EvoGraph: Hybrid Directed Graph Evolution toward Software 3.0

EvoGraph, from AutoHand AI, proposes Hybrid Directed Graph Evolution as a closed-loop modernization system that treats code, build pipelines, documentation, tickets, schemas, logs, and runtime telemetry as one typed directed graph, then mutates that graph with small-language-model-driven operators and a multi-objective safety gate. The framework combines weight merging, AST patching, documentation sync, build weaving, and cross-language transmutation, while ranking candidates with a contextual bandit and Pareto-plus-novelty selection. On seven legacy-style benchmarks, EvoGraph fixed 83% of known security vulnerabilities, translated COBOL to Java with 93% functional equivalence, and kept documentation freshness within two minutes; across multi-language modernization it reached 82–96% semantic equivalence on COBOL, .NET, Lisp, CGI, ColdFusion, legacy Python, and C, while reducing computational cost by 90% versus large language models. The system also cut p95 latency by 40% and feature lead time by 7× relative to strong baselines, and on 1.2K COBOL programs its transmutation step reduced manual review edits by 58%. In resource terms, processing 10K lines of code required 12 GPU hours and 45 kWh, compared with 120 GPU hours and 450 kWh for a GPT-4-based approach, highlighting the paper’s central claim that specialized small language models make production-safe Software 3.0 style evolution practical rather than purely aspirational.