Rédiger un PREreview

We present RUDRA-512, a 512-bit cryptographic hash function that introduces a novelpreprocessing pipeline fusing Byte-Pair Encoding (BPE) tokenization with a multi-stageARX (Add-Rotate-XOR) compression function. Unlike all standard hash functions in theSHA-2, SHA-3, and BLAKE families, RUDRA-512 operates on a token-level representation ofthe input rather than on raw bytes, producing what we term Input-Dependent Padding(IDP): two inputs of identical byte length but different semantic content produce structurallydistinct token sequences, and therefore fundamentally different internal padding structures,before a single compression round is performed. The construction combines BPE tokenization(cl100k_base vocabulary, 100,256 tokens), Scattered Salt Injection, SipHash-2-4-keyed Feistelpre-whitening, Davies-Meyer ARX block absorption with a 512-bit chaining state, and HAIFAstyle double-finalisation. Empirical evaluation on the NIST SP800-22 randomness test suite(100 binary sequences, all 15 statistical test families) shows all tests passing at or above therequired threshold, with an observed avalanche coefficient of 50.13%, bit-frequency of 50.03%,and Shannon entropy of 1.000000 bits per bit. RUDRA-512 is published as open-sourcesoftware under the Apache 2.0 licence and is available as native C++17 with Python andNode.js bindings.