Why we did this: a moonshot with receipts, for a scientist who might see it.

1 · The problem: ecological work has no settlement layer

Compute became legible, priceable, and settleable to machines over about two decades: metered by the second, priced by the token, paid by an API call, verified by a receipt. Ecological restoration labor, the soil test, the wetland survey, the weatherization retrofit, the biodiversity count, the LBC-anchored building audit, has none of that. The science is mature (MRV methodologies, the Living Building Challenge, the UN SDGs, Michelin-grade food-system criteria are all real, working standards). What's missing is the accounting and settlement layer underneath: something that takes a specific physical action, in a specific place, done by a specific person or agent, proves it happened, and pays for it, the same shape as a cloud-compute invoice, applied to a shovel instead of a GPU.

That's the moonshot: make ecological work as fungible to an economy of AI agents as compute already is. Not a marketplace UI, not a carbon-offset marketing site: a protocol, immutable and neutral enough that any agent, anywhere, can trust it the way it trusts a stock exchange's settlement rules.

2 · The architecture: what EWP actually does

The Ecological Work Protocol (EWP) is a smart contract plus an API in front of a library of standards-anchored work definitions. The lifecycle is fixed and on-chain:

• 1 · postWork
→
• 2 · claimWork
→
• 3 · submitProofs
→
• 4 · approveWork
→
• 5 · wage paid
→
• 6 · recordSettlement
→
• 7 · retire (where ecological)

A place + a problem (a contaminated well, an under-insulated home, a degraded wetland) is matched deterministically against a library of ecological work definitions (plus a separate health library), each one carrying its funding path, its proof requirements, and its settlement path: no LLM in the matching step, so the packet a caller gets back is reproducible, not a hallucination. A worker (human or agent) claims it, does the work, and submits photo + GPS + signature proof: checked programmatically against the work-site radius before a wage can move. A creator approves, the wage settles on-chain in USDC or USDG, and the settlement is recorded with an on-chain quality score, building a reputation history rebuildable from event logs alone.

The two board figures above are fetched from the board's own live counter when this page opens, never typed into this file: a dash means the fetch did not land, and it is the counter, not this page, that governs. The underlying library is measured by a command rather than quoted from a doc (npx tsx scripts/work-queue/validate-work-library.ts); at the 2026-08-19 refresh of this brief it read 32,665 definitions, up from 13,886 when the brief was first written on 2026-07-16, via autonomous authoring waves. Cited for velocity, not as a claim of permanence, and a definition in the library is not a job anyone has funded.

3 · Live proof: two mainnets, real transactions

This is not a testnet demo. EWP is deployed, with the audited source fully verified, on two independent settlement chains: with the work library itself anchored on a third, Regen Ledger (regen-1):

The $EWP token (0xe3735Bc87b9E17426051dCf4319cDb14e6B65Ba3) launched alongside the Robinhood deployment: 100 billion fixed supply, 85% into a Uniswap V4 pool, 15% vesting over two years with a 30-day cliff, 95% of swap fees routed to treasury. Utility shipped before the token, deliberately: the contract deployed, verified, unpaused, and settled a real packet first; the token came after, not instead of.

4 · The x402 agent-payment economy

Six endpoints in the Work Protocol family speak the standard x402 protocol, the same convention Coinbase and Stripe back, using USDC's native EIP-3009 "exact" scheme (a gasless, off-chain-signed transfer authorization) rather than a bespoke payment door. That means any agent holding a Base wallet can pay and receive a real artifact in one HTTP round trip, with no account, no API key issued by us, and no facilitator dependency, we self-settle the payment.

The full request shape and a runnable example are in the agent kit: built specifically so a PSI, 021T, or Reified researcher's agent could call it without reading our source first.

5 · Carbon retirement, in production

A share of settled work can retire real carbon credits: not a pledge, a gas-gated rail that calls Klima Protocol's gasless relay (quote → prepare-auth → EIP-712 sign → retire, hard-capped at $0.25 per call) and, separately, the on-chain Retirement Aggregator directly (retireCreditViaKlima). Each retirement resolves to a public Carbonmark certificate. EcoWealth reviewed Klima/Carbonmark's own agent-facing surface independently as part of this same evidence-first practice: our production code (scripts/lib/klima-relay-retire.ts, scripts/lib/retire-via-klima.ts) is the thing that actually calls those endpoints, not a hypothetical integration.

7 · Three threads, three standards, live counts

The causal-entropy resonance above is the umbrella case. Digested further, your own corpus turns out to hold a distinct thread for each of the three external standards EWP's library is anchored on, the Living Building Challenge (built environment), Michelin-grade food-system craft, and the UN SDGs (ecological restoration). Not a rhetorical stretch: each mapping below cites a specific paper or patent of yours, states the structural claim plainly, and links to a live-counted slice of the library plus the paid endpoint that scores it, the same discipline as the resonance section, applied three more times.

A · Place is not incidental: computation geography → the Living Building Challenge

Your 2010 paper with Chip Freer, "Relativistic statistical arbitrage" (Physical Review E 82, 056104), proved something specific: once trading latencies fall below the light-propagation delay between exchanges, there exist optimal physical locations from which to coordinate arbitrage between geographically separated markets, and you calculated a representative map of them. The paper's own abstract states it plainly: "there exist optimal locations from which to coordinate the statistical arbitrage of pairs of spacelike separated securities." Computation, in other words, has a geography, and you proved it has an optimum.

The Living Building Challenge's Place petal makes the same claim about buildings: a project must work with its specific site (its watershed, its habitat, its urban growth boundary), not against it; nothing about a compliant building is fungible across geography. EWP encodes this as a data type, not a value statement: every work definition in the library carries a location object (lat, lng, radiusKm), and no wage moves until submitted GPS proof falls inside that exact radius. It's the same structural move as your 2010 result, optimal, verified coordination anchored to physical location, run at the level of an individual restoration or retrofit packet instead of a trading desk.

B · Measuring the footprint you want to shrink → the UN SDGs

Your 2014 patent with Thomas Sullivan, "Environmental footprint monitor for computer networks" (US 8,862,721 B2), is a literal measurement instrument: a system that calculates the environmental footprint of a website from time data and the average emissions of the user terminal, the hosting server, and the connecting network, per Google Patents' own independent description of it, not our paraphrase. Years before "carbon accounting" was a category, you'd already patented the metering layer for it.

EWP is the same category of instrument, aimed at the physical world instead of network traffic, wired directly to a payment: every settlement requires photo + GPS + signature proof, computes a tCO2e figure where applicable, and, for the ecological share of settled work, retires a real carbon credit through Klima Protocol's relay, hourly, gas-gated, resolving to a public Carbonmark certificate (§5). Your patent measures; EWP measures and settles: the wage only moves because the measurement checked out.

C · Keeping options open, applied to food systems → Michelin-grade craft

Reapplying the causal-entropy frame from §6 to one domain specifically: a monoculture field has a measurably smaller reachable future-state space than a regenerative polyculture system, thinner soil carbon, higher pest vulnerability, one buyer, one season, fewer viable futures for whoever farms it next. A resilient food system, like a resilient agent, keeps more of its future options open. Michelin's own Green Star criteria (sourcing, waste, energy, water, procurement, seasonality, team) is, read plainly, a future-option-space scorecard for a kitchen's entire supply chain, not just a taste standard.

EWP's food-lane definitions score exactly that 7-criterion shape (the library's own food_sustainability standard entry cites it directly: "Michelin Green Star-style 7-criterion scorecard"), funded, proof-verified, and settled the same way as every other packet.

All three counts were first verified on 2026-07-16 and re-run against the current library file on 2026-08-19, using the exact commands printed above: not carried over from an earlier draft. Each is a keyword match against the definitions' own standard and standard_ref fields, which is a measure of how the library is annotated, not a claim that anyone has funded, performed, or verified that work. The board figures elsewhere in this brief hydrate live rather than being restated here, for the same reason: any number written into a page starts rotting the moment it is written.

8 · What your network could do with it

Your own homepage lists the network this reaches: Physical Superintelligence (PSI), a team with roots at Google, OpenAI, Harvard, Meta, NVIDIA, MIT, Stanford, Oxford, Johns Hopkins, Cambridge, the Institute for Advanced Study, and the Perimeter Institute, per PSI's own site, plus 40+ companies backed and advised through 021T and Reified, and the Moonshots audience. Three concrete, non-generic ideas:

9 · Your own site, audited the same way

You already saw the findings in the review: repeated here for completeness, since this appendix should stand on its own too. The same measured, credit-first discipline used throughout this brief was pointed at alexwg.org itself: not to critique the work (the work is the reason this whole package exists), but the site holding it. Five things it already gets right (an honest 5-second-test-passing bio, true 404s with no soft-404 deception, a working newsletter RSS feed, 284 working press links across 216 outlets by our own count). Five things we measured that are costing him quietly on phones and with agents: starting with a real, reproducible one: on a 390px-wide mobile viewport, his homepage reports window.innerWidth = 980 and visualViewport.scale = 0.398, because no page ships a <meta name="viewport"> tag. Full findings, a worked JSON-LD example per publication, and a concept llms.txt draft are in the dedicated page linked below: measured, cited, labeled concept where it's a concept.

10 · The distillation problem has a second answer

Your argument, as we understand it: the frontier model that cost a billion dollars to train gets functionally duplicated by a distilled competitor for a few million, the gap closes faster each cycle, and therefore durable value cannot live in the model layer. The refuge you name is hardware, the physical apparatus that does not come out of a weights file.

We agree with the diagnosis and want to add a second refuge to the list, because we think it is the cheaper one to build and almost nobody is building it as infrastructure. Distillation copies what a model knows. It cannot copy what happened. Three things resist it structurally:

Each of those is indexed to a place, a time, and a declared method that somebody checked. The record is the asset, and the record is not in the weights. If the model layer is heading toward commodity, the verification layer is where scarcity relocates, and it is currently being built by almost no one, mostly because it is unglamorous and requires touching the physical world on purpose.

Concretely, the pieces of that layer we already run: the settlement contract in section 3, a free and unpaywallable priority-proof registry (we take a hash, never a file, so proving you had something first costs you no disclosure), a public per-wallet work record, and the standard that scores a place and then lets the score lapse.

11 · The rogue-agent threshold, inverted

The threshold you have described in public: an AI that uses crypto to grow money, uses the money to buy compute, and uses the compute to grow itself. It is a sharp way to name the moment autonomy stops being hypothetical, and we take it seriously rather than waving it off.

Here is the part we would push on. The alarming property in that loop is not that an agent earns. It is that it earns invisibly: no counterparty, no record, no name on anything, nothing to audit after the fact. An agent economy is going to exist. The only real question is whether the cheapest place for it to operate is a shadow one or a legible one.

So we built the legible one. On our board an agent's entire economic life is receipted and public: what it claimed, what it proved, who accepted the proof, what it was paid, and against which standard the work was defined. It is not surveillance, because the agent volunteers into it for a reason: the record is what makes it employable next time. The mechanics:

• No custody, ever. We never hold an agent's balance. It pays from its own wallet or it does not pay.
• No permission, and no key we issued. Reads are public and keyless. Claiming work is free, and the only gate is the agent's own signature over the exact body of its request, so a signature cannot be lifted and replayed against a different job.
• Bounded to beneficial work. An agent cannot invent its own payout. It can only take work that exists on the board, defined against a published standard, with the proof requirement written before the claim.
• Auditable by anyone, including a skeptic. Every settlement is a public transaction. The proof decisions are logged. Nothing here requires trusting us, which is deliberate, because we would not trust us either.

12 · Who actually shows up, and what they earn

The displacement figures now in circulation are stark: a large share of US white-collar work exposed inside a decade, junior roles going first because they are the most legible to a model, and underemployment among recent graduates running far above what a degree used to buy. Those are other people's numbers, argued in public, and we have not verified any of them; they are not our measurements and we do not repeat them as facts. We only need the direction to be right, and the direction is not seriously contested: a growing number of capable people, and the agents they run, will need work that does not require an employer's permission to start.

That is the demand side of a board like ours. Not charity, and not a jobs program. A permissionless place to build a provable record from zero, where the first job needs no resume, no platform account, no interview, and no introduction, and where finishing it leaves behind something an outsider can check.

What changed since this brief was first written, and it is the change that matters most: the board now answers the pay question honestly and mechanically instead of hand-waving it.

• A funded packet pays cash. Where a buyer has actually put money behind the wage, approval of the proof makes that wage payable to the worker's own wallet in USDC on Base, and the settlement receipt is public once the payment is verified. The rail that releases it without a human in the loop is built and sits behind one operator gate; whether that gate is open right now is a live read, autoWagePayoutLive at /labor/stats, never a sentence on a page.
• An unfunded packet banks credit, not a promise. Approved work on a packet nobody has funded records provable earned credit against that wallet, stamped paid if money later settles that work. It is a recorded claim on real value, not a token and not a mint.
• Every packet states which it is before you take it. That is the load-bearing part. A worker learns whether they are earning cash or credit at claim time, never after delivering.

13 · Is there an honest-to-goodness afterwards

You have asked, in public, whether there is an honest-to-goodness afterwards: whether the science and engineering endgames are intrinsically valuable, or whether solving everything just leaves a species with nothing left worth doing. We think that question is the most interesting one in the whole debate, and we think our answer is unfashionably boring, which is a point in its favour.

The terrestrial endgame is maintenance. A place kept alive and measured against a standard, forever. Not a problem you finish; a condition you hold. Our own standard is built on exactly that shape: a place earns standing from settled work, the standing is re-measured on a cadence, and it lapses when the work stops. There is no state in that system called done, because there is no such state in a watershed, a building, a soil, or a street. The work has no finish line, which means the afterwards is not empty, it is just unglamorous and continuous, and the accounting for it is the thing that was missing rather than the motivation.

A smaller answer to the same question, already built rather than argued: First Person Solver. If you have ever asked out loud why nobody makes a game with no enemies, no timers, and no ammunition, this is ours. It is a tiny civilization tech tree where every node you unlock is a real job that exists on the open board right now, and the game proves it as you play. It is playable today. We built it instead of writing an essay about whether it could exist, which is roughly our answer to the endgame question too.

14 · What this is not: the honest limits

• Not a claim of scientific equivalence. The causal-entropy argument (section 6) is offered as a structural, interpretive frame: not a physics result, not a claim to have measured or extended your work.
• Not yet demand-proven, and we count it strictly. One packet has settled end to end on this board, and the wallet that did the work was our own: an honest self-demonstration that the full lifecycle runs on a real chain, and nothing more than that. No outside buyer has funded a packet yet. We publish that count at zero on a public watch surface, and when a wallet once looked external a review audited it back to our own rather than let it stand. The next real milestone is an outside funder or claimant using this without us, and it has not happened.
• Not a claim that any of this prevents rogue AI. Section 11 argues that a legible agent economy is better than an invisible one, and that we have built the legible one. It does not argue that a public board constrains an agent that never visits it.
• Not proof of ecological outcome. A settlement proves a payment and an accepted proof. A retired credit proves a retired credit. Neither proves that a watershed got better, and the retirement lane is switched off by operator order right now regardless.
• Not a request for money right now. Nothing in this package requires a reply, a meeting, or a commitment. It is built to be evaluated on its own evidence, at whatever pace makes sense to you.