Four steps, all of them from public records, none of them requiring anything from a utility.
For 44 interconnection-relevant transmission nodes in Northern California, we read two complete calendar years of CAISO day-ahead settlement data — 17,544 hours per node — and isolate the congestion component of the price. Exposure turns out to be intensely locational: the worst node in the atlas sits north of Sacramento, not in the Bay Area where the load wants to be.
Congestion is not spread evenly through the year. It clusters at midday in the shoulder seasons — the signature of a grid long on solar. That matters commercially: overnight training load is barely touched, so a flexibility window can be narrow and still be worth something.
Not a price signal. We take CAISO's own constraint record and identify the transmission elements that bind disproportionately during a given node's congested hours, across 625 named elements. A planning engineer can check that against their own model, which is the point.
History is the easy half; anyone can explain last year afterwards. So the model is tested walk-forward, scored only on months it never trained on: 31% median error on annual congested hours against 65% for the standard month-and-hour benchmark, better at 38 of 44 sites.
It is given only what a forecaster could actually know a day ahead — CAISO's own demand and renewable forecasts, satellite irradiance at the substation's own coordinates, and the node's own congestion history lagged two days rather than one, because a day-ahead forecast is made while yesterday is still settling.
A congestion price is a market outcome. What actually curtails a flexible load is a utility's contingency analysis — an N−1 event out of their own study process. Those two things correlate, and we have measured how strongly. They are not the same thing.
We cannot close that from outside, and it is the reason the first thing we ask a utility for is a scored test rather than a contract.