How the data is collected.

Everything on the bench comes from the collector at workers/latency-bench in the route99 tree. This page is generated from that code: every number, address and provider row below is read out of the collector's source, and the test suite fails if this page and the code ever disagree. What you read here is what runs.

1 · The edge series

A Cloudflare Worker cron fires every 30 min. Each firing serves one tick group of providers (the panel is split so a run stays inside the platform's per-invocation subrequest budget): each configured group is probed once per rotation, one timed quote request per (provider × chain). Before the timed probes, each provider's hosts get one untimed warm-up request, so a sample measures the service rather than TLS setup. The worker's own location is not pinned; every run records the Cloudflare colo it measured from, and requests carry Mozilla/5.0 (compatible; route99-latency-bench/1.0; +https://route99.xyz) with Referer: https://route99.xyz.

• Timing. ms is request-out to last body byte in. Timeout: 10 s.
• Validity. A sample counts as latency only if the response is a real quote: HTTP 2xx, parseable, positive amountOut at the documented path. Anything else is stored with its ms and an err code — a fast 429 is an error, never a fast quote.
• Politeness. Sequential per-provider requests are spaced by that provider's gap (default 300 ms; see the panel table) sized to documented rate limits.
• Provider-side location. Where a response names its serving site (CloudFront, Fastly, Fly markers), it is stored per sample as via. Cloudflare strips its own cf-ray from Worker subrequests, so CF-fronted providers stay unattributed.

2 · The canonical probe

One trade per chain, wrapped native → the chain's dominant USD stable, ~$1–2k notional — big enough for real routing, small enough to be polite. The taker 0x1111111111111111111111111111111111111111 never holds or receives anything: quotes are priced and simulated, never executed.

HyperEVM buys USD₺0 — the chain has no canonical USDC. A measured swap-mix draft (workers/latency-bench/mix.draft.json) exists to widen this matrix by pair class and size once per-swap distributions land.

3 · The panel

Providers whose secrets are not set are recorded as skipped in every run of their tick — a stated absence, never a silent hole or a fake failure. No public quote API at all: Barter. The rotation covers only ticks that have a configured provider, so setting a secret activates its tick with no deploy.

4 · Simulation (re-validation)

Where a quote carries executable calldata, the sample is re-validated through Nordstern's swap-simulation API: an eth_call with state overrides plants a proxy at the taker, funds and approves it synthetically, executes the exact quoted calldata against the latest block, and reports success, the real output delta and gas. The sim call is never timed into ms. Stored per sample: the quoted amount q and sim (ok / revert with reason / error).

• Attribution rule. A revert (ok:0) is a fact about the provider's quote. A simulator failure (sim.err) is a fact about the simulator and is never counted against the provider. 502s retry from a per-run budget of 2.
• Stated caveat, and its audit. The primary simulator is operated by Nordstern, itself a panel member — the same shape as Enso Shield measuring its competitors on route99's other series. To convert that conflict into mutual auditing, every simulated Base sample is also simulated through a second independent engine, Enso Quoter (run by another panel member), with both simulators fired at the same instant — sequential firing would measure block drift, not instruments. Wei-exact agreement stores as sim2:{agree:1}; anything else keeps the second verdict for examination. sim2 is an instrument audit and never counts against the provider; a systematic discrepancy now requires two competitors' infrastructures to agree.
• HyperEVM caveat. Routers that read HyperEVM's Bbo oracle precompile false-revert in simulation at ~full gas (the environment lacks that precompile). A full-gas revert on chain 999 is a simulator fact; read the reason field.

5 · The twelve datacenters (geo sweep)

Every fourth edge slot is borrowed for a sweep from twelve deliberately picked AWS datacenters, each running our own probe endpoint — one Lambda function deployed identically to every region (workers/latency-bench/lambda/), gated by a shared token. DNS resolves locally at each endpoint, which is how a real integrator in that region experiences a DNS-geo-routed provider. The sweep is transport only (dns / tcp / tls / ttfb per region, cumulative ms from request start, one Base quote per provider; each invocation is a cold-ish environment, so dns includes the resolver's cold path — consistently across regions) and covers keyless providers only: not for trust — the fleet is ours — but because per-key rate limits make twelve concurrent same-key requests a violation, where twelve IPs sending one keyless request each is not. Currently that panel is Nordstern, KyberSwap, LiFi, OpenOcean, SushiSwap, Velora. Unlike the Worker vantage, Lambda egress keeps cf-ray intact, so Cloudflare-fronted providers are attributable per cell here.

6 · Storage, health, and honesty rules

• Every run and sweep appends to day-keyed KV records (UTC): run at + tookMs bound every sample in time; raw days are served verbatim by the API the charts read (/api/latency), so the site can never show what storage does not hold.
• Health is written on every attempt, success or failure, separately per series — a stale timestamp (cron dead) is distinguishable from a recent failure.
• Aggregation discipline: medians and p95 are computed over valid samples only; error counts travel beside them; n travels with every figure so thin data reads as thin.
• Known limits are displayed, not hidden: providers whose WAF blocks a vantage render as the error they return; a datacenter that cannot be reached marks every provider's cell with its reason, never a silent hole; pre-fleet substitute cells keep their historical mark.