01 / THE GEOMETRY

No external border.
Enough structure to be found.

Tiles can touch edge to edge. A 9 × 9 internal area contains a guarded, two-cell-thick L finder. Three 5 × 5 guarded anchors help recover the tile’s projection. These 156 reserved cells—including their white guards—are counted as overhead, not free space.

Each remaining cell carries one binary bit. Data is whitened with a fixed mask; shortened Reed–Solomon blocks are byte-interleaved. Tile dimensions are 24, 32, 40, 48, or 64 cells per side. No fixed minimum printed size is claimed: optics and cell contrast set the physical limit.

02 / TWO LAYERS OF RECOVERY

Bad pixels and missing tiles
are different problems.

Inside a tile: choose 0, 8, 16, 24, 32, 48, or 64 parity bytes per RS block. A block with p parity bytes corrects up to ⌊p / 2⌋ unknown erroneous bytes. That is not a percentage of damaged surface. A packet CRC-32 rejects failures before fountain decoding.

Across tiles: the first K packets are systematic source blocks. Later packet IDs deterministically select XOR combinations using a robust soliton degree distribution (c = 0.1, δ = 0.05). A GF(2) elimination decoder accumulates independent equations in any order.

You need rank K, not just K scans. Repeated or dependent packets do not add rank. More distinct repair IDs can always be generated, within the 32-bit packet-ID space. This is an LT-style prototype, not RaptorQ.

04 / THE REAL-WORLD LIMITS

Experimental means
there is work left to do.

Automatic detection is implemented, including rotation and anchor-based perspective sampling. It is not validated across physical phones, printers, light conditions or motion. Scanning a finder alone cannot recover its tile’s payload. A damaged locator can make an otherwise correctable tile unreadable.

A transfer is limited to 1,024 source blocks; the maximum encoded stream is 420 KiB at 64 × 64 with 24 parity bytes per block. The website caps the generated set at 4,096 tiles and exports pages of at most 64 tiles. Raw input and decompression are bounded at 8 MiB.

Progress is local and in memory; session export is explicit. CRC-32 and SHA-256 check integrity, not identity. The format provides neither encryption nor signed authentication. Do not scan untrusted transfers into sensitive workflows.

05 / THE COMPARISON

“More efficient than QR”
needs a fair benchmark.

Conventional QR requires a four-cell quiet zone. Removing per-tile external spacing can help a dense mosaic, but does not prove better total capacity or scanning reliability. Data Matrix already uses an L-shaped finder; it is an important baseline, not a novel idea to claim here.

Compare identical payloads, correction and erasure budgets, printed cell pitch, total printed area, and measured successful reconstruction rates. Tiny independent tiles pay significant per-tile header and locator costs. Larger tiles increase payload density but demand more resolved pixels in each camera view.