# Robust QAOA Weight Ablation and Size Stratification

This report reuses the full candidate table from the expanded RA-QAOA suite.
It does not rerun QAOA state evolution; it recomputes selector scores from stored candidate metrics.

## Weight Profiles

| profile | degradation | success | shot variance | noise sensitivity |
| --- | ---: | ---: | ---: | ---: |
| default | 0.350 | 0.200 | 1.000 | 0.350 |
| no_sensitivity | 0.350 | 0.200 | 1.000 | 0.000 |
| half_sensitivity | 0.350 | 0.200 | 1.000 | 0.175 |
| double_sensitivity | 0.350 | 0.200 | 1.000 | 0.700 |
| no_degradation | 0.000 | 0.200 | 1.000 | 0.350 |
| double_success | 0.350 | 0.400 | 1.000 | 0.350 |

## p=2 Robust Selector Summary

| weight_profile | graph_count | mean_noisy_expected_cut | ci95_noisy_expected_cut | mean_success_probability | mean_noise_sensitivity | ci95_noise_sensitivity |
| --- | --- | --- | --- | --- | --- | --- |
| default | 60 | 7.750276 | 0.859378 | 0.127260 | 1.079351 | 0.232858 |
| double_sensitivity | 60 | 7.386141 | 0.905820 | 0.054527 | 0.032561 | 0.022331 |
| double_success | 60 | 7.796574 | 0.862513 | 0.129304 | 1.212512 | 0.219968 |
| half_sensitivity | 60 | 7.860022 | 0.870598 | 0.125977 | 1.394998 | 0.191146 |
| no_degradation | 60 | 7.806815 | 0.862043 | 0.129388 | 1.241965 | 0.217020 |
| no_sensitivity | 60 | 7.876849 | 0.869920 | 0.121379 | 1.443395 | 0.181728 |

## p=2 Robust vs Ideal-Expected Sensitivity Delta

| weight_profile | graph_count | mean_delta_robust_minus_baseline | ci95_low | ci95_high | robust_better_rate |
| --- | --- | --- | --- | --- | --- |
| default | 60 | -0.393734 | -0.521034 | -0.266434 | 0.783333 |
| double_sensitivity | 60 | -1.440525 | -1.622549 | -1.258501 | 1.000000 |
| double_success | 60 | -0.260574 | -0.361591 | -0.159557 | 0.733333 |
| half_sensitivity | 60 | -0.078087 | -0.114589 | -0.041585 | 0.633333 |
| no_degradation | 60 | -0.231120 | -0.321252 | -0.140989 | 0.733333 |
| no_sensitivity | 60 | -0.029690 | -0.046874 | -0.012506 | 0.433333 |

## p=2 Default Robust vs Additional Baselines

| comparison_method | metric | graph_count | mean_delta_robust_minus_baseline | ci95_low | ci95_high | robust_better_rate |
| --- | --- | --- | --- | --- | --- | --- |
| grid_center | noise_sensitivity | 60 | 0.011273 | -0.103216 | 0.125763 | 0.216667 |
| grid_center | noisy_expected_cut | 60 | 0.022131 | -0.020504 | 0.064766 | 0.366667 |
| grid_center | success_probability | 60 | 0.009089 | 0.003560 | 0.014618 | 0.383333 |
| random_candidate | noise_sensitivity | 60 | 0.551560 | 0.233145 | 0.869974 | 0.266667 |
| random_candidate | noisy_expected_cut | 60 | 0.441077 | 0.322647 | 0.559507 | 0.866667 |
| random_candidate | success_probability | 60 | 0.077644 | 0.055757 | 0.099531 | 0.916667 |

## p=2 Size-Stratified Default Deltas

| num_nodes | comparison_method | metric | graph_count | mean_delta_robust_minus_baseline | ci95_low | ci95_high | robust_better_rate |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 10 | ideal_expected | noise_sensitivity | 20 | -0.559158 | -0.807464 | -0.310851 | 1.000000 |
| 10 | ideal_expected | noisy_expected_cut | 20 | -0.194412 | -0.280744 | -0.108079 | 0.000000 |
| 10 | ideal_expected | success_probability | 20 | 0.004543 | 0.000138 | 0.008947 | 0.600000 |
| 6 | ideal_expected | noise_sensitivity | 20 | -0.253609 | -0.433626 | -0.073592 | 0.550000 |
| 6 | ideal_expected | noisy_expected_cut | 20 | -0.088177 | -0.150766 | -0.025587 | 0.000000 |
| 6 | ideal_expected | success_probability | 20 | 0.020595 | 0.001795 | 0.039395 | 0.300000 |
| 8 | ideal_expected | noise_sensitivity | 20 | -0.368436 | -0.585855 | -0.151018 | 0.800000 |
| 8 | ideal_expected | noisy_expected_cut | 20 | -0.128100 | -0.203694 | -0.052507 | 0.000000 |
| 8 | ideal_expected | success_probability | 20 | 0.020771 | 0.009637 | 0.031906 | 0.600000 |

## Interpretation

Use the ablation table to check whether the stability claim depends on a single hand-picked weight.
Use the size-stratified table to check whether the aggregate result is driven by only one graph size.
If the sign of the noise-sensitivity delta changes for larger graphs, the paper claim should be narrowed.