Proof of reserves

Attestation and audit are different products

An attestation engagement and a financial statement audit run under different standard sets and deliver different things. The AICPA's Statements on Standards for Attestation Engagements govern attestation reports for nonissuers, and the current codification, effective April 2026, divides into AT-C Section 100 on common concepts, Section 200 on level of service and Section 300 on subject matter.² An attestation reports on a defined subject matter against stated criteria. An audit opines on whether a complete set of financial statements is fairly presented, for a period.

Proof of reserves is the first thing, sold with the vocabulary of the second. Note that the standards themselves separate level of service from subject matter, which is exactly the pair to check on any report you are handed.² Two engagements can share a subject matter and deliver completely different assurance: one produces an opinion, another produces findings against procedures the engaging party chose, with no opinion at all. A heading reading "Independent Accountant's Report" tells you neither. The scope paragraph does.

What a Merkle tree actually establishes

The construction is straightforward. Take every customer account at one instant, hash an identifier and a balance into a leaf, then hash leaves in pairs upward until a single root hash represents the entire book. Publish the root. Give each customer the sibling hashes along their path and they can recompute the root themselves, confirming their balance was in the total without seeing anyone else's. Sum-tree variants carry a running subtotal at each node, so the root also carries the claimed total liability.

One named provider states the deliverable plainly: a CPA attestation report covering reserve assets greater than or equal to customer liabilities, reporting made available to customers, and subsequent customer verification through the Merkle tree.¹ That is an honest and precise description of the engagement, and it is a far smaller claim than the word audit implies.

So what does it establish? Inclusion. Your balance was one of the leaves rolling up into the published root. What it does not establish: that the root covers every liability, that no leaf was fabricated, or that the assets on the other side of the comparison belonged to the firm on any day other than the snapshot date.

Four holes on the assets side

Borrowed coverage. If the snapshot instant is announced or negotiable, coverage can be arranged for it. Securities regulation reached for the fix decades ago: the examination date is chosen by the accountant, without notice, and is irregular from year to year.³ Almost no crypto proof of reserves works that way. The date is usually month end and usually known in advance.

Signatures that prove less than they look like. Signing a message from an address shows that somebody can sign. It does not show the address is exclusively controlled, unencumbered, or not being counted simultaneously by a second firm's report. Ruling out the second claim takes control evidence, and a signature is not control evidence.

Assets that are not really assets. Reserves carried at mark rather than at what they would fetch under stress, illiquid tokens the firm itself issued, and receivables from affiliates all land on the assets side of an inequality that then reads as satisfied.

And the snapshot itself. Twelve monthly attestations observe twelve instants in a year. Generously counting each as a full hour, that is twelve hours of observation against 8,748 hours nobody looked at, roughly a seventh of one percent of the calendar. Everything in the gap is outside the engagement by construction, and a well written report will say so.

The liabilities problem, which is the whole problem

An exchange's liabilities are not a wallet balance. They are a database maintained by the firm being examined, and the practitioner can test only what the firm produces. Omitted accounts, negative balances netted against real ones, and obligations recorded outside the customer ledger entirely all shrink the claimed liability without touching a blockchain.

Merkle verification is the intended fix, and it works only if customers actually check. If one customer in a thousand verifies their leaf, an issuer can omit accounts with a high probability of not being caught. The cryptography is sound. The sampling is not. So the design consequence is a product decision rather than an accounting one: publish the verifier, put it one click from the account page, and expect to be judged on the verification rate rather than on the root hash.

What the report covers, and what it silently excludes

Six things to locate before you read the conclusion: the subject matter, the criteria it was measured against, the as-of instant, who the responsible party is, the level of service, and the use restriction. Management asserts and the practitioner reports on the assertion. If those six are not on the page, you are holding a letter rather than an attestation.

Assume the following are absent unless the report names them: liabilities not recorded in the customer ledger, encumbrances and rehypothecation, related party balances, controls over a period rather than at an instant, going concern, and everything happening between snapshots. None of these are defects in the practitioner's work. They are the scope the engagement was bought with.

The Tether order is where the gap between marketing and reserve composition got priced. The CFTC ordered Tether and Bitfinex to pay $42.5 million in October 2021, $41 million of it against Tether, over claims the stablecoin was fully backed by US dollars.⁶ The representation was the violation. Whatever documentation existed behind it did not carry the claim being made in public.

What FTX changed, and what it did not

FTX Trading and its affiliates filed for Chapter 11 in November 2022, and the attestation market repriced itself within weeks. In December 2022 Mazars halted proof of reserves work for crypto clients, and the reason reported at the time was that such reports "only provide a brief snapshot and do not show a firm's liabilities."⁴ That is a practitioner's own account of the limitation, from a firm that had been doing the work.

Test any report against the FTX fact pattern rather than against a checklist. The CFTC alleges that from at least May 2019 through November 2022, FTX customer assets were commingled with Alameda Research's assets and used by Alameda for its own trading, investment and expenses.⁵ Forty two months. An assets-side snapshot is not built to surface an encumbrance sitting behind assets that are genuinely present, and no increase in cadence changes that.

What actually replaced the pre-FTX practice was mostly the same product, published more often. Merkle roots became normal. Liabilities scope moved very little. The distinction the market still does not price is between an issuer picking the snapshot date and an accountant picking it unannounced.³ In our view that single variable separates a control from a marketing artifact, and it is the cheapest thing on this page to fix.

What we specify when a client needs proof of reserves

Start from the liabilities side and work back. The engagement has to include a liability figure the practitioner can independently reconcile, or it is not proof of anything. That means a customer ledger open to the practitioner, a Merkle root published with every report, a verifier customers can reach without asking for it, and a published verification rate.

Then set cadence against mint velocity. If backing is confirmed daily and the contract can mint continuously, the exposure window is a full day of issuance at whatever velocity the market allows. We treat the onchain mint gate as a control rather than as assurance: a feed pausing issuance when posted backing falls below supply is a circuit breaker, and a circuit breaker is not an accountant's opinion. Neither one substitutes for the other, and the oracle behind the feed is its own dependency to design.

Then the custody dependency, which decides whether any of this is independent. An attestation over assets held by a custodian that is a related person of the issuer is an attestation of the group's own numbers. Separating custodian, transfer agent and issuer is what makes the reporting layer worth reading, and it is settled in the custody arrangement long before an attestation scope gets drafted.

The honest version of this product, in our view, is narrower and more frequent than what most issuers publish, and it states out loud what it does not cover. Harder marketing page. Much better disclosure. Whether a specific reserve arrangement satisfies any particular regulator is fact-specific and belongs with counsel; nothing here is legal advice or a recommendation to buy, sell or hold any asset.