When OCR Is Not Enough: Semantic Limits in the Digitization of Historical Computer Music Code

The preservation of early computer music faces a critical bottleneck caused by the inability of standard digitization workflows to recover executable code from degraded physical media. This paper presents a first study on the recovery of source code from the Music-N family, contributing a new dataset of historical computer music scores from the Centro di Sonologia Computazionale (CSC) archive. We evaluate the performance of four recognition architectures spanning two paradigms: Tesseract, EasyOCR, and TrOCR as representative OCR systems, and Qwen2.5-VL as a modern Vision-Language Model (VLM) applied to documents characterized by high symbolic density, obsolete syntax, and physical degradation. Our methodological contribution consists of a specialized preprocessing pipeline involving binarization, skew correction, and rescaling to mitigate the noise inherent in source code printouts. The results show a clear improvement in representative OCR systems performance following the application of the pipeline. However, the analysis also identifies a significant divergence between visual accuracy and syntactic validity. This demonstrates that visual digitization alone cannot guarantee the functional preservation of software-based heritage. This work represents an initial step toward the recovery of this heritage, highlighting the need for future systems to incorporate structural constraints, such as Large Language Model (LLM) fine-tuning or Retrieval Augmented Generation (RAG) integrations, to handle the specific syntax of historical programming languages.