What finally moved composition
Same loop, same model, same compute, same ~600 practice tasks. The only difference: how many distinct compositions those tasks are spread over.
Run A — low diversity
12 structures · context multiplicity 2.0
+3.4
novel-composition pts, not monotone
Run B — high diversity
48 structures · context multiplicity 8.0
+15.8
novel-composition pts (2.11×), monotone 5/5
Run B · high diversity
Run A · low diversity
solid = novel composition · dashed = same structure
novel composition
same structure
this run's own novel-composition base
Run A · low diversity
12 structures · mult 2.0 · base 0.083 / 0.100
Run B · high diversity
48 structures · mult 8.0 · base 0.142 / 0.125
4
Honesty note.
Specialisation note.
Rounds note.