Shorten paper title to When Does a Perturbation Model Know Enough?

Add descriptor-non-redundancy vs task-non-additivity distinction to halt-head section

Scrub dataset names from card; remove base-model from title

Retitle: lead with method + separate working frozen heads from collapsed CD4-native negative

Update model card: honest pseudobulk-collapse framing, non-redundancy vs non-additivity, corrected cell counts, paper title

Add README.md

initial commit