Every market, every cadence

How the scan works

Every price on this page is a Uniswap V3 Swap event, and its value is computed from the sqrtPriceX96 the event carries. There is no oracle in this repository and no vendor anywhere in it.

The census reads every Swap on the chain over 40,000 blocks — 1,125 pools traded in that window — then keeps the 40 busiest whose pair is an issuer equity against USDG. 82 pools were rejected, and the build fails if that number is zero: a filter that has stopped running looks exactly like a clean chain.

The equity test is anchored and bounded, not a substring. On a sibling chain a substring test admitted a token whose symbol was 9,575 characters and whose name was 34,090, and it laid out one card 25,198 pixels wide.

Three things this node does that look like data

• An unpaced burst answers 429 and keeps answering 429. The pacer is why the scan finishes; concurrency is not.
• eth_getLogs over a wide range answers log query timed out with code -32000, which is the code a revert uses. Read as an answer it is an empty range, which reads as “this market did not trade”.
• A query matching more than 10,000 logs is refused outright. That is a refusal, not a throttle, and retrying it nine times gets nine refusals; the range is halved instead. It had to be halved 5 times, and the smallest chunk that answered was 1,000 blocks.
• A browser cannot see the throttle at all. The rate limiter answers with access-control-allow-origin: *,* — it adds its own copy on top of the backend’s, and a duplicated value is a CORS violation — so Chrome discards the response and fetch rejects with the same bare TypeError an unplugged cable produces. Every live panel here therefore treats anything that is not a JSON-RPC error object as transient, and says so when the retries run out rather than showing a zero.

Historical state is not available on this node either — an eth_call more than a few thousand blocks back fails with metadata is not found — so a price series cannot be sampled by asking a pool what it cost. It has to be rebuilt from logs, which is why a path here is its swaps rather than a grid.

Does the real chain obey the theory?

Partly, and the site says which part. The direction is not in question: the identity makes it a theorem that looking more often can never charge less, and across 240 cadence pairs on real markets that holds, with 2 pairs dipping by at most 9.81 parts in ten thousand.

Those dips are not noise and they are not an error. The identity is exact in price space; the vault pays its fee by dilution, which is multiplicative, while the give-up bookkeeping is additive, and the two differ at second order — of order f times the gross move. Each dip is held against its own market’s bound rather than against a tolerance somebody chose: the worst is 9.81 parts in ten thousand against a predicted 30.58.

The shape is a different question. Fitting each market's ladder to the predicted 1 − β/(√(2/π)√n) curve gives a median R² of 0.385 across 35 markets, with 14% above 0.9. Real prices are not Brownian motion — a few hours of a tokenised equity is long flat stretches and discrete jumps — and nothing on this site depends on that fit holding. It is published because a page that only shows the fits that worked is not showing a measurement.