Accepted sends per second, higher is better. Same machine, same method, fresh server per level.
Milliseconds per send, log scale, lower is better.
Milliseconds, log scale. Dot is the median; the bar spans p50 to p99.
Lower is better on every panel.
Each change was measured, kept only if it won. Throughput at 16 concurrent clients; CPU per message.
Measured from a residential connection: Vercel edge (Cleveland) → Go function (iad1) → Neon Postgres.
Attacker compute needed to keep one stranger out of a recipient's request queue, versus what that stranger pays once. Invited and trusted senders skip the queue entirely (0 bits).
Milliseconds to solve by stamp size in bits, log scale. The relay verifies any stamp in .
Bytes to prove one entry is on the ledger, log–log.