Benchmarks

Generation and rating times for every genre, measured on one machine. Times are comparable within a run. Operation counts are deterministic and the same on every machine.

Runs

Genres	Code	Date
fillomino, hashi, hitori, kakurasu	pencil-puzzles 0.2.0	2026-11-06
kakuro, nurikabe	pencil-puzzles 0.3.0	2026-11-08
sudoku	pencil-puzzles after 0.3.1	2026-11-09

All runs used the same setup:

• Machine: Intel Core i7-8700K, Windows 11 Pro, otherwise idle

• Rust: rustc 1.99.1, release profile

• WebAssembly: Node.js 24.17.0

Native

cargo run --release -p pencil-puzzles --example benchmark -- <genre> 20

Each genre’s sizes are listed in benchmark.rs. For each size and level, the benchmark generates seeds 0-19 with the command-line tool’s limits (10,000 attempts and 100,000,000 operations), then rates every corpus puzzle. <genre>.csv has one row per size and level:

Column	Meaning
shape, requested	Size and requested level (any for none)
success	Seeds that produced a puzzle
median_ms, p95_ms	Generation time over all seeds, failures as infinite
mean_attempts	Attempts per puzzle produced
mean_operations	Operations per seed, including failures
easy … expert	Levels of the puzzles produced

The rows after corpus,level,rating_ms give the time to rate each corpus puzzle.

WebAssembly

node web/build.mjs
node web/benchmark.mjs <genre> <default size> 20

This times the web app’s engine with its limits (10,000 attempts and 100,000,000 operations). Each genre page gives the WebAssembly median divided by the native median, as a geometric mean over the levels where both found all 20 puzzles and the native median is at least 10 ms.