Stostone
Shade some cells. Each region holds exactly one stone — one connected group — and stones from different regions may not touch. A number gives the size of its region's stone. Then drop the stones: each falls straight down, rigid, and they must come to rest exactly filling the bottom half. Click a cell to cycle stone → blank → empty; right-click cycles the other way. The right-hand board is the same stones, dropped.
the board
dropped
The falling rule is not a physics
Rule 5 is written as a simulation: drop the stones, check where they
land. It never has to be run. Falling moves cells straight down, so it
cannot change how many stone cells a column holds — and the bottom half
ending up full means every column ends with H/2 of them.
So every column starts with H/2 of them,
and a rule about a final configuration turns out to be a local count on
the given one. That is the col rung, and it has a strange
shape: on an empty board it settles nothing at all,
because no column is finished either way yet. Give it one cell from
anywhere else and it writes about a sixth of the grid — and taking it
out of the solver costs 350,027 assumptions across the
77 shipped boards, against 4,917 for taking out the geometric half.
The half of rule 5 that cannot open a board is 71 times the more
expensive half to do without.
The rest of rule 5 is one invariant. A cell at row r with
b blanks under it in its column has to land on row
H−1−b, so it wants to travel H−1−b−r. A stone
is rigid, so all of its cells must want to travel the same distance:
blanks-below is the same for every cell of a stone, and that common value is how far the stone falls. No stone moves, no order of falling has to be chosen, nothing is simulated. Two engines in this repository check it both ways and agree on every grid up to 4×5 and 6×2 — every one of the 220.
The count is necessary and not sufficient
The counting half looks like it might be the whole rule, and on short boards it is: on a two-row board the two are literally the same condition, and a single column packs whenever it is balanced. It stops being the whole rule at six rows. Here is the smallest witness — a 4×3 grid where every column carries 2 of 4, neither stone has a vertical hole, and the arch still hangs one row too high:
Counting the region-free grids — no regions, no numbers, just the
shadings an H×W board admits — puts numbers on the gap.
count is exactly C(H, H/2)^W; the real rule
keeps 77.8% of them at 4×2, 49.4% at 4×4 and
7.0% at 6×6. The 4×n counts are 6, 28,
134, 640, 3058, 14612, 69822, 333640, 1594282, 7618204 — not in OEIS,
but they satisfy
a(n) = 6a(n−1) − 5a(n−2) − 4a(n−3) for every term
computed.
What the middle reading buys
Between the two sits a rule you can see: the invariant applied to two
cells of one stone in one column says that
a stone never has a vertical hole. That is what the
drop rung propagates first and what the fit
rung uses to throw candidate stones away. It is vacuous below six rows
— with only two stone cells per column a stone cannot straddle a gap —
and from six rows on it starts to bite: at 6×6 it throws away 16.4
million of the 64 million balanced grids, and the real rule then throws
away 43.1 million more. As a rule of the puzzle it is worth
less than that sounds. Swap exact for convex
and 10 of the 44 shipped 6×6 boards stop having a unique answer — and
exactly the same 10 fall if you swap it for plain counting instead. At
8×8, 31 of the 33 lose uniqueness under either, with a
median of 37 answers where there had been one.
Five rungs
col is the counting half of rule 5.
clue is the numbers, the one-stone-per-region rule and the
stones-do-not-touch rule. fit lists every way a region's
stone could still be drawn and keeps what they agree on.
drop is the invariant. probe is singleton
consistency on top.