This programme asks when the dispatch from a linear power-flow model, the kind electricity markets clear on, survives the real AC physics, and how cheaply it can be repaired when it doesn't. The repair is a successive linearisation: linearise the AC power flow at the current point, solve a small QP over generator outputs and voltage setpoints with every limit pulled slightly inside, run an AC power flow, and repeat. Until today it had a budget of 10 rounds and was fully feasible on 21 of the 34 PGLib-OPF benchmark cases that have an AC optimum. The last milestone found that cases finishing late were the fragile ones. So this step asked a different question: when a round fails, why does it fail?
I wrote six predictions down before running anything. Then I re-ran the frozen repair with a 30-round budget and logged every round. For each round I recorded the QP kind, the step, how far each limit's real AC value landed from the linear model's prediction (measured in units of that limit's margin), and how well-conditioned the power-flow Jacobian was. The first 10 rounds reproduce the earlier run exactly.
The budget was the binding choice. With 30 rounds the repair alone is fully feasible on 31 of 34 cases. All 31 finish by round 20, and 28 finish by round 12. The ten cases that finish after round 10 cost a mean 0.41 % more than the AC optimum (worst 1.02 %). The cases that finish early cost 0.38 % more. The slowest repair takes 105 seconds. I had predicted at most 3 more cases; there were 10. That puts the repair alone level with what the repair plus a separate voltage stage reached at 10 rounds.
Most misses are curvature, not a model with no answer. A round can fail in two ways. In the first, the linear model has no feasible point inside the trust region, so the QP falls back to its elastic form and accepts some violation. In the second, the QP finds a point the linear model calls feasible, but the AC power flow disagrees by more than the margin. Of the 224 misses in rounds 1 to 10, 168 are the second kind. They break thermal limits and generator reactive limits far more often than angle or voltage limits. A simple check holds on all 190 such misses: a limit breaks only when the linearisation's second-order error on it exceeds its margin. That is the textbook one-step condition, and here it is exact.
The convergence bound in our own proposal is refuted. The proposal predicted the round count would be bounded by the initial violation divided by the smallest per-round improvement, within a factor of 10. That holds on only 6 of the 31 finishing cases, and on one it is off by a factor of 55,000, because some accepted rounds improve the merit by a hundred-thousandth of where it started. The curvature estimate also varies by more than a factor of 10 along the path on half the cases, so no single per-case constant describes it. And one case, case73 of the congested set, stalls even though its Jacobian is well-conditioned throughout. Every one of its 30 QPs is elastic. From round 14 its merit sits flat at two bus voltages 0.002 pu out of band, while the trust region shrinks to nothing. That is a point of local infeasibility: the linearised limits cannot all be met near where it is standing. Successive linear programming theory allows this outcome, and the proposal's statement did not. The AC OPF solver reports an optimum for that case, so a feasible point does exist.
What this changes is narrow but concrete. The held-out runs on larger cases will use a 20-round budget fixed now, not 10. The convergence statement for the paper needs an extra assumption, that the QP stays feasible inside the trust region, and should be written in terms of step size rather than merit decrease. The misses point to margins scaled to each limit's curvature as the next thing to try, but only on cases the rules were not set on. All of this is on the same 34 cases the repair was designed on, from one starting point, so it says how the method behaves, not yet how it generalises.
Code, tables and log: research/infra-dcopf/ (experiments/run_converge.py, src/infra_dcopf/converge.py, results_converge.txt); programme: research/grid-linopf/. Written with AI assistance (Claude Code).