Local signature performance reference

Observed 2026-09-12 on Apple M4 Max, 16 logical CPUs, 48 GiB memory, macOS arm64, OpenSSL 3.6.2 default provider. The full OpenSSL build output and raw host values are in openssl.txt and host.txt.

This measures conventional key generation/signing/verification through OpenSSL. It does not measure Quantus's Rust provider, a Mudra adapter, circuit execution, private proofs, blockchain throughput, confirmation latency or phone behavior.

Method

Executed three sequential trials of:

openssl speed -elapsed -seconds 1 ML-DSA-65 ML-DSA-87 SLH-DSA-SHA2-128s SLH-DSA-SHA2-128f SLH-DSA-SHA2-256s SLH-DSA-SHA2-256f

Each trial gives each algorithm one second per keygen/sign/verify phase. OpenSSL selects its benchmark message and signing path; this experiment does not sweep payload sizes, contexts, keys or deterministic versus hedged mode. No -multi worker count or CPU affinity was set. The machine was not isolated from other workspace activity and no power/thermal controls were applied.

1000 / operations_per_second gives average milliseconds per operation within one phase. The table uses the median of the three phase averages. The min/max columns range over those three averages; they are not individual-operation latency percentiles. OpenSSL's printed rates are rounded, especially for slow SLH variants with only a few signatures per phase. These are short reference measurements, not a robust tail-latency or sustained-throughput characterization.

Observed results

All 18 algorithm/trial combinations completed successfully. This verifies that the selected benchmark paths ran; it does not independently audit cryptography. The separate Quantus/OpenSSL cross-check was executed earlier the same day and establishes bounded interoperability.

Artifacts and reproduction

Raw trial-{1,2,3}.stdout.txt files contain OpenSSL's result tables; trial-{1,2,3}.stderr.txt contain phase counts and elapsed times. samples.csv preserves all 18 computed observations; summary.csv preserves the median and signing ranges.

From the Mudra repository root:

nu audit/signature-optimality/benchmarks/summarize.nu
nu audit/signature-optimality/benchmarks/run.nu

The first command regenerates the checked-in CSVs from the original raw logs. The second runs new measurements into ignored generated/ without replacing the original evidence. It requires Nushell, OpenSSL with these algorithms and macOS sysctl; adapt host metadata collection explicitly on other platforms. The saved summarizer intentionally reads the original logs; to summarize a new run, copy it into generated/ and run that copy there.

The summarizer was executed successfully against these logs. No additional production source or tests were changed for this benchmark.