We publish the tail, not just the headline
Every competitor on this chain advertises a p50 and hides the p99. A published tail is a claim only a confident operator makes, so it is one we intend to make.
These are targets, not measurements
The fleet is not yet serving production traffic, so there is nothing honest to plot. This page states what we commit to and becomes a live readout the day the probes have data — the layout stays the same, the labels change from “target” to “measured”. We would rather say that than publish a synthetic uptime chart.
The numbers we are held to
Uptime
Credited against the SLA
99.99%
Matches the best published figure on this chain
RPC latency p50
From the nearest region
< 8 ms
Goldsky publishes sub-10 ms edge
RPC latency p99
Published continuously
< 50 ms
No competitor on this chain publishes a p99 at all
Indexing latency
Chain tip to queryable
< 25 ms
Ormi claims sub-30 ms
Stream delivery p99
gRPC, end to end
< 100 ms
Bitquery publishes ~320 ms WebSocket
Webhook delivery p99
With replayable log
< 1 s
Nobody publishes this
What this page will show
Not a green tick and a percentage. The things an operator actually needs to size their own risk.
Per-region, per-interface latency
p50, p95 and p99 for JSON-RPC, gRPC, REST and GraphQL, from independent synthetic probes rather than our own infrastructure.
Historical incident log
Every incident with a written post-mortem, kept permanently. No quiet deletions once a quarter closes.
Reorg depth and frequency
Arbitrum Nitro reorgs are rare but not impossible. We publish how often, and how deep, so you can size your own confirmation policy.
Sequenced → included delta
The measured gap between a transaction appearing on the sequencer feed and landing in a published block. This is our latency edge, in the open.
Invariant suite results
All 9 correctness assertions, run continuously against production, pass or fail.
Publishing incident post-mortems and reorg statistics is not generosity — in a market where the officially recommended provider quietly lacks tracing, being the only participant whose numbers can be checked is the fastest route to being the one people trust.
Hold us to it.
The invariant suite is open source and runs against production. Point it at us, or at anyone else implementing the same endpoints.