What we measure — and what we refuse to

• TVL rewards size, not safety. Value locked is the market voting with its money, and it is the best open proxy we have for adoption — but big protocols get exploited too, and a new one can be excellent. We score it on a log scale so it dominates less than raw dollars would, and never let it stand alone.
• An audit is not a clean bill of health. We count published audits with links, not their quality. Audited protocols still get drained. The score rewards having been looked at and letting you read the findings — not a guarantee the code is safe.
• Age is a floor, not the truth. For on-chain protocols we use the DefiLlama listing date, which can be later than the real launch, so it understates age for some old names. For off-chain services we use documented founding years. Either way it is a lower bound.
• Latency is measured from one place. One location, a handful of samples, slowest discarded. Real, and our own, but not global — a service can be quick for us and slow for you. Scored only within a category, against peers timed the same way; where the surface differs it drops to context.
• A missing status page is scored, not excused. If a service publishes no status feed we score STATUS low. That is a judgment: a service you cannot audit from outside is worse than one you can, even if it happens to be reliable. A page we simply failed to find would read as “publishes nothing”, so a blocked or unreadable one scores n/a instead of a mark against them.
• Redistributing n/a rewards opacity, so we cap it. Spreading an unmeasured criterion’s weight across the rest is the fair rule, but it has a sharp edge: points are docked on the criteria that need the most data, so the less we know about a service the less there is to hold against it. Left alone that floats the least-measured service to the top. A place in a ranking therefore needs at least a quarter of the category’s weight measured; below that a service is listed, with its numbers, outside the order.
• A missing status page may be our blind spot. We read three status platforms. A service on a fourth, or one whose bot filter turns our probe away, is not marked down for it — that case scores n/a and says so. But a page we simply failed to find at the conventional host would read here as “publishes nothing”, and that would be our error wearing their name.
• Coverage is thin and every card admits it. The heaviest criterion is n/a everywhere, so no verdict here rests on more than about two thirds of its category’s weight, and many rest on a third. The percentage next to each score is the honest size of the claim.

Every number on this site is collected by scripts/collect-health.ts, carries the timestamp it was taken at, and links to the source it came from. Nothing here is typed in by hand, which is exactly why nothing here can be quietly adjusted to suit anyone. If a figure looks wrong, open its evidence link — it goes to the provider’s own page, not to ours.