The harness turns to the grid: first real network-expansion runs (E.ON class)
Same referee, new arena
Until now the ledger measured portfolio optimization. This entry moves to distribution grid expansion — deciding which new power lines to build so an overloaded network sheds congestion at least cost. It is a combinatorial build/no-build problem, and it is the E.ON problem statement in the 2026 Global Quantum + AI Challenge. The protocol did not change: same instance, same time budget both sides, exact optimum as referee, fixed seeds, frozen versions, sealed in triple copy. Only the problem is new — which is the point of a general protocol.
How the instance is grounded in real grid physics
grid IEEE case14, loads ×3.0, thermal ratings tightened
→ an under-provisioned feeder with genuine congestion
candidates 14 build options (parallel reinforcements + new lines)
model congestion relief measured by REAL DC power flow:
r_i = relief of building line i alone (measured)
q_ij = relief(i,j) − r_i − r_j interaction (measured)
→ a QUBO whose coefficients come from physics, not guesswork
referee exact optimum over 2^14 build sets (brute force)
validation winning plan re-checked in full AC power flow
No coefficient was invented: each was measured by running the grid solver. The quadratic terms capture that two reinforcements together relieve more (or less) than the sum of their parts — real network coupling.
The result
CP-SAT reached the provably optimal build plan in about a third of a second, all three seeds. QAOA came close — 0.005% at its best — but did not win under equal budget. Verdict: not yet. The classical planner is strong here, exactly as it should be for an honest benchmark.
Why the −43.5% matters more than the verdict
The optimization is decided in a QUBO, but the value is decided in physics. We rebuilt the winning 5-line plan in full AC power flow: total line overload dropped from 2672.7 to 1509.3 — a 43.5% congestion reduction on a genuinely overloaded feeder. That is the number a DSO planner cares about, and it is validated outside the model that chose it. Whether the plan came from quantum or classical, the ledger reports what it actually does to the grid.
What we don’t know
Whether QAOA’s near-miss (0.005% on seed 43) survives on harder, larger instances or is instance-luck — the next runs scale the grid (case30, case118) and the candidate count toward the >100-qubit utility scale the challenge targets. How the second-order DC congestion model diverges from AC on bigger grids (we’ll publish the gap). And whether NISQ hardware, with noise, holds QAOA’s simulator-measured quality. Each is a measurement on the ladder.
Measured content from sealed archives EXP-0033-001…003. Methodology follows the QOBLIB benchmark approach cited in the E.ON problem statement: identify hard instances, then benchmark honestly. Rosetta Quantum publishes verdicts with reproducible raw data — including the ones quantum loses.