Measured on developer hardware, not on a deployed network. Where a target appears it is our own engineering goal, not a commitment or a forecast. Where none appears we have not set one, and a blank is more useful than a number invented to fill it.
Transparent throughput
2,161 tx/s
no target set
1,000 independent-UTXO transactions, 200 ms blocks, release build, macOS arm64. A lab burst on one machine, not a sustained network figure. Re-measured after the move to ML-DSA-44: the same fixture gave 2,425 tx/s on HAWK-512, so the migration cost about 11%, consistent with ML-DSA-44 verify being dearer per call. Sustained figures on a strictly-chained workload are far lower, around 200 tx/s, where the ceiling is block policy and admission rather than signing.
measured 2026-08-23
Transparent admission latency
p50 75 ms, p99 140 ms
no target set
The same 1,000-transaction run as the row above, under ML-DSA-44. HAWK-512 measured p50 66 ms and p99 132 ms on the identical fixture, so the migration cost about 13% at the median. Latency degrades sharply under saturation.
measured 2026-08-23
Privacy-lane batch verify
121 - 168 tx/s
target 100 tx/s · met
Eight cores, release build, M1 Pro. The consensus-path cost of verifying private transactions in a batch. The conservative endpoint is used for the percentage. This is CPU throughput and it is not the lane's ceiling; see the relay row.
measured 2026-08-12
Privacy-lane ceiling, 1 Gbit relay
40.8 tx/s
target 100 tx/s · 41% of target
Every validator ingests the whole proof before verifying it, so the real ceiling is link bytes per second divided by proof bytes: 4.1 tx/s on 100 Mbit, 408 tx/s on 10 Gbit. This is the number that binds the lane, not batch verify, and it is why proof size is the open item on the private-amounts track.
measured 2026-08-12
Client proving, 2 in / 2 out
6.2 - 6.5 s
no target set
Single-threaded, measured natively on an M1 Pro, not in the browser. In-browser wasm has not been measured and will differ; treat this as a floor for the browser case, not an estimate of it.
measured 2026-08-12
Privacy-lane proof size
2,988.7 KiB
no target set
A 2-in / 2-out transaction. The proof is not carried in the block; see the next row.
measured 2026-08-12
In-block cost per private transaction
28.6 KiB
no target set
The bundle only. Proofs live in separate storage, so roughly 143 private transactions fit a 4 MiB block and proof size does not cut block capacity.
measured 2026-08-12
Tree node hash
1.45 ms
no target set
Eight cores, release, three runs within 3%. A number-theoretic transform is structurally unavailable at this prime, so the speedup came from parallel matrix-vector work rather than a faster transform. The two parallelism axes contend rather than compose: the hash is already internally row-parallel, so fanning out across a block's inputs gives 1.7x on eight cores, not 8x. One lever remains, a larger gadget base, worth roughly 1.8x once the binding parameter is re-derived.
measured 2026-08-10
Membership proof, depth 32
1.426 MB, 1.87 s prove, 437 ms verify
no target set
Two internal tables record the verify figure as 437 ms and 427 ms. The slower number is published.
measured 2026-07-30