Gen 0 | Score: 100 | Best: 100 | Status: baseline | Change: gemini2.5pro_4.cpp baseline
Gen 1 | Score: 100 | Best: 100 | Status: improved | Change: Switch to gemini3pro_2.cpp (unbounded 136.36 vs 134.03)
Gen 2 | Score: 100 | Best: 100 | Status: baseline | Change: Revert to gemini2.5pro_4.cpp (gemini3pro_2 TLEs)
Gen 3 | Score: 100 | Best: 100 | Status: same | Change: Rewrite with faster DP (no vector copy), add fallback. 12ms vs 25ms
Gen 4 | Score: 33.33 | Best: 100 | Status: reverted | Change: Displacement-based weights - WA (stuck)
Gen 5 | Score: 100 | Best: 100 | Status: same | Change: Scaled weights + swap bonus - no improvement
Gen 6 | Score: 33.33 | Best: 100 | Status: reverted | Change: Augmented matching - caused oscillation
Gen 7 | Score: 100 | Best: 100 | Status: same | Change: Multi-root matching - no improvement, 3x slower
Gen 8 | Score: 100 | Best: 100 | Status: same | Change: Max cardinality matching - no improvement
Gen 9 | Score: 100 | Best: 100 | Status: reverted | Change: Flow-based weights - slower and worse unbounded (133.05)
Gen 10 | Score: 33.33 | Best: 100 | Status: reverted | Change: Next-on-path weights - WA (gets stuck)
Gen 11 | Score: 100 | Best: 100 | Status: improved_unbounded | Change: Randomized trials with noise - unbounded 134.77
Gen 12-17 | Score: 100 | Best: 100 | Status: improved_unbounded | Change: Randomized trials with time-based budgeting - unbounded 135.23
Gen 18 | Score: 100 | Best: 100 | Status: reverted | Change: Multi-strategy - slightly worse (135.08)
Gen 19 | Score: 100 | Best: 100 | Status: reverted | Change: Root randomization - overhead too high (135.09)
Gen 20-21 | Score: 100 | Best: 100 | Status: reverted | Change: Various noise experiments - none beat 135.23
Gen 22-23 | Score: 100 | Best: 100 | Status: reverted | Change: Adjacency shuffle + other tweaks - no improvement
Gen 24-26 | Score: 100 | Best: 100 | Status: same | Change: Various optimizations, cycle-based - no improvement over 135.23
=== Final Summary ===
Best: Score 100, Unbounded 135.23
Algorithm: Greedy max-weight matching with subtree heuristic + randomized noise trials