THE GRID

Measured, before the page was drawn

Every one of the 46 transmission lines was failed in turn, under 6 random seeds, at 100% and at 115% of nominal load. 552 scenarios per stress level, run headless through the exact WebAssembly module you just watched. Nothing below is hand-picked.

1 · The head-to-head

Deep = horizon 14, 64 rollouts. Myopic = horizon 3, 24 rollouts. Same code, same seed, same eight periods of advance warning, same grid. “No operator” is the grid left completely alone: relays and automatic balancing only.

At nominal load the New England system is genuinely robust — 249 of 276 scenarios are a dead tie, because almost nothing bad happens and holding still is the right answer. Push it to a realistic summer peak and the picture inverts: the myopic room keeps the lights on for 5,653.3 MW, barely above the 5,642.8 MW you get with nobody in the room at all, while looking further ahead holds 5,903.3 MW. That gap — 261 MW, averaged over every single-line outage on the system — is roughly a large city. It is bought with nothing but imagination: no new wires, no new generation, no learning, no training data.

2 · The scaling law

This is the result that matters. Hold everything constant — same 276 scenarios at 115%, 64 rollouts — and turn one dial: how far forward the machine is allowed to imagine.

Horizon 1 lands on 5,642.8 MW — exactly the no-operator number, to the decimal. That is the sanity check: a machine that cannot see one step past its own action can never find a reason to deviate from holding still, so it holds still, so it is the empty control room. Everything above that line is purchased purely by looking further. The curve is not clean — horizon 16 is worse than horizon 14 — and it is printed here unsmoothed because that is what came out. Search over a chaotic cascading system is noisy; deeper is better on average, not on every run.

The cost column is the part utilities should read twice. A full horizon-14 decision — both control rooms and the untouched baseline — takes 4.18 milliseconds on one core of a laptop. Real control centres re-dispatch on a five-minute cycle. This is about seventy thousand times faster than the clock it would have to keep.

3 · How many futures

Same thing, other dial: horizon pinned at 14, vary how many futures get sampled per decision. One rollout is again exactly the empty room.

It saturates. Past about 64 futures per decision the returns are inside the noise — 80 rollouts scored 5,884.8, below 64’s 5,903.3. Depth of imagination buys more than breadth of it.

4 · Every line, at summer peak

All 46 branches of the case, each failed in turn, mean of 6 seeds at 115% load. Sorted by how much load the deep room saves over an empty one. Nothing removed.

5 · Where looking further made it worse

7 of 276 runs at summer peak, and 9 of 276 at nominal, the myopic room beat the deep one. Here is every one of them, unedited. In most cases the deep room spent a few megawatt-hours probing a redispatch that did not pay off; in run 13–14 at seed 7 it lost 460 MW outright by committing to a path the shallow room never explored. Search is not a guarantee. It is a better bet.

What this is, and what it is not