Scored against what actually happened
1 · The near term: the schedule we source
Refueling dates are not public. Utilities do not announce them, no regulator filing carries them, and they cannot be scraped. We source them through the outage network NukeWorker has run inside this industry since 1999, corroborate them across independent sources, and publish them as a firm schedule. Published here means published by us.
Scored against the day each reactor actually dropped power, the dates we published run 99.7 out of 100 fleet-wide. That is the front of the curve, and it is the part no one else can rebuild from public data.
2 · Once it starts: how long it really runs
Outages run past their published end date, and the published number does not tell you which ones or by how much. NukeGryd projects the extension per unit, in days, from fleet-wide history blended with that unit's own recent overrun record.
Our projected end date beats the utility's published one 59.6% of the time, and scores 83 out of 100 on overrun fleet-wide. That is the difference between knowing a unit is scheduled back and knowing when it is actually coming back.
See which units are projected to run long| NukeGryd overrun score | 83 / 100 |
| We are closer than the utility | 59.6% of outages |
| Start-date score | 99.7 / 100 |
Rescored 14 Sep 2026.
3 · Out to 18 months: the model
Beyond the dates we already hold, the model takes over, and it is not a fuel-cycle countdown. It scores every US reactor on more than thirty signals: its own cycle length and how tightly it holds to it, the month it habitually goes down and even the day of the week it prefers, how its owner sequences outages across the rest of their fleet, reactor supplier and containment type, capacity factor, forced-outage and scram rate, unit age and license runway, and how long that unit's recent outages have actually run. Owner-specific weights are fitted on top of that.
That is why it still holds up a year and a half out, where a cycle calculation does not. The forward curve publishes it as a P10 to P90 band, so the uncertainty is on the screen rather than hidden inside a single date.
| Start date | 4.1d |
| Duration | 7.5d |
| Start date, all years on record | 5.4d |
| Duration, all years on record | 7.3d |
The first two rows cover outages from 2023 onward; the two beneath them cover every outage on record. The recent window is better because the model is, and both are published so the difference is ours to explain rather than yours to discover. Scored out-of-fold: every projection graded by a model that never saw that outage.