A foundation model earns its place only where it beats what the plant already runs; its failures teach as much as its wins.
Out of the box, are Timestral's bands honest on an ordinary hourly electricity series, and how fast is each size on a laptop CPU?
The bands are honest: on an hourly electricity series the smallest size's 80 % band covers 85 % of the truths, and the largest size answers in 137 ms on a laptop CPU.
Used zero-shot as a normal-behaviour model on a simulated turbine, which of fourteen injected faults does Timestral see, how fast and at what false-alarm cost, against a ridge regression with a CUSUM and two one-line witnesses?
Every size raises its first alert on the frank jump within 2.6 days, but stops, curtailment and the onset of winter, which are not faults, read as faults: up to 25 false alarms a year on a healthy turbine. Beyond its own noise the 209M sees 5 of the 11 faults; a ridge regression with a CUSUM sees 8.
Used zero-shot as a normal-behaviour model on the temperatures of real turbines, how many of the 45 documented faults of CARE to Compare does Timestral see before they happen, how early, at the same false-alarm rate as a ridge regression?
On the temperatures of real turbines (45 documented faults in CARE to Compare), at the same false-alarm target, each column detects about what chance alone would at its own alarm rate: the 123M 18 against 16.3, the ridge 22 against 29.7. Timestral's quantiles are calibrated on real temperatures; the replayed threshold does not hold on new weeks.
How every number here was made: the hypotheses are committed before any run, and CI checks the order; every result is re-run from scratch by a session that did not write the code; make reproduce J=<n> rebuilds it.