A DNA-Local, Constraint-Aware Dual-Head Transformer for Pseudorandom Stream Generation

Pseudorandom number generators (PRNGs) used in deoxyribonucleic acid (DNA)-oriented computational workflows often generate outputs in the bit domain and then map them to DNA symbols. This indirect strategy may treat DNA-specific constraints, including GC balance, homopolymer limits, and short-range sequence dependencies, as separate from generation. This study proposes a constraint-aware, dual-head decoder-only Transformer framework for DNA-local PRNG generation directly in the adenine/cytosine/guanine/thymine (A/C/G/T) alphabet. The model generates the next DNA base and derives the bitstream through dynamic selection among eight equivalent DNA-to-bit coding rules. The framework was evaluated under R1 based on real genomic data, R1-ext as independent validation, R2 based on synthetic data, and R3 without training or reference data. For each setting, 10 independent runs were performed, each producing a 500,000-base DNA sequence and a 1,000,000-bit stream. Bit-level evaluation used NIST SP 800-22, SP 800-90B-inspired min-entropy/health indicators, and ENT, while DNA-level evaluation used GC balance, homopolymer control, and symbolic structural metrics. The reported NIST tests satisfied the acceptance criterion, t-tuple min-entropy lower bounds ranged from 0.9955 to 0.9964 bit/bit, and core DNA-compatibility constraints were preserved. Multi-stream and exact-match k-mer leakage analyses indicated no systematic bit-level dependence or direct long-fragment copying. Overall, the framework supports reproducible DNA-local PRNG generation and multilayer validation.