The shared engines under Raphaël Anjou's experiments. The named experiments are studies that run on these; this is the apparatus they share. The CSP presets and the Verhaard reimplementation are documented in full, alongside the reference engine, the JIT backtracker, a field guide to speed, and the scaling-ladder harness; the constructive engines are not yet published.
The experiments are studies: each one asks a question and reports the board it reached. But most of them are not built from nothing. They run on a small set of shared engines, and the same engine turns up under experiment after experiment. This page documents that apparatus once, so the studies can point here instead of re-explaining the machine each time. The reference engine that powers this site's live demos is a different machine and has its own page; it appears below in its apparatus role.
A single constraint-propagation engine with interchangeable orderings and propagators: arc-consistency, colour-graph matching, several fill orders. Faithful reimplementations of known community techniques, run under a dozen presets on the ten variants so the exact cost of each knob is on the leaderboard rather than argued about.
A from-scratch reimplementation of Louis Verhaard's eii method, since his own binary ships no source and will not run here. Set-composition swap-annealing under the 2×2-tiling metric; on the real five-clue puzzle it reaches 438 of 480, single core.
The wiki's verification and demo machine, distinct from the shared search engines: the Rust-to-WebAssembly backtracker that runs every live demo and checks every number quoted on these pages, ported four times and parity-tested byte-for-byte.
A safe, portable Rust depth-first backtracker specialised at runtime by emitting and compiling per-puzzle Rust, taken from 43 to 123 million search-nodes per second on one core and measured against Peter McGavin's hand-tuned C on the same machine.
A field guide to speed-first solvers: what raw throughput buys, the three different things people mean by "fast", and why the fastest engine ever built still cannot solve the puzzle.
The harness that runs any solver unchanged on planted, fully solvable boards at N = 8, 10, 12, 14, so a method's collapse size is measured before weeks are spent on the real 16×16.
The two constructive engines these studies also lean on, the beam producer and the ALNS polish stage, do not yet have their own write-ups here; those pages are held back for now, and neither is on the benchmark leaderboard. Where an experiment steers one of them, its own page says what that engine does at the level the study needs, so nothing below depends on reading a page that is not here yet. The two engines documented and measured in full are the CSP presets and the Verhaard reimplementation.
A study and its engine answer different questions. The engine answers how the search works, the data structures, the fill order, the propagators, and it is reused unchanged across many studies. The study answers what happens when you point that engine at a particular idea: a learned prior, a fixed border, a break mask. Keeping them apart means a reader who wants to understand PRIOR's prior does not have to re-read how a beam works, and a reader who wants the engine gets it in one place, current and complete.
Where these engines race head to head under a fixed budget, they appear on the single-core benchmark. The general theory behind each method lives in the build-a-solver section; these pages are the specific engines as built and measured here.