For one stretch of this project I put the solvers down and asked a different
question: not "what score can I reach" but "what can I prove". The plan was a
sweep. Take the puzzle into every branch of mathematics that plausibly has
something to say about it (flows and cuts, extremal combinatorics, statistical
physics, algebra, CSP theory, automata, proof complexity, the complexity of
local search) and push each branch until it either yields a theorem or explains
exactly why it cannot. The sweep produced thirteen families of results. This
page is the map: a few sentences per family, with links to the full write-ups
where they exist. Four families have dedicated pages today; the others will
get theirs as their reproductions land in the repository.
One note on flavor before the list. Some of these results are laws the piece
set obeys, some are impossibility theorems about the score ladder, and some are
clean negatives: proofs that a standard tool from another field, applied
correctly, certifies nothing here. The negatives are stated with the same care
as the positives. Knowing that a door is locked, and why, is how you stop
paying rent on it.
Ring purity. The five colors that appear only on border pieces are scarcer
than they look: their 5×24=120 half-edges exactly saturate the 120
ring-facing slots of the 60-piece border, with zero slack. In every solution
each edge piece is forced to point its single non-frame color inward, and the
whole border problem collapses to finding an Eulerian circuit in a 5-vertex,
60-edge multigraph. Full write-up:
ring purity.
Frame-ring counting. This family prices what the ring law buys. Color
matching alone cuts the first border placement from 59 candidates to about 21,
and a legally built random ring still sits near 10−27 on the probability
scale: 53 orders of magnitude better than a uniform ordering (10−80), and
still astronomically short of certainty. The frame also hands the interior a
fixed bill: each frame color closes exactly 12 ring joints, and the 56
inward-facing edges carry a color demand, determined by the piece set alone,
that every solution must reproduce. A dedicated write-up will follow.
Flux invariants. Give every color a numeric weight and every piece the
vector of its east-west and south-north weight differences; summed over any
region, this telescopes to a boundary flux, and a quarter turn acts on the
vector as multiplication by i. Decomposing along the four characters of the
rotation group yields the complete lattice of linear piece-intrinsic
invariants: a color census, a Gaussian-integer flux law of full rank 22 on the
real piece set, and a checkerboard parity coupling rotation to cell color. The
flux law doubles as a sound incremental certificate that catches endgame
placement errors. Write-up to follow.
A quasi-random piece set. Every second-order statistic audited (color-pair
frequencies, adjacency matrices, the spectrum of the induced transition graph)
is indistinguishable from a random control with the same color counts; the one
deliberate signal is the known absence of rotation-duplicate pieces. On the
generation side the evidence supports exactly enforced color counts over an
otherwise uniform matching-consistent coloring, and a theorem closes the loop:
reconstructing the hidden layout from the bag of pieces is exactly as hard as
solving the puzzle. Write-up to follow.
The parity floor: 479 is impossible. In any legal placement, each color's
one-sided appearances come in pairs. A single mismatch would leave two colors
odd, so no board scores 479: the ladder jumps from 478 to 480. The minimum
nonzero defect is 2, realized by swapping near-twin pieces, and around any
solution there are at most 76 one-move boards at that defect, so near-misses
are provably sparse rather than abundant. Full write-up:
the parity defect floor.
The annealed landscape and the 470 wall. Treat boards uncorrelated with
the built-in solution as a random ensemble and count them by score: the count
is astronomical up to roughly 465 to 470 and collapses beyond. The community's
twenty-year plateau then reads as a phase boundary, not an engineering
failure. The same analysis prices the instance at about 10 to 20 mutually
near-orthogonal perfect boards, isolated needles with an empty overlap band
around them; this is exact on small planted instances and a stated conjecture
at full size. The community's best board stands at 470 on the open track and
464 on the strict five-clue track; the conventions and the full table live on
the records page. Full write-up:
the 470 wall.
The entropy area law. The puzzle has two rules: edges must match, and each
piece is used once. The measured entropy budget shows the first rule is
generous and the second carries essentially all the hardness, with piece
distinctness collapsing the count of legal blocks at a measurable scale. This
family already has its page:
entropy and the area law.
CSP width. The plain 16 by 16 grid has treewidth exactly 16, which sounds
exploitable until the global all-different constraint over all 256 cells enters
and makes every width-based tractability argument vacuous. The propagation
picture matches: plain arc consistency reduces 48 of 196 interior domains,
while matching-based global consistency reduces 191. The constraint that hurts
is the one no decomposition can cut. Write-up to follow.
Convex relaxations. The standard SDP lift and a correctly derived LP
relaxation, constructed exactly on small planted sub-instances with known
optima, certify only the trivial ceiling and miss obstructions that elementary
counting resolves immediately. At this instance's scale, convexity buys
nothing. Write-up to follow.
Algebraic certificates. In the bounded-degree Nullstellensatz over GF(2),
everything a good solver's counting propagation already does has a degree-2
certificate, and nothing else is cheap: refuting a piece-reuse defect in a
K×K window needs degree growing with the area, so no global algebraic
certificate exists at board scale. The tensor-network alternative dies by a
rank computation: the per-cell tensor has effective bond dimension near 289
with no spectral gap to truncate against. Write-up to follow.
No lossless compression of the frontier. An exact dynamic program over
scan frontiers compresses only if two distinct used-piece sets can safely
merge, and a Myhill-Nerode argument shows they never can, under any scan
order; measured on the real instance, the exact top-row frontier multiplies by
about 8.9 per column. Meet-in-the-middle splits fail for a complementary
reason: both halves draw from the same finite piece pool, so the two-way
optimum is degenerate and no sound interface signature exists short of literal
equality. Write-up to follow.
Local search is PLS-complete. For the natural family of edge-matching
instances containing this puzzle, the improvement landscape is PLS-complete
(proved for a generalized palette, with the bounded-palette refinement stated
as a conjecture), and deciding whether a specific better board is reachable by
improving moves alone is PSPACE-complete. A years-long plateau is the expected
behavior of such a landscape, not the signature of a badly tuned solver.
Write-up to follow.
The proof complexity of exhaustion. Every "no completion exists here"
subtree a backtracker closes is a treelike resolution refutation, and its cost
is bounded below by resolution width, governed by the cut around the open
region. The piece-uniqueness constraint contributes no pigeonhole hardness for
extended resolution to attack, because edge matching sparsifies the
piece-to-cell compatibility graph to near-permutation. The consequence is
sharp: clause learning and better encodings buy polynomial factors, and no
resolution-family method exhausts super-polynomially faster than what we
already run. Write-up to follow.
The point was never to lower expectations. A perfect board exists by
construction, and nothing in these thirteen families touches that fact. What
the sweep does is replace folklore with priced statements: the wall has a
mechanism, the plateau has a complexity class, and each missing shortcut has an
impossibility proof instead of a vague reputation. Every route that remains
open now comes with the bill it must pay, and that is a far better starting
point for the next attempt than a blank map.