Public technical results where the problem, evaluation contract, and accepted improvement can be inspected together.
A five-node integration rule is compared against a fixed baseline, with both the acceptance score and direct residual errors exposed.
Across two governed campaigns, a deterministic dispatch rule lowers the frozen PSPLIB J30 score while preserving feasibility on every instance.
Three tolerance-aware corpus leaders plus an exact rational certificate of strict local optimality for one 78-contact configuration.
A routing policy is replayed against LightSABRE on a frozen 72-case portfolio across Q20, Willow, and Heron-FEZ targets.
A BESS dispatch policy is challenged offline against commercial baselines, oracle headroom, downside risk, degradation proxy, and hard constraint checks.