The Inversion
Tenth entry in this series, and the one the whole Australia thread was actually building toward. The previous post found perfect recall and collapsing precision past onset — the opposite failure mode from Germany. This post explains why, and it's not a subtle difference. It's the mechanism running backwards.
What the German series actually found
Worth restating plainly, since this post is the direct comparison: in the German redispatch data, background operating state dominates and event magnitude is a poor proxy — r=0.06 for magnitude versus r=−0.99 for loading. That correlation gap is the paper's central finding. Event information alone tells you almost nothing about whether a redispatch instruction will cause a voltage violation; the network's background load state tells you almost everything.
The natural next question, and the reason this Australia thread exists at all: does that hold generally, or is it specific to German redispatch physics?
Running the same regression here
Exact replication of the German design: , fit on a 70/30 split of converged critical-path solves, refit fresh at each severity past onset, applied only to the population it was never fitted on, catch rate reported with a Wilson-score 95% CI.
At every severity tested — 5.5×, 6.0×, 6.5×, 7.0× — there was nothing left in the non-critical population for the state-aware rule to catch. Zero real violations remained uncaught, at any severity. That's not a null result to shrug off — it's the direct, mechanistic consequence of the previous post's finding: the plain magnitude selector already has 100% recall here, so there's nothing left over for a second rule to recover.
The actual measurement
The real question wasn't whether the state-aware rule added value — it was why the magnitude selector was already sufficient when it demonstrably wasn't in Germany. Measuring the same two correlations, at the same severity as the table above:
Germany: r(voltage, magnitude) = 0.06, r(voltage, loading) = −0.99 Australia (this work): r(voltage, aggregate injection) = 0.90, r(voltage, loading) = −0.47
The relationship is inverted, not replicated. In Germany, event magnitude is a poor proxy and operating state is the dominant signal — the paper's central thesis. Here, event magnitude is a good proxy, and background load is real but secondary. That's exactly why the magnitude-only selector already catches everything, and exactly why a state-aware rule that deliberately excludes magnitude by design has nothing left to add.
Why this isn't a replication failure
This is the finding worth sitting with: whether "operating state beats magnitude" holds depends on whether the phenomenon is generation-driven or load-driven. German redispatch is fundamentally about managing load variability against fixed generation events. Australian rooftop-solar overvoltage is fundamentally about managing generation variability — aggregate solar injection — against comparatively stable background load. The direction of the dominant driver flips because the physical mechanism itself flips.
The German paper alone couldn't have established this — a single case, however rigorously audited, can show a correlation exists, not that it generalizes in one particular direction. This comparison is what turns "operating state dominates event magnitude" from a single result into a boundary condition with a stated domain of validity: true when the phenomenon is load-driven, not true when it's generation-driven. That's a genuinely new, falsifiable claim, not present in the original paper, and it only exists because the same rigor got pointed at a second, structurally different problem instead of stopping after the first result held up.
One honest caveat on the regression itself, not swept under anything: multiplier and P_load are measured as correlated 0.99999+ here, since both derive from the same real household load shape — individual β coefficients aren't cleanly separable as a result, though this doesn't affect the catch-rate methodology, which only uses the fitted prediction. The next post covers what happened when that specific limitation got checked formally, not just flagged.
Reproducibility: state-aware rule and the correlation measurement
# Exact replication of the German regression design, refit fresh at
# each severity, applied strictly out-of-sample:
# V ~ b0 + b1*multiplier + b2*P_load, 70/30 split, Wilson-CI catch rate
for severity in [5.5, 6.0, 6.5, 7.0]:
# zero non-critical real violations exist to catch at any severity --
# direct consequence of the magnitude selector's own 100% recall
pass
# The measurement that explains why:
# r(voltage, aggregate injection) = 0.90 (Germany: r(voltage, magnitude) = 0.06)
# r(voltage, background load) = -0.47 (Germany: r(voltage, loading) = -0.99)
Full script: au_state_aware_rule.py in the Australia repository.
