Observatory
Live · fleet quietEvery hour, each US power grid publishes a day-ahead demand forecast — and then the demand actually arrives. We monitor the error with the same anytime-valid engine we sell, and raise a page only when it genuinely changes regime, inside a stated false-alarm budget. It’s a real deployed model we neither built nor own — standing in for yours.
updated Jul 6, 22:08 UTC
Live board
Every card below is this run’s real state, straight from the artifact the pipeline rewrites on each poll. Green is the win: it means no regime change was detected on that grid’s forecast error — not that the grid is fine, just that the evidence hasn’t moved.
The budget is the page tier’s guarantee, not a wish: across all 18 grids the engine expects only about 18 false pages per year total. A quiet grid means no regime change was detected in its forecast error — never proof the forecast is fine.
Every grid publishes an embeddable status badge that updates on each poll — open any grid to copy the snippet for your README, status page, or dashboard.
Why your dashboard cries wolf
A fixed threshold you check every hour is a coin you flip every hour. Left up long enough, noise alone will trip it — so the more often you look, the more it cries wolf. That is not a tuning bug; it is what fixed thresholds are. An anytime-valid e-process instead spends a stated false-alarm budget, so checking as often as you like cannot inflate its error rate.
Below is ~5 months of the Southwest Power Pool’s real day-ahead demand-forecast error — the central-US grid across Kansas, Oklahoma, Nebraska and the Dakotas — run through three alarm policies at the product’s default parameters. Same numbers into all three.
Checking hourly for ~5 months, on the same real stream: the static threshold and the sensitive CUSUM page dozens to hundreds of times — unbudgeted, mostly noise. The anytime-valid e-process pages 3 times, each at a genuine winter cold snap that pushed the forecast into a new regime, every one inside a budget of about one false page per grid-year. Sensitive is easy; the product question is how many pages you can ask a human to trust. We detect that the operator’s own forecast changed regime — not that the grid failed.
Ledger highlights
The engine has written 107 dated pages across the fleet since 2021. The ones below overlap documented grid-stress events — the page fired when the operator’s own forecast error entered a new regime, not when the grid failed. Each links to that grid’s full history.
Same engine, same two-tier alarms, on your metrics — LLM-judge scores, drift, latency, any KPI. Connect a stream and reproduce a “caught it on day X” verdict on your own history, free, in minutes.