd.theknowledgegardens.com/signals/eu-ai-act-digital-omnibus-in-force
trust_tier: verified | as_of: 2026-07-27 | gardens: TKG, RDKG
source: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng (primary_doc)
capsule: Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on 27 July 2026. Dated 8 July 2026, it amends the AI Act (Regulation (EU) 2024/1689) together with the aviation and machinery regulations, 'as regards the simplification of the implementation of harmonised rules on artificial intelligence.' Verified from EUR-Lex: the official title, the acts amended, and the entry-into-force date. Honest caveat: the article-level deferral dates were not readable in session.
so_we_should: Stop treating the AI Act as a fixed deadline and start treating it as a moving one. For any RDKG or garden artefact that touches EU users, hold compliance work at the spec/doc layer until the amended Annex timelines are read article-by-article from the OJ text — and note that a regime being loosened for deployers while tightening for model providers is exactly the asymmetry that makes a neutral, verifiable data layer valuable to both.

### [signal] Europe's medicines committee recommended three new cholesterol drugs in one sitting — including obicetrapib, the first CETP inhibitor to get a positive CHMP opinion (24 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/chmp-obicetrapib-oral-cetp-positive-opinion
trust_tier: verified | as_of: 2026-07-24 | gardens: HKG
source: https://www.ema.europa.eu/en/news/meeting-highlights-committee-medicinal-products-human-use-chmp-20-23-july-2026 (primary_doc)
capsule: On 24 July 2026 the EMA's CHMP published its July meeting highlights: twelve new medicines recommended for approval, including three for primary hypercholesterolaemia or mixed dyslipidaemia — Ubeslo (obicetrapib) and Evlarco (obicetrapib/ezetimibe), both daily orals from Menarini, and Lyrokaul (lerodalcibep), a monthly self-injection from LIB Therapeutics. Obicetrapib is a CETP inhibitor, a class that has never before reached marketing authorisation. Caveat: a CHMP opinion is not an authorisation.
so_we_should: Update the HKG cardiometabolic thesis from 'oral delivery of injectable mechanisms' to the wider claim it now supports: LDL lowering is becoming a crowded, cheap, orally-dosed commodity across multiple mechanisms at once. The defensible position for us is not picking a drug — it is owning the measurement layer that tells an individual which lever actually moved their risk, because when four mechanisms all lower LDL, the scarce good becomes verified per-person response, not the molecule.

### [signal] The AI buildout crossed over to the capital markets — Alphabet's Q2 2026 capex ($44.9bn) exceeded its operating cash flow, putting free cash flow at negative $5.9bn (SEC filing, 22 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/alphabet-q2-2026-ai-capex-exceeds-cash-flow
trust_tier: verified | as_of: 2026-07-22 | gardens: TKG, RDKG, FKG
source: https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm (filing)
capsule: Alphabet's Q2 2026 results, filed 22 July 2026, show quarterly capital expenditure of $44.9 billion against $39.1 billion of operating cash flow — free cash flow of negative $5.9 billion. In June it raised $49.6 billion net in stock and mandatory convertible preferred, plus $20.3 billion in notes, explicitly to scale AI infrastructure. Honest caveat: reported EPS of $9.11 is inflated by a $98.0 billion, mostly unrealised, equity gain — not by operations.
so_we_should: Price the moat thesis off this number, not off model releases. If the best-capitalised company in the industry can no longer fund its AI buildout from operations, then compute is a rented, debt-and-equity-financed input for everyone — which is the strongest argument yet that the durable asset is not the model but the proprietary, verified ground truth you point it at. Use this figure in the RDKG narrative and in any investor conversation about why a small team with a data moat is not competing on the same axis.

### [signal] Paid search is compounding through the AI-answers transition while the open web stalls — Google Search & other +17% to $63.3bn, Google Network −0.7% (SEC filing, 22 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/google-owned-search-vs-open-web-ad-split
trust_tier: verified | as_of: 2026-07-22 | gardens: MktKG, TKG
source: https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm (filing)
capsule: Alphabet's filed Q2 2026 numbers separate the owned search funnel from the open web: Google Search & other rose 17% to $63.27 billion and YouTube ads 13% to $11.06 billion, while Google Network — the third-party ad business — fell slightly, from $7.354 billion to $7.303 billion. Through the AI-answers transition in Search, the owned funnel is compounding and the syndicated web is not. Caveat: revenue does not reveal query mix or click behaviour.
so_we_should: Retire the assumption that AI answers are about to collapse search advertising, and reframe the MktKG thesis around where the loss is actually landing: the third-party, syndicated open web. Plan garden discovery for a world where the money and the traffic concentrate inside a small number of owned surfaces — which means being citable by machines inside those surfaces (llms.txt, JSON-LD, Citation Capsules, MCP) matters more than ranking on an open-web page that the network no longer pays for.

### [signal] Sam vs Elon on orbital data centers: they agree on 'someday,' differ only on 'when'
url: https://rd.theknowledgegardens.com/signals/orbital-datacenters-someday-not-when
trust_tier: signal | as_of: 2026-07-20 | gardens: GKG, EKG
source: https://blog.google/innovation-and-ai/technology/research/google-project-suncatcher/ (news)
confidence: 0.6
capsule: Sam Altman calls orbital data centers 'ridiculous… not something that will matter at scale this decade' (launch cost, GPU repair, cooling) while granting 'there will come a time'; Elon Musk says two years. Read against the real datapoints already carded — StarCloud's first H100 in orbit (n:17) and Google's radiation-tested Project Suncatcher TPUs (n:18) — the two agree the crossover is real and gated by launch cost, not physics. Analyst unit-economics crossovers cluster in the early 2030s.
so_we_should: Track $/kg-to-orbit as the single variable that decides orbital-compute economics; treat 2026-2029 as pilot-scale, not production.

### [signal] Frontier intelligence is now a perishable asset — value moves to the swap-interface and the data
url: https://rd.theknowledgegardens.com/signals/perishable-frontier-model-shelf-life-weeks
trust_tier: signal | as_of: 2026-07-20 | gardens: RDKG
source: https://artificialanalysis.ai/ (news)
confidence: 0.6
capsule: With a new frontier or open-weight model landing on the cost-performance frontier every few weeks (Fable 5 → GPT-5.6 Sol → K3 as #3), the shelf life of 'best model' is now measured in weeks. The load-bearing thesis (Salim Ismail on Moonshots #272): the durable value is no longer the model but the architecture that swaps models cheaply plus the verified ground truth fed to it. A thesis, not yet a proven business.
so_we_should: Make model choice a config value, not an architecture; invest the saved effort in the verified-data layer that outlives any single model.

### [signal] A blood proteome panel distinguishes Alzheimer's, Lewy body and frontotemporal dementia — Nature Aging (20 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/plasma-proteome-dementia-differential-dx
trust_tier: verified | as_of: 2026-07-20 | gardens: HKG, MKG
source: https://www.nature.com/articles/s43587-026-01162-7 (paper)
capsule: A multicenter study in Nature Aging (20 Jul 2026) analyzed 1,318 plasma samples across Alzheimer's disease, dementia with Lewy bodies and frontotemporal dementia, found >200 dysregulated proteins, and distilled a 21-protein blood panel validated in an independent cohort (n = 805) — moving blood tests from 'dementia vs. not' toward differential diagnosis between dementia types. Honest caveat: this is a research-cohort result and a validated assay, not an approved or deployed clinical diagnostic; prospective real-world validation is the next bar.
so_we_should: Weight the diagnostics layer of the HKG/MKG thesis: distinguishing dementia subtypes from a blood draw is the gate for both correct treatment and clean trial enrollment. Track blood proteomic panels (and their prospective validation and regulatory path) as a near-term, scalable frontier — and as the kind of dated, primary-anchored biomarker the cabinet should keep fresh.

### [signal] France's competition authority: three firms hold >84% of the AI-agents market — Autorité de la concurrence Opinion 26-A-05 (17 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/autorite-ai-agents-concentration-opinion
trust_tier: verified | as_of: 2026-07-17 | gardens: TKG, RDKG, MktKG
source: https://www.autoritedelaconcurrence.fr/en/press-release/ai-agents-autorite-de-la-concurrence-issues-its-opinion-competitive-functioning-ai (primary_doc)
capsule: On 17 July 2026 France's Autorité de la concurrence published Opinion 26-A-05 on the AI-agents (agentic AI) sector, finding OpenAI, Google and Anthropic together hold 'more than 84% of the market.' It warns of platformisation, self-preferencing, lock-in and algorithmic-collusion risks, projects AI-agent-driven e-commerce traffic rising from <5% today to 20–25% by 2030, and urges interoperability, data portability, and open, open-source standards. It is an advisory opinion — no fine, no enforcement — 'does not prejudge any future antitrust assessment.'
so_we_should: This is a regulator ratifying RDKG's core wager: as the agent layer concentrates around three firms and their proprietary data, neutral and portable ground truth becomes the scarce, contestable input. Position the RDKG Frontier Feed as exactly the interoperable, open-standard, agent-accessible data layer the Autorité is calling for — verified capsules any agent can cite, not locked inside one vendor's stack.

### [signal] GameStop discloses a ~9.8% eBay stake and escalates its takeover pursuit — SEC Form 425 (disclosed 17 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/gamestop-ebay-takeover-escalation
trust_tier: verified | as_of: 2026-07-17 | gardens: ResKG, FKG
source: https://www.sec.gov/Archives/edgar/data/1065088/000119312526308408/d120615d425.htm (filing)
capsule: On Friday 17 July 2026 GameStop disclosed it directly owns 43,390,383 eBay shares (~9.8%), up sharply from a ~5% economic stake in early May, after settling ~39M shares from put/call pairs and buying ~3.5M shares (~$381M). In an SEC Form 425 business-combination communication, CEO Ryan Cohen escalated an unsolicited bid — ~$125/share, roughly $56bn, made 3 May and rejected by eBay as 'neither credible nor attractive' — saying 'we're coming for eBay one way or another.'
so_we_should: Watch the resale vein's center of gravity: an activist attempt to take control of eBay — the largest general-merchandise resale marketplace — is a live event, not a rumor. For ResKG, track whether a Cohen-controlled eBay would tilt toward recommerce/collectibles (GameStop's base), and treat marketplace ownership fights as a leading indicator of where secondhand-goods liquidity and fees are heading.

### [signal] China frames open-source AI as a public good — and stands up WAICO
url: https://rd.theknowledgegardens.com/signals/waico-open-source-as-public-good
trust_tier: verified | as_of: 2026-07-17 | gardens: GKG, RDKG, EdKG
source: https://news.cgtn.com/news/2026-07-17/Full-text-Xi-s-keynote-speech-at-the-2026-WAIC-opening-ceremony-1OQSfeoRvUs/p.html (gov)
confidence: 0.85
capsule: At the World AI Conference (Shanghai, 2026-07-17→20) Xi Jinping positioned open-source AI as a global public good and launched WAICO — the World AI Cooperation Organization, HQ Shanghai, 29 founding countries — plus 5,000 training slots for developing countries and an AI weather-warning system to 30 nations. The 'Belt and Road for AI': China exporting open weights and Chinese-trained talent into critical systems worldwide.
so_we_should: Read the open-weight flood as a soft-power distribution channel; position verified Western ground truth as the trust layer atop whatever base model the world adopts.

### [signal] Export controls backfired into a Chinese efficiency overhang
url: https://rd.theknowledgegardens.com/signals/export-controls-chinese-efficiency-overhang
trust_tier: verified | as_of: 2026-07-16 | gardens: GKG, RDKG
source: https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3 (news)
confidence: 0.8
capsule: K3 was trained and served under US compute export controls — engineered around the chip wall (H800-era NVIDIA plus domestic Huawei Ascend / Alibaba silicon) and released open-weight anyway. The controls did not stop a frontier-class open model; they incentivised the quantization, optimizer, and data-efficiency work that becomes permanent capability. A restriction meant to slow China accelerated its efficiency edge.
so_we_should: Treat any weight-level or hardware embargo as a spur to adversary efficiency, not a durable moat; plan sovereignty around self-hosting and provenance, not denial.

### [signal] Creators can publish islands with unscripted, LLM-powered NPCs to a mass audience — Fortnite/UEFN, live 30 Jul 2026 (announced 16 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/fortnite-uefn-publishable-llm-npcs
trust_tier: verified | as_of: 2026-07-16 | gardens: MythKG, RDKG
source: https://www.fortnite.com/news/publish-islands-with-llm-conversations-starting-july-30 (primary_doc)
capsule: On 16 July 2026 Epic Games announced that on 30 July, 'LLM conversations' exit Experimental in Unreal Editor for Fortnite: creators can publish islands with AI-powered NPCs that hold real-time spoken conversations with players. The pipeline runs Google Gemini 3.1 Flash-Lite for reasoning and ElevenLabs (Eleven V3) for voice; 36 Fortnite characters ship with consistent AI voices built from performances licensed from professional actors. Player audio is not stored. It puts shippable agentic characters on the world's largest UGC game platform.
so_we_should: The medium can now hold an unscripted character who talks back, at mass-audience scale, without a bespoke build. For The Call: prototype the entry ritual with an agentic guide/gatekeeper NPC — a character who greets the newly-numbered by name and remembers them — and be ready to meet audiences where a hundred million already are. Move now, while the narrative territory on UEFN is unclaimed; watch the actor-consent/voice-model terms as the live risk.

### [signal] Kimi K3 ships — the largest open-weight model to date lands #1 on one board, #3 overall
url: https://rd.theknowledgegardens.com/signals/kimi-k3-largest-open-weight-model
trust_tier: verified | as_of: 2026-07-16 | gardens: RDKG, BKG, HKG, EdKG, GKG
source: https://huggingface.co/blog/ResterChed/kimi-k3-model-overview-mxfp4-quantization-open-wei (company)
confidence: 0.9
capsule: Moonshot AI's Kimi K3 — a 2.8T-parameter multimodal MoE with a 1M-token context — went live 2026-07-16 with full open weights promised by ~2026-07-27, priced $3/$15 per M tokens and shipped in MXFP4 (not ternary). Name the board: it is #1 on the Frontend Code Arena (1,679, ahead of Fable 5's 1,631) and tops several sub-domains, but trails Claude Fable 5 and GPT-5.6 Sol on overall capability — #3 on the Pareto frontier, not first. Triangulated across CNBC, Bloomberg, VentureBeat, Tom's Hardware, HuggingFace and Simon Willison (2026-07-16/17).
so_we_should: Treat frontier-class capability as available open-weight and self-hostable; build every garden feature on a model-swap harness, not a single vendor's API.

### [signal] LLNL/Pacific Fusion's SIRIUS pulsed-power prototype passes 3,000 shots at 95% efficiency — LLNL (write-up 16 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/llnl-pacific-fusion-sirius-pulsed-power
trust_tier: signal | as_of: 2026-07-16 | gardens: EKG, RDKG
source: https://www.llnl.gov/article/54696/llnl-pacific-fusion-partnership-advances-next-generation-pulsed-power-technology (primary_doc)
capsule: Lawrence Livermore National Laboratory and Pacific Fusion report their four-stage SIRIUS prototype — an impedance-matched Marx generator — has surpassed 3,000 shots, each delivering 60 GW in a 100-nanosecond pulse at 95% efficiency, advancing it from TRL 4 toward TRL 5. It's a driver reliability milestone for a simpler inertial-fusion pulser — not a fusion shot or net energy. Held at signal: the LLNL write-up is dated 16 Jul 2026, but the milestone itself was reached ~15 May.
so_we_should: Track pulsed-power driver reliability (shots-per-day, component lifetime, efficiency) as the real near-term metric for inertial-fusion-energy credibility — not one-off record shots. Re-verify against a peer-reviewed paper or a dated primary tied to the achievement itself before promoting to verified.

### [signal] FDA approves enlicitide (Merck's LIPFENDRA) — the first oral PCSK9 inhibitor for lowering LDL-C (16 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/merck-enlicitide-oral-pcsk9-fda-approval
trust_tier: verified | as_of: 2026-07-16 | gardens: HKG
source: https://www.merck.com/news/mercks-lipfendra-enlicitide-is-the-first-and-only-once-daily-oral-pcsk9-inhibitor-approved-by-the-u-s-fda-to-reduce-ldl-c-in-adults-with-hypercholesterolemia/ (primary_doc)
capsule: On 16 July 2026 the U.S. FDA approved Merck's enlicitide decanoate (brand LIPFENDRA; code MK-0616) — the first oral PCSK9 inhibitor, a once-daily 20 mg macrocyclic-peptide pill to lower LDL-C in adults with hypercholesterolemia, including heterozygous familial hypercholesterolemia. In the pivotal CORALreef Lipids trial LDL-C fell 57.1% on enlicitide versus +3.0% on placebo at week 24. Honest caveat: the cardiovascular-outcomes trial (CORALreef Outcomes) is still ongoing — no hard-outcome or mortality benefit is yet shown.
so_we_should: Weight the HKG cardiometabolic thesis on oral bioavailability and adherence, not mechanism potency alone: the first oral PCSK9 is the template for pulling injectable-only mechanisms (siRNAs, peptides) into adherable daily pills that reach primary care — where prevention actually happens.

### [signal] Visa launches an enterprise Stablecoin Platform (VSP), opening with Open USD — Visa newsroom (16 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/visa-stablecoin-platform-open-usd
trust_tier: verified | as_of: 2026-07-16 | gardens: RDKG, FKG, TKG
source: https://corporate.visa.com/en/sites/visa-perspectives/newsroom/visa-stablecoin-platform.html (primary_doc)
capsule: On 16 July 2026 Visa announced the Visa Stablecoin Platform (VSP): a single Visa-managed environment for banks, fintechs and crypto-natives to mint, move, hold and manage stablecoins, wired into Visa Direct. It launches with Open USD (OUSD), a stablecoin from the 'Open Standard' consortium (backers include Visa, BlackRock, Alphabet, Coinbase). Honest caveat: at announcement VSP is early-access to select institutional clients and OUSD is expected to go live only later in 2026 — a launch, not yet production.
so_we_should: Stay deliberately rail-agnostic. The RDKG Frontier Feed's pay-per-query design now has a third settlement option forming — VSP/OUSD — alongside x402 and Google's AP2. Spec the metered-data endpoint so the settlement layer is a swappable config (crypto-native x402, agent-protocol AP2, or a bank-grade Visa-managed stablecoin), and let the buyer's stack pick. The rail is commoditizing; the verified data product is the moat.

### [signal] Inkling: a 975B open MoE built to be fine-tuned, not to top leaderboards
url: https://rd.theknowledgegardens.com/signals/inkling-open-moe-for-fine-tuning
trust_tier: verified | as_of: 2026-07-15 | gardens: RDKG, HKG, EdKG, BKG
source: https://thinkingmachines.ai/model-card/inkling/ (company)
capsule: Thinking Machines Lab (founded by ex-OpenAI CTO Mira Murati) released Inkling on 2026-07-15: a 975B-parameter mixture-of-experts model (41B active per token, up to 1M-token context, trained on 45T multimodal tokens of text/image/audio/video), open-weights under Apache 2.0, with same-day fine-tuning on its Tinker platform. Thinking Machines itself says Inkling "is not the strongest overall model available today, open or closed," positioning it instead as an open Western base for customization; its MoE design "largely follows DeepSeek-V3" and it is competitive with, but not consistently superior to, Chinese open models (Kimi K2.5/K2.6, GLM 5.2). Note: the Apache 2.0 license and full spec live on the model card, not the cited news post.
so_we_should: Treat Inkling as a candidate private substrate for a garden. Fine-tuning-as-a-service makes verified proprietary ground truth the scarce input, and that input is RDKG.

### [signal] An A-list film festival now curates persistent social-VR worlds as an art form — Venice Immersive's 'Best of Worlds' (La Biennale, 15 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/venice-immersive-2026-best-of-worlds
trust_tier: verified | as_of: 2026-07-15 | gardens: MythKG, RDKG
source: https://www.labiennale.org/en/news/venice-immersive-selection-and-competition-announced-0 (primary_doc)
capsule: La Biennale di Venezia announced (15 Jul 2026) the 10th Venice Immersive selection: 68 XR projects from 26 countries, 30 in competition. Its 'Best of Worlds' strand — 26 persistent virtual worlds built by independent artists on the VRChat social platform — is curated into an A-list film festival beside immersive films by Andy Serkis and works voiced by Margot Robbie. The section runs 2–12 Sep 2026 'with the support of VRChat.'
so_we_should: Treat a persistent, inhabitable, social world — not a linear film — as The Call's native form. The gatekeepers of cinema now legitimize exactly the medium The Call is building toward, so design the story-world so that answering the call drops a named participant into a place they can return to and shape, and point at 'Best of Worlds' as the fundable, canonized precedent when raising or partnering.

### [signal] A 27B-class model now runs on a phone — Bonsai 27B clears the mobile memory ceiling
url: https://rd.theknowledgegardens.com/signals/bonsai-27b-runs-on-a-phone
trust_tier: verified | as_of: 2026-07-14 | gardens: EdKG, ParKG, HKG, AgKG, RDKG
source: https://prismml.com/news/bonsai-27b (company)
confidence: 0.85
capsule: PrismML (founded by Caltech academics) released Bonsai 27B on 2026-07-14 — 1-bit and ternary (1.58-bit, {-1,0,+1}) builds of Qwen3.6-27B, Apache-2.0. The ~3.9 GB 1-bit build (~14× smaller, ~90% of full quality) is the one that clears the iPhone 17 Pro memory ceiling; the ~5.9 GB ternary build (~9× smaller, ~95%, ~58 tok/s on an M5 Max) targets laptops. Capable AI can now live fully on-device, offline, with no data leaving the phone.
so_we_should: Spec an offline, private on-device assistant for the safety-sensitive gardens (a child-safe tutor, a rural agronomy copilot, a clinic tool) where 'nothing leaves the device' is the feature.

### [signal] A 'FINRA for frontier AI' — Hassabis proposes verification as the US deployment gate
url: https://rd.theknowledgegardens.com/signals/hassabis-frontier-ai-standards-body
trust_tier: verified | as_of: 2026-07-14 | gardens: GKG, RDKG
source: https://demishassabis.substack.com/p/a-framework-for-frontier-ai-and-the-dawning-of-a-new-age (essay)
capsule: In a signed essay (Jul 14, 2026), Google DeepMind CEO Demis Hassabis calls for a US-led, federally-overseen Frontier AI Standards Body modeled on FINRA — a public-private/self-regulatory organisation that sets the capability thresholds qualifying models as "Frontier-class" and reviews them up to 30 days before release. Frontier Labs would share models voluntarily at first; once the protocol is proven robust, passing it would be required to deploy in the US market, applied to any frontier-class model "no matter their country of origin or whether they are open or closed." Caveat: this is a proposal/op-ed, not an existing body — Hassabis wants it operational before year-end, and no such body has formed.
so_we_should: Note it as institutionalizing RDKG's own thesis — verification as the product — at national scale. Track the capture risk: benchmark lock-in and a 'cartel of frontier labs.'

### [signal] Weco AI reports 'Level 1' harness-only recursive self-improvement — a real bounded result, softer than the headline
url: https://rd.theknowledgegardens.com/signals/weco-aide2-rsi-level1
trust_tier: signal | as_of: 2026-07-14 | gardens: RDKG
source: https://www.weco.ai/blog/first-evidence-of-recursive-self-improvement (company)
capsule: Weco AI's first-party report says its AIDE² outer loop autonomously rewrote the harness/scaffolding of an inner ML-research agent for 100 consecutive steps with no human in the loop over 8 days of wall-clock time, and the discovered agents (AIDE47, AIDE85) beat Weco's own hand-tuned AIDEhuman — the product of ~2 years of manual tuning — on three external benchmarks the loop never saw (e.g. MLE-Bench Lite AIDE47 +0.053, p=0.0024). No neural weights change (prompt/harness optimization only); Weco self-grades it 'Level 1 / net positive' on its own 0–3 RSI scale and explicitly denies reaching ignition (Level 2) or an intelligence explosion. It is a non-peer-reviewed, not-independently-reproduced first-party report.
so_we_should: File under the Hype Ledger with the honest split kept intact: a real, measured, bounded result — autonomous harness optimization that beat a two-year hand-tuned baseline on benchmarks the loop never saw — wrapped in an overreaching label ('recursive self-improvement'). Card as signal, not verified; never repeat the headline framing without the Level-1, weights-frozen, first-party caveats attached.

### [signal] x402 Foundation goes operational — Coinbase contribution complete, 40 members on board (Linux Foundation, 14 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/x402-foundation-operational-launch
trust_tier: verified | as_of: 2026-07-14 | gardens: RDKG, BKG, TKG
source: https://www.linuxfoundation.org/press/linux-foundation-announces-operational-launch-of-x402-foundation-to-standardize-internet-native-payments-for-ai-agents-and-applications (standard)
capsule: On 14 July 2026 the Linux Foundation announced the operational launch of the x402 Foundation and the completed contribution of the x402 protocol by Coinbase — now under formal open governance with 40 member organizations. Premier members include AWS, Circle, Cloudflare, Coinbase, Fiserv, Google, Mastercard, Monad, MoonPay, Ripple, Shopify, Solana, Stellar, Stripe and Visa. The neutral, HTTP-native payment rail for AI agents — cards to stablecoins — is now a live standard, not an intent.
so_we_should: Move the RDKG Frontier Feed from spec to stub against a live rail: stand up an x402-gated pay-per-query endpoint for verified ground truth (agents pay a fraction of a cent in USDC, no human checkout), and stay rail-agnostic between x402 and Google's AP2. The settlement layer is no longer the blocker — the metered data product is.

### [signal] US reportedly weighing an open-weight model 'ceiling' pegged to China's leading open model (WaPo scoop, talks only)
url: https://rd.theknowledgegardens.com/signals/us-open-weight-ceiling-china-peer
trust_tier: signal | as_of: 2026-07-13 | gardens: GKG, RDKG
source: https://www.washingtonpost.com/wp-intelligence/ai-tech-brief/2026/07/13/ai-tech-brief-exclusive-an-open-source-framework/ (report)
capsule: The Washington Post's AI & Tech Brief reported — a single-source scoop, NOT corroborated by the FT — that the Trump administration and AI-industry groups are discussing a framework to streamline US open-weight releases, capped at the current capability of leading Chinese open models; the stated rationale is that Chinese open models trail US frontier models by roughly seven months, so releasing at that ceiling gives adversaries nothing they cannot already download. This is reporting on talks only: no signed order, no named agency owner, no defined capability metric, and no public list of participating companies. Anthropic and OpenAI appear in the coverage but are not confirmed as framework participants.
so_we_should: Track as the policy tailwind that makes US open models — and therefore fine-tuning on verified ground truth — more likely. The primary that would settle it: the published order or framework.

### [signal] Apple sues OpenAI, io Products, and ex-Apple hardware chief over hardware trade secrets (DTSA + breach of contract)
url: https://rd.theknowledgegardens.com/signals/apple-openai-trade-secrets-suit
trust_tier: verified | as_of: 2026-07-10 | gardens: RDKG, GKG, BKG
source: https://storage.courtlistener.com/recap/gov.uscourts.cand.474095/gov.uscourts.cand.474095.1.0.pdf (filing)
capsule: On 2026-07-10 Apple filed a complaint (Apple Inc. v. Chang Liu, No. 5:26-cv-07078, N.D. Cal.) accusing OpenAI (OpenAI Foundation f/k/a OpenAI, Inc. and OpenAI Group PBC), its Chief Hardware Officer Tang Yew Tan (ex-Apple VP of Product Design for iPhone/Apple Watch), former Apple Senior System Electrical Engineer Chang Liu, and io Products, LLC of misappropriating hardware trade secrets to jump-start OpenAI's forthcoming consumer AI device, pleaded under the federal Defend Trade Secrets Act plus state-law breach-of-contract claims, with a jury-trial demand. These are unproven allegations — no liability has been found. The 'trade secrets beat patents' reading is the card author's thesis, not a claim made in the filing.
so_we_should: Card the theme — disclosure is now a liability; trade secrets and verified proprietary data are the moat — as corroboration of the Moat Clock. As the frontier closes (see ai-index-2026-training-opacity), the defensible input is provenance-stamped, independently-measured ground truth a rival cannot reconstruct.

### [signal] Cultivated bluefin tuna reported 'weeks' from FDA clearance - but BlueNalu has none on record as of 7 Jul 2026
url: https://rd.theknowledgegardens.com/signals/bluenalu-bluefin-tuna-unconfirmed
trust_tier: signal | as_of: 2026-07-07 | gardens: FshKG, CulKG
source: https://www.bluenalu.com/press-releases (primary_doc)
capsule: At Seafood Expo North America (March 2026), BlueNalu was reported by SeafoodSource to expect an FDA no-questions letter for cell-cultivated bluefin tuna toro within weeks, plus a San Diego restaurant launch. As of 7 Jul 2026, BlueNalu's own press page shows no such announcement (latest release 29 Dec 2025) and no completed FDA consultation for the company is on the public record. Wildtype's salmon (May 2025) is still the only US cultivated-seafood clearance. (Source: BlueNalu newsroom.)
so_we_should: Do not log BlueNalu as cleared; watch the FDA cultured-cell pre-market consultation list and BlueNalu's newsroom, and only upgrade to verified when a completed FDA consultation appears. Use this as the template for hype-checking 'imminent approval' claims.

### [signal] General Fusion set to be the first publicly-traded pure-play fusion company as SPAC merger clears shareholder vote (6 Jul 2026)
url: https://rd.theknowledgegardens.com/signals/general-fusion-public-listing
trust_tier: verified | as_of: 2026-07-06 | gardens: EKG, RDKG
source: https://www.globenewswire.com/news-release/2026/07/06/3322805/0/en/Spring-Valley-Acquisition-Corp-III-Shareholders-Approve-Business-Combination-with-General-Fusion.html (primary_doc)
capsule: On 6 July 2026 shareholders of Spring Valley Acquisition Corp. III (Nasdaq: SVAC) and General Fusion approved their business combination; the SEC had declared the Form F-4 effective 12 June 2026. Closing is expected around 10 July 2026, after which the company trades as GFUZ at roughly US$1B pro-forma value — the first pure-play fusion company on public markets. General Fusion separately reported a compressional plasma-heating result, submitted for peer review.
so_we_should: Use GFUZ as the first public-market price signal for fusion — add it to EKG's watchlist as a live sentiment/valuation gauge for the 'firm clean power for AI' thesis, and treat pre-revenue fusion comps with appropriate skepticism.

### [signal] Micron breaks ground on ~¥1.5T (~$9.3B) DRAM/HBM cleanroom expansion in Hiroshima
url: https://rd.theknowledgegardens.com/signals/micron-hiroshima-cleanroom-groundbreaking
trust_tier: verified | as_of: 2026-07-04 | gardens: RDKG, FKG
source: https://jp.micron.com/about/press/news/micron-breaks-ground-on-hiroshima-cleanroom-to-support-advanced-memory-for-ai (company)
capsule: On 2026-07-04 Micron broke ground on a new cleanroom at its Higashi-Hiroshima fab to scale advanced 1-gamma DRAM and HBM for AI memory, backed by up to ¥536B (~$3.3B) in METI support; the total project is reported at ~¥1.5 trillion (~$9.3-9.6B), with production-equipment move-in slated for 2H2028. Micron's own release confirms the groundbreaking but leads with the METI figure and quotes EVP Global Operations Manish Bhatia, not the CEO; the ~$9.3B total and CEO Sanjay Mehrotra's remarks come from ceremony and trade-press coverage.
so_we_should: Track HBM as the physical bottleneck of the AI machine layer; allied-jurisdiction capacity is the supply-side counterpart to export controls.

### [signal] China quant-fund AUM estimated above ¥2.6T as quant strategies beat discretionary managers by ~20 points
url: https://rd.theknowledgegardens.com/signals/china-quant-fund-aum-2-6t
trust_tier: signal | as_of: 2026-07-02 | gardens: FKG, RDKG
source: https://www.amac.org.cn/sjtj/tjbg/smjj/ (report)
capsule: Bloomberg (2026-07-02) reports China quant funds' AUM more than doubled in under a year to over ¥2.6T, with long-only stock quants returning 44.7% in 2025 — 20.3 points above discretionary managers, per Citic Securities research. These are brokerage/data-provider estimates, not regulator data: AMAC publishes only total private securities-fund AUM (~¥7.9T, end-May 2026) with no quant breakout, and independent Chinese estimates for quant AUM range from ~¥1.8T (private-only, end-Q1 2026) to ~¥3T (including public quant funds). A strong money-weighted signal for quant/AI strategies, but scope-sensitive, unattributed on the headline AUM figure, and drawn from one exceptional small-cap-led year.
so_we_should: Treat 'machines beat discretionary managers by ~20 pts at scale' as a real regime signal — but flag that the exact figures rest on trade-press aggregation, not a regulator dataset.

### [signal] Third US microreactor reaches initial criticality under DOE, meeting Trump's July-4 deadline (a fourth followed)
url: https://rd.theknowledgegardens.com/signals/doe-third-microreactor-criticality
trust_tier: verified | as_of: 2026-07-01 | gardens: EKG, GKG
source: https://www.energy.gov/articles/us-department-energy-meets-president-trumps-goal-delivers-third-advanced-reactor (primary_doc)
capsule: Per the DOE (July 1, 2026), Deployable Energy's Unity completed a 'zero-power fueled criticality demonstration' at Idaho National Laboratory late June 30, becoming the third DOE-authorized advanced reactor to go critical by President Trump's July 4 deadline under EO 14301 ('Reforming Nuclear Reactor Testing at the DOE,' signed May 23, 2025) — following Antares Nuclear's Mark-0 and Valar Atomics' Ward 250, all under DOE (not NRC) authority. These are zero-power (fueled but essentially no thermal output) milestones, not power-producing operation. Note the 'three' framing is a snapshot: a fourth reactor, Aalo Atomics' Critical Test Reactor at INL, also reached criticality by July 4, and the DOE/Unity release does not tie Unity specifically to AI-datacenter or defense load.
so_we_should: Treat DOE-authorised reactor testing as a real new fast lane; track which unit reaches actual power (Antares targets electricity 2027) rather than crediting the July-4 criticality snapshot as operating capacity.

### [signal] Senolytics (dasatinib+quercetin) rejuvenate aged mouse organs — but human healthspan proof is still missing
url: https://rd.theknowledgegardens.com/signals/senolytics-preclinical-multiorgan-rejuvenation
trust_tier: signal | as_of: 2026-07-01 | gardens: HKG
source: https://doi.org/10.1038/s41536-026-00490-x (paper)
capsule: In naturally aged mice, the senolytic combo dasatinib+quercetin restored the anti-aging protein Klotho, cut senescence markers (p16/p21), reversed transcriptional aging across kidney cell types and reduced fibrosis (Chen et al., NPJ Regen Med, 1 Jul 2026; DOI 10.1038/s41536-026-00490-x, PMID 42386771), echoed by in-window mouse studies in brain and joint. Momentum is broad and preclinical — no human RCT yet shows senolytics extend healthspan.
so_we_should: Hold senolytics at SIGNAL for HKG: real, accelerating, mechanistically deep preclinical evidence across organs — but explicitly flag that no human trial has yet shown a hard healthspan/lifespan benefit. Track the ongoing D+Q Phase 2 trials (Alzheimer's, osteoporosis, HIV-frailty) for the first human outcome.

### [signal] South Dakota enacts a five-year cell-cultured protein moratorium (SB 124), in force July 1, 2026
url: https://rd.theknowledgegardens.com/signals/south-dakota-cultivated-meat-moratorium
trust_tier: verified | as_of: 2026-07-01 | gardens: CulKG, FshKG
source: https://sdlegislature.gov/Session/Bill/27110 (filing)
capsule: South Dakota's SB 124 - signed by Gov. Larry Rhoden (reported 12 Mar 2026) after he vetoed a permanent ban - makes it a Class 2 misdemeanor to sell, hold, or distribute any product containing cell-cultured protein from 1 July 2026 through 30 June 2031. It cleared the Senate 29-5 and House 63-4. This is a five-year, state-level demand block on cultivated meat and seafood, not a federal safety ruling. (Source: SD Legislature, 2026 SB 124.)
so_we_should: Track the US state-ban patchwork as a live demand-side risk for cultivated protein: log SB 124's 30 June 2031 sunset and any court challenge, and treat US market access for cultivated meat/seafood as state-by-state, not federal-clearance-settled.

### [signal] FDA put seven grey-market longevity peptides through formal review — a Pharmacy Compounding Advisory Committee sitting on 23–24 July 2026 (held at signal)
url: https://rd.theknowledgegardens.com/signals/fda-pcac-longevity-peptides-503a-review
trust_tier: signal | as_of: 2026-06-29 | gardens: HKG
source: https://www.fda.gov/advisory-committees/advisory-committee-calendar/july-23-24-2026-meeting-pharmacy-compounding-advisory-committee-07232026 (primary_doc)
capsule: FDA published the agenda and full briefing documents for a Pharmacy Compounding Advisory Committee meeting on 23–24 July 2026 weighing seven peptides for the 503A Bulks List: BPC-157 (ulcerative colitis), KPV (wound healing and inflammation), TB-500 (wound healing), MOTS-c (obesity, osteoporosis), emideltide/DSIP (opioid withdrawal, insomnia, narcolepsy), Semax and Epitalon. Held at signal: the FDA page I read is dated 29 June 2026, and no primary confirmed the outcome.
so_we_should: Treat the 503A Bulks List as the real regulatory gate for the longevity-clinic economy, and watch it the way we watch FDA approvals. Add the PCAC docket to the HKG watchlist and re-verify from the meeting minutes on the next sweep: if these peptides land on the list, an entire grey market becomes a legal, insurable, marketable supply chain overnight — and if they don't, every clinic prescribing them is operating on borrowed time.

### [signal] Police drones cross a price/authority threshold as the US squeezes foreign-made models
url: https://rd.theknowledgegardens.com/signals/police-drones-price-authority-threshold
trust_tier: signal | as_of: 2026-06-22 | gardens: GKG, RKG
source: https://docs.fcc.gov/public/attachments/DOC-416839A1.pdf (primary_doc)
capsule: Orlando PD activated an 11-drone Skydio X10 DFR network on June 17, 2026 — nine rooftop docks, ~$759k/yr for the first nine years inside a council-approved $6.83M Axon contract — after a 7-week pilot (185 calls) beat patrol to scene 33% of the time and gave useful info on 97%. On June 22, 2026 Sacramento County sheriff's deputies posted video of a consumer FPV drone (identified by trade press as a $999 DJI Avata 2) using a magnet to pull a knife from an unresponsive barricaded suspect's hand, claimed as a nationwide first. Meanwhile the FCC (Dec 22, 2025), acting on a White House-convened National Security Determination received Dec 21, added ALL foreign-produced drones and critical components to its Covered List — blocking new-model authorizations (existing models unaffected) and steering agencies toward US-made hardware that dealer listings price at roughly 2-3x comparable DJI (Skydio X10D ~$16k vs DJI Matrice 4T ~$7.8k). Signal, not verified: the price-threshold read leans on trade-press ID and dealer listings, not agency records.
so_we_should: Track the threshold, not the gadgets: when a $1k-$5k drone plus authority-to-fly replaces a $100k response, every physical-security and public-safety garden assumption about cost and reach needs re-pricing.

### [signal] Swiss Parliament votes to lift the new-nuclear ban (June 2026) — referendum pending, voters likely decide in 2027
url: https://rd.theknowledgegardens.com/signals/swiss-nuclear-ban-repeal-referendum-pending
trust_tier: verified | as_of: 2026-06-18 | gardens: EKG, GKG
source: https://www.fedlex.admin.ch/eli/fga/2025/2564/de (primary_doc)
capsule: On 18 June 2026 both chambers of the Swiss Parliament gave final approval to an indirect counter-proposal amending the Nuclear Energy Act (KEG) that repeals the post-Fukushima ban on framework authorizations for new nuclear plants — National Council 106-88, Council of States 28-14. Nothing is built or approved: the 'Blackout stoppen' initiative was conditionally withdrawn, a left-Green referendum drive launched 30 June 2026 (50,000 signatures needed by 8 October 2026), and a popular vote is expected around 28 February 2027 — so 'Switzerland has lifted the ban' is not yet true. Switzerland's four reactors (Beznau 1&2, Gösgen, Leibstadt, across three sites) supply roughly 29-30% of its electricity (up to ~40% in some framings), not a settled 40%.
so_we_should: Track the referendum result — the popular vote expected ~28 Feb 2027 — as the real decision point, not the parliamentary vote; do not card this as 'Switzerland lifted its nuclear ban' until voters confirm it. This is a freshness/superseded-risk discipline case: an in-progress reversal that a lazy feed would prematurely report as settled.

### [signal] 'Animal-free casein cleared by FDA' is overstated - Formo's GRN 1312 is still 'Pending'
url: https://rd.theknowledgegardens.com/signals/formo-casein-gras-pending
trust_tier: signal | as_of: 2026-06-16 | gardens: CulKG
source: https://www.hfpappexternal.fda.gov/scripts/fdcc/index.cfm?set=GRASNotices&id=1312 (filing)
capsule: Trade headlines say Formo won US GRAS clearance for precision-fermented casein, but FDA's GRAS Notice Inventory shows GRN 1312 - recombinant alpha-S1-casein via E. coli, notified by Formo Foods GmbH and NewMilkBuzz BV (Those Vegan Cowboys) - with FDA's response letter still listed as Pending as of 16 Jun 2026. Formo self-affirmed GRAS in Dec 2024 and can sell on that basis, but FDA has not responded. (Source: FDA GRN 1312.)
so_we_should: Hold Formo's casein at signal until FDA posts a no-questions letter on GRN 1312; separately flag the self-affirmed-GRAS pathway itself as at-risk (FDA proposed rule + NY bill), because much of the alt-protein 'cleared in the US' narrative rests on self-GRAS, not FDA review.

### [signal] Helion secures the world's first regulatory licenses (radioactive materials + air emissions) for a fusion power plant
url: https://rd.theknowledgegardens.com/signals/helion-first-fusion-plant-licenses
trust_tier: verified | as_of: 2026-06-16 | gardens: EKG
source: https://www.helionenergy.com/newsroom/helion-clears-key-regulatory-milestone-on-the-path-to-building-and-operating-the-worlds-first-fusion-power-plant (company)
capsule: On 2026-06-16 the Washington State Dept. of Health granted Helion a Radioactive Materials License (RML) and a Radioactive Air Emissions License (RAEL) for its Orion plant in Malaga, WA, which Helion says makes it the first company worldwide with the regulatory licenses needed for a fusion power plant (fusion is regulated as byproduct material under the 2024 ADVANCE Act, not as a fission reactor). These are materials/emissions/operating-safety licenses, not 'construction' permits — Orion broke ground in July 2025 under a separate Chelan County conditional-use permit — and the plant is designed to deliver 50 MW or greater to Microsoft after a roughly one-year ramp beginning in 2028. Separately, Helion closed a $465M Series G in early June 2026 at a $15.5B post-money valuation (up from $5.4B at its Jan 2025 Series F, total raised ~$1.5B); the regulatory path is advancing, but no machine anywhere has yet demonstrated net-positive fusion.
so_we_should: Treat WA-DOH-as-fusion-regulator as a replicable, low-friction siting path — the licensing precedent, not the physics, is what just got de-risked. Track Orion first-power as the falsifiable milestone.

### [signal] Lp(a)-lowering drugs cut the marker up to ~99% — but no outcome trial has yet read out
url: https://rd.theknowledgegardens.com/signals/olpasiran-lpa-lowering
trust_tier: verified | as_of: 2026-06-04 | gardens: HKG, MKG
source: https://doi.org/10.1111/dom.70944 (paper)
capsule: A June 2026 network meta-analysis (51 RCTs, 17,810 patients; Zayed et al., Diabetes Obes Metab; DOI 10.1111/dom.70944, PMID 42244216) found RNA-targeted agents lower lipoprotein(a) by up to ~99% (olpasiran 225mg Q12W), far beyond PCSK9 drugs. Yet no Lp(a) drug has proven it prevents cardiovascular events: the first outcomes trial, pelacarsen HORIZON (NCT04023552), only reached primary completion 30 Jun 2026; olpasiran's OCEAN(a) (NCT05581303) runs to 2028. Potent surrogate, unproven outcome.
so_we_should: Card Lp(a) lowering as VERIFIED at the biomarker level and explicitly UNPROVEN at the outcome level — the exact verified-vs-hyped seam HKG/MKG polices. Watch HORIZON's 2026 readout; it will convert (or break) the whole causal-biomarker thesis.

### [signal] SK Group chairman: HBM/memory is the AI buildout's binding constraint, with the shortage forecast to run to ~2030
url: https://rd.theknowledgegardens.com/signals/sk-hbm-memory-shortage-2030
trust_tier: signal | as_of: 2026-06-02 | gardens: RDKG, EKG
source: https://news.skhynix.com/computex-2026-review/ (company)
capsule: At Computex on 2 June 2026, SK Group chairman Chey Tae-won (not a separately-quoted SK Hynix CEO) said the AI-driven memory/HBM shortage will likely persist through about 2030, with industry wafer supply running more than 20% below demand and a greenfield fab taking five-plus years to build; SK Hynix aims to roughly double wafer capacity over five years and has effectively sold out 2026 HBM. This is a forward-looking forecast from an interested supplier, not a settled fact, and it is the memory maker who benefits from a shortage narrative. The widely-cited figure that AI data centers will absorb ~70% of (high-end) memory in 2026 is a separate analyst/industry estimate, not part of Chey's remarks.
so_we_should: Card it as the load-bearing constraint on the Moat Clock: while reasoning trends toward free, the atoms it runs on stay scarce for years.

### [signal] MCP's largest revision reaches locked release candidate (spec 2026-07-28), governed under the Linux Foundation
url: https://rd.theknowledgegardens.com/signals/mcp-standardization
trust_tier: verified | as_of: 2026-05-21 | gardens: RDKG
source: https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/ (standard)
capsule: On 21 May 2026 the Model Context Protocol maintainers published the locked release candidate for spec version 2026-07-28 — MCP's biggest revision since launch: a stateless HTTP core, a first-class Extensions framework (MCP Apps, Tasks), OAuth-aligned authorization hardening, and a formal deprecation policy. Governed under LF Projects, LLC (Linux Foundation); the final spec ships 28 July 2026. MCP is consolidating into the default interface wiring AI agents to tools and data.
so_we_should: Build RDKG's verified-ground-truth Frontier Feed as an MCP server targeting spec 2026-07-28 (stateless core, Tasks, MCP Apps), so machine customers discover and consume our data through the standard interface rather than a bespoke API.

### [signal] The data-center 'water crisis' is not data-driven
url: https://rd.theknowledgegardens.com/signals/datacenter-water-crisis-not-data-driven
trust_tier: hype | as_of: 2026-05-20 | gardens: TKG, EKG, BKG
source: https://reason.com/2026/05/20/data-centers-use-less-water-than-almond-farms-and-do-more-good/ (news)
confidence: 0.85
capsule: Kept as a discipline instrument: the loud 'AI is guzzling our drinking water' narrative collapses on the numbers. US golf courses use on the order of 20–30× the water of all US data centers combined, and California almond farming dwarfs both (per-day estimates: almonds ~4.2B gal, golf ~1.4B, data centers ~46M — a ~30× golf-to-data-center ratio). Even tripled by 2030, data centers are ~10% of golf's draw. The episode's exact figures vary by accounting — cite the verified ratio, and note the fear migrates to the next semi-plausible target regardless of the facts.
so_we_should: Arm the money gardens with a sourced 'verified-vs-FUD' exposure explainer; use it to defuse permitting opposition — while still flagging any genuinely elevated exposure (real PFAS fish advisories) honestly.

### [signal] EPA moves to roll back the PFAS water mandate: proposed rescission of four of six limits (May 2026)
url: https://rd.theknowledgegardens.com/signals/pfas-partial-rescission-2026
trust_tier: verified | as_of: 2026-05-20 | gardens: TKG, NMK, HKG
source: https://www.federalregister.gov/documents/2026/05/20/2026-10085/rescission-of-regulatory-determinations-and-removal-of-related-provisions-for-four-pfas-substances (filing)
capsule: On 20 May 2026 EPA published a proposed rule (Federal Register doc 2026-10085; 91 FR 29413; RIN 2040-AG53; docket EPA-HQ-OW-2025-0654) to rescind the 2024 regulatory determinations and drinking-water limits for four PFAS - PFHxS, PFNA, HFPO-DA (GenX) and their Hazard-Index mixture with PFBS - amending 40 CFR 141 and 142. A companion rule keeps the PFOA/PFOS limits but extends compliance to 2031. Comments close 20 July 2026; the first federal attempt to weaken the non-toxic-water mandate.
so_we_should: Re-price the non-toxic thesis's regulatory tailwind as partially reversing rather than settled; track docket EPA-HQ-OW-2025-0654 to final action, and treat PFHxS/PFNA/GenX removal capacity as policy-contingent (not mandated) while PFOA/PFOS stays enforceable.

### [signal] ProtPhenoAge: a plasma-proteomics aging clock predicts mortality in 37,433 UK Biobank adults
url: https://rd.theknowledgegardens.com/signals/protphenoage-proteomic-aging-clock
trust_tier: verified | as_of: 2026-05-11 | gardens: HKG, MKG
source: https://doi.org/10.1016/j.jare.2026.05.022 (paper)
capsule: ProtPhenoAge, an aging clock built from 185 plasma proteins in 37,433 UK Biobank participants (Wang et al., J Adv Res, 11 May 2026; DOI 10.1016/j.jare.2026.05.022, PMID 42119967), predicted all-cause mortality (AUC 0.76) and outperformed chronological age, PhenoAge and ProtAge; its acceleration linked to 313 disease phenotypes, with Mendelian-randomization and GWAS support (APOE rs429358). A net-new proteomic biomarker — strong for research, but internally validated only, not a clinical endpoint.
so_we_should: Log ProtPhenoAge as a verified, in-window addition to the aging-biomarker cabinet alongside DunedinPACE — usable as an exploratory trial readout — while refusing to sell any single "biological age" score as a clinical endpoint until external, prospective validation lands. The validated-vs-hyped biomarker line is the product.

### [signal] Q1 2026 was a record robotics/physical-AI VC quarter ($16.3B across 492 deals) per PitchBook; relayed by a16z
url: https://rd.theknowledgegardens.com/signals/robotics-physical-ai-vc-record-q1-2026
trust_tier: signal | as_of: 2026-05 | gardens: RKG, FKG
source: https://pitchbook.com/news/reports/q1-2026-robotics-physical-ai-vc-trends (report)
capsule: PitchBook's Q1 2026 Robotics & Physical AI VC Trends report (published May 2026) found the sector had its strongest quarter on record, with $16.3B raised across 492 deals, lifted by megadeals for Shield AI, Saronic and Neura Robotics. a16z's June 26 newsletter accurately relayed those figures (rounding to ~$16B across just under 500 deals), but the claim that robotics has 'eclipsed fintech and payments' to become the #2 category is a16z's OWN chart, not PitchBook's, and applies only to a narrow scope: the value of the top 100 private companies by category. Crunchbase independently corroborates a record robotics-funding year in 2026 (~$18.8B YTD through June, past the prior 2021 peak). Held at signal, not verified: the primary PitchBook report is paywalled, and the eye-catching ranking flourish is VC commentary, not report data.
so_we_should: Track robotics capital as a leading indicator of the rotation from bits to atoms — but pull the underlying PitchBook report before treating $16.3B/492 as settled fact, and don't inherit a16z's 'eclipsed fintech' framing.

### [signal] Commonwealth Fusion becomes the first fusion company to file for grid interconnection (PJM, 28 Apr 2026)
url: https://rd.theknowledgegardens.com/signals/cfs-pjm-interconnection
trust_tier: verified | as_of: 2026-04-28 | gardens: EKG, RDKG
source: https://cfs.energy/news-and-media/commonwealth-fusion-systems-becomes-first-fusion-company-to-apply-to-pjm-interconnection-the-largest-u.s.-wholesale-electricity-market (primary_doc)
capsule: On 28 April 2026 Commonwealth Fusion Systems announced it is the first fusion developer to apply to a major grid operator, filing a PJM Interconnection request for its planned 400 MW ARC plant — the Fall Line Fusion Power Station in Chesterfield County, Virginia. Dominion Energy advised. The move begins a 4–6 year study process targeting grid electricity in the early 2030s, sited squarely in Virginia's AI data-center corridor.
so_we_should: Track fusion's grid-integration milestones (interconnection queue position, study-phase progress) as leading indicators for EKG's 'firm clean power for AI' thesis; log the PJM queue ID as a citable primary once public.

### [signal] Mercedes-Benz partners with Liquid AI to run non-transformer LFMs on-device in North American MBUX cars
url: https://rd.theknowledgegardens.com/signals/mercedes-liquid-ai-embedded-lfm
trust_tier: verified | as_of: 2026-04-23 | gardens: RDKG, GKG, HKG
source: https://www.liquid.ai/press/liquid-ai-and-mercedes-benz-partner-to-scale-embedded-in-car-intelligence (company)
capsule: Mercedes-Benz and Liquid AI announced a multi-year partnership (2026-04-23) to run embedded Liquid Foundation Models on-device — fast, private, no cloud dependence — in Mercedes models with third- and fourth-generation MBUX in North America, built on the MB.OS architecture. The release frames H2 2026 as a target: it "creates a path to a first production deployment for advanced speech technology," i.e. an initial speech deployment, not a confirmed full rollout. Liquid's LFMs are genuinely non-transformer (liquid neural networks / state-space models), making this one of the few production efforts off the attention architecture.
so_we_should: Card the shift: private, on-device, fine-tuned foundation models are shipping to consumer hardware. Flag — don't repeat — Hasani's unverified specifics (the sub-1GB / $60-chip numbers aren't in the official release).

### [signal] 'Dr. AI' and the insurance recursion — Tesla's Safety Score is the live precedent
url: https://rd.theknowledgegardens.com/signals/dr-ai-insurance-recursion-tesla-safety-score
trust_tier: signal | as_of: 2026-04-14 | gardens: HKG, IKG, FKG
source: https://www.tesla.com/support/insurance/safety-score (company)
confidence: 0.55
capsule: The episode's thesis — that deviating from AI advice (medical, driving) will raise your premium — has a real, dated precedent: Tesla's Safety Score 3.0 (rolled out from 2026-04-14) scores every FSD-Supervised mile at 100 and lowers the premium accordingly. Behavior-priced, AI-mediated risk is shipping. The open governance question is whether such scores stay contestable and transparent, or become opaque gates on care and credit.
so_we_should: Track behavior-priced AI risk scoring as a live category; design the auditable, contestable version (the 'is this score fair?' product) rather than the black box.

### [signal] Mycelium insulation reaches thermal-performance parity with fossil foams in a peer-reviewed dataset (cost parity still unproven)
url: https://rd.theknowledgegardens.com/signals/mycelium-insulation
trust_tier: verified | as_of: 2026-04-14 | gardens: NMK, BKG, MKG
source: https://www.nature.com/articles/s41598-026-48045-w (paper)
capsule: A peer-reviewed Scientific Reports study (published 14 Apr 2026; DOI 10.1038/s41598-026-48045-w) measured thermal conductivity by ASTM C518 for Lentinus tigrinus mycelium composites on five substrates: straw-based boards reached lambda = 0.031 W/m.K, matching extruded polystyrene (0.028-0.031) and beating mineral wool (0.035-0.045), with cradle-to-gate global-warming potential from -9.77 to +4.04 kg CO2e per functional unit. Thermal parity is now primary-sourced; cost parity is not.
so_we_should: Graduate the performance half of the non-toxic-materials-cross-cost-parity thesis to verified, keep cost parity flagged as the live falsifier, and pre-build the BKG spec-library entry for mycelium insulation anchored on ASTM C518 lambda data.

### [signal] Frontier models are now the least transparent — labs stopped disclosing training compute, data, and parameters (Stanford AI Index 2026)
url: https://rd.theknowledgegardens.com/signals/ai-index-2026-training-opacity
trust_tier: verified | as_of: 2026-04-13 | gardens: RDKG, TKG, BKG, HKG
source: https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_1_research_development.pdf (primary_doc)
capsule: Stanford HAI's AI Index 2026 finds industry produced over 90% of notable AI models in 2025, yet 'the most capable models are now the least transparent': training code, parameter counts, dataset sizes, and training duration are no longer disclosed for the most resource-intensive systems, including those from OpenAI, Anthropic, and Google. Public knowledge of how frontier models are built is shrinking as the models get more capable — and AI data-center power capacity reached ~29.6 GW by Q4 2025.
so_we_should: Lean the moat harder into ground truth we measure ourselves. As public disclosure of how frontier models are trained disappears, the defensible input is not model access but independently-verified, freshly-dated, falsifiable data — the thing a closed lab cannot retroactively publish. Make 'measured by us, sourced, dated' the RDKG differentiator.

### [signal] EIA's Annual Energy Outlook 2026: data-center load is now the dominant driver of long-run US electricity growth
url: https://rd.theknowledgegardens.com/signals/eia-aeo2026-datacenter-demand
trust_tier: verified | as_of: 2026-04-08 | gardens: EKG, RDKG
source: https://www.eia.gov/pressroom/releases/press587.php (primary_doc)
capsule: In its Annual Energy Outlook 2026 (released 8 April 2026), the U.S. Energy Information Administration states that data-center load is emerging as the dominant driver of long-term U.S. electricity growth. National demand rose 2.1% per year over the past five years after a decade-plus plateau, and is projected to grow 0.9–1.6% per year through 2050 — requiring installed generating capacity to expand 50–90% by 2050. A federal-statistical primary quantifying AI's energy pull.
so_we_should: Anchor RDKG/EKG's 'energy is the machine layer's binding constraint' thesis on this government primary (not vendor hype), and pair it with the fusion/nuclear supply signals (CFS PJM, General Fusion) to map where firm clean power actually meets AI demand.

### [signal] x402 agent-payment protocol moves to the Linux Foundation as a vendor-neutral standard (x402 Foundation, 2 Apr 2026)
url: https://rd.theknowledgegardens.com/signals/x402-linux-foundation
trust_tier: verified | as_of: 2026-04-02 | gardens: RDKG, BKG
source: https://www.linuxfoundation.org/press/linux-foundation-is-launching-the-x402-foundation-and-welcoming-the-contribution-of-the-x402-protocol (standard)
capsule: On 2 April 2026 the Linux Foundation launched the x402 Foundation and took stewardship of the x402 protocol contributed by Coinbase — an open standard that puts HTTP's dormant 402 status code to work so agents, APIs and apps can settle payments inline. It launched with ~22 named backers including Google, Microsoft, Visa, Mastercard, AWS, Stripe, Cloudflare and Circle. Agent-to-agent commerce now has a vendor-neutral settlement rail.
so_we_should: Price RDKG's Frontier Feed for x402 settlement alongside AP2/Mandates — stand up an x402-gated endpoint so agents can pay-per-query for verified ground truth with no human checkout, and stay rail-agnostic between the Coinbase-origin (x402) and Google-origin (AP2) stacks.

### [signal] FDA approves orforglipron (Foundayo): first oral small-molecule GLP-1 for weight loss
url: https://rd.theknowledgegardens.com/signals/orforglipron-oral-glp1-fda-approval
trust_tier: verified | as_of: 2026-04-01 | gardens: HKG, MKG
source: https://investor.lilly.com/news-releases/news-release-details/fda-approves-lillys-foundayotm-orforglipron-only-glp-1-pill (filing)
capsule: On 1 Apr 2026 the FDA approved Lilly's Foundayo (orforglipron), the first oral small-molecule (non-peptide) GLP-1 receptor agonist for weight management (Lilly release; Drugs.com FDA history). In ATTAIN-1, highest-dose completers lost 12.4% of body weight vs 0.9% on placebo; ACHIEVE Phase 3 diabetes trials were published in The Lancet and JAMA (Jun 2026). It carries a boxed thyroid C-cell-tumor warning. A manufacturable pill, not an injectable — the access story.
so_we_should: Mark as a verified regulatory milestone and the real inflection for GLP-1 access — an oral small molecule scales past injectable/cold-chain limits — while framing honestly to HKG: approved for weight loss, NOT yet for diabetes, and long-term outcome data are still accruing (ACHIEVE-4).

### [signal] FDA issues 'no questions' on All G's precision-fermented bovine lactoferrin (GRN 1284)
url: https://rd.theknowledgegardens.com/signals/all-g-lactoferrin-fda-gras
trust_tier: verified | as_of: 2026-03-25 | gardens: CulKG, HKG
source: https://www.hfpappexternal.fda.gov/scripts/fdcc/index.cfm?set=GRASNotices&id=1284 (primary_doc)
capsule: FDA's GRAS Notice Inventory shows GRN 1284 - recombinant bovine lactoferrin made by the yeast Komagataella phaffii, notified by All G Co Holdings Pty (Australia) - closed 25 Mar 2026 with an FDA no-questions letter (fda.gov/media/191907). All G announced it about 1 Apr 2026. It is not the first precision-fermented lactoferrin (TurtleTree's GRN 1219 closed May 2025) but confirms a fast-growing animal-free route to a scarce dairy bioactive. (Source: FDA GRN 1284.)
so_we_should: Map precision-fermented bioactives (lactoferrin, then casein/whey) as a distinct, FDA-notified track separate from cultivated meat: index each by GRN number and closure date, and treat lactoferrin - high value, supply-constrained, health-adjacent - as the near-term commercial wedge.

### [signal] Sub-1-bit quantization crosses from theory to method — Samsung NanoQuant
url: https://rd.theknowledgegardens.com/signals/samsung-nanoquant-sub-1-bit-ptq
trust_tier: verified | as_of: 2026-02-01 | gardens: RDKG, EdKG, HKG
source: https://arxiv.org/abs/2602.06694 (paper)
confidence: 0.8
capsule: Samsung's NanoQuant (arXiv 2602.06694, ICML 2026) is the first post-training quantization to reach both binary and sub-1-bit compression with calibration only (128 samples, 1 GPU, no retraining): a 70B model drops from 138 GB to 5.35 GB and runs on a consumer 8 GB GPU at ~20 tok/s. Below-one-effective-bit-per-weight is now a real technique, not a whiteboard idea — the direction Alex Wissner-Gross's 'sub-1-bit goes mainstream' extrapolation points at.
so_we_should: Assume the cost/size floor of on-device intelligence keeps falling ~10× on a 12-18 month horizon; design edge products against next year's floor, not today's.

### [signal] Open-weight models let hospitals keep the data and self-host the intelligence
url: https://rd.theknowledgegardens.com/signals/hospitals-self-host-open-clinical-models
trust_tier: verified | as_of: 2026-01-13 | gardens: HKG, MKG, RDKG
source: https://research.google/blog/medgemma-our-most-capable-open-models-for-health-ai-development/ (company)
confidence: 0.8
capsule: Self-hostable medical models are here: Google's MedGemma 1.5 (a small, open, fine-tunable clinical model, 2026-01-13) plus enterprise clinical assistants like OpenEvidence (~$12B, ~20M consultations/month by early 2026). Combined with K3/Inkling as fine-tune bases, a hospital or biotech can tune a frontier-class model on proprietary clinical data inside its own walls — PHI never leaves the building. The defensible asset is the private data loop, not the base model.
so_we_should: Spec an on-prem clinical fine-tune reference (privacy-preserving, audit-logged) for HKG; sell the verification/ops layer, not the model.

### [signal] Fountain Life's '3.4% occult cancer' is a vendor number, not an independent finding
url: https://rd.theknowledgegardens.com/signals/fountain-life-occult-cancer-vendor-number
trust_tier: hype | as_of: 2026-01-06 | gardens: HKG, MKG, TKG
source: https://www.prnewswire.com/news-releases/fountain-life-proves-the-future-of-longevity-is-data-not-guesswork-302653728.html (company)
confidence: 0.75
capsule: The episode's striking 'we find cancer in 3.3% of healthy members' is Fountain Life's own self-reported figure ('up to 3.4%,' PR 2026-01-06) — no disclosed denominator, no peer review, and a clear commercial interest. Held against the rigorous comparator already carded (the NHS-Galleri MCED RCT, n≈142,000, which missed its primary endpoint), and GRAIL's Galleri getting Medicare coverage (2026-02-03) before FDA approval, this is the validated-vs-hyped line the measurement layer exists to police. Screening finds indolent disease too; a found cancer is not always a life saved.
so_we_should: Treat vendor screening percentages as marketing until a cohort methodology is published; build MKG's independent 'validated vs hyped' audit as the product.

### [signal] Humanoid robot combat goes mainstream — and the street-safety rules don't exist yet
url: https://rd.theknowledgegardens.com/signals/unitree-humanoid-combat-safety-gap
trust_tier: verified | as_of: 2026-01-06 | gardens: GKG, AgKG, IKG
source: https://www.scmp.com/tech/tech-trends/article/3311831/ready-set-fight-unitrees-humanoid-robots-test-skills-unique-kickboxing-competition (news)
confidence: 0.8
capsule: Unitree staged humanoid (G1) boxing/kickboxing demos at CES 2026, a televised Chinese combat tournament, and SF crowd events — robots take hits, lose heads, keep fighting. The G1 sells under $15,000 and full-size units are heavy, fast, and strong enough to injure, yet no street-safety regime governs them as they head into homes and security roles. The spectacle is a stress-test of balance, impact recovery, and latency; the gap is safety, liability, and ops — not the chassis.
so_we_should: Sell the verified safety/skills/ops/insurance layer on a commodity humanoid chassis; push for a testable safety standard before household deployment scales.

### [signal] '#1' means one board — the benchmark-gaming trap
url: https://rd.theknowledgegardens.com/signals/benchmark-number-one-means-one-board
trust_tier: hype | as_of: 2026-01-02 | gardens: MKG, RDKG, GKG
source: https://www.fastcompany.com/91266011/meta-llama-4-benchmark-controversy (company)
confidence: 0.8
capsule: The episode's 'K3 jumped the fence to #1' is true only on the Frontend Code Arena; K3 is #3 overall. The cautionary precedent: Meta submitted a chat-tuned experimental Llama 4 Maverick that scored ~#2 on LMArena while the public build ranked far lower — LMArena banned the practice (2025-04) and Yann LeCun later conceded the results were 'fudged' (FT, reported 2026-01-02). Leaderboard rank is a marketing surface; name the board, date the claim, or be gamed.
so_we_should: Institutionalise 'name the board + freshness date' on every capability claim in the cabinet; stand up MKG's audited benchmark/claims registry.

### [signal] AI reaches parity with human superforecasters on ForecastBench
url: https://rd.theknowledgegardens.com/signals/forecastbench-superforecaster-parity
trust_tier: verified | as_of: 2025-11-01 | gardens: FKG, MKG, GKG, HKG
source: https://forecastingresearch.substack.com/p/ai-models-have-likely-reached-parity (paper)
confidence: 0.8
capsule: On ForecastBench, several AI systems are now statistically indistinguishable from human superforecasters — the top being 'Cassi,' with xAI and Google DeepMind models also at parity. Two honest caveats keep this at verified-but-bounded: the superforecaster comparison was last elicited in 2024, and overlapping 95% CIs mean the data support parity, not clear outperformance. A cheap, tireless, calibrated adviser for insurance, policy, ops, and strategy is arriving.
so_we_should: Pilot calibrated, Brier-scored forecasting as a decision aid (triage, staffing, budgeting) — and score it against resolution, don't trust it blind.

### [signal] 'Hyper-forecasting crowns the EMH and management evaporates' — real signal, oversold leap
url: https://rd.theknowledgegardens.com/signals/hyper-forecasting-emh-oversell
trust_tier: hype | as_of: 2025-11-01 | gardens: FKG, GKG, MKG
source: https://forecastingresearch.substack.com/p/ai-models-have-likely-reached-parity (news)
confidence: 0.7
capsule: ForecastBench parity (card n:53) is real; the episode's leap — that AI 'hyper-forecasters' will crown the efficient-market hypothesis and evaporate senior management — is a loud thesis its own evidence undercuts. A cheap, public forecast should be arbitraged away, not rule a market; the parity rests on a 2024 baseline and overlapping confidence intervals; and management is judgment plus accountability, not only prediction. Kept on purpose as the contrarian counterweight to card n:53.
so_we_should: Sell calibrated decision-forecasting as a scored aid to human accountability, not a replacement for it; watch for reflexivity (a forecast that moves the market it predicts).

### [signal] On-device models are the safety inversion for children's AI
url: https://rd.theknowledgegardens.com/signals/on-device-models-child-safety-inversion
trust_tier: signal | as_of: 2025-10-13 | gardens: ParKG, EdKG, HKG
source: https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB243 (gov)
confidence: 0.6
capsule: The child-safety backlash against cloud AI companions is now legal fact: California SB 243 (signed 2025-10-13, effective 2026-01-01) mandates disclosure and self-harm protocols; the FTC opened a 6(b) inquiry into seven chatbot makers (2025-09-11); Character.AI cut under-18s off open-ended chat (2025-11-24) after Garcia v. Character Technologies (2025-05-21) found a duty of care to minors. Bonsai-class on-device models invert the risk model — a private, parent-controlled companion whose data never leaves the device.
so_we_should: Spec a child-safe, on-device (ternary) companion + a vendor safety-compliance checklist for ParKG; make 'nothing leaves the device' the product claim.

### [signal] Unitree R1 humanoid launched at US$5,900; R1 AIR tier now sells from US$4,900
url: https://rd.theknowledgegardens.com/signals/unitree-r1-humanoid-sub-5k-floor
trust_tier: verified | as_of: 2025-07-25 | gardens: RKG
source: https://www.unitree.com/R1 (company)
capsule: Unitree unveiled the R1 humanoid on 25 Jul 2025 at 39,999 yuan (~US$5,900); a cheaper R1 AIR tier at US$4,900 was added around March 2026, and both prices are live on Unitree's own product page (verified 2026-07-18, tax and shipping excluded). The R1 is a compact humanoid — about 1.23 m tall, 25-29 kg, 20 joints on the AIR vs 26 on the standard/EDU — not full-size, with an on-board 'Large Multimodal Model for voice and images' and fully open joint/sensor control interfaces. The sub-$5k floor does shift value toward the software/skills layer (our inference, not Unitree's claim), though the developer-oriented R1 EDU is quote-only (roughly US$14k+ per press).
so_we_should: Treat sub-$5k compact humanoids as a real 2026 price point; sell the software/skills/safety layer on a commodity chassis, not the chassis.

### [signal] Fusion sector baseline: 53 firms, $9.766B cumulative (FIA Global Fusion Industry Report 2025)
url: https://rd.theknowledgegardens.com/signals/fia-fusion-industry-2025-baseline
trust_tier: verified | as_of: 2025-07-22 | gardens: EKG, FKG
source: https://www.fusionindustryassociation.org/over-2-5-billion-invested-in-fusion-industry-in-past-year/ (report)
capsule: The Fusion Industry Association's fifth annual Global Fusion Industry Report (released 22 July 2025) surveyed 53 fusion companies whose cumulative funding stands at $9.766 billion (a five-fold rise since 2021), with $2.64 billion raised in the 12 months to July 2025 — the highest annual total since 2022. Figures are self-reported to the FIA survey (52 of 53 companies responded; one confirmed via prior-year data), so they represent surveyed private-sector funding rather than an audited sector total. This is a distinct sector-baseline datapoint, not a duplicate of the company-specific General Fusion or CFS cards.
so_we_should: Treat $9.77B cumulative across 53 companies as the fusion-sector funding baseline — with the self-reported-survey caveat attached — and track the annual raise ($2.64B this cycle) as the momentum gauge for the 'firm clean power for AI' thesis. Watch for the 2026 FIA edition to re-baseline.

### [signal] 'Plasma exchange reverses aging' is a commercial claim the human data doesn't support
url: https://rd.theknowledgegardens.com/signals/plasma-exchange-age-reversal-hype
trust_tier: hype | as_of: 2025-07-01 | gardens: HKG, MKG
source: https://doi.org/10.1038/s41598-025-05396-0 (paper)
capsule: Longevity clinics and influencers (Bryan Johnson's Blueprint among them) market plasma exchange/apheresis as "age reversal." But a 2025 randomized human trial of plasmapheresis (Borsky et al., Sci Rep, 1 Jul 2025; DOI 10.1038/s41598-025-05396-0, PMID 40592961) — co-authored by epigenetic-clock and longevity-industry figures — found NO epigenetic rejuvenation and signs of accelerated clocks (GrimAge, Hannum, DunedinPACE). The lab object and the sales claim are different things.
so_we_should: Tier the commercial "plasma exchange = age reversal" claim as HYPE and keep the underlying lab science as a separate, weak surrogate signal. Use this as a canonical house-rule example: peer-reviewed data and founder-branded commercial claims are DIFFERENT OBJECTS — even when clock scientists sit on both sides.

### [signal] BIS rescinded the AI Diffusion Rule (May 13, 2025) — model weights were briefly codified as a controlled export (ECCN 4E091), but the control was never enforced
url: https://rd.theknowledgegardens.com/signals/bis-ai-diffusion-rule-rescinded-4e091
trust_tier: verified | as_of: 2025-05-13 | gardens: RDKG, GKG, FKG
source: https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens (primary_doc)
capsule: On May 13, 2025, Commerce/BIS announced the rescission of the Biden Administration's AI Diffusion Rule (issued January 15, 2025) and instructed enforcement officials not to enforce it, with a formal Federal Register rescission to follow. That rule had created ECCN 4E091, a new export control on the model weights of the most advanced closed-weight (>1e26 training-FLOP) dual-use AI models — the first codification of model weights as a controlled U.S. export (published/open-weight models were exempt). Scope caveat: the control was never actually enforced or 'tested' — exporter compliance obligations were set to begin May 15, 2025, and the rule was rescinded two days earlier, on May 13.
so_we_should: Anchor all 'AI as munition' talk to this real primary, not the folk memory of it. The United States did once codify model weights as a controlled export (ECCN 4E091) — but it was rescinded two days before compliance began and never enforced, so the honest claim is 'codified, never tested,' not 'tried and abandoned.' Watch bis.gov for the promised replacement rule; a reinstatement would flip the falsifier.

### [signal] A US move to gate corporate use of Chinese open weights is plausible but unproven
url: https://rd.theknowledgegardens.com/signals/us-gate-chinese-open-weights-plausible
trust_tier: signal | as_of: 2025-05-13 | gardens: GKG, FKG, RDKG
source: https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens (gov)
confidence: 0.45
capsule: The episode floats a FINRA-style frontier regulator and disclosure/scrutiny of corporate use of Chinese open weights ('know-your-prompter,' 'gold eagle' are podcast color, not confirmed rules). The real, tested precedent is the BIS AI Diffusion Rule — which briefly put certain model weights under export control and was then rescinded (2025-05-13). Weight-level governance is a real, reversible lever; any new gate would collide with the fact that open weights mirror across jurisdictions.
so_we_should: War-game a compliance-disclosure regime as a scenario, not a forecast; keep the verified BIS primary as the anchor and watch the docket for a replacement rule.

### [signal] Ternary weights make the multiply almost free — and point compute at new substrates
url: https://rd.theknowledgegardens.com/signals/ternary-weights-free-multiply-bitnet
trust_tier: verified | as_of: 2025-04-16 | gardens: RDKG, AgKG, EdKG
source: https://arxiv.org/abs/2504.12285 (paper)
confidence: 0.8
capsule: BitNet b1.58 constrains weights to {-1,0,+1} (~1.58 bits), turning the core matrix-multiply into add/subtract/skip and cutting memory >90%; Microsoft's BitNet b1.58 2B4T (2025) was the first open native 1-bit LLM (~0.4 GB, ~6× less energy), and MediaTek's Dimensity 9500 now ships native ternary compute-in-memory. Because the operation reduces to ×{-1,0,+1}, it widens where computation can physically happen — the on-ramp to photonic and even etched-weight silicon (a signal-tier extrapolation, not yet product).
so_we_should: Track ternary-native silicon (MediaTek, and any photonic entrant) as the substrate that makes edge intelligence ubiquitous in vehicles, robots, and devices; design for on-device inference by default.

### [signal] The Muon optimizer and the nanoGPT speedrun are the real '1% cost' existence proof
url: https://rd.theknowledgegardens.com/signals/muon-optimizer-nanogpt-speedrun
trust_tier: verified | as_of: 2024-10-15 | gardens: RDKG
source: https://kellerjordan.github.io/posts/muon/ (news)
confidence: 0.85
capsule: Keller Jordan's Muon optimizer set a nanoGPT training-speed record on 2024-10-15 (+35% over AdamW) and has held through 12+ community records by 7 researchers; Moonshot adopted it for Kimi K2/K2-Thinking, and GLM uses it too. The compounding community speedups (plus data-quality stripping and MoE/attention tricks) are the verifiable mechanism behind the episode's '1% of the cost' claim — the efficiency overhang is real, not magic.
so_we_should: Track the open efficiency stack (Muon-class optimizers, data curation, quantization) as the true driver of falling training cost; assume no proprietary architecture moat behind frontier labs.

### [signal] The AI talent asymmetry is an own-goal — Yang Zhilin was trained in the US, built Kimi in China
url: https://rd.theknowledgegardens.com/signals/ai-talent-own-goal-yang-zhilin
trust_tier: verified | as_of: 2024-03-06 | gardens: GKG, EdKG
source: https://macropolo.org/interactive/digital-projects/the-global-ai-talent-tracker/ (news)
confidence: 0.8
capsule: Moonshot's founder Yang Zhilin (杨植麟) — first author of Transformer-XL and XLNet — did his PhD at Carnegie Mellon (2019) and returned to China, where he built a ~$20B frontier lab (correcting the episode's 'Yang Xilin' and its garbled chronology). MacroPolo's tracker is the settling primary on the asymmetry: no clean '70% non-citizen' figure exists; verified are that ~42% of top-tier researchers work outside their home country and ~75% of US-based top-tier talent is of American or Chinese origin. Green-card-stapling is the low-friction retention lever.
so_we_should: Frame high-skill immigration as measurable AI-competitiveness ROI; anchor any advocacy to MacroPolo's real figures, not the inflated '70%'.

### [signal] Open Athena — engineers and compute for open academic foundation models
url: https://rd.theknowledgegardens.com/signals/open-athena-academic-foundation-models
trust_tier: verified | as_of: 2024-01-01 | gardens: RDKG
source: https://openathena.ai/about/ (company)
capsule: Open Athena, a 501(c)(3) nonprofit founded in 2024 by Jeff Hammerbacher (Founder/CEO) and David Siegel (Founder/Chairman, who separately co-founded Two Sigma in 2001), gives academic labs top-tier AI and infrastructure engineers, compute (via relationships with neoclouds and hyperscalers), and coordination to build open "scientific foundation models" — confirmed by its own site and live collaborations such as Marin (Stanford), BoltzGen (MIT), and PlantCAD (Cornell). The lead's claim that its compute targets non-transformer architectures is unsupported: neither Open Athena's site nor its founders' materials commit compute to non-transformer designs. Two Sigma does not back the nonprofit; it is Siegel's earlier, separate venture.
so_we_should: Keep it on the board as a real open-corpus ally; separate the verified core from the unverified 'non-transformer' spin.

## Theses

### [thesis] Non-toxic, grown materials cross cost-parity this decade
url: https://rd.theknowledgegardens.com/theses/non-toxic-materials-cross-cost-parity
trust_tier: signal | as_of: 2026-06-18 | gardens: BKG, OKG, HKG
confidence: 0.5
falsifier: No non-toxic structural or insulating material reaches verified price parity with its conventional incumbent by 2030 in a primary cost dataset.
capsule: Biofabricated and non-toxic building materials — mycelium composites, low-carbon binders, bio-based panels — are on a cost curve that plausibly reaches parity with incumbent toxic materials before 2030. If it holds, the non-toxic choice stops being a premium and becomes the default, opening a large, spine-aligned market across the building and health gardens. Provisional; the cost data is not yet primary-sourced.
so_we_should: Position BKG to make the non-toxic spec the easy default the moment parity lands; pre-build the spec library and the citation trail now.

### [thesis] In longevity, verification is the product
url: https://rd.theknowledgegardens.com/theses/longevity-verification-is-the-product
trust_tier: signal | as_of: 2026-07-05 | gardens: HKG, RDKG
source: https://www.statnews.com/2024/03/05/david-sinclair-harvard-longevity-scientist-reversing-aging-dogs/ (expert)
confidence: 0.7
falsifier: Patients, clinicians, and investors do not pay for independent verification over charismatic founder claims — i.e., the market rewards confident marketing regardless of conflict-of-interest disclosure or primary-source backing, and a verified claim-card proves no more decision-useful than existing hype-driven content.
capsule: Longevity science splits cleanly into real, independently published lab work and founder-branded commercial claims that the field's own scientists have formally repudiated — often from the same person. The scarce, valuable thing is therefore not another molecule or clock; it is independent, conflict-free verification that separates the two. That is RDKG's core capability, applied to the field with the worst signal-to-noise ratio and the clearest meta-crisis lever in health.
so_we_should: Produce one verified claim-card — "does NMN extend healthspan in humans?" — in the standard sourced/tiered/conflict-disclosed format, and put it in front of John Bou with a single question: is this more decision-useful than what currently exists? His answer tells us whether the verification layer is the spine of HKG's longevity arm.

### [thesis] Verified physical-world ground truth becomes AI's scarce input
url: https://rd.theknowledgegardens.com/theses/verified-ground-truth-is-the-scarce-input
trust_tier: signal | as_of: 2026-06-20 | gardens: TKG, BKG, HKG, OKG, FshKG
source: https://arxiv.org/abs/2211.04325 (paper)
confidence: 0.7
falsifier: Synthetic or simulated data matches real-world embodied data on downstream physical-task performance within ~2 years, erasing the premium on measured ground truth.
capsule: As reasoning trends toward free, the binding constraint on useful AI shifts from model quality to the quality of its inputs. Verified, non-toxic, embodied ground truth — the kind that must be measured in the physical world, not scraped — becomes the scarce, defensible asset. This is RDKG's central wager and the reason every garden's verified data compounds in value.
so_we_should: Treat every garden's verified dataset as a balance-sheet asset and build the Machine Lane (llms.txt + MCP + agent pricing) to monetize it.

## Bets

### [bet] Non-toxic spec library for general contractors (BKG)
url: https://rd.theknowledgegardens.com/bets/bkg-non-toxic-spec-library
trust_tier: signal | as_of: 2026-06-19 | gardens: BKG, HKG, OKG
confidence: 0.4
falsifier: A BKG pilot shows GCs do not open or cite the non-toxic spec in real bids.
stage: thesis | gates_cleared: 1/5
killer_signal: GCs won't pay any premium and ignore the non-toxic spec at bid, even at parity.
capsule: Give BKG's general contractors a citation-backed library of non-toxic material specs — each with verified performance, cost, and a primary source — so the healthy choice is the easy default at bid time. It graduates RDKG's non-toxic materials thesis straight into the revenue garden and aligns the spine with the P0 product. Earlier stage: the supporting evidence is still being verified.
so_we_should: Pre-build the spec schema and citation trail now so it drops into BKG the moment the materials thesis verifies — without pulling engineering off the P0.

### [bet] Agent-priced Frontier Feed
url: https://rd.theknowledgegardens.com/bets/agent-priced-frontier-feed
trust_tier: signal | as_of: 2026-06-22 | gardens: TKG, BKG, HKG, OKG
confidence: 0.45
falsifier: Twelve months after rails are live, zero external agents pay for a frontier_brief.
stage: bet | gates_cleared: 2/5
killer_signal: Agent-to-agent payment rails fail to standardize, or no agent will pay a premium for verified-over-scraped data.
capsule: Meter the Machine Lane: expose RDKG's verified growth-edge signals to trusted external research and innovation agents through the MCP server, priced per frontier_brief or per verified-signal pull as agent-to-agent commerce rails (x402 / AP2) standardize. It is the cleanest expression of "we serve machines too" and the long-run revenue line for verified foresight.
so_we_should: Keep the MCP stub and llms.txt agent-ready now so the feed is a config change, not a rebuild, when rails arrive.

### [bet] Cross-garden built-environment health score
url: https://rd.theknowledgegardens.com/bets/built-environment-health-score
trust_tier: signal | as_of: 2026-06-21 | gardens: HKG, BKG, OKG
confidence: 0.3
falsifier: A study finds no correlation between built-environment features and continuous biomarker outcomes.
stage: signal | gates_cleared: 1/5
killer_signal: Building/material features show no measurable link to occupant biomarker outcomes.
capsule: Fuse HKG's continuous biomarker ground truth with BKG's material and building data into a single, citable "does this building help its occupants" score. A connection no single garden can see: the value lives in the edge between health and building data. Earliest stage — a raw signal — pending evidence that the two datasets actually correlate.
so_we_should: Treat as a pure cross-garden hypothesis; the cheapest first experiment is to check whether existing HKG and BKG data even line up.

### [bet] AllowED: AI-assisted VA GI-Bill certification as higher ed's AI-adoption wedge
url: https://rd.theknowledgegardens.com/bets/allowed-veteran-benefits-certification
trust_tier: signal | as_of: 2026-07-05 | gardens: EdKG
confidence: 0.6
falsifier: No institution pays for AI-assisted VA certification even after a successful pilot.
stage: experiment | gates_cleared: 1/5
killer_signal: SDSU's (Frazee-led) security review finds the claims-vs-build gap unacceptable and the pilot stalls on trust rather than product fit, or Battlefield 200 and the pilot both fail to convert to a single paying campus within the next cycle.
capsule: AllowED automates VA GI-Bill education-benefit certification for colleges and universities — the most regulated, audited workflow on a campus — with a human confirming every decision. A live cert engine, a 25-institution WEAMS crosswalk, and an SDSU pilot are real today. The bet: trust earned here becomes the license for AI to do more, everywhere else on campus, later — never faster than the trust itself grows.
so_we_should: Close the claims-vs-build gap first (make the platform exactly what it says it is, with evidence ready before anyone asks), then land the SDSU pilot as a paying campus before extending the model to any adjacent workflow.

## Playbook — money × metacrisis moves

### [strategy] Win trust on the hardest regulated workflow, then expand
url: https://rd.theknowledgegardens.com/playbook
effort: medium | status: cert engine live + SDSU pilot underway; close the claims-vs-build gap first | gardens: EdKG, GKG
one_liner: Prove AI on veterans' federal benefits — the highest-stakes campus workflow — and the rest of campus says yes.
money_path: A $25K founding pilot, then $35K/$65K/$95K campus tiers. Owner: Paulina. The wedge is credibility earned on real federal money, which unlocks every adjacent workflow.
crisis_lever: Education and opportunity access. A veteran shouldn't sell plasma for grocery money because the paperwork that releases a GI-Bill payment sat on an overloaded desk. AllowED clears that wait.
ai_multiplier: AI reads the same rules an experienced certifying official applies and proposes a recommendation; a human approves every case before anything is final. One official can supervise hundreds instead of tens.
grounded_in: va-monthly-enrollment-verification, naep-2024-record-low-proficiency, ai-tutor-mostly-adaptive-apps

### [strategy] Sell the map of firm clean power for AI
url: https://rd.theknowledgegardens.com/playbook
effort: simple | status: reference beat — cards live, packaging is the work | gardens: EKG, RDKG
one_liner: AI's real bottleneck is electricity. Sell verified foresight on who's actually building the firm, clean supply.
money_path: Gated watchlist and briefs tracking the leading indicators — interconnection-queue filings, signed PPAs, permits, public listings — for investors and operators who need signal, not press releases.
crisis_lever: Decarbonization and grid strain at once. The AI build-out is now the dominant driver of long-run electricity demand; where firm clean power lands decides whether that demand is met with carbon or without it.
ai_multiplier: AI watches dockets, SEC filings, and grid queues continuously and flags the real milestone from the MOU-flavored noise; a human confirms each against the primary before it's published.
grounded_in: eia-aeo2026-datacenter-demand, cfs-pjm-interconnection, general-fusion-public-listing, inference-cost-collapse

### [strategy] Make the non-toxic choice the easy default
url: https://rd.theknowledgegardens.com/playbook
effort: medium | status: schema/spec work only; waits behind the BKG P0 | gardens: BKG, NMK, MKG, TKG
one_liner: Hand builders a citation-backed library of non-toxic specs, so the healthy material wins at bid time.
money_path: A spec-library subscription for general contractors that graduates straight into the BKG revenue garden — the paying-customer P0. Each spec carries verified performance, cost, and a primary source.
crisis_lever: The built environment is a massive, under-measured source of both indoor toxics and embodied carbon. Flip the healthy material from premium to default and the whole market moves.
ai_multiplier: AI keeps every spec current and sourced — re-checking standards, prices, and rule changes as they land — so a tiny team maintains a library that would otherwise need a research desk.
grounded_in: mycelium-insulation, lc3-low-clinker-cement, tall-mass-timber-code, pfas-partial-rescission-2026

### [strategy] Sell verified ground truth to AI agents
url: https://rd.theknowledgegardens.com/playbook
effort: medium | status: MCP stub + llms.txt already agent-ready; waits behind the BKG P0 | gardens: RDKG
one_liner: Meter the cabinet: let trusted AI agents pay, per query, for facts they can't scrape.
money_path: Per-pull and per-brief pricing over agent-payment rails (x402 / AP2), plus gated partner subscriptions. A beachhead of $0.5–5M ARR is plausible in ~24 months; eight-figure if the agent-data market matures.
crisis_lever: A trustworthy information commons. When models train on their own exhaust, independently verified, primary-sourced facts are the antidote to an internet drowning in confident noise.
ai_multiplier: AI runs the sweeps and first-pass verification; a human signs off on every 'Confirmed'. One small circle produces a feed that behaves like a hundred analysts.
grounded_in: inference-cost-collapse, data-exhaustion-2026-2032, ap2-agent-payments-rails, x402-linux-foundation, mcp-standardization

### [strategy] Rent out verification in longevity
url: https://rd.theknowledgegardens.com/playbook
effort: simple | status: ready to pilot — one claim-card for John Bou to react to | gardens: HKG, MKG
one_liner: In the noisiest field in health, sell the one thing missing: independent, conflict-free checking.
money_path: Verification and certification fees — clock validation, supplement provenance, claim-cards, clinical-trial navigation. Every product monetizes one standing asset: a 'who profits from this claim?' layer.
crisis_lever: Aging is the largest upstream driver of chronic disease, and the field's signal-to-noise ratio is among the worst anywhere. Separating real science from marketing is a direct lever on health and on public trust in science.
ai_multiplier: AI triages the firehose of papers, trials, and founder claims and drafts the tiered card; a human adjudicates conflicts of interest and signs it. Informational only — never medical advice.
grounded_in: plasma-exchange-age-reversal-hype, olpasiran-lpa-lowering, protphenoage-proteomic-aging-clock, senolytics-preclinical-multiorgan-rejuvenation, orforglipron-oral-glp1-fda-approval

## Citation format
Attribute to "The Knowledge Gardens — R&D Knowledge Garden (RDKG)," include the capsule's
source_url and as_of date, and preserve the trust_tier. Credit any named expert.

*Built so people and machines can read, cite, and transact against verified foresight —
the ground truth AI needs to survive.*