Case note · Technology disputes · AI risk
Quoine v B2C2 [2020] SGCA(I) 2: algorithms, mistake and reversed trades
- Case
- Quoine Pte Ltd v B2C2 Ltd [2020] SGCA(I) 02
- Court
- Court of Appeal of Singapore (on appeal from the Singapore International Commercial Court)
- Judges
- Sundaresh Menon CJ (for the majority), Andrew Phang Boon Leong JA, Judith Prakash JA and Robert French IJ; Lord Mance IJ dissenting
- Decided
- 24 February 2020
- First instance
- B2C2 Ltd v Quoine Pte Ltd [2019] SGHC(I) 03, [2019] 4 SLR 17 (Simon Thorley IJ)
- Judgment
- Read on eLitigation
- A crypto exchange could not reverse trades made by a market maker's algorithm at about 250 times the market rate. Nothing in its terms allowed it, and its own terms said executed orders were irreversible.
- Where deterministic algorithms make contracts, knowledge of a mistake is judged by the programmer's state of mind, from programming up to the moment the contract is formed.
- The programmer didn't know of any mistake, so the defence of unilateral mistake failed at common law and in equity.
- The trust claim failed. The court left open whether crypto is property, though it said there was much to commend that view.
- Lord Mance's dissent would have asked what a reasonable trader would have understood. It is the starting point for arguments about AI agents.
The facts
Quoine operated a cryptocurrency exchange, QUOINExchange, and was also the main market maker on it, using its own "Quoter Program". B2C2, a UK-based electronic market maker, traded on the exchange through its own algorithmic trading software.
On 13 April 2017 Quoine changed the passwords to critical operating systems but didn't make the matching changes to its Quoter Program. The program could no longer pull external market prices, and the order book for the ETH/BTC pair thinned out over the following days ([27]–[28]).
B2C2's software had a "deep price" fallback of 10 BTC for 1 ETH, designed to keep its strategy running when the order book offered little to work with ([18], [29]). On 19 April 2017 that fallback was triggered. Thirteen trades were executed at 9.99999 and 10 BTC per ETH, roughly 250 times the going rate of about 0.04 BTC ([2]). The counterparties were margin traders whose positions the platform had force-closed. B2C2 sold 309.2518 ETH and received 3,092.517116 BTC ([2]).
On 20 April 2017 Quoine cancelled the trades and reversed the settlements ([30]). B2C2 sued for breach of contract and breach of trust. The SICC found for B2C2 on both. Quoine appealed.
The issues
- Did Quoine's terms allow it to reverse the trades, expressly or by implication?
- Were the trading contracts void or voidable for unilateral mistake, at common law or in equity, and whose knowledge counts when algorithms do the contracting?
- Did Quoine hold the crypto in B2C2's account on trust, so that the reversal was a breach of trust?
The decision
1. No right to reverse: breach of contract
The trading contracts were made between B2C2 and the counterparties, with Quoine as the platform operator ([50]–[51], [55]). Quoine's terms said that once an order was filled the action was irreversible. The majority held that this bound Quoine as well as its users ([73]–[74]), and that the implied terms Quoine proposed were inconsistent with it ([71]–[72]).
Quoine also relied on a clause about trades at "aberrant" values. That clause wasn't incorporated, because users hadn't been given notice of it ([66]), and wouldn't have applied on the facts anyway ([68]–[69]). The appeal on breach of contract was dismissed.
2. Unilateral mistake: the programmer's knowledge
At common law, a unilateral mistake avoids a contract only if it concerns a term of the contract and the other party actually knew of it ([82]). The hard question was how to apply a knowledge test when no human was involved in the trades themselves.
The majority's answer was that, for deterministic algorithms, knowledge must be assessed "by reference to the state of mind of the programmers of the algorithms at the time of the programming" ([97]). The inquiry "cannot be directed at the parties themselves" ([98]). The relevant period runs "from the point of programming up to the point that the relevant contract is formed" ([99]). The majority rejected an approach based on a hypothetical meeting between the parties at the time of the trades ([97], [104]).
B2C2's programmer had written the deep price to keep the software running, not to exploit a malfunction elsewhere. He had no actual knowledge of any mistake, and that finding was upheld.
3. Unilateral mistake in equity
Equity can intervene on constructive knowledge, but it also requires unconscionable conduct by the non-mistaken party. The majority held that both elements "simply did not exist" here ([92]), and left open whether equity can reach mistakes that aren't about the terms of the contract.
4. No trust over the Bitcoin
Quoine's appeal on the trust claim succeeded. Even assuming Bitcoin can be held on trust, there was no certainty of intention to create one ([144]). Segregating customer assets is necessary but not sufficient ([145]). Quoine's cold wallet wasn't segregated per customer, account balances didn't have to match the wallet, and the arrangement looked more like a bank deposit ([146]–[147]). Quoine's risk disclosure statement also warned that customers could lose assets if it became insolvent, which pointed away from a trust ([148]).
On whether cryptocurrency is property, the majority said there "may be much to commend the view" that it is, but that there are difficult questions about what type of property it is, and left the point open ([144]).
5. Lord Mance's dissent
Lord Mance would have allowed the appeal on the contract claim. In his view the test should ask what a reasonable trader would have understood, knowing the circumstances as they actually occurred ([194], [198]–[200]). No seller would knowingly have agreed to those prices ([182]). He would have granted relief in equity and held the contracts voidable, so Quoine could reverse them ([183], [195]). As he put it, "Computers are outworkers, not overlords" ([193]), and the law must adapt to "the new world of algorithmic programmes and artificial intelligence" ([193]).
Why it matters
- The terms decide most cases. An exchange or platform that wants to unwind trades after a malfunction needs a clear, properly incorporated right to do so. A clause users were never shown won't help.
- An irreversibility clause cuts both ways. Quoine drafted it to protect itself, and it ended up binding Quoine.
- Mistake is a narrow escape route. For deterministic code, the question is what the programmer knew, from coding up to the trade. Keep design records that show why fallback logic exists.
- AI agents raise the next question. The majority's test was framed around deterministic algorithms, whose behaviour traces back to a programmer's choices. Agents built on large language models don't work that way, and Lord Mance's reasonable trader approach may get a second look. I discuss this in who pays when an AI trading agent goes wrong.
- Segregation alone doesn't create a trust. Customers who want proprietary protection on an exchange's insolvency need terms that show an intention to hold assets on trust, not just separate wallets.
Quoine sits alongside CLM v CLN as one of the starting points for digital asset disputes in Singapore. For how exchange terms are tested against users more generally, see "You clicked I accept".
Frequently asked questions
Whose knowledge counts when an algorithm makes a contract?
Can an exchange reverse trades made at an obviously wrong price?
Did the court decide that crypto is property?
Facing a dispute over automated trades or a platform's decision to unwind them? Get in touch.
This case note is general information and is not legal advice. It summarises a published judgment; read the full judgment for the court's reasoning.
