One baseline.
Nineteen branches.

The search began broadly: change the allocation, change the order, or import new data. The first useful result—upweighting scarce categories—became the trunk for most of the work that followed.

Two deep families emerged. One investigated short-answer stages and curricula. The other removed generic vision and caption data, then became a test bed for wrap position, filtering, visual dependence, and science data.

The branching is the important result. Several ideas that looked promising in one recipe failed when transplanted to another. Quality filtering, terse endings, and staging were not independent knobs; their effect depended on the branch around them.

The final run kept nearly the same ingredients but reordered them. Its large change made the campaign’s cleanest warning visible: when order and run noise can move outcomes this much, small differences between recipes should not be over-read.

Source: the public tracker’s completed-run table, event timeline, and 64-round logbook for ndrugov/vlm-data-autoresearch. The map includes the cleaned baseline and all 19 completed post-clean experiments; failed launches and pre-clean runs are excluded. Position encodes conceptual family only. The visual-max parent is revealed before visual-anneal so the lineage reads in causal order.