Public benchmark
When one provider says a wallet holds four million and another says three point eight, somebody is wrong. This page asks five questions the chain can settle, puts them to every provider nightly, and keeps score. The methodology is published in full, the sample counts travel with every rate, and nothing here is behind a paywall.
Read this before the numbers
etherscan (ethereum, base, arbitrum, optimism, polygon) and helius (solana) are the chain readers everyone else is measured against. These providers are the chain readers this benchmark scores everyone else against. Measuring them against themselves would return 100% and mean nothing, so they are excluded on the chains where they are the instrument.
kaito sells attention and mindshare data keyed on social handles, not addresses. None of the five questions below are ones it offers an answer to, so it is listed with no score rather than quietly dropped — a board that omits a provider it named has a thumb on the scale.
Cluster membership is scored as agreement, not accuracy. Nobody outside an exchange knows which addresses it controls, so there is no ground truth — only our own heuristic’s second opinion. Agreement is reported separately and never enters the headline number.
Turning on-chain movement into a dollar figure needs a licensed price series at each block, which we do not have. A provider that answers in dollars is recorded as having answered — it counts toward coverage — and then left unscored, because grading it against a price series we invented would publish our error as theirs.
Every rate on this page is null until it has observations, and renders as a dash. A provider we could not reach is never shown as a provider that was wrong. Where a cell is blank, the reason is printed in it.
The board
Nothing measured yet
No credentialed benchmark run has been recorded yet, so there is nothing to rank.
This is the honest state of a board on the day it ships: the methodology, the query classes, and the sample floors are published; the measurements are not, because they do not exist.
The nightly job begins producing scores once a provider API credential is configured. Until then this page shows what will be measured and how, and no numbers.
An empty board is the honest state before the first credentialed run. The alternative — plausible-looking numbers with a small grey label — is the screenshot that ends up somewhere without the label.
engine unreachable: GHOSTCOPY_API_URL unset
The questions
Is this address a CEX, an MEV bot, a whale, a fund, or a contract?
Exact match on the canonical label after vocabulary normalization. Ground truth comes from a published address registry, or from bytecode when a label asserts the address is a contract.
Ranked at 20+ settled samples
How much of this token does this wallet hold?
Balance in raw base units, correct within 0.5% of the chain reading taken in the same run. A zero chain balance requires an exact zero from the provider.
Ranked at 30+ settled samples
How much did this wallet move in the last 30 days?
Native-token volume in base units, or transaction count, over a 30-day window, correct within 2%. USD-denominated answers are not scored: this deployment has no licensed price source, and scoring them would publish our pricing error as the provider's.
Ranked at 30+ settled samples
When did this address first appear on chain?
Within 1 day of the earliest transaction on the sampled chain. An answer earlier than ours is treated as a possible cross-chain difference and left unscored rather than counted wrong.
Ranked at 30+ settled samples
Do these two addresses belong to the same entity?
Agreement with our own published clustering heuristic (shared first funder, excluding exchange funders). There is no ground truth for entity attribution, so this class measures divergence between two methods and is excluded from the headline accuracy number.
Ranked at 20+ settled samples
How to check us
The methodology is published in full, including the parts that make the board weaker. Competitors should be able to reproduce it; that is the point of publishing a number about somebody else.
/docs/methodology/provider-reliability-v1.md — scoring rules, ground-truth sources, and every known limitation.If a number here is wrong, it is a bug in this benchmark and not in the provider that disagreed with it. The registry and the scoring rules take pull requests.