PReRAG v0.1.0

PReRAG v0.1.0

Initial public reference release of PReRAG: path-aware retrieval reward shaping
for controlled retrieval-augmented policy learning.

Included

• Global exact-page and strict step-aligned document reward contracts.
• GlobalRAG process reward composition and answer-gated curriculum.
• Controller-trusted retrieval metadata and information-token loss masking.
• Minimal Search-R1 Controlled comparator using normalized outcome-only EM.
• Portable configurations, evaluation aggregation, documentation, and 20 CPU
  contract tests.

Scope

This release is a lightweight reference implementation. It does not distribute
model weights, adapters, private datasets, retriever indexes/corpora, raw
trajectories, or a paper-scale Search-R1 reproduction. See the README and
reproducibility documentation before attempting end-to-end training.

Release commit: 0d8815bc8fc5d5b96bd18ef10b05173f8eae6c2c.