neural-storage — self-healing vault with a verified GF(256) core
Repositories: GitHub · 🤗 HuggingFace
A content-addressed, deduplicating, self-healing storage vault. Any k of n
Reed-Solomon shards reconstruct the whole, so pieces can be lost or corrupted and
the data survives. The erasure-coding arithmetic — GF(2⁸) multiply — is provided
by neural LOG/EXP units verified bit-exact over all 65,536 (a,b) pairs,
the same N/N discipline as
neural-aarch64-units.
Honest by design: dedup removes redundant data; RS adds redundancy for resilience. Incompressible data never shrinks — you cannot beat entropy. Chunk hashes are real SHA-256, not a neural emulation.
What's verified
GF256(LOG/EXP): GF(2⁸) multiply — composed multiply 65536/65536- Reed-Solomon any-k-of-n recovery (every loss pattern)
Vault: dedup + RS shards on disk, survives deleted/corrupted shards,heal()sVaultFS+ optional WinFsp drive-letter mount
Use
pip install torch
python step1_storage.py # verified GF(256) + dedup chunk store
python step2_rs.py # Reed-Solomon recovery
python step3_vault.py # self-healing vault
python cli.py store <vault> <folder>
python cli.py export <vault> <folder> # RS-healed reconstruct
python cli.py mount <vault> X: # requires WinFsp
# image a whole drive/partition into a self-healing .pt (may need admin):
python cli.py image \\.\C: diskC.pt
python cli.py image-verify diskC.pt
python cli.py image-restore diskC.pt out.img
Weights: GF256.pt.
Create your own verified unit (template: storage/gf256.py): write the exact
golden finite function → enumerate the domain (decompose big/linear ones into
bit/byte slices, see storage/rs.py) → common.train → common.verify must be
bit-exact on 100% of inputs → compose. step1_storage.py shows the full loop.
Citation
@misc{byrne2026neuralstorage,
title = {neural-storage: Self-Healing Erasure-Coded Vault with a Verified GF(256) Core},
author = {Byrne, Dean (Quazim0t0)},
year = {2026},
howpublished = {\url{https://huggingface.co/Quazim0t0/neural-storage}}
}
Dean Byrne (Quazim0t0) · 2026