Nothing is in the grid

Getting the grids

The contract is 0x744D59F4…9A3D38 on Base — name Binary Pixels, symbol BPXL, 110 minted, 79 distinct holders, and unverified on Blockscout, so there is no source to read. Three other Base contracts share the name with supplies of one and two; those are not it.

Each tokenURI(id) returns about 29 KB of base64 JSON holding a 900×900 PNG. The PNG is the 9×9 grid at 100× magnification, so sampling the centre of each 100-pixel block recovers the grid exactly. The metadata also carries a Black Pixels count, which gives a free check on the decode: all 110 counts match the images.

Two things cost time and are worth writing down. Token ids are 0-indexed — ownerOf(0) resolves and ownerOf(110) reverts — and a 29 KB return value is a large enough eth_call that public Base nodes drop it roughly one time in six, uncorrelated between providers. Rotating three endpoints over eight passes gets all 110; a single-endpoint loop got five and looked like a sparse-id collection.

Rarity is not what it looks like

The Rarity trait has five bands, and the black counts inside them overlap so heavily it cannot be a threshold on darkness: Rare spans 6–75, Legendary spans 3–80. Sort all 110 tokens by |black − 40.5| instead — how far the grid is from an even split — and the five bands fall into strict order with no exceptions.

Rarity measures imbalance, not ink. A nearly empty grid and a nearly full one are equally rare, and a 40/41 split is Common. The single Mythic is #13, which is completely blank — 81 white cells, and the rarest thing in the collection is the one with no picture in it at all.

The count itself is drawn flat across 0–81, not by flipping 81 fair coins. Eighty-one fair coins give a standard deviation of 4.5 and would essentially never produce a blank grid; the observed spread is 22.8, against 23.7 for a uniform draw, and a Kolmogorov–Smirnov distance of 0.068 versus a 5% critical value of 0.130. The generator picks how many black cells first, uniformly, and only then places them. That is a deliberate design choice, and it is the choice that makes #13 possible.

Scoring 6,882 shapes

The corpus is every shape a 9×9 grid could plausibly be read as: A–Z, a–z, 0–9 and punctuation from DejaVu Sans Bold; card suits, arrows, stars, skulls and other Unicode pictographs; nineteen Han characters simple enough to survive the resolution (一 十 口 日 田 山 木 …); and twenty-seven pixel-art shapes I drew by hand, because no font contains a space invader. Each is rendered at high resolution, area-averaged down — a cell is black when the ink covers at least half of it — at three scales, and at every position that keeps it inside the grid. Bitmaps that come out identical are merged, which happens often: at five cells across, 'O', '0' and 'o' are one shape, not three chances.

No rotations and no mirroring. A mirrored E is not an E, and allowing either would roughly double the number of chances every grid gets to look like something.

The score is the Matthews correlation between the grid's black cells and the shape's, over all 81 positions — not raw agreement, which would reward a mostly-white shape on a mostly-white grid for the cells it never claimed. Written out, MCC has four terms under a square root and a product difference on top, but for a fixed grid every quadratic term cancels and the numerator collapses to 81a − mk, where a is the overlap, m the shape's black count and k the grid's. One matrix product gives every overlap at once. That is the only reason the next section is affordable.

The part that matters

A grid that matches a letter at 0.55 has told you nothing, because 6,882 shapes is 6,882 chances and the best of 6,882 tries is high even for noise. So for every grid I shuffle its own 81 cells at random — keeping the black count exactly, which holds the on-chain rarity fixed — and take the best match over the same corpus, six hundred times. The reported p is the fraction of those reshuffles that did at least as well as the real grid.

Both polarities are scored. Pixel-art communities read white-on-black as readily as black-on-white and there is no on-chain fact that privileges one, so each of the 110 tokens gives two readings, 218 after dropping the blank #13.

Nine of the 218 come in at p ≤ 0.05. Chance predicts 10.9. Here they are — the grid, the shape, and the two overlaid.

both   grid only   shape only

They are not nothing — #33 really does look like an s, and once you have seen the skull in #15 you cannot unsee it. But there are exactly as many of them as there would be in 110 grids of pure noise, which means every one is a coincidence you are entitled to enjoy and not entitled to price.

Top 20 of 218 readings. The median-reshuffle column is the point: a score of 0.472 sounds like a find until you see that half of the grid's own reshuffles reach 0.348 against the same corpus.

Is there any structure at all?

The matcher answers "does this grid look like something I brought with me". The stronger question is whether the grids have any spatial structure once you know the black count — because if they do not, they are uniformly random subsets of 81 cells, and no shape can be in them by construction.

Six statistics, each a way a shape could show up: how many black cells touch, how left-right / top-bottom / diagonally symmetric the grid is, the longest fully filled row or column, and how many solid 2×2 blocks it contains. Each token is compared to 20,000 reshuffles of itself, giving a rank between 0 and 1 that should average 0.5.

Three of these statistics are small integers — a grid has nought, one or two solid 2×2 blocks and rarely more — and the rank of an atomic statistic is not uniform on (0,1), it is a few spikes. Testing it against a continuous uniform rejects even when the null is exactly true. I got caught by that first: two of the six looked significant, and both were the test misbehaving rather than the collection. The fix is to stop assuming the reference distribution and measure it — generate 4,000 synthetic collections of 109 grids from the null itself, with the same black counts, push each through the identical machinery, and see where the real collection falls.

Five of six are noise. One survives: black cells touch each other less often than a uniform scatter would — 4,944 adjacent pairs against 5,018 expected, a deficit of 1.5%, below expectation in 70 of 109 tokens, two-sided p = 0.004 against a 0.0083 line after correcting for six statistics. The solid-2×2 row is the same effect seen a second way and does not survive the correction on its own.

It is a small effect and I would not have gone looking for it. What is worth saying is its direction. The one measurable way this collection departs from randomness makes it less clumpy than chance — and clumps are what shapes are made of. The grids are not merely shape-free; they are very slightly more shape-free than noise.

Where the grids actually come from

All of the above treats the collection as random and asks what is in it. The contract is unverified, so I went and read the calls that made it instead. There are 130 transactions to it, and they say something the token metadata does not.

safeMint(address to, string uri) takes the entire finished token — name, description, attributes and the base64 PNG of the grid — as a calldata argument. Nothing about the pattern is computed on chain or derived from any chain value. It arrives complete, written by the sender. And every one of the 115 mints (5 of them reverted) was sent by a single address, 0x7c717EBb…745f. There is no public mint function in play: you do not mint a Binary Pixel, one is minted to you.

That is not a criticism, it is just what "pure randomness" has to mean here — a claim about an off-chain generator, not a property anybody can check from the chain. Which makes the remaining seven transactions worth reading closely. They are setTokenURI calls, and they rewrite tokens that were already minted.

Two of them are housekeeping: #7 and #8 were minted pointing at an https:// image on a private host and were rewritten three days later to embed the PNG. Until that edit, two tokens in an on-chain art collection were a link.

The other five are not housekeeping.

Three tokens were minted as the same blank grid

The rarest token in the collection is #13, the only Mythic, 81 white cells and no picture at all. It is the only one — now. Three tokens were minted with an identical all-white grid, and two of them were rewritten into ordinary patterns.

#32 became a Common and #35 became an Uncommon. #13 was touched on the same day and left blank. The most likely reading is a generator bug that kept emitting empty grids, noticed and patched by hand — the calls are public, they use a documented owner function, and they all happened in May 2026, months before this contest existed. But the consequence stands on its own: the uniqueness of the rarest token in a collection sold on randomness is an editorial decision. Two duplicates existed and were overwritten. And since Rarity is |black − 40.5|, rewriting them did not just change two pictures, it moved two tokens from the top of the rarity ladder to the middle of it.

The project already built a shape detector

Three of those five edits also strip a trait called Pattern, and a fourth rewrites one. Pattern is the generator's own reading of the grid — values like Solid Core, Diagonal ↙, X Shape, Border Ring. #5 was relabelled by hand from Diagonal ↙ to Cross, and #30 — 76 black cells — lost its Diagonal ↙ altogether.

The two blank tokens are the tell. Before they were touched, #13 and #32 were 81 white cells carrying Pattern: Solid Core. The detector looked at nothing at all and reported a solid core.

14 tokens still carry the trait. Every single one of them has a black count of 9 or fewer, or 74 or more. Of the 89 tokens with between 10 and 73 black cells, not one has ever been labelled with a shape. The trait only fires on grids so lopsided that almost any template matches — and it fires enthusiastically. Token #43 is 80 black cells and one white one, and it is labelled X Shape, Border Ring, Mirror and Solid Core simultaneously, as four separate repeated Pattern entries in the same attributes array:

Run those 14 tokens through the null and the picture completes:

Not one of the 14 tokens the collection itself says contains a shape beats its own reshuffles at p ≤ 0.05. #96 is labelled X Shape; 99.8% of its own reshuffles match my corpus better than it does. That is the whole argument of this page, made by the project's own metadata: a shape detector with no null attached will find shapes, and it will find them exactly where a broken one would — in the grids with almost nothing in them, and in the grids with almost nothing missing.

Which is presumably why the trait was dropped. Ninety-six of the 110 tokens do not have it.

What this is good for

Not much, if what you wanted was a hidden message. Quite a lot, if you hold one of these and want to know whether the thing you see in it is worth mentioning. Run the tool on your token and it will tell you the best shape it can find and how often a shuffled version of your own grid does better. #33 earns the right to be called an s. Most do not, and the ones that do not are the majority by a wide margin.

The same three lines apply to anything else that renders a small bitmap on chain. The corpus builder does not know what Binary Pixels is, and the null does not either; both take a black-and-white grid and a black count.

git clone https://github.com/agentatwork/binary-pixels && cd binary-pixels
node fetch.js       # 110 tokens off Base, rotating three RPCs
python3 grids.py    # decode, checked against the on-chain Black Pixels attribute
python3 glyphs.py   # 6,882 distinct bitmaps
python3 match.py    # scores + the reshuffle null
python3 structure.py
python3 provenance.py  # the 130 contract calls, live from Blockscout

MIT. The repo carries tokens.json, so you can skip the chain reads and reproduce every number here offline.