EPRAG v1.0.0

This is the first archival release of the reproducibility repository accompanying the manuscript:

EPRAG: An Epistemic Policy Framework for Action Selection in Multi-Source Enterprise Retrieval-Augmented Generation

The repository contains the code, configurations, prompts, public evaluation artifacts, aggregate results, and documentation required to inspect and reproduce the main analyses reported in the study.

Included materials

• Reference implementation of the EPRAG decision-policy pipeline
• Public BIRD-derived evaluation artifacts and predictions
• Taxonomy-blind public audit materials and trusted annotations
• Structured GPT-5 judge evaluation outputs
• Counterfactual-answerability candidate and threshold sensitivity results
• Baseline and ablation outputs
• Non-sensitive helpdesk schema, synthetic workbook template, label-derivation documentation, and aggregate results
• Manuscript-aligned configuration files
• File manifest and SHA-256 checksums

Evaluation subsets

• Public BIRD-derived benchmark: n = 720
• Taxonomy-blind public audit: n = 200
• Structured GPT-5 judge:
  • BIRD-derived subset: n = 240
  • Helpdesk subset: n = 100
• Answerability sensitivity subset: n = 96

Privacy and availability

The private university helpdesk records are not included because they contain operational and potentially sensitive information. Only non-sensitive schemas, synthetic templates, label-derivation rules, prediction-level metadata where permissible, and aggregate results are provided.

This release does not contain the manuscript, LaTeX sources, reviewer correspondence, cover letters, or journal-submission files.

Version

Version: 1.0.0

This release represents the frozen archival version associated with the manuscript.