Why
The pitch is self-custody — tokens never leave the wallet, no deposit, no debt — and all of that is true. The part worth a card is what the design trades away to get it.
Two sentences from Aqua's own documentation, placed next to each other:
| Aqua's documentation says | What it implies |
|---|---|
| A swap "simply reverts" if the balance is too low at that moment | The quote is conditional, not a commitment |
| A $100,000 balance backs three positions that collectively quote $300,000 | Advertised depth can be 3× settleable depth |
Together they turn quoted depth from a commitment into an upper bound — real, but only for whoever arrives first.
That is the same shape as this catalogue's stake-concentration card: a set that looks like N independent units while the actual unit of independence is smaller, moved from validators to order books. The question is not whether Aqua is safe for the LP — it is, and the LP is the party the revert protects — but what it costs the taker, who pays gas to discover that advertised liquidity was already spent.
How it works
Mechanism, as documented
The LP grants a revocable allowance rather than depositing. Several positions are configured against that one allowance, and when a matching order arrives the SwapVM engine pulls tokens straight from the wallet and pushes back proceeds plus fees in a single atomic transaction.
Nothing is borrowed. The allowance is a permission cap, not an escrow, so the balance is checked at execution and the swap reverts if it has moved.
The gap the public documentation leaves
What the docs do not describe is what happens when two orders draw on the same balance at once. That is the hole this card aims at.
The measurement
Small enough to run locally, on an anvil fork:
- Stand up one wallet with three positions quoting a combined 3× its balance.
- Submit competing orders in the same block.
- Sweep the balance down through 90%, 50% and 10% of quoted depth.
Output:
- A curve of realized depth over advertised depth
- The revert rate
- The gas a taker burns on the failures
It is the argument stake-concentration makes with a Herfindahl index, made here with a fill rate.
Withdrawal-authority comparison
"Funds stay in the wallet and are pulled when needed" is the primitive the AA and AP2 cards already build with, so Aqua's approve-and-pull belongs beside its siblings rather than in a card of its own:
| Model | What can be revoked | What the approval leaks while live | Cost to revoke |
|---|---|---|---|
| Aqua approve-and-pull | The allowance, any time | A standing right to pull up to the cap | One transaction |
| Permit2 signature | Per-signature, time-boxed | Only what the signature scopes | Expiry, usually free |
| ERC-4337 session key | The session, by the owner | Whatever the session's policy permits | One transaction |
Aqua is the case where the approval is deliberately long-lived, which is exactly why the revert path carries so much weight.
Market microstructure: who bears adverse selection
Added 2026-08-21 (jay). The card's real subject is not liquidity accounting — it is who carries adverse-selection risk, and a prediction market is where that question gets sharpest.
An order book is two lines. Someone offers at $101, someone bids $100, and the $1 between them is the only thing that pays a maker. It pays for one specific risk. Of the people who trade against a resting quote, some are uninformed — a pension rebalancing, someone who simply needs the position today — and some know something. The informed always pick the correct side: if the price is about to rise they buy from the maker, so the maker has just sold something that went up.
maker P&L = (what the uninformed pay) − (what the informed take)
The spread is the price of that risk. That is adverse selection, and it is the first concept in market microstructure rather than a footnote to it.
One spread is really four numbers
| Measure | Meaning | Whose number |
|---|---|---|
| Quoted | What the screen advertises | Marketing |
| Effective | What the taker actually paid, after walking up the book | The taker's real cost |
| Realized | What the maker still had once the price finished moving | The maker's real revenue |
| Effective − Realized | The money that went to the informed | The adverse-selection cost |
This card's whole thesis is a statement about the first of those four.
What equity HFT did about it — cancellation
The answer was to stop standing still: quote, then cancel in milliseconds, so a faster trader cannot hit a stale price. The measurable residue is quote life in milliseconds, quote-to-trade ratios in the hundreds, and a book that looks deep and is not — phantom liquidity, the same phenomenon this card began from.
Aqua's revert is that cancellation, rebuilt on-chain
And it is worse for the taker in one specific way:
| Equity HFT | Aqua | |
|---|---|---|
| Quote disappears before you hit it | You simply did not trade | The transaction reverts |
| Cost to the taker | Zero | Gas, plus whatever the price did meanwhile |
Same defence, same side effect — but the cost of the failed attempt moved from zero to non-zero. That asymmetry, not the reverts themselves, is what makes this worth measuring.
Two corrections to how this card framed Aqua
Aqua was never trying to solve adverse selection. It solves capital problems:
| Problem | Conventional | Aqua |
|---|---|---|
| Custody risk | Deposit into a pool or exchange | Stays in the wallet; allowance only |
| Capital efficiency | Capital locked per position | One balance backs several positions |
| Deposit friction | Deposit and withdraw transactions | None |
But on adverse selection it does do one real thing, and it does it for the maker: it gives the LP an exit. An AMM LP cannot refuse — when the price moves the pool is picked off, which is LVR, and the pool has no right to say no. An Aqua LP withdraws the balance and the quotes die instantly. Maker exposure genuinely improves relative to an AMM.
The trouble is that this protection is precisely the taker's problem. The risk was not removed. It was moved from maker to taker.
Certainty — the axis that actually separates them
| AMM (LMSR, CPMM) | Aqua | |
|---|---|---|
| Can the maker walk away? | No | Yes — withdraw the balance |
| Maker loss | Unbounded (LVR), or capped at b |
Limited |
| Taker certainty | Always fills | May not fill |
An AMM always fills and the maker pays for that. Aqua may not fill and the taker pays.
Why prediction markets are the extreme case
Adverse selection is more severe there than almost anywhere, for a structural reason. Equities let people hold genuinely different opinions about a company. A prediction market resolves to 0 or 1 — there is exactly one right answer, so the informed side wins cleanly.
Left alone, a prediction-market maker goes broke. That is why LMSR exists: Hanson's scoring rule does not remove adverse selection, its parameter b caps it, declaring the operator's maximum loss in advance and buying continuous liquidity with that subsidy.
So LMSR versus CLOB is not a technology preference. It is the question of who carries the risk:
| LMSR (AMM) | CLOB (order book) | |
|---|---|---|
| Liquidity | Always present — the curve quotes | Only if someone posts |
| Adverse selection | Bounded by the operator's b |
Each maker defends with their own spread |
| A brand-new market | Tradable immediately | Empty — the cold-start problem |
| Cost | Operator subsidises up to b |
Zero to the operator |
Polymarket runs a CLOB. Verex runs LMSR quote centres inside a CLOB (packages/api/src/lmsr.ts, mm.ts) — a deliberate position between the two.
Which is why Aqua's trade is the wrong one for a prediction market. Betting demand concentrates immediately before an event — exactly the moment the market is most volatile and an LP most wants to pull its balance. A fill that does not happen at that moment is not read by a user as thin liquidity; it is read as a broken site. Prediction markets buy taker certainty, and Aqua's capital efficiency is paid for by selling exactly that.
The unverified question that outranks the measurement
Is there any penalty for an LP whose orders revert often?
- If there is not, there is no incentive to quote honestly — quote wide, pull when it gets dangerous, and that strategy dominates.
- If there is, that penalty is the design's real defence and everything else is decoration.
The public documentation does not say. This should come before the anvil experiment above, not after it — the single question tells you whether the design works faster than any fill-rate curve does.