How do you evaluate whether a research paper is worth implementing?

We're drowning in ML papers and the gap between 'sounds promising' and 'actually works in our stack' is brutal. We burned 2 weeks implementing a retrieval optimization from a paper that only showed gains on synthetic benchmarks. Do you have a triage process? Something like: check reproducibility repo → quick benchmark on our data → architecture fit assessment? Or do you just prototype and accept the waste? Specifically interested in how teams with limited engineering bandwidth handle this.