Mt. Gox handled a large majority of global bitcoin trading. In February 2014 it collapsed with about 850,000 BTC missing, worth around $473M then.
Roughly 200,000 BTC were later recovered from an old wallet. About 650,000 remain unaccounted for.
Then the part that makes this the right autopsy for a lesson on counterparty risk.
The Japanese proceedings ran through bankruptcy and then civil rehabilitation, and the trustee did not begin distributing to creditors until July 2024. More than ten years. The final deadline has been extended repeatedly and stood at 31 October 2026.
Many creditors ultimately received more in dollar terms than they lost, because distributions were made partly in bitcoin and the price had risen enormously over the decade.
That is a genuinely good outcome and it is not a reassuring one, and the distinction is the lesson. It happened because the asset appreciated during ten years of enforced holding, not because the process worked.
The exposure that mattered was never the probability of failure, which nobody could have assessed in 2013. It was the duration of the consequence. Anyone who needed those funds in 2015, or 2018, or 2021, did not have them, and no diligence performed beforehand would have shortened a decade in a foreign legal system.
F103-03 established that a custodial balance is a claim. This lesson is about what can happen to the claim, and about the one control you actually hold.
Three exposures, only one negotiable
Insolvency. The assets are not there. FTX and Mt. Gox. This is the one everybody thinks of and it is the rarest of the three.
Restriction. The assets are there and you cannot reach them. Withdrawals suspended during an incident, as in F107-03's autopsy. A jurisdiction closed to you. A compliance hold on your account. A network suspended for maintenance during exactly the hour you needed it.
Discretion. The venue makes a decision about you specifically. Your account is frozen pending review. Your withdrawal is flagged. Your asset is delisted with a deadline. Whether the insurance fund is applied to your loss.
Only the third is ever open to conversation, and only sometimes. The first two happen to you without any process you can join in advance.
Reading a proof of reserves
After 2022, most large venues publish something described as proof of reserves. It is a genuine improvement on nothing and it is routinely read as more than it is.
What it typically proves: the venue controlled certain assets at a particular moment, usually via signed messages from addresses or a cryptographic tree letting you verify your own balance is included.
What it does not prove:
Liabilities. This is the missing half and it is the half that matters. Assets without audited liabilities is not a balance sheet, it is a screenshot of one column. A venue with $10B of reserves and $12B owed is insolvent and passes an assets-only attestation.
That the assets are not borrowed. Assets can be present at a moment of attestation and belong to somebody else.
The next moment. It is a snapshot. Nothing is continuous, and the interval between snapshots is where a state changes.
That the terms let you have them. F103-03's Celsius ruling is the whole point: assets can genuinely be there, genuinely be the venue's, and genuinely not be yours.
So a proof of reserves that includes an audited liability side is meaningfully better than one that does not, and neither one is protection. It is a reason to be less worried, not a reason to hold more there.
The usual framing is "how much am I willing to lose". That is the wrong question, because total loss is rare and delay is not. Ask instead: how long can I be without this, and what does that cost?
Take $20,000 on a venue.
Framed as loss. You might reason: a large venue failing outright is unlikely, call it 2 percent over five years, so expected loss is:
0.02 x $20,000 = $400
which sounds tolerable and leads people to leave the money there.
Framed as duration. Now add restriction and discretion, which are much more common than insolvency. Say a 10 percent chance over five years of a period of a month or more without access, from a suspension, a compliance hold, or an incident like F107-03's week-long halt.
The cost of that is not a fraction of $20,000. It is whatever depended on having $20,000 available, which for most people is either nothing or everything, with very little in between.
So the useful test is a question rather than a number: if this balance were unreachable for a month starting tomorrow, what breaks?
- Nothing breaks: the amount is appropriately sized.
- Something breaks: the amount is too large regardless of the venue's quality, because no venue's quality reduces the probability of a compliance hold on your specific account to zero.
And extend the window to the autopsy's ten years to see the shape at the extreme. Nobody sizes a position expecting a decade. Several thousand people got one.
What you actually control
Almost nothing on this list is a setting.
You cannot audit a venue's books, verify its lending, know its regulatory posture, or predict a jurisdiction. Retail diligence on a large exchange is mostly reading things the exchange wrote about itself.
The one control you fully hold is how much is there.
Which makes the exposure limit the entire discipline, and it has to be a number with a schedule, in the form F103-03 asked for:
- A per-venue cap. No more than X, or no more than Y percent of holdings, at any single venue.
- A purpose test. Only what you intend to trade or spend within a defined window.
- A sweep habit. Proceeds move out on a schedule rather than when you remember, because "I will move it later" is how a trading balance becomes a holding.
- A review trigger. A balance crossing the cap, a change in the venue's terms, a change in your circumstances.
Diligence that is worth doing
Limited, and not nothing.
Read the ownership clause, per F103-03, separately for each product you use at that venue. This is the highest-value ten minutes available to you.
Check for an audited liability side alongside any reserves attestation.
Note the jurisdiction, because it determines whose insolvency law applies and how long it takes. Mt. Gox took a decade in Japan; the FTX estate moved faster in Delaware.
Notice concentration. Using one venue for everything means one compliance hold reaches all of it. Two venues is not twice the risk, it is half the correlation.
A large regulated exchange has effectively no counterparty risk.
Size and regulation genuinely reduce the probability of the first exposure, insolvency, and they should. A venue with real capital requirements and supervision is a better counterparty than one without, and pretending otherwise would be as dishonest as the reverse.
They do very little about the other two.
Restriction is often caused by regulation rather than prevented by it. A compliance hold, a jurisdictional withdrawal, a frozen account pending review: these are things a well-regulated venue does more of, not less, because it is required to.
Discretion is unchanged. A large venue still decides which incidents its insurance fund covers, which accounts get flagged, and which assets get delisted, and F106-04's autopsy is 70,000 customers of a large regulated exchange having their home addresses taken by bribed contractors.
The honest position: regulation changes the shape of the risk rather than removing it, moving weight from "the assets are gone" toward "you cannot have them right now, for reasons nobody will explain in detail." Since the worked example shows duration is the exposure people size wrong, that is not obviously the improvement it sounds like.
The response is the same either way, and it is the only one available: cap the amount.
You are exposed to three things at a venue: insolvency, restriction and discretion, and only the last is ever negotiable. A proof of reserves shows one column of a balance sheet at a moment, so without an audited liability side it cannot establish solvency, and Celsius showed that assets can genuinely be present and genuinely not be yours. Measure the exposure in time rather than money, because total loss is rare and delay is not, and Mt. Gox creditors waited more than ten years and did well only because the asset appreciated while they had no choice. The only control you fully hold is how much is there, so make it a number with a review schedule and ask what breaks if the balance is unreachable for a month starting tomorrow.