Crypto exchanges and wallets hold customer funds. A rating measures the experience and conduct we can observe — it is not, and cannot be, a guarantee that a brand is solvent. So custodial brands carry a separate solvency flag, kept deliberately apart from the number. This page sets out — in advance — what that flag is, what moves it, and what it explicitly cannot see.
The solvency flag is a lattice with three states. It can add context or change a brand’s placement, but it never alters the score itself.
The flag reads only externally observable signals, and we state what we observed, never a diagnosis. The kinds of signal that earn a watch or an alert include: independent reports of delayed or paused withdrawals across distinct sources; a brand stopping publication of a reserve or asset attestation it used to publish; a large, sustained net outflow of customer-attributable assets on-chain; and a regulator or court taking public action specifically about customer funds. A watch needs one sustained signal; an alert needs a confirmable event — a public withdrawal halt, an order touching customer funds. The exact thresholds are published configuration, set conservatively because a wrong flag is unfair to a brand.
The thresholds are calibrated against two well-documented collapses, not invented. Before Celsius paused withdrawals in June 2022, there were weeks of escalating user reports of slow withdrawals, a steep sustained on-chain outflow, and a lapse in its published asset disclosures — the kind of signal that earns a watch days to weeks before the pause, which is itself the alert-grade event. Mt. Gox in 2014 followed the same shape: mounting reports of stuck withdrawals, then a full halt. We tune the watch and alert bounds so that, replayed against records like these, the flag would have lit on the observable signal — and we publish those bounds rather than keeping them secret.
This is an evidence surface, not a solvency oracle. It reads only what is externally observable. A failure driven by concealed fraud — a balance sheet that is false with no observable tell until the run, an FTX-class failure — is outside its reach, and we do not pretend a threshold would have caught it. “No flag” is never a guarantee that a brand is solvent. Crypto custody carries risk that no public-evidence rating can fully price; we surface what we can see and are explicit about what we cannot.
For how a crypto score is computed, see /methodology/crypto. For what we do when a brand actually fails, see When we are wrong. Every flag and every change to it is recorded in our audit trail.