Module mask 

Client-side keyed obfuscation of sparse hypervectors.

A VectorMask is a secret permutation of the coordinate space, derived from a passphrase with Argon2id. Applying it before vectors leave the client lets a remote store index and rank them without learning which coordinates were originally active.

This is obfuscation, not encryption, and it is not homomorphic encryption. It is deterministic and preserves every pairwise similarity by construction, so a store holding masked vectors still learns the full similarity structure of the collection, equality of vectors, and coordinate activation frequencies. Each known (plain, masked) pair reveals part of the permutation. Use AES-GCM storage encryption for confidentiality at rest.

Set algebra (bind, bundling, similarity) commutes with the mask. The cyclic permute used for ordered binding does not, so sequence operations must be composed before masking.

Structs

VectorMask

A secret permutation of [0, dim).

Constants

MAX_MASK_DIMENSIONS

Largest coordinate space a mask may be derived for (bounds the allocation).

MIN_SALT_LEN

Minimum salt length accepted by VectorMask::derive.