Uncertainty-Aware Construction of Semantic Scheduling Instances from Incomplete Software Artifacts

Abstract

Small software teams without a dedicated project manager repeatedly determine task assignments, execution sequences, and schedules from noisy and incomplete process artifacts. Existing multi-skill project scheduling methods assume an already-verified scheduling instance; however, in this setting, the main challenge is constructing that instance. This study formulates semantic scheduling as an uncertainty-aware framework that builds an admissible, quality-controlled scheduling instance from incomplete artifacts and optimizes it into feasible calendar-resource schedules. A Large Language Model (LLM) structures raw artifacts under a canonical vocabulary gate and Human-in-the-Loop (HITL) review; the surviving competences become continuous eligibility thresholds; a Machine Learning (ML) execution-time prior, trained under a leakage-controlled temporal split of a public issue corpus, is incorporated as a bounded relative multiplier; and a standard Genetic Algorithm (GA) optimizes a calendar-aware Multi-Skill Resource-Constrained Project Scheduling Problem (MS-RCPSP). The language model does not participate in scheduling decisions. Skill extraction achieved an F1-score of approximately 0.65 against reference labels, with inter-annotator agreement of approximately 0.66 (Cohen's kappa ) from an independent blind annotator. The public-corpus prior achieved a median of 20% of assignments. On a retrospective replay of an anonymized industrial project of 656 tasks and six developers, a same-project effort prior changed the modeled median makespan from 86.9 to 72.2 days, whereas semantic competence thresholds trade makespan and load balance for assignments closer to the historical team, increasing agreement from 0.28 to 0.35. These results demonstrate feasibility and measured decision shift, not improved delivery outcomes or cross-organizational generality.