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 method is Louis Verhaard's.
The code here is a from-scratch reimplementation by
Raphaël Anjou. It is filed here, in Verhaard's
section beside the method it rebuilds, as a reconstruction: a reading of his
method, not his program. The byline stays Raphaël's because the code is his;
the section is Verhaard's because the idea is.
Verhaard's own eii is a source-less Windows binary that
will not run here. The only way
to run his approach is to rebuild it, and this is that rebuild: the same engine
that appears as verhaard on the
grid.
Verhaard's documented method: set-composition swap-annealing under the
2×2-tiling metric. Pick a roughly 180-piece interior subset, swap-anneal its
composition until the number of achievable 2×2 sub-tilings is locally maximised,
front-load the worst performers, then search the rest under that scaffold. The
engine's numeric constants were recovered byte-exact from the strings inside
eii.exe, because that binary is the only surviving record of them.
Calling this "running Verhaard" is a stretch worth naming: it is a reading of his
method, not his code. That it is the only way to run his approach at all is part
of the finding.
Single core, 120 seconds, all five official clues pinned:
438 / 480 matched edges, a fully placed board (256 of 256 pieces).
All five clues respected (a strict clue solution), 42 broken edges, at roughly
40 million nodes a second.
The run and its output board are committed in the engine's
backing directory,
and just experiments single-core-benchmark reruns it, so the 438 is checkable
rather than asserted. This is the five-clue instance, a harder puzzle than the
corner-pinned variants the leaderboard
scores, where the same engine reaches a best of 451.
It found 438 within about twelve seconds and then plateaued for the rest of the
budget: a genuine local optimum for that seed and time. It is the only one of the
three community reimplementations here that solves the real, five-clue puzzle well
rather than an easier version of it, but the three are not measuring the same
thing, so the comparison needs care.
Blackwood is measured on the
same footing, matched edges with the clues pinned, and stalls near 45.
McGavin is not: its figure
is a placementdepth (about 204 of 256 pieces reached), not a matched-edgescore,
so it cannot be read on the same axis as the 438 here, and the two numbers are not
directly comparable. What the three share is only the verdict that the constrained
puzzle is far harder than the unconstrained one; the numbers behind that verdict
are on different scales.