A proof's cost is dominated by the backtest run itself; canonicalization and the blake3 hash are a thin, near-constant tail on top. The benchmarks here measure prove end-to-end (backtest + canonicalize + hash) and canonicalize in isolation, so you can see how the determinism layer compares to the work it wraps.

The proof-bench crate (criterion) covers:

• prove across a matrix of candle count (200, 1 000, 5 000 bars) × indicator count (specs referencing ~2 and ~10 indicators). Throughput is reported in bars per second.
• canonicalize on a small and a large BacktestReport, both taken from genuine proofs so the JSON shapes match production output exactly.

Run against fixed, in-process synthetic universes so the numbers are reproducible and contain no I/O variance:

cargo bench -p proof-bench

Measured with cargo bench -p proof-bench (criterion) on a Windows x86-64 laptop. Figures are the median estimate; treat them as orders of magnitude, not guarantees — they vary with CPU, toolchain and the linked wickra-backtest version.

The takeaway: per-bar throughput stays roughly constant as history grows (~150–177 K bars/s), so prove cost scales linearly with candle count and only mildly with the number of indicators a spec references. Canonicalization of even a large report is a few milliseconds — a small fraction of the prove that produced it — so the determinism guarantee costs little over the backtest it commits to. The nightly bench.yml workflow reruns this on a clean Linux runner for tracking over time.

These figures bound the prove / canonicalize work only. End-to-end time in a real run also depends on loading the candle series from disk, which these in-process benchmarks do not capture. verify recomputes prove, so its cost tracks the prove figures above plus a second canonicalize.