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 2: source data and frozen settings

The fifteen aggregate CSV tables and both configuration exports retain the original
measured values. Question-level CSVs contain eight methods for each of 1,000 questions
per cohort (16,000 rows total); paired-ratio CSVs contain two balanced observations
per question (4,000 pairs total).

Data dictionary:
sample_id: original benchmark question identifier.
dataset: hotpotqa or triviaqa.
shard: evaluation-session stratum.
method: frozen implementation label, mapped in method_labels.json.
tier: final depth, 0 direct, 1 first retrieval, 2 escalation.
status: recorded execution status.
quality_em, quality_f1: scores on [0,1].
cost_*_ms: measured milliseconds for the named component or complete path.
cost_*tokens: newly processed input/output token counts.
cost_retrieval_calls, cost_generation_calls: physical call counts.
repeat: balanced paired-observation identifier.
ratio: relative on/off increment. Average the two observations within each question
before question-level resampling. The recorded bootstrap uses seed 42 and 20,000
resamples; session strata follow the estimator described in the article.

Aggregate tables use manuscript rounding; per-question measurements retain their
original precision. Benchmark text, reference answers, model weights, and source
passages are obtained from their original providers under the applicable terms.

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