HN user
You're right to be skeptical. The 'Restricted Access' was an automated runtime guard meant for our enterprise internal builds that leaked into the public demo. I’ve just disabled it. You should be able to inspect the network and WASM instantiation now.
A bit more context for those interested in the implementation:
The core of PXG v3.0 treats the bitstream as a 2D topology. Traditional LZ-based methods rely on a sliding dictionary window (e.g., 128MB in Zstd), which works for local redundancy but fails to capture global spatial patterns in complex files like 3D vertex clouds or deep JSON hierarchies.
On Searchability: We achieve the ~0.04ms seek time by using an in-memory succinct data structure. Essentially, during the "Decomposition" phase, we build a bit-vector index that remains compressed. This allows us to perform Rank/Select operations directly on the tiles without decompressing the surrounding blocks.
The 13505 Grid: The grid resolution was chosen to optimize cache-locality on modern x86/ARM architectures. Each 4KB tile is designed to fit within the L1 cache during quantization, minimizing the I/O bottleneck that usually kills high-ratio compressors.
I’ve kept the demo strictly browser-side for the initial pass—I want people to see the byte-perfect reconstruction for themselves. I’m especially curious to see how the engine handles edge cases in your specific datasets.