How the EmberCheck score is calculated
EmberCheck gives your home a wildfire risk score that changes with the weather. This page explains what goes into that number, what we have measured about how well it works, and the things it cannot tell you.
What the score is
A number from 0 to 100 for your address. It is recalculated through the day as new conditions arrive, so it moves with the weather rather than describing a fixed characteristic of your property. Most of it reflects conditions that change day to day: how dry the fuels around you are, how hot and windy it is, and whether a fire is already burning nearby. The rest reflects things about the property that change slowly.
The score is an index for ranking days and places, not a probability. A 74 does not mean a 74 percent chance of anything. It means today at this address ranks high against the same address on other days, and against other addresses under the same model.
That is the difference between EmberCheck and a hazard map. A hazard map describes the long-run character of a place and says the same thing in January that it says in September. EmberCheck answers a narrower question: is today unusual here.
What it is computed from
Four kinds of information: measured weather and fuel-dryness conditions at the property, the terrain it sits on, the vegetation around it, and whether fire is already burning nearby.
We publish those categories, not the recipe. The specific products behind each one, and the weighting, combination and thresholds that turn them into a number, are not published. What is published is the result and how it was tested, which is the part you can hold us to. The known limitations of the terrain and vegetation terms are listed with the validation result below.
Two things are deliberately missing. We do not score your defensible space or how your home is built, because neither can be measured remotely from satellite or federal data, and pretending otherwise would be guessing at the two things you most control. We also do not use your insurance history or any personal data.
Computed, not generated
The score is the output of a fixed formula applied to measured readings. Those readings come from instruments and satellites operated by government agencies, the same public observations fire agencies work from. The same formula is applied at every address.
No language model writes this number. No part of it is invented, and nothing is filled in by inference when a reading is missing: when data we rely on is unavailable, the product says so rather than estimating around the gap.
It is also what makes the result testable. A number that was generated could not be checked against what actually happened. This one was, against 87 real fires, and the misses are published below beside the catches.
The California validation
The test we hold ourselves to is not whether the model looks sensible. It is whether, on the day a destructive fire actually started, the model had already flagged that location, and whether it managed that without flagging everywhere else.
We took every California fire from 2015 to 2025 that destroyed at least 25 structures, as recorded in CAL FIRE's damage inspection data: 87 events. We scored each on its day of ignition and compared against a standard fire-weather-index threshold, holding both to the same false-alarm rate so the comparison is fair.
In a California validation of 87 destructive-fire events, PRECEDE flagged 73.6% (64/87) versus 50.6% (44/87) for a standard fire-weather-index threshold at the same 6.27% false-alarm rate. PRECEDE scores point-precise weather at each location while the index uses 0.25-degree reanalysis; the asymmetry is a property of the comparison.
Caveats that travel with this result
These are quoted verbatim from our internal claims register, including its instructions to ourselves. Where the register's wording is technical, our plain-English restatement follows it.
- CV-ASPECT-UPSLOPE
The v13 terrain aspect term is applied to an upslope-convention field (180° from the convention its bands assume) and the slope input is inflated ×1.3–2; the validated performance includes both as built. Correction is a post-season preregistered change; the served EmberCheck per-property modifier was corrected 2026-09 (ASPECT-FIX-1) without touching v13.In plain terms: the terrain part of the model reads the direction a slope faces on the opposite convention to the one its own rules assume, and it treats slopes as steeper than they are. The validated result was produced with both of those in place, not corrected for them. The model is frozen for this season and the correction comes after it, under its own preregistration; the separate per-property part of EmberCheck has already been corrected. - CV-FUEL-VINTAGE
The static fuel-type term in the validated CA scores uses LANDFIRE LF2022 labels, which post-date some fires in the 2015–2025 validation panel; a pre-era (LF2016) re-labeling changes no catch and no miss at the T65 operating point (bounded re-scaling, not a re-score). The served grid's fuel labels are a LF2022/LF2024 mix scheduled for post-season homogenization.In plain terms: the vegetation labels used in the validated scores were published after some of the fires in the test had already happened, so for those fires a label can describe the ground as it was afterwards. Re-running the check with labels from before that period changes none of the catches and none of the misses, though that check was a bounded re-scaling, not a fresh scoring run. The labels on the live grid come from more than one edition and are scheduled to be made consistent after the season. - CV-PROXIMITY-EXCLUDED
The validated anchor EXCLUDES the active-fire-proximity term that the served scorer carries: the frozen backtest scorer has no proximity parameter in its signature, so the term is absent BY CONSTRUCTION — never state that the term was zero-valued. The gap is one-directional: the anchor can only UNDERSTATE the served score on a day with a nearby active fire.In plain terms: see “Two asymmetries we will name ourselves” above, which states this one without the register’s shorthand.
Exactly what this covers
- Where. California only.
- Which fires. Destructive fires, meaning at least 25 structures destroyed. Not all fires, and not fires measured by acreage.
- When. 2015 to 2025, each fire scored on its day of ignition.
- Operating point. A fixed threshold of 65, model v13-lightning, frozen.
- Reproducibility. The PRECEDE result reproduces from the frozen code and inputs. The comparator figure is cited from the same locked record and was not re-run.
- What it is. An index, not a calibrated probability.
Two asymmetries we will name ourselves
PRECEDE scores point-precise weather at each location while the comparison index uses 0.25-degree reanalysis. That asymmetry is a property of the comparison, and part of why the two numbers differ.
Separately, the frozen benchmark excludes the proximity term that the live product carries. Because that term only ever adds to a score, the published figure can only understate what the served model does on a day when a fire is already burning nearby, never overstate it.
Where it works, and where it is only watched
EmberCheck monitors wildfire conditions across 11 western states: California, Oregon, Washington, Arizona, Idaho, Montana, Nevada, New Mexico, Utah, Colorado, and Wyoming. Coverage describes where we monitor — it is not, on its own, a promise of detection or accuracy for any specific place or fire.
The validation above is California only. We have not published an accuracy result for any other state, and we will not imply one. Scoring runs the same way in all 11 states.
PRECEDE's approach has been evaluated most rigorously in California, where the historical fire record is richest. Any accuracy figures we publish are specific to that California evaluation and do not transfer to other states.
What EmberCheck does not do
- It is not an evacuation system. Evacuation orders come from your county and your fire agency. Follow them, and follow them ahead of anything on this site.
- A low score is not an all-clear. Destructive fires start on ordinary days. The model missed 23 of the 87 California fires it was tested on.
- It does not know your house. Defensible space and home hardening are the two biggest things you control, and neither is in the score.
- It is not insurance advice. It does not determine whether you can get a policy, or what you will pay for one.
- When a data feed goes down, we say so. An empty fire-activity panel means we have no current detection data, not that there are no fires near you.
For live information during an active fire, Watch Duty is free, fast, and run by people who do it well. We would rather point you there than pretend to be it.
Attribution
These sources are named because their licences require attribution, and because EmberCheck already credits them where their data is shown.
- NASA FIRMS: satellite fire detections
- NIFC: confirmed wildland fire incident records
- CAL FIRE: damage inspection (DINS) records, used as the validation labels above
This is an attribution list, not a description of the model. The weighting, combination and thresholds are not published.
Corrections
If you find an error on this page, tell us. We will fix it and say what changed. Every figure here traces to a frozen, dated record, and the result above is the only accuracy figure we publish anywhere. See also where EmberCheck works and how PRECEDE works.
Methods last updated September 2026 · Model v13-lightning