What we measured, and how to attack it.
The two-tier architecture is a design decision. Feedforward is textbook. Autocorrelation is standard signal processing. None of that is a moat. The apparatus that tries to break them is, because it cannot be shortcut, and no competitor in this market publishes one.
What we measured, and against what.
Eight plant scenarios run under fixed KPI gates. Each one carries a counterfactual: the same plant, the same load, the same generation, with the storage assets removed entirely. That is the baseline. On top of it sits a sweep of 1,188 points across the design space.
The gates are seed-locked. They are never updated automatically. Recalibration happens for a deliberate model change and for nothing else. When it happens, the old and the new values are published side by side with the mechanism that moved them.
How the rig is built.
Two lanes. The regression lane is seed-locked with fixed gates. The exploration lane is randomised and checks invariants only. A finding graduates from exploration into regression by explicit decision, never automatically.
Every assertion carries a sensitivity proof: a deliberate mutation of the model that turns that assertion red. An assertion that cannot be made to fail is not testing anything.
Every measurement point declares what it observes. The world, the plant, or a command. A quantity the plant itself influences cannot say whether the plant is meeting its objective.
The rig exists to break this company's own claims. It has.
The number.
Median 7.0% lower delivered energy cost per MWh served, measured against the same plant with the storage assets removed entirely. Band 4.0% to 8.6% at the 5th and 95th percentile across 1,188 unweighted points. Cost is fuel, grid and ageing in EUR. Served energy is load less unserved energy, floored at zero. Capex, fixed O&M, the value of unserved energy and state-of-charge correction are excluded.
This is a modelling result under a stiff grid assumption. It is not a contractual figure and is not offered as one.
What the number does not cover.
-
It measures the storage layer as a whole.
Hardware and controller are not separable in it. A run with PlantOS frozen against identical hardware is what separates them, and it is scheduled.
-
The cost denominator prices served energy only.
A baseline that sheds load is not penalised by it, which makes the measured benefit conservative. A second KPI that prices unserved energy is in build.
-
The grid model behind the figure is stiff.
A non-stiff fixture runs alongside it from the next release. A rig finds defects in code, never in the model it was handed.