The Knowledge Gardens · RDKG — codename "The Meristem"
Turning loud claims into verified ground truth.
RDKG is the foresight arm of The Knowledge Gardens. It sweeps the frontier, checks every claim against a primary source, and turns the survivors into plain-language findings that bridge two things at once — a way to make money and a lever on a real global problem. Start with the big picture; click anything to drill into the research behind it.
Internal · confidential · for the Circle and teammates · updated 2026-07-06 · thinktank.theknowledgegardens.com
The big picture — from macro forces to what we build
The whole strategy is one chain of cause and effect. Two verified macro forces make a bet true; the bet tells us what to grow and sell. Click any box to see the evidence.
1 · Macro forces (Confirmed)
→
2 · So the value moves upstream
→
3 · So we grow & sell it
Every box opens the primary source, the trust tier, and the one thing that would prove it wrong.
How to read a finding
Every finding carries a trust tier and a date. The tiers are honest on purpose — we keep the overhyped ones and label them, because a field full of confident claims is exactly what a verified feed arbitrages.
Confirmed
A definitive primary — a granted patent, a published rule, a peer-reviewed result — or the same fact from several independent sources.
Early signal
One credible primary, or a real-but-partial result. We name the source that would confirm it.
Overhyped
The loud version of a claim the evidence doesn't support. Kept as a caution and a contrarian tell.
Tier ≠ how sure
Tier = how well-sourced. Confidence (a %) = how much we'd bet it holds up. Usually aligned, not always.
The gardens
Money gardens fund mission gardens. Two are far enough along to anchor real products — one live, one clinical. Click a garden to drill into the research and the roadmap.
EdKG · The Education Garden
Building
Anchor: AllowED · lead: Paulina
Verified content a parent can trust and a kid can't cheat with — and AI adoption without rupture. Not "an AI tutor" (everyone will have one), but verified, cheat-resistant learning plus a coaching layer that makes a learner do the work.
Its live anchor, AllowED, automates the most regulated workflow on campus — VA GI-Bill certification — with AI proposing and a human deciding every case. Trust earned on veterans' federal money is the license to bring AI to the rest of campus.
Click to drill into AllowED, the subcells, and the next step →
HKG · The Health Garden
Building
Longevity arm · partner: John Bou
Verification is the product. In the most hype-saturated field in health, the scarce asset isn't a molecule or a clock — it's independent, conflict-free verification that separates real science from marketing.
The teaching case is David Sinclair: real, field-defining science sitting beside a "reverses aging in dogs" claim that cost him a scientific presidency. One person, two tiers, weighted separately.
Click to drill into the Sinclair case, subcells, and the money →
The rest of the gardens (mapped, feeding the findings) — and a note on codes to confirm
BKG
Builder's / construction — the current paying-customer P0.
NMK
NatureMark — non-toxic, regenerative materials.
TKG
Toxicology — what's harmful, and the safe alternative.
MKG
Materials + biomarker "Marker" work.
OKG
Orchid — growing / agriculture + operations.
CulKG
Cultivar — food and cultivated protein.
FshKG
Fish / aquaculture / seafood.
GKG
Government & defense.
ParKG
Parenting — the parent is the blocker, not the child.
EKG
Energy.
RDKG
The machine / meta-layer — this think tank itself.
To confirm (founder): the exact scope of OKG and whether a dedicated agriculture garden exists; the FshKG (Fish vs Fashion) and PKG (Parenting vs Pharma) code collisions. The cabinet uses ParKG for parenting to avoid the clash.
The strategy layer — theses & bets
A thesis is a claim about where the world is going. A bet is something we could build or sell because of it. Click any card for the full detail — confidence, gates, the killer signal, and the falsifier.
Theses
Verified ground truth is AI's scarce input
Early signal
As reasoning trends to free and cheap training text runs out, the differentiator moves upstream to inputs that must be measured, not scraped.
How sure: 70% · Touches: TKG·BKG·HKG·OKG·FshKG
Click for the full thesis →
In longevity, verification is the product
Early signal
The field splits into real science and founder-branded claims its own scientists repudiate — often from one person. The scarce thing is the layer that tells them apart.
How sure: 70% · Touches: HKG·RDKG
Click for the full thesis →
Non-toxic materials cross cost-parity this decade
Early signal
Mycelium composites, low-carbon binders, bio-based panels may reach parity with toxic incumbents before 2030 — flipping the healthy choice from premium to default.
How sure: 50% · Touches: BKG·OKG·HKG
Click for the full thesis →
Bets
Agent-priced Frontier Feed
Early signal · 2/5 gates
Meter the Machine Lane: sell verified signals to trusted AI agents over x402/AP2 rails, per brief. The long-run revenue line for verified foresight.
How sure: 45% · Est: $0.5–5M ARR beachhead in ~24mo
Click for gates, killer signal, the experiment →
AllowED — VA GI-Bill certification
Early signal · 1/5 gates
AI-assisted VA certification as higher ed's AI-adoption wedge. Live cert engine, 25-institution crosswalk, SDSU pilot — real today. Owner: Paulina.
How sure: 60% · Pricing: $25K pilot; $35/65/95K tiers
Click for current state and the one risk →
BKG non-toxic spec library
Thesis stage · 1/5 gates
A citation-backed library of non-toxic specs so the healthy choice is the easy default at bid — RDKG foresight graduated into the BKG revenue garden.
How sure: 40% · spec work only, waits behind the BKG P0
Click for the full detail →
Built-environment health score
Raw signal · 1/5 gates
Fuse HKG biomarker data with BKG building data into one "does this building help its occupants?" score — value in the edge no single garden sees.
How sure: 30% · cheapest test: do the datasets even correlate?
Click for the full detail →
The findings
The verified core — twelve seed findings, each primary-sourced and dated, across the four spines (materials/building, health/biomarkers, growing/robotics, the machine layer). Filter by tier.
EPA enforceable PFAS water limits
Confirmed
First federally enforceable PFAS limits in drinking water.
Touches: TKG·NMK·HKG
US EPA / Fed. Reg. 89 FR 32532 (2024). Live risk: 2026 proposed rescind of co-contaminant limits.
Tall mass timber in the model code
Confirmed
Wood buildings up to 18 stories allowed in the model code.
Touches: BKG·NMK·MKG
ICC, 2021 IBC §602.4.
LC3 low-clinker cement cleared
Confirmed
A lower-carbon cement passes standards-body qualification.
Touches: MKG·BKG·NMK
RILEM TC-282, Materials & Structures (2023).
Semaglutide cuts cardiovascular events
Confirmed
The SELECT trial: a GLP-1 drug reduced heart attacks and strokes.
Touches: HKG
NEJM 2023 / NCT03574597.
DunedinPACE clock predicts dementia
Confirmed
A pace-of-aging clock associated with later dementia risk.
Touches: HKG·NMK
Neurology 2022 (PMID 35794023).
"Taurine deficiency drives aging"
Overhyped
A splashy claim the follow-up evidence undercuts; human causality unproven.
Touches: HKG·TKG
Science 2023 vs Aging Cell 2025 (contrarian read).
US cultivated-meat clearance
Confirmed
Lab-grown meat cleared for US sale — though retail presence is near-zero and several states banned it.
Touches: CulKG·FshKG·HKG·OKG
FDA "No Questions" + USDA-FSIS GOI (2023).
Autonomous laser weeding
Confirmed
A granted patent for killing weeds with lasers instead of chemicals.
Touches: OKG·NMK
USPTO US 11,602,143 B2 (2023).
"Vertical farms will feed cities"
Overhyped
The grand claim, against a marquee bankruptcy — economics don't hold at scale yet.
Touches: OKG·CulKG
US Bankruptcy Ct, D. Del. 23-10737 (2023).
Google AP2 agent-payment rails
Confirmed
A standard for AI agents to make payments — the rails the Frontier Feed rides.
Touches: machine layer
Google Cloud (16 Sep 2025).
Inference cost fell ~280× / 18mo
Confirmed
The price of a unit of AI reasoning collapsed — the "AI goes free" leg.
Touches: machine layer
Stanford HAI, AI Index 2025.
Training text runs out ~2026–2032
Confirmed
Estimates that the internet's high-quality text is nearly exhausted — the "data runs dry" leg.
Touches: machine layer
Villalobos et al., Epoch AI, arXiv 2211.04325.
Coverage: 3 findings in each of the four spines · 10 Confirmed, 2 Overhyped. A researched bench is ready to promote: olpasiran >95% Lp(a) reduction, CRISPR high-yield soybean deregulated, MCP, mycelium insulation at lab parity.
This season's sweeps
Each sweep takes a source — a podcast, a news window, a field — and runs every claim through the discipline. Click one to drill into the split between what held up and what didn't.
All-In #279 — the AI sovereignty fight
7 findings
The real story: who owns the model is now a live commercial fight — and it validates our founding bet. Three same-direction events in one fortnight. Confirmed
Click for the findings →
Moonshots — Alpha School & AI education
7 findings
Crisis real (NAEP lows); Alpha's outcomes self-reported; its "AI tutor" is mostly adaptive apps. The tell: it bans chatbots because kids cheat. Early signal
Click for the findings →
David Sinclair & the longevity frontier
7 findings
Reprogramming real & now clinical (ER-100, Jan 2026); resveratrol/SIRT1 overhyped after a $720M deal; a dog-supplement claim cost him a presidency.
Click for the findings →
Broader corpus sweeps
reference
The 90-day Open Frontier Sweep (18 cards) and the Innermost Loop + Moonshots #26 sweep (Sonnet 5, a $4,900 humanoid, Helion fusion, orbital data centers).
Click for the summary →
How we work — and the discipline we hold
The loop
Sweep wide, verify deep, then card it. The frontier-research skill scans a source or field and pulls every distinct signal; the thinktank skill takes the strongest, verifies each against a primary, tiers it, and writes a plain-language finding with a money angle, a meta-crisis lever, one action, and one falsifier. The survivors land here. Longevity cards also get a conflict-of-interest check; every card carries a freshness date and gets re-verified, because claims decay.
The guardrails
Internal-first. Nothing here diverts engineering from the company's P0 — the BKG paywall and first paying customer. Anything that becomes a build starts as a spec or a stub.
Say only what is true. Trust is the entire business. We label the overhyped, credit the people behind each finding, link primaries rather than reproduce them, and reconcile to canonical docs the moment they arrive.
Health cards are informational only, not medical advice.
AI reasoning is racing toward free
Macro force · Confirmed · machine layer
The price of a unit of AI reasoning fell roughly 280× in 18 months. When thinking is nearly free, no model keeps an edge on cleverness alone — the advantage moves to what you feed it.
Source
Stanford HAI, AI Index 2025 (verified in-session against the primary page).
Why it matters
This is one of the two legs the entire RDKG thesis stands on. Cheap reasoning + scarce data = value moves upstream to verified inputs.
What would prove it wrong
The cost curve reverses or plateaus such that model quality, not input quality, stays the differentiator.
The public text to train on is running out
Macro force · Confirmed · machine layer
Credible estimates put the exhaustion of high-quality public human text at roughly 2026–2032. The cheap substrate that trained today's models is drying up.
Source
Villalobos et al., Epoch AI — arXiv 2211.04325.
Why it matters
The second leg of the thesis. If reasoning is cheap and text is scarce, the premium shifts to freshly-measured, hard-to-synthesize, physical-world data.
What would prove it wrong
Synthetic or simulated data matches real embodied data on downstream physical-task performance within ~2 years, erasing the premium.
Verified physical-world ground truth becomes AI's scarce input
Thesis · Early signal · 70% sure
The thesis follows directly from the two macro forces: if reasoning gets cheap while cheap training text runs dry, models converge and the differentiator moves upstream to the 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.
Touches
TKG · BKG · HKG · OKG · FshKG
What we should do
Treat every garden's verified dataset as a balance-sheet asset, and build the Machine Lane (llms.txt + MCP + agent pricing) to monetize it.
What would prove it wrong
Synthetic/simulated data matches real-world embodied data on physical tasks within ~2 years. Held at 70% because that falsifier is live.
The gardens produce the ground truth
The model · money gardens fund mission gardens
Each Knowledge Garden is a source of verified, physical-world data in one domain — building materials, health, growing things, energy, government. The revenue gardens (starting with BKG, the paying-customer P0) fund the mission gardens. Every output has to bridge both: a way to make money AND a lever on a real global problem.
Live anchors
EdKG (AllowED — live cert pilot) and HKG (longevity — clinical frontier cell). The rest are mapped and feeding the findings.
Discipline
Internal-first. Nothing diverts engineering from the BKG P0; any new build starts as a spec/stub.
Agent-priced Frontier Feed
Bet · Early signal · 45% · gates 2/5
Meter the Machine Lane: expose RDKG's verified signals to trusted research and innovation agents through an MCP server, priced per brief or per verified-signal pull, as agent-to-agent commerce rails (x402 / AP2) standardize. The cleanest expression of "we serve machines too," and the long-run revenue line for verified foresight.
Value estimate (not verified)
$0.5–5M ARR beachhead in ~24 months from gated partner briefs + metered agent pulls; eight-figure potential if the agent-data market matures.
Killer signal
Agent payment rails fail to standardize, or no agent will pay a premium for verified-over-scraped data.
Smallest experiment
One external/partner agent completes one paid frontier_brief pull over an x402/AP2 paywall — the first non-zero agent-settled transaction proves the thesis one level down.
What we should do
Keep the MCP stub and llms.txt agent-ready now, so the feed is a config change, not a rebuild, when rails arrive.
EdKG · The Education Garden · AllowED
Building · anchor platform live · lead: Paulina
It starts with a veteran in line at a plasma center, selling plasma for grocery money, because the paper that would release his GI-Bill payment sat on someone's desk — waiting, because the person certifying enrollment was buried. That wait is why AllowED exists.
AllowED automates VA GI-Bill certification: it reads the same rules an experienced School Certifying Official applies, runs each case through a seven-step decision process, and hands back a recommendation — a human always approves before anything is final. The engine proposes; the SCO decides. An SCO who could carefully certify 10–20 students at peak can now supervise ~800 on day one, and spend the reclaimed hours on the cases that actually need a person.
The wedge
AI doing the most regulated, audited workflow on campus, correctly, with a human in the loop. Once a campus trusts AI with veterans' federal money, "can we let AI touch student data at all?" is already answered yes.
Subcells
AllowED (building) · verified cheat-resistant K-12 content — the "anti-cheatbot," the open flagship vacancy (scoped) · motivation/coaching layer (scoped) · independent edtech-outcome verification (scoped) · adult reskilling on verified content (frontier).
Honest next step
Close the gap between what AllowED's materials claim and what the build can demonstrate today — so any security review can simply be shown the evidence. The three-pillar expansion vision is DRAFT, under Paulina's review.
Cross-links
RDKG · ParKG · GKG. Teaches via the blossom-education philosophy: meet people where they are, small wins first, humans direct AI.
HKG · The Health Garden · Longevity
Building · clinical frontier cell live · partner: John Bou
The longevity arm is not a bet on a molecule, a clock, or a founder — it's a bet on verified, conflict-free ground truth in a field that runs almost entirely on self-reported, founder-conflicted, over-hyped claims. Aging is the largest upstream driver of chronic disease and the capital chasing "cures" is enormous, but the signal-to-noise ratio is among the worst anywhere. That combination is the whole opportunity.
The house rule: peer-reviewed lab science and founder-branded commercial claims are different objects, weighted separately — even from the same person. David Sinclair is the teaching case: partial reprogramming reversed vision in mice (Nature 2020) and reached the first human trial (ER-100, Jan 2026) — real, load-bearing biology. In the same period a Sinclair-affiliated "reverses aging in dogs" claim triggered a scientist revolt and cost him the presidency of the Academy for Health and Lifespan Research.
Independent verification/certification of longevity claims (flagship), clock validation, supplement provenance, clinical-trial navigation — all monetize one asset: a standing "who profits from this claim?" layer.
Guardrail
Informational only, not medical advice; RDKG endorses no supplement or protocol. Cross-links: RDKG · MKG · EKG.
In longevity, verification is the product
Thesis · Early signal · 70% sure
The scarce asset in longevity is not another supplement, clock, or founder claim — it's independent, conflict-free verification that separates real science from marketing. The field has the clearest meta-crisis lever in health and the worst signal-to-noise ratio this cabinet has swept, which is exactly what makes verification, not discovery, the higher-leverage product.
The evidence split
Sinclair's own record holds both poles at full strength: OSK reprogramming (Nature 2020) and ER-100's first human trial vs. the dog-supplement revolt; the resveratrol/Sirtris saga (a $720M GSK deal on a SIRT1 mechanism Amgen and Pfizer couldn't replicate).
Smallest experiment
Produce one verified claim-card — "does NMN extend healthspan in humans?" — in the standard tiered, sourced, conflict-disclosed format, and ask John Bou: is this more decision-useful than what exists?
What would prove it wrong
Patients, clinicians, and investors don't pay for verification over charismatic founder claims. Held at 70% on that untested assumption.
Non-toxic materials cross cost-parity this decade
Thesis · Early signal · 50% (a coin-flip on purpose)
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.
Why only 50%
The supporting signal (mycelium insulation at lab parity) is itself only signal-tier, and the cost data is not yet primary-sourced. We are allowed to be wrong here; we are not allowed to be vague.
To raise it
Graduate the mycelium signal to verified with a primary cost/performance dataset, then add two more material classes on the same curve.
What would prove it wrong
No non-toxic structural or insulating material reaches verified price parity with its conventional incumbent by 2030 in a primary cost dataset.
AllowED — VA GI-Bill certification as higher ed's AI-adoption wedge
Bet · Early signal · 60% · gates 1/5 · owner: Paulina
AI proposes, a human decides: the decision that used to take an SCO an hour now takes a few minutes of review. The wedge isn't "AI for schools" in the abstract — it's AI proven on the single hardest workflow a campus runs: real federal money, real student data, real audit trail, zero tolerance for error. Practitioner-first design is what makes it credible — the rules came from an actual SCO's years of judgment, not a generic template.
Current state, plainly
A seven-step decision process passing its full regression suite; a crosswalk covering 853 programs across 25 institutions; three working frontends (SCO, student, admin); a security one-pager citing 206/206 automated checks passing; an SDSU pilot underway; Battlefield 200 submitted May 2026. Not aspiration — the current build.
The one risk
The materials describe some security features (SSO, MFA, access controls, audit logging) in language that outpaces what's confirmed live. That claims-vs-build gap is the single biggest risk — a platform whose pitch is "trust us with your most regulated workflow" can't afford daylight between the one-pager and the code.
What we should do
Close the claims-vs-build gap first, then land the SDSU pilot as a paying campus before extending to any adjacent workflow.
Pricing
$25K founding pilot; $35K/$65K/$95K campus tiers; ~$50K/campus system blend (Battlefield numbers, built on labor-replacement math, flagged for a value-created rebuild).
Non-toxic spec library for general contractors (BKG)
Bet · thesis stage · 40% · gates 1/5
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.
Why thesis stage, not a bet
Pattern-fit and clearly spine-aligned, but the materials thesis is only 0.5 confidence and still signal-tier, and no pilot has run.
Discipline
Spec/schema work that waits behind the BKG P0 — not engineering pulled off it. Pre-build the schema and citation trail now so it drops in the moment the thesis verifies.
What would prove it wrong
A BKG pilot shows general contractors do not open or cite the non-toxic spec in real bids, even at parity.
Cross-garden built-environment health score
Bet · raw signal · 30% · gates 1/5
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. This is RDKG doing its actual job — proposing value in the edge between two gardens that neither could surface alone.
The de-risking move
Don't build the score. First check whether built-environment features and continuous biomarker outcomes correlate at all in data we can already touch. Foresight that doesn't change a decision is entertainment.
What would prove it wrong
A study finds no correlation between built-environment features and continuous biomarker outcomes.
All-In #279 — the AI sovereignty fight
Sweep · 2026-07-04 · 7 findings
The real story: who owns the model is now a live commercial fight — which validates the founding bet that control of verified capability is the scarce thing. Three same-direction events inside one fortnight.
Confirmed
Palantir–Nvidia sovereign-AI product (Nemotron, June 29). Anthropic's Claude Design blindsiding Figma (Apr 17; Figma CPO had resigned its board days prior). An 18-day worldwide export suspension of Fable 5 / Mythos (June 12–30).
Also confirmed
A Ramp / Revelio study of 21,500 firms found heavy AI adopters hired more, not fewer (correlation, caveated).
Parked
A SCOTUS birthright ruling and the CA budget — real but off-spine — and one founder's self-reported benchmark. We also corrected our own earlier error that had tagged the real Fable 5 export saga as "confirmed fiction."
The crisis is real and documented; the school's own numbers are not independently verified; and the "AI" is more modest than the framing. All three matter for EdKG.
Confirmed
Only 22% of US 12th-graders are proficient in math and 35% in reading (NAEP 2024, lowest on record); recent-grad unemployment now runs above the national rate (NY Fed).
Early signal
Alpha's "10x faster / top-1% / 1535 SAT" outcomes are self-reported; its "AI tutor" is largely adaptive apps (IXL/Khan) plus an attention coach — "take AI out of it, we could still function."
The tell
The most-hyped AI school bans general chatbots in class because ~90% of kids use them to cheat — which is precisely the EdKG opening: verified content + coaching, not "an AI tutor."
Three piles, weighted separately — the whole reason "verification is the product" in HKG.
Real & now clinical
OSK partial reprogramming reversed vision in mice (Lu et al., Nature 2020). Life Bio's ER-100 IND cleared the FDA (Jan 28 2026); Phase 1 in glaucoma + NAION (NCT07290244). The Information Theory of Aging (Cell 2023) is a serious hypothesis — with a published rebuttal, so cite both.
Overhyped
Resveratrol as a SIRT1 "longevity switch" — the mechanism didn't replicate (Pfizer/Amgen assay-artifact); GSK bought Sirtris for ~$720M and shut it in 2013.
The credibility event
A Sinclair-affiliated "reverses aging in dogs" claim (2024) drew an open revolt — Kaeberlein called it "a lie" — and cost him the Academy presidency. Weight lab science and commercial claims as different objects.
Broader corpus sweeps
Sweep · reference · in the project files
Two larger sweeps feed the same library and are available in full in the project files.
Open Frontier Sweep — last 90 days
18 cards (11 Confirmed / 7 Early signal) across all four spines, adversarially re-audited with in-window primaries only.
Innermost Loop + Moonshots #26
Sonnet 5; a ~$4,900 Chinese humanoid; Fable 5; Helion fusion; StarCloud (orbital data centers) — each checked against a primary.