| """The MergeBench arm: the Beetle mergeability metric suite, computed over MergeBench's 40 checkpoints. | |
| `suite` enumerates the HF org (never a hand-written list) and forms the mergeable pairs. | |
| `sweep` computes the F8 property panel per pair, one base family at a time, streamed. | |
| `panel` joins those properties to MergeBench's OWN published merge scores and draws the figures. | |
| Read `RESULTS_MERGEBENCH.md` at the repo root first: the outcome column here is NOT ours. We do not | |
| re-evaluate merged models; we reuse the numbers MergeBench published, and those numbers cover a | |
| different population (5-expert merges, one per base family) than the pairwise properties we measure. | |
| That mismatch is the headline limitation and it is recorded, not worked around. | |
| """ | |