Stability-Screened Source-Pipeline Selection for Later Transfer Validation in Wi-Fi Sensing

Wi-Fi sensing experiments often leave the researcher with several promising source-side pipelines, while time and measurement budget allow later validation of only one or two of them. In this paper, a pipeline denotes a concrete processing chain: for example, CSI packets may be segmented into short windows, simple statistical descriptors may be computed, and the resulting feature vectors may be submitted to a classifier. Another pipeline may use longer windows or a different descriptor set. Selecting the next candidate only by the highest source accuracy is risky because the apparent winner may depend on a favorable split, unstable sessions, or degraded data quality. This paper presents a stability-screened rule for source-pipeline selection before later transfer validation. Each candidate is evaluated by its source-side task score together with three penalties for temporal variation, split sensitivity, and data-quality degradation. The scientific novelty is a source-only screening protocol that treats candidate selection itself as a reproducibility-controlled decision stage, combines these penalties with a frozen no-retuning transfer check, and keeps target data out of the selection step; it does not propose a new sensing model or claim broad transfer robustness. Only the retained candidate is then checked under limited target shifts such as a closed-door setting or a laterday repeat, without target-specific retuning. This helps reduce wasted follow-up experiments, improves reproducibility in early comparison stages, and is applicable to both single-link and cooperative Wi-Fi sensing pipelines.