Convergence Curves

Quality improvement follows a logarithmic curve. Round 1 captures 60-70% of the total possible improvement. Round 2 captures another 15-20%. Round 3 captures 5-8%. After round 3, you are spending exponentially more tokens for linearly smaller gains.

We track this with convergence velocity — the delta between quality scores at round N and round N+1. When convergence velocity drops below 0.50 points per round, the loop has reached diminishing returns. At that point, continuing iteration is a cost decision, not a quality decision. In our production system, the average convergence profile is: Round 0 (raw) = 5.84, Round 1 = 7.92, Round 2 = 8.61, Round 3 = 8.93. The jump from 5.84 to 7.92 justifies the first round every time. The jump from 8.61 to 8.93 rarely justifies a third round except on high-stakes deliverables.