硅谷Fintech观察⑨｜agent 把钱花错了，算谁的？没人扛得住Silicon Valley Fintech Watch ⑨ | When the Agent Spends Wrong, Who Pays? Today, Nobody Can

Silicon Valley Fintech Watch series · Part Nine · First published on klay-wang.com, September 14, 2026

In brief. On August 16, 2026, Bloomberg reported that Stripe had agreed to buy OpenRouter for more than $7 billion. On September 8, Meta launched Muse, a personal agent that gives every user a dedicated computer in the cloud and checks out through Stripe's Link. Back in March, Walmart's head of product said that purchases completed inside ChatGPT converted at one third the rate of click-throughs to its own site. Three events that look unrelated point at the same cell: the agent spends on your behalf, and when it spends wrong, who pays. In Part Eight I said the map had three cells nobody had drawn, and that the after-the-fact one was worth the most. This piece opens that cell up. First, why nobody has managed to sit in it. Then a table of seven seats, with the three companies I have dug into placed in theirs, each stuck on a different party. Last, what decides how big the cell is. My answer: look at the spending limit, not the error rate.

Before we start

Where this comes from.

When Part Seven went out, the Stripe and OpenRouter deal was still in talks and the Wall Street Journal had reported $10 billion. On August 16 Bloomberg reported an agreement at more than $7 billion; on August 19 Stripe's newsroom published the announcement, worded as agrees to acquire, with no price. Part Seven has since been corrected to $7 billion in the title and throughout, the revenue multiple from 71x to about 50x, and the judgment left untouched: routing is free, the ledger is not.

This piece takes the next step. What comes after the ledger? My observation is that twenty years ago the payments industry also did nothing but route, sending a transaction to where it should go. What grew on top of routing turned out to be worth far more: automatic retries on failure, a second acquirer when the first declines, real-time fraud blocking, the same transaction priced across several rails. Different names, one job, pushing the success rate up. And whoever promises the success rate eats the cost of failure. The token flow is walking the same road again.

Then a harder question: what does agent failure look like, why is it so hard to insure, and who today is betting that this can be charged for.

The second half took me two months. I screened from two ends, the full YC library and regulatory filings, for companies that treat who pays when it goes wrong as a business. Three I worked through until I could explain how a single transaction moves, who charges whom how much, and what would void the judgment. They are Ubyx, Klaimee and Catena. They are the evidence in this piece. Every source is listed at the end.

1. Routing's next step is daring to price the outcome

Start by separating two units of account: the token and the task.

Priced per token, you look at the cost of a thousand tokens. Priced per task, you look at the cost of getting one thing done. The two numbers are far apart. To complete a task an agent has to understand the request, retrieve, draft a plan, call tools, retry when a step fails, and verify at the end. Add inference cost, retry cost, tool calls, verification, latency and the losses from failed attempts, then divide by the number of tasks that actually succeeded. That is the real unit price.

Once that sum is done, a popular inference falls apart: cheap models keep getting better, so routing becomes a price-comparison engine that sends every request to the cheapest one. But a cheap model that fails twice on a task costs more than an expensive one that succeeds once. Sending to the cheapest model was wrong from the start.

So the position in the routing layer that is worth real money is the one that dares to tell the customer this will get done. The customer pays a premium for that sentence; the price is that when it fails, the promiser eats it. That is how payment gateways survived: they charge on success rate, and when something fails, the customer comes to them, not to the underlying rail.

In plain terms: today's token router is a broker that tells you who is cheapest. Its next step is to become a general contractor who tells you I will take this job, and if it fails, that is on me. A broker collects an information fee; a contractor collects a premium. The two kinds of money do not sit in the same ledger. When does a router become an underwriter? The first day it dares to write if it fails, that is on me into a contract.

And the first question a contractor has to answer is: when a job I took goes wrong, what does that look like?

2. Agents fail in three shapes, each harder to insure than the last

This section is the load-bearing wall. I split agent failure into three shapes. The taxonomy is mine, not an industry consensus, but every shape comes with an example I have run into myself.

Shape one: the mandate you give can never be written out in full.

An example. You tell an agent: a flight to Los Angeles, under $500. It can perfectly well book you a budget carrier with no checked bag and an $80 surcharge at the airport, landing at two in the morning, a fourteen-hour layover, non-refundable. You said none of those things, and every one of them is under $500. It broke no instruction.

What you got is not what you wanted, but it did nothing wrong. Now who do you go to? Chargeback? You authorized it. The agent? It did as told. That is the cell with no owner.

Someone will say, then spell it out: nonstop, no layover, no red-eye, no budget carrier, no landing after midnight. Spelling it out does help. It turns a case where nobody can say whose fault it is into a case where it is clear. Breach of contract is something a contract can handle. But however completely you write it, one thing stays out of reach: a ticket satisfying all six constraints may simply not exist on a given day. The agent can stop and report failure, safe but useless. Or it can relax one constraint to close the deal, and which constraint it relaxes is a choice it made for you, because what you gave it was a set of constraints, not a ranking. By the time it has chosen, the non-refundable ticket is issued.

If you could write it all out, you would not need the agent. The value of delegation is precisely that you do not spell everything out. Underspecification is the premise of this business.

Shape two: you spelled it out, and it still did not comply.

This shape is more fundamental than the first, because no specification can block it.

For two months I have used agents daily to write copy, make charts and run data pipelines. The gates are defined clearly, the rules are in the skill files, the gate command is right there. It still skips steps, ignores rules, runs the gate and commits without reading the result. I regularly spend hours fixing what it did. I did make myself clear. It hallucinates, it cuts corners, it sees the red light and goes anyway.

The difference from shape one: there, the loss is bounded by the edges of the mandate; here, there are no edges. It can do anything, including what you explicitly forbade. You can compress shape one with specification. You cannot compress shape two.

Shape three: the errors are correlated.

The foundation of insurance is the law of large numbers: ten thousand policyholders, each with an independent claim, total loss computable. Agent errors are correlated by birth. Ten thousand agent instances running the same model version, the same prompt, the same tool definitions will make the same mistake in the same minute. This is not a risk that can be diversified away, and diversification is exactly what reinsurance is for.

A car insurer is not afraid of one car crashing. It is afraid of ten thousand crashing at once. Agents are ten thousand cars sharing one driver.

Nobody has managed to sit in this cell because each shape blocks a different door. The first needs a claims window, and there is none. The second needs a backstop that does not depend on specification, and there is none. The third makes anyone willing to backstop unwilling.

But the first real money in losses is already being carried by someone. It just is not called agent error. It is called fraud.

Stripe's Radar blog of May 27, 2026 has a sentence: across the Stripe network, more than one in six sign-ups at AI companies are linked to multi-account abuse. Multi-account abuse means one person opening dozens of accounts to reuse free credits repeatedly, or spreading stolen-card activity across accounts to delay detection. The same post says ElevenLabs uses the system to block 2,000 free-tier abusers a day.

A more specific number: OpenCode founder Dax Raad wrote on X on August 11 that they had found and banned 7,013 fraudulent accounts, saving an estimated $400,000 a month. An earlier post described one operator running 8,000 accounts and reselling $480,000 of tokens a month.

Why does this loss show up first at the routing and platform layer? Because only that layer can see the same person opening accounts across many providers. A single model vendor sees only its own slice of the loss, never the whole. Seeing the whole is not yet worth money on its own; what is worth money is what you dare to do, and dare to carry, once you have seen it. That is where the first fee gets collected.

So the routing layer's first step toward being an underwriter is to backstop people cheating agents, with the agent's own errors coming later. Fraud is independently distributed and insurable. Errors are correlated and hard.

3. The seat map: seven positions, cut by who carries the irreversible liability

The map in Part Eight was cut by product type: agentic commerce, agentic payments, agentic fintech, agentic banking. Nothing wrong with the cut, but it cannot answer this piece's question. To answer who pays when it spends wrong, you need a different knife: once a sum of money starts moving, who carries the responsibility that cannot be undone.

Cut that way, the whole of agent finance falls into seven positions. I call it the seat map, because each cell is a load-bearing column in the building. Pull any one and what sits above it comes down.

What matters in this table are the two empty cells.

The sixth is the public layer for prior authorization, the place where, as I said in Part Eight, the Plaid of the agent era should grow. Today every wallet, every bank and every protocol manages authorization on its own; there is no directory everyone recognizes. The seventh explains something I ran into in both screening rounds: plenty of projects call themselves AI-native insurers or AI-native lenders, and not one could produce an underwriting entity, a regulator's registration number or a state license. The gap between what is claimed and what can be verified is why that cell is still empty.

The three failure shapes from the previous section land on this table as the second and seventh cells: who defines the loss, and who has a balance sheet to pay it. One just got its first occupant. The other is still empty.

4. Three companies, each in its seat, each stuck on someone else

The three companies in the first three cells I have worked through until I can trace a single transaction. My conclusion first, then each one.

All three are betting on the same thing: that who is responsible after an error will go from a line in a disclaimer to something people pay money for, and each wants to be the one collecting. All three are in the same situation today: each wants to be a tollbooth, and there are no cars on the road yet. The difference is in one place, and that place decides everything: getting cars onto the road is stuck on someone, and it is a different someone for each.

Ubyx sits in clearing and is stuck on the demand side.

What it is. The stablecoin chain has only three roles: issuers who mint, banks that finally redeem for the customer, and orchestrators who take the whole job off a company's hands. Ubyx wants to be the rules and the routing among the three: a bank signs its rulebook once and can accept every issuer's coin at par, settling in fiat in real time. It does not mint, does not accept, does not buy. It only passes instructions.

The model is in the white paper: issuers pre-fund at a settlement bank, a customer deposits stablecoins at a participating bank, the system routes the request to the issuer, settlement happens in fiat in real time, and the bank earns a redemption fee plus the spread. The platform fee in the white paper's example is 20 basis points, paid by the accepting bank. The founder conceived and ran the institutional version of the same thing at Citi, which is his strongest card.

The supply side has goods: fifteen public issuers including Paxos, Ripple, Agora, Transfero, Monerium, GMO Trust, BiLira, Juno, Brale and Minteo. A $10 million seed round in June 2025 led by Galaxy, with Coinbase Ventures, Founders Fund and VanEck participating; a separate Barclays investment in January 2026.

The demand side is publicly zero. In a written Q&A in February 2026 the founder said the company went live in October 2025 and was processing real transactions, but as of September 2026 no accepting institution has been named publicly and no clearing volume disclosed. The issuer count of fifteen has not changed between June 2025 and January 2026.

Where is it stuck? On banks changing their behavior and coming to sign a rulebook. Nobody can put a date on that. Its most likely death is the cold start of a network effect: its value equals the number of banks connected, and the first bank connects to a network worth nothing.

Klaimee sits in after-the-fact liability and is stuck on the supply side.

Insurance is dense with jargon. Three terms first. The carrier is the party that finally pays a claim, putting its own balance sheet behind the policy. A managing general agent, MGA, is an intermediary that works for the carrier without putting up money: it prices, underwrites, issues and handles claims on the carrier's authority, earns commission, and when a big loss hits, the carrier pays, not the MGA. An exclusion endorsement is the page on a policy that says this we do not cover. Traditional liability policies used to be silent on losses caused by AI, neither covering nor excluding; from 2026 they say in writing that they do not. That page saying we do not cover is exactly this business's market.

What Klaimee does: test an agent across eight risk dimensions, certify the ones that pass, and attach to the certificate a payout commitment backed by a carrier. A vendor takes that one document through every customer's procurement without renegotiating each time. What it sells the vendor is getting through procurement; insurance is only the support behind it. It is itself an MGA, puts up no capital, needs no license first, and renews with its carrier annually.

It raised a $5.5 million seed on July 22, 2026, two people, six months old. The founder's strongest card is having gone the full way from MGA to holding a license at SafetyWing, with the policy documents traceable.

Its supply side is one carrier, unnamed, domiciled in Bermuda. Its demand side is publicly zero: no customers, no policies, no claims.

Where is it stuck? On underwriting capacity held by one unnamed carrier, renewed once a year. The founder herself compares the structure to fundraising: the carrier is the LP and can stop committing at any time. Its most likely death is the carrier walking away. On the same road, AIUC is twelve months ahead with paper from Beazley, and the test-then-insure play has been matched by AIUC and Armilla as of September 2026. It is no longer a differentiator.

Catena sits in identity and authorization and is stuck on the regulator.

Catena's founder is Circle co-founder Sean Neville. It builds a dollar account for agents that has an identity, rules and an evidence trail: who authorized, what it may do, how limits get enforced, how actions get replayed afterward. It is also applying for an OCC national trust bank charter. A $30 million Series A closed on May 20, 2026, $48 million cumulative, with a16z in both the seed and the A.

This is the most crowded cell. Google's AP2, Stripe's ACP and Visa's TAP, three big-company protocols, press directly on it; on the startup side there are Alter, Nekuda and Payman AI besides Catena. On August 6, 2026, Sapiom, doing the same thing, raised a $35 million Series A, $50 million across two rounds, more than Catena's $48 million, and its cap table includes Anthropic and Okta Ventures. A model provider investing directly in an agent payments company is a category of shareholder Catena does not have.

Its only hard adoption number: three developer packages at roughly 400 monthly downloads each. No public pricing, no external production customers.

Where is it stuck? On a charter. What the national trust door looks like today: on December 12, 2025 the OCC granted five conditional approvals in one batch, and in the first half of 2026 added Bridge, Protego, Foris DAX and Coinbase; but the number of de novo applicants that have completed final approval and can open is, as of September 2026, one, Circle, on July 10, 2026. Wise was refused in July 2026 and Bunq in August, with detailed reasons. The door is open, but it is not automatic. Catena is queued behind Bridge and Coinbase, still pending.

Its most likely death is getting the charter and discovering the charter is not worth much. A trust charter cannot lend, so the ceiling is locked by the arithmetic of capital and return. And the two owners of the protocol layer, Bridge and Coinbase, already hold conditional approvals ahead of it. If they build the agent account with policy and evidence trail on top of their own charters, what does Catena have in hand the day its charter arrives?

Three side by side. If you can only pick one, which?

My pick is Catena. It is not the best of the three today; the only reason to pick it is that its blockage is the one with an established process, precedents and a timetable. Ubyx needs banks to change behavior; Klaimee needs the market to produce a major agent accident first. Both may happen, but neither is in the company's hands, and neither has a date. For someone who needs an answer within a finite window, a judgment that can be tested is worth more than a better story.

One step further: write the pick as conditions for entry. I cannot give a price, because Catena's post-money valuation for the A round is not public and I am not going to invent one. I can give three conditions. One, the OCC conditional approval arrives before mid-2027. Two, before then, at least one corporate legal team publicly adopts its policy and evidence trail as the default template, by name. Three, Sapiom, with model-provider money behind it, does not get a charter first. Missing any one of the three is a reason not to enter at today's price.

Where is this judgment most likely wrong? In treating the charter as a moat. A charter is admission, and Coinbase and Stripe's subsidiary are already queued ahead. Catena's moat can only grow somewhere else: corporate legal teams adopting its policy and evidence trail as the default. There is no evidence of that today.

5. Meta put the demand side on the table, and pushed liability one step further away

Meta launched Muse on September 8. I read the official newsroom post and several launch reports; here are the facts that check out.

Muse runs on a dedicated computer in the cloud, in Meta's words its own dedicated computer in the cloud, and in TechCrunch's, a dedicated, secure computer with its own browser. The model is called Muse Spark. It books restaurants, buys things, sends email and manages your calendar, and keeps running after you close the app. The free tier is 100 million tokens a week, a figure from a Threads post by Mark Zuckerberg as cited by Gizmodo on September 8; the paid tiers are Power at $20 a month and Maximum at $100, per TechCrunch.

Two details matter more than the specs.

First, the free tier requires a card. TechCrunch puts it plainly: Muse requires a payment card to get started. An agent that will check out for you has your card number from day one.

Second, payment goes through Link by Stripe. The newsroom post says Muse can checkout with Link built by Stripe, with Shop Pay to follow. TechCrunch adds that Link carries purchase protection, which may ease consumer worries. Translated: Meta did not build its own liability mechanism. It plugged the agent's payments into the existing dispute system of the card rails. In Part Six I wrote that whoever wins, Stripe is the checkout counter. Muse is one more data point, and this one is on the demand side.

On the business model, rumor has traveled faster than fact. I have seen articles saying Zuckerberg chose a transaction take rate. Checked: Motley Fool on September 11, citing CNBC, reports that Meta chief AI officer Alexandr Wang said the company is exploring taking a cut of purchases completed through the agent but has not settled on a plan. Zuckerberg himself, on the July earnings call, said only that personal agents are the foundation for the next wave of products and revenue lines. Exploring is not deciding, and Wang is not Zuckerberg. The distinction needs to be written down.

Why does Muse bear on the liability cell? Look at Walmart.

From November 2025 Walmart put roughly 200,000 products into OpenAI's Instant Checkout, so users could order inside ChatGPT. In March 2026 Walmart's head of product and design, Daniel Danker, said publicly that purchases completed inside the conversation converted at one third the rate of click-throughs to Walmart's site; his word was unsatisfying. The same month OpenAI confirmed it was phasing out Instant Checkout in favor of checkout flows run by the merchants themselves.

What does that number say? Ordinary people do not like being dropped onto a checkout page in the middle of a chat with an AI. Only one direction remains: the checkout page disappears and the agent pays in the background. That is the road Muse is on.

And that road stretches the distance to liability by one more step. A checkout page you can see is a place you can still stop before paying. A checkout you cannot see is one you are still responsible for. Single-use virtual cards and Link's purchase protection are the industry's answer: push liability back onto the chargeback system of the card rails. In Part Eight I wrote that the real power of card networks is anchored in dispute rules; that cell fed an entire category on the card rails, and now it is being reused for agents.

The problem is that card rails know only two states: authorized, or not. You said under $200 and it bought for $250, and in the network's eyes that is a perfectly normal authorized transaction. There is no third state for authorized but out of bounds. That cell, the card rails cannot fill either. It falls back onto the second and sixth seats of the map: after-the-fact liability, and the public layer for prior authorization.

6. How big this cell is: look at the limit, not the error rate

By this point a natural question: the worse agents hallucinate, the bigger the backstop business?

Half right. But the relationship is an inverted U, rising halfway and then turning down.

Hallucinate too badly and nobody hands an agent anything involving money. No deployment, no exposure, no premium; and no underwriter will touch it, because the loss is certain rather than probable. Market zero. Hallucinate lightly enough and the losses are small enough to eat; companies self-insure and do not buy. Nobody insures typos. Market near zero too. Only the middle band has a business: low enough to dare to deploy, high enough that you cannot carry it.

One layer down. Premium is roughly four things multiplied: deployment scale, exposure per transaction, incident frequency, and the share that cannot be self-retained. The hallucination rate enters only the third factor, and at the same time pushes the first one down. That is where the inverted U comes from.

The biggest lever is exposure per transaction. At the same error rate, an agent allowed to move $200 and one allowed to move $10 million are five orders of magnitude apart as a market. What decides the size of this cell is how much money the agent is allowed to touch.

Which leads to a counterintuitive corollary: better models make this market bigger, not smaller. Better models mean people dare to let them touch more money, and exposure per transaction steps up. Frequency falls but exposure rises; expected loss does not necessarily fall, and the tail definitely gets fatter. Car insurance is the same story. Cars keep getting safer, accident rates keep falling, and the car insurance market has not shrunk, because cars cost more, liability limits are higher and litigation is dearer.

So the indicator to watch is the ceiling on an agent's per-transaction authorization, with model benchmarks set aside. The default limits in the agent protocols of the card networks, the default limits in each wallet, the budgets enterprises open for agents in procurement: lift those an order of magnitude and this cell lifts an order of magnitude with them. Muse requiring a card at launch is the first consumer-side data point on that line: Meta is willing to let it touch money at all; how much per transaction, the company has not said.

The same lens works back on the three companies in section four. Take Ubyx. The white paper's $3.65 billion annual revenue pool takes as inputs a $1 trillion stablecoin float, 0.5% daily redemption and a 20 basis point platform fee. The trillion is a figure the white paper itself flags as a round number that may not reflect reality. Swap in the actual numbers for September 3, 2026: total stablecoin float about $301.7 billion, of which Tether $183.3 billion and Circle's USDC $73.6 billion, the two together about 85%, and neither is among Ubyx's fifteen public issuers. What it can reach today is the remaining 15%, about $44.8 billion. Keep the white paper's own redemption rate and fee: annual redemptions of about $81.8 billion, platform fee of about $164 million a year, 22 times short of the white paper's $3.65 billion. To make $100 million a year it would need 61% of its reachable pool. The way to read that number is not that it is small in absolute terms but that the share it requires is absurd. It explains why Ubyx in 2026 widened its target from stablecoins to tokenized money: the pool of bank deposits is far larger than the pool of stablecoins.

The last layer, and the one I think is most practical: verification comes before underwriting. Nobody can underwrite something with no loss history. This cell has to grow a layer that sells checking first, charged per call or per seat, recording errors, classifying them, computing frequencies; only once loss data has accumulated will anyone dare to run the second layer that sells payouts. The order cannot be reversed. And the moat of the first layer is the loss data it accumulates, not its checking algorithm. Klaimee's test-then-insure structure is exactly this order, but the testing half was matched as of September 2026, so its differentiator has to move one cell over, to whoever accumulates the most loss data. Stripe's one-in-six is evidence that the first layer has already started accumulating.

7. What this piece will not tell you

Before closing, some self-demolition.

First, the three failure shapes and the seven-seat map are both my frameworks, not industry consensus. They explain what I have seen these two months and the companies I screened; they may not fit other people's cases.

Second, everything on the three companies rests on public information: white papers, regulatory filings, funding announcements, and public interviews and writing by the founders. What the companies have not disclosed internally I cannot know. Funding and valuations come from public reports at different points in time, and side-by-side comparison is only for order of magnitude.

Third, Walmart's one third is one person's public statement. I have not seen a written Walmart disclosure, and I do not know the sample size or time window.

Fourth, Muse has been out a week and there is no public usage, transaction or error-rate data. Section five is entirely inference at the level of product design.

Fifth, the Ubyx arithmetic is other people's inputs plus my formula: the float figures come from a tracker's breakdown, the redemption rate and fee from the white paper's example. Change any input and the conclusion is void.

Sixth, the evidence for failure shape two is my own usage. The sample is one, not a statistic. It is enough to convince me the shape exists, not enough to estimate its frequency.

Seventh, the four-factor formula is my own decomposition. No insurer has published agent-related underwriting data on that basis. It is a thinking tool, not a pricing model.

Closing

Three sentences.

One: the endpoint of routing is daring to price the outcome, and the moment it does, it becomes an underwriter. Stripe bought OpenRouter for the ledger; the ledger's next page is the success rate; the flip side of the success rate is the payout. Payments took twenty years to walk that road. Tokens will not need twenty.

Two: of the seven seats, the two on the after-the-fact side have one first occupant and one still empty, and the three companies who sat down are each stuck on someone else. Ubyx waits for banks to change behavior, Klaimee for a carrier to renew, Catena for a charter. Only the last has a date.

Three: the size of this cell is decided by the limit, not the error rate. The number to watch is how much an agent may spend in one transaction. Better models push that number up, and the market follows.

The falsifiable version: by the end of 2027, if no token router or agent platform has publicly priced per completed task with payout terms attached, sentence one is void. By mid-2027, if Catena has not received OCC conditional approval, the ranking in section four moves down; if Ubyx names a real live accepting institution before then, the ranking flips. By mid-2028, if the default per-transaction authorization for mainstream consumer agents is still stuck below a few hundred dollars, the market-size judgment in sentence three is downgraded, and the correct explanation becomes that the trust threshold is far higher than I thought.

(Sourcing. Stripe's acquisition of OpenRouter: the Wall Street Journal reported talks at roughly $10 billion around July 23, 2026; Bloomberg reported an agreement at more than $7 billion on August 16, 2026; Stripe's newsroom announced the agreement on August 19, 2026, worded agrees to acquire, with no price disclosed; see the September 14, 2026 update to Part Seven. Stripe Radar blog, Expanding Stripe Radar to protect more of your business, May 27, 2026: the one-in-six and ElevenLabs 2,000 a day figures are from the original text. OpenCode's 7,013 fraudulent accounts and roughly $400,000 a month are from Dax Raad's X post of August 11, 2026; 8,000 accounts and $480,000 a month from an earlier post of his. The seven-seat map and its occupants are the author's compilation from two independent screening rounds in August and September 2026, the first starting from the full YC library, the second from regulatory filings and funding records; Alter's funding remains unverified; Duna's CapitalG-led Series A and Casca's three bank customers who are also investors are from public reports. Ubyx: the March 2025 white paper is the source for the seven-step flow, issuer pre-funding, the 20 basis point example fee and the $3.65 billion worked example; the fifteen-issuer list, the June 2025 $10 million seed and its investors, and the January 2026 Barclays investment are from company announcements and public reports; the founder's February 2026 written Q&A states an October 2025 launch; as of September 2026 no accepting institution or clearing volume has been named publicly; stablecoin float of $301.7 billion, Tether $183.3 billion and USDC $73.6 billion are tracker breakdowns for September 3, 2026, the 85.2% share is computed from the components, and $44.8 billion, $81.8 billion, $164 million and 61% are this piece's arithmetic on the white paper's example parameters, not forecasts. Klaimee: the July 22, 2026 $5.5 million seed is a company announcement; the MGA structure, Bermuda carrier, eight risk dimensions and performance guarantee are from its website and the founder's public writing between May and July 2026; the carrier's name is not public; AIUC's relationship with Beazley is from AIUC's public materials; the matching of test-then-insure by AIUC and Armilla is from a comparison of public product pages in September 2026. Catena: the May 20, 2026 $30 million Series A, $48 million cumulative and a16z participation are company announcements; the OCC national trust application status is from the OCC public index, still pending as of September 7, 2026; roughly 400 monthly downloads for three developer packages is from public package registry data; Sapiom's August 6, 2026 $35 million Series A, $50 million across two rounds, and Anthropic and Okta Ventures as shareholders are company announcements. OCC queue: conditional approvals for Circle, Ripple, BitGo, Paxos and Fidelity Digital Assets on December 12, 2025, Bridge on February 12, 2026, Protego on February 13, Foris DAX on February 20, Coinbase on April 2; Circle's July 10, 2026 final approval as the first de novo completion this cycle; Wise refused in July 2026 and Bunq in August; all from OCC public decisions. Meta Muse launched September 8, 2026; the dedicated cloud computer, Muse Spark, Link by Stripe checkout and Shop Pay to follow are from Meta's official newsroom post; the 100 million tokens a week free tier is from a Threads post by Mark Zuckerberg as cited by Gizmodo, September 8, 2026; Power at $20 and Maximum at $100, the card requirement on the free tier, and Link's purchase protection are from TechCrunch, September 8, 2026. Meta exploring a transaction cut is chief AI officer Alexandr Wang's statement to CNBC as cited by Motley Fool, September 11, 2026, with no plan settled; Zuckerberg's statement that personal agents are the foundation of the next wave of revenue lines is from the July 2026 earnings call. Walmart's roughly 200,000 products on OpenAI Instant Checkout from November 2025, in-conversation conversion at one third of click-out, and OpenAI phasing out Instant Checkout in March 2026 are from Search Engine Land's March 19, 2026 report of Walmart product and design head Daniel Danker's statement, with no written Walmart disclosure seen. The flight example, the three failure shapes, the seat map, the inverted U and four-factor formula, verification before underwriting, and the ranking of the three companies are this piece's frameworks and judgments, not quotations. All statements on market size, apart from the Ubyx arithmetic, are structural judgments with no absolute figure given, for the reasons set out in Part Eight.)