After moving the rubric into the cached prefix, scores got more consistent. Too consistent. We think the model stopped re-reading the rubric for edge cases because the prefix is treated as 'known'. Scores on the hard cases drifted up by about half a point. Anyone measured this?
not measured, but we saw caption quality drop on unusual clips after caching. put the unusual-case instructions in the suffix and it recovered.
Half a point on hard cases is exactly what you would expect if the hard-case guidance is in the part that is no longer attended to freshly. Move the exceptions to the suffix.
Exceptions in the suffix, rubric in the prefix. Scores on hard cases back down. Posting the before/after in the kit changelog.