A Trustworthiness-Oriented Approach to Story Point Estimation for Agile Planning: Worthy-SP as a Task-Specific Estimator Built on Frozen Compact LLM Backbones

Story point estimation supports agile planning, but point accuracy alone does not show whether an estimator provides useful confidence information, behavior-relevant rationales, or practical deployment-cost evidence. This paper proposes a trustworthiness-oriented approach to story point estimation for agile planning. At the center of this approach is Worthy-SP, a task-specific story point estimator built on frozen compact LLM backbones. Within the proposed approach, Worthy-SP operationalizes legal-class distributional prediction, parameter-efficient task adaptation, rationale generation, deletion–sufficiency behavioral diagnostics, and cost-aware evaluation under a frozen-backbone compact-LLM setting. The approach is evaluated on five cleaned TAWOS-derived project-level datasets using prompt-only controls, head-only trained baselines, Worthy-SP, and additional prediction baselines including TF-IDF+LightGBM, MiniLM+MLP, and CodeBERT-FT under a shared label space and chronological evaluation protocol. Across paired project–backbone–seed scenarios, Worthy-SP reduces mean MAE relative to the head-only trained baseline by approximately 4.4% and reduces the reported snapped-ECE after temperature scaling by approximately 16% under this proxy. The results indicate that task-specific training is necessary for compact-LLM-based story point estimation and that the proposed approach remains competitive in prediction while adding distributional outputs, rationale diagnostics, and deployment-cost evidence.