Cost-Aware Gated Retrieval: Decoupling Retrieval Initiation from Evidence-Conditioned Escalation
Journal of Intelligent Information Systems
Zida Hao and Yubo Li
School of Software, Xinjiang University, Urumqi, China
Corresponding author: Zida Hao; 20242501431@stu.xju.edu.cn

Online Resource 3: frozen scientific source code

This archive contains the Python, C++/CUDA, and configuration sources used for the
two test cohorts. Scientific source bytes match the original execution archives;
source_sha256.json records their hashes. Method labels map to CAGR-R and CAGR-F
as described in the data-and-settings resource.

Environment and inputs:
- FP16 Qwen2.5-3B-Instruct, revision aa8e72537993ba99e69dfaafa59ed015b17504d1.
- Python 3.13.15, PyTorch 2.11.0/CUDA 12.8, Transformers 4.51.3.
- One T4 device, batch size 1, concurrency 1.
- A 483,690-passage SQLite FTS5 index and the article's generation budgets.
- run_test1000.py and run_heldout_v2.py are the cohort controllers;
  v04_v3_native_state_* implements persistent native state.

Reproduction requires the original benchmark inputs, model weights, passage index,
and the dependencies listed above. Controllers retain their original relative paths
and input-integrity checks. Statistical checking uses the scalar question records
and paired observations in the data-and-settings resource. Third-party assets
retain their original terms.

Reproducibility repository: https://github.com/human-com-h/CAGR
The repository is publicly accessible.
