Benchmark methodology

How every performance figure in reports/ is produced, so it can be checked. Master Prompt 19 §6.

1. Hardware record with every number. CPU model and count, memory, disk, kernel, toolchain (reports/12-baseline.md is the template).
2. Profile stated. ci (dependencies optimized, node crate not) or optimized (release, and any environment overrides such as thin LTO written out). A figure without its profile is not comparable.
3. The exact command, and its raw output. Numbers are copied from the output, never retyped from memory (Standing Order 1).
4. What the benchmark cannot see is printed by the benchmark. Examples: apply_pipeline prints “NOT MEASURED” where /proc is missing; certificate prints “NOT BUILT” for options (b) and (c); the DA dispersal comparison prints “MODEL”.
5. Targets are not tuning goals (Standing Order 2). A result below a brief’s number is reported as below it.
6. Simulations say so ([SIM] in their output) and state their model: tick length, cost per operation, what stands in for what.
7. Comparisons on one machine only. The regression gate (scripts/bench_gate.py) compares a run only with the same runner’s history, and it is not live until a fixed-spec runner exists.

Reproducing

Figure	Command
Node verify/apply, allocations, write amplification	cargo bench -p custom-l1-node --bench apply_pipeline
Parallel execution speedup	cargo bench -p maya-parallel-exec --bench speedup
Signature and certificate costs	cargo bench -p maya-crypto-pq --bench certificate
DA encoding	cargo bench -p maya-da --bench da_bench
Handshake reject/accept	cargo bench -p maya-dos-guard --bench handshake_gate
RPC throughput	maya2c-rpc-load --targets … --threads 16 --secs 10
100M-account state sync	cargo run --release -p maya-state-sync --bin state-sync-sim