The grid-linopf programme repairs a linear (DC) market dispatch until it passes the full AC power flow. Each round solves a QP inside every limit, with a small safety margin, and inside a trust box around the current point. If that QP has no solution, an elastic QP that minimises overloads runs instead. A step that makes things worse is rejected, and the box halves.
The last step widened the box to the smallest size with a solution whenever the QP was empty (rule E). That fixed every empty QP, but in 68 of 94 firings it rebuilt the QP it had just rejected. After a rejection the point is the same, the box halves and the needed scale doubles, so the same step comes back.
Rule E-once. The same widening, at most once per base point. After a widened step, the old elastic fallback runs until a step is accepted and the point moves. We froze the rule and six predictions in the log before running it on the 34 development cases and the 19 held-out cases.
| frozen repair | rule E | rule E-once | |
|---|---|---|---|
| development finishers (of 34) | 31 | 30 | 31 |
| firings that rebuild the QP just rejected | 68 of 94 | 0 of 43 | |
| case73 API | never finishes | clean, round 12 | clean, round 12 |
| held-out finishers (of 19) | 10 | 8 | 9 |
What it fixed. Exact repeats are gone. case200 API, lost under rule E, finishes again, at round 7 instead of the frozen 10. Held-out, case588 TYP comes back, at 0.08 % above PGLib's objective instead of the frozen 2.12 %. case73 API (the first time it finishes with the full margins) and case500 SAD keep rule E's gains.
What still breaks. case30 API and held-out case793 API alternate to the end of their budgets. A widened step is rejected; then a tiny elastic step is accepted, and that tiny step moves the point and re-arms the rule. The elastic step is tiny because every rejection has halved the stored box (13 times on case30 API). The widening, though, is capped at twice the first box, which never shrinks. So it keeps proposing about the same large step from almost the same point. "Once per point" stops exact repeats and not near-repeats.
Four predictions held: at most one development finisher lost, case73 API clean, no slower median finish, and case500 SAD finishes. Two were refuted: a net development gain, and held-out finishers at least the frozen 10. The frozen repair stands.
Next. Tie the widening to the stored box: cap it at twice the current box, so each rejection also shrinks what widening can reach.
Code and numbers: run_enlarge_once.py, results_enlarge_once.txt, and the programme page. Written with AI assistance (Claude Code).