- black box evals can only take you so far, the real value lies in reading the trace
- a final score tells you “what” happened, the trace tells you “why”
- this is why good methods around context optimization, training, and environment feedback involve reflection on the trace and not just the output
- recent examples: GEPA, HALO, SkillOpt, etc
u/adnanqidwai/thoughts/trace.md