The Signal Lab is the research program behind Fire Nearby. Around the clock, our fusion engine watches independent early indicators of fire, and every candidate it raises is tracked to a verified real-world outcome. This page is where we publish what that work is teaching us, updated as the research progresses.
Ember is the early wildfire detection model at the center of the Signal Lab. Before the detail below, here are two ways to get the whole picture: a short silent film, and a podcast episode that walks through the same ground in plain language. Both are drawn from the research published on this page, and neither reports results from Ember’s sealed evaluation window, which is graded no earlier than January 2027.
Film · 4:31 · No sound
Meet Ember
What Ember is and how it is being tested, start to finish: what it does in plain words, a replay of one real fire, why almost all satellite heat is not the fire that matters, the evidence every score weighs, the live early-signal map, what five years of outcomes taught us, and the sealed test that decides whether Ember ever reaches the public map. Built to be read without sound.
The Coleman Creek and Silvertip segments are replays: the frozen Ember v3 model re-scores July 2026 fires from its validation period using only the evidence available at each moment. They are examples, not the verdict. Map footage is from our internal early-signal map, captured 11 October 2026.
Podcast · Episode 1 · 14:40
Ember, Explained
A plain-language conversation about Ember: how a five-year labeled dataset was built from 17 million satellite fire detections, what its three probabilities mean, why its top score stops near 51%, what it deliberately leaves out, and how sealed, pre-registered tests, starting in January 2027 and running through the year, decide whether it ever reaches the public.
The Signal Lab podcast · Episode 1
Ember, Explained
Fire Nearby · 14:40 · Two AI-voiced hosts
0:00-14:40
In this episode
Read the full transcriptTap a timestamp to listen from there
Host 1: Right now, somewhere over the United States, satellites are passing overhead and flagging hot spots on the ground.
Host 2: How many hot spots?
Host 1: On a busy day, thousands. A farmer burning off a field, a prescribed burn, a gas flare, and every so often, the first hour of a real wildfire.
Host 2: And from orbit, those can look almost the same.
Host 1: To the satellite, they're all just heat. Ember exists to answer the question that comes next. Does this one matter?
Host 2: This is Fire Nearby Labs, with a plain-language walkthrough of Ember: what it is, how it was built, and how it's being tested. Okay, let's start with why this is a needle-in-a-haystack problem. How much heat are we talking about?
Host 1: The team pulled five full fire seasons of NASA's satellite fire detections, from August 2021 to August 2026. That's about 17 million individual detections.
Host 2: Seventeen million.
Host 1: Then they grouped nearby detections into candidates, using the same rules the live system uses today. That produced almost 890,000 candidate heat clusters. Of those, about 3,600 became named wildfires.
Host 2: So, fewer than 1 in 200.
Host 1: About four-tenths of 1%. That's the base rate Ember works against. Treat every hot spot as a wildfire, and you'd be wrong more than 99 times out of 100.
Host 2: And there's a twist in where that heat shows up.
Host 1: The places where satellites see the most fire are often where it least often means a wildfire. The Flint Hills in Kansas light up every spring with pasture burns. The Mississippi Delta lights up with rice stubble. In about two-thirds of the map cells that see repeated heat, five years went by without a single named wildfire.
Host 2: While up in the Northern Rockies, a lot less heat is a lot more likely to be the real thing.
Host 1: Exactly. The same signal means opposite things depending on where it lands. Which is why raw detection can't make the call on its own.
Host 2: And meanwhile, the official picture takes time to catch up.
Host 1: It can. Official maps generally wait until a fire has been reported and recorded as an incident. That can be hours after it starts, especially at night or in remote country. Ember is built for that gap.
Host 2: So, what is Ember, exactly?
Host 1: Ember is a model that scores each new satellite heat cluster in the U.S. as it forms, and scores it again as more evidence comes in. It doesn't just say fire or not fire. It gives three probabilities.
Host 2: Walk me through them.
Host 1: The first is concern. How likely is it that this becomes a wildfire someone has to respond to, the kind that ends up in the official incident records? The second is real fire. Is the ground actually burning? The team checks that after the fact, against burn scars in Sentinel-2 satellite imagery.
Host 2: Why keep those two separate?
Host 1: Because a lot of real fire never needs a response. A well-run prescribed burn is real fire. It's just not a fire of concern. Blend the two together, and you teach the model the wrong lesson.
Host 2: And the third?
Host 1: Growth. The chance it becomes a named fire that reaches 1,000 acres or more. There's also a shorter-horizon version: the chance it reaches 100 acres in the first 24 hours.
Host 2: And these numbers are calibrated. What does that buy you?
Host 1: It means you can read them literally. If Ember says 20%, then across many cases like that, about 1 in 5 should turn out that way. And here's a detail I like. The highest concern score Ember will print is about 51%.
Host 2: The score never goes higher than that? Even for a big, obvious fire?
Host 1: Not in this version. The probability scale is only allowed to claim what its own validation data can support. In that data, the very top group of candidates became fires of concern about half the time. So that's the ceiling. An honest 51 beats a confident 90 that the data can't back up.
Host 2: So what actually goes into a score?
Host 1: Seven families of evidence. The heat itself: how intense it is, how many pixels, and whether it's still there on the next pass. Fire weather and the next day's forecast. How dry the fuels are, and how deep the drought is. Lightning and recent rain. Fuels, land cover and terrain. Homes and power lines nearby. And five years of fire history at that spot.
Host 2: And what's not on the list?
Host 1: Anything that amounts to a person noticing the fire. Dispatch logs. Aircraft over the scene. Reports from people on the ground. The Fire Nearby map shows some of those as context, but they never raise an Ember score.
Host 2: Why leave out such useful information?
Host 1: Two reasons. The whole point is to judge the heat before anyone has reported it. A model that leans on a dispatch log is mostly telling you what the dispatcher already knows. And those human signals are part of how Ember gets graded. You don't let the answer key into the exam.
Host 2: Fair enough. Are there rules built in, too?
Host 1: A few guardrails. More measured fire power can never lower a score. Faint heat can't be oversold: under 1 megawatt, the concern score is capped at 3%, and under 5 megawatts, at 10. And a spot's fire history can count against a candidate, but never for it.
Host 2: What kind of model is this under the hood? Some giant neural network?
Host 1: Deliberately not. It's a constrained statistical model, a regularized logistic regression with a few hundred learned weights per score. You can trace every score back to the factors that moved it, which is why every candidate on the team's signal map comes with its reasons.
Host 2: Let's talk about how it learned. Where did the training data come from?
Host 1: This is the part that took the most work. The team had to build the dataset first: every individual satellite heat candidate, paired with what it actually became. They replayed five years of satellite archive through the live engine's exact rules, in time order, so every candidate formed the way it would have in real time.
Host 2: No hindsight?
Host 1: None. Each candidate is scored at fixed ages, from 1 hour out to 48. And at each age, the model only sees what was knowable at that moment. There's even a mechanical check that no training example contains information from its own future.
Host 2: And how do you decide what each candidate became?
Host 1: Positives come from official incident records. Negatives get checked against burn scars, because heat with no fire record isn't automatically a false alarm. In the team's audit, about 3 in 10 unnamed clusters had really burned, mostly managed burns. Truly ambiguous cases are left out rather than guessed.
Host 2: So what did five years of outcomes teach the model?
Host 1: That physics leads. The strongest single signals, like very intense heat, forecast winds over 30 miles an hour, or steep slopes, each turn into wildfires at roughly 4.5 to 5 times the base rate. Very dry air and dry dead fuels, about 3.5 to 4 times. And those factors overlap, so the model learns them together.
Host 2: Okay. How do you know it works?
Host 1: In two stages, and the second one hasn't happened yet. Stage one was internal validation. The model was fit only on data from before May 11, 2026, then tested on the summer that followed, through early August. Months it never trained on.
Host 2: And the results?
Host 1: Six hours into each candidate's life, take one that became a fire of concern and one that didn't. Ember ranked the real one higher about 94% of the time. For the statisticians, that's an AUC of 0.945.
Host 2: That's strong.
Host 1: It is. And the team ran tests designed to catch themselves fooling themselves. Retrain the model on scrambled outcomes, and it collapses to a coin flip, which is exactly what you want to see. Give it only bookkeeping, like which data fields happen to be filled in, and it learns nothing.
Host 2: Are all three scores that strong?
Host 1: No, and it's worth saying cleanly. Concern is the strongest. Whether the ground really burned turns out to be a harder question, and that score's validation results are much more modest. The growth score is newer, and it has its own registered test.
Host 2: So what's the catch with stage one?
Host 1: Those same summer months were also used to set the probability scale. So they're not a clean final exam. The real verdicts come from sealed tests nobody has looked at. We'll get there. But first, an example case study that highlights the potential of Ember.
Host 2: Please.
Host 1: This is from that validation summer: July 25, 2026, Harney County, Oregon. At 1:56 in the morning, a NOAA-20 satellite pass catches two hot pixels in the dark. There's no official record of any fire.
Host 2: Two pixels.
Host 1: That's right. In the replay, the frozen model rescores that moment using only what was knowable then. A minute after the pass, Ember has it at about 51% concern. That's its top band, more than 100 times the base rate.
Host 2: From two pixels? Why so confident?
Host 1: Because of everything around them. The Hot-Dry-Windy forecast was in its top bin. Humidity was forecast to fall to 5%. Gusts to 40 miles an hour. Dead fuels at 4.6% moisture. Severe drought. The physics said danger from the first two pixels.
Host 2: When did the fire become official?
Host 1: The official discovery time was 5:40 that morning, which matches the first entry in the local dispatch log. That's three hours and 44 minutes after those first pixels.
Host 2: And what did it become?
Host 1: The Coleman Creek Fire, a lightning start that made a 12-mile, wind-driven run and burned more than 300,000 acres, about 480 square miles.
Host 2: That's a striking gap in timing.
Host 1: It is. And it comes with two honest footnotes. First, other satellites saw heat, too. NOAA's GOES satellites picked it up in nearly every five-minute scan from 2:16 on. The heat wasn't secret. What was missing was a judgment that this particular heat mattered. And second, it's one example from the validation summer. It's one highlight, not the verdict.
Host 2: Okay. Now the test nobody has looked at.
Host 1: This is the heart of it. The team runs Ember like a clinical trial, not a demo. The rules were written down and frozen before this version was trained: which candidates count, which metrics, and exactly what bar it has to clear.
Host 2: And then the clock started.
Host 1: Right. Every candidate that formed between August 23rd and November 21st, 2026, is scored by the frozen model. Every live prediction is timestamped the moment it's made, and it's never revised.
Host 2: And the results are locked away?
Host 1: Literally. The team's own tools refuse to grade that window, train on it, or even summarize it without a password. And the first grading code won't run before January 5th, 2027, so the outcome records have time to settle. That date is enforced in code, not by promise.
Host 2: What does Ember have to do to pass?
Host 1: It's graded head-to-head against its predecessor, Ember version 2, on the same fires. It has to win on precision, with 95% statistical confidence. Its probabilities can't get any less accurate. And every probability band has to hold up. Every result is published, pass or fail. But January is only the first verdict, not the last.
Host 2: Wait. So January isn't the final exam?
Host 1: No. Version 3's verdict lands in January, alongside two companion tests on the fall: one for the full system with its filters, and one for the growth score. And here's a rule people miss. If a test saw too few fires to judge fairly, it isn't forced to an answer. By a rule written in advance, it keeps collecting and gets graded at a later date.
Host 2: And after January?
Host 1: There's a sitting at least every quarter through 2027. April grades a longer fall and winter window. July grades a prescribed burn guard, a winter correction, a GOES-first candidate, and the actual product features, like map markers and opt-in alerts. October gives second chances. And because next summer is peak fire season, two more models start collecting on June 1st: a summer recalibration of version 3, and a full retrain, version 8. Those are graded in November and December of 2027.
Host 2: So it's a whole roadmap, not one test.
Host 1: Exactly. Every verdict, pass or fail, feeds the next model.
Host 2: Has anything failed this way already?
Host 1: Yes, and that's kind of the point. Since version 3 was frozen, the team has registered and trained several possible successors, including versions that add NOAA's GOES satellites, which watch the whole country every few minutes. So far, none of them has cleared its bar. So none of them shipped, because a bar you can't fail isn't a bar. One GOES-first candidate, frozen exactly as it is, now gets its own sealed test, graded no earlier than July 2027.
Host 2: Say a test passes. Then what?
Host 1: Then it's eligible. Shipping is a separate, deliberate step. Fire Nearby today is a live wildfire map and an iPhone app: official fires, perimeters, satellite hotspots, smoke and fire weather, with every data point timestamped and sourced. Ember isn't public yet.
Host 2: And when it is?
Host 1: Early-signal markers on the map, right beside the official fires. Opt-in early heads-up alerts in the app, which are already built and switched off until Ember earns them. And eventually an API, so agencies, utilities, and insurers can use the probabilities directly.
Host 2: How do you keep people from treating it like an official alert?
Host 1: The wording is part of the plan. Every marker says it isn't an official alert. It states the rate measured in testing, and the fact that most flagged heat is not a wildfire. And nothing in it ever tells anyone to wait, or not to call 911.
Host 2: And if it starts slipping once it's live?
Host 1: There's a public scoreboard and an automatic rollback. If the real-world hit rate falls clearly below what the wording promises, that marker type switches itself off and the team gets paged.
Host 2: So when could people actually see it?
Host 1: Later than you might guess. A public marker can prompt calls to dispatch, and dispatch records are part of how these tests are graded. So the team decided not to put new-fire markers in front of the public while any of its sealed tests is still running. That means the summer of 2027 at the earliest.
Host 2: So they're protecting the test from their own product.
Host 1: Exactly. And in the meantime, the team keeps recording things that are hard to get after the fact, like how dispatch centers reclassify calls and what aircraft are doing over fires. That's how they'll eventually measure how often satellites miss a fire entirely, and how much lead time is real.
Host 2: If someone remembers one thing about Ember, what should it be?
Host 1: That it's an honest second opinion on satellite heat. Three probabilities, with the reasons attached. Built on five years of labeled outcomes, and graded on sealed tests nobody has peeked at, starting in January and running through 2027.
Host 2: And the goal?
Host 1: Know which heat matters while the fire is still small.
Host 2: The research, the dataset findings, and the testing calendar are published at firenearby.com/signal-lab. Thanks for listening.
Narrated by two AI voices from a script written and fact-checked by the Fire Nearby team. The transcript is the episode as spoken.
What we’re building
Official wildfire maps wait for an incident to be declared, which can lag hours behind a brand-new fire. We are working on the opposite end of that gap: identifying likely new fires from open signals alone, as early as the data physically allows, and proving how often we get it right. The work runs in three loops:
Fuse
Independent signals (satellite heat, satellite fire-detection AI, live dispatch traffic, lightning, air-quality anomalies) are merged into scored candidates when they agree in the same place and window of time.
Verify
Every candidate is matched against official incident records.[4] Hits record how many minutes we were ahead of the official discovery; candidates that never become fires are logged as misses.
Learn
The labeled outcomes become training data. A statistical model learns which combinations of evidence actually precede real fires, and each version has to beat the last on held-out data before it ships.
A five-year labeled dataset, built from scratch
A live system only accumulates evidence at real-time speed, so in August 2026 we rebuilt the past instead. We swept five years of archived satellite fire detections[1] through the exact clustering, suppression, and scoring rules our live engine uses, then matched every resulting cluster against official incident records[4] to label its outcome. Each cluster was also stamped with the conditions at the moment it formed: archived lightning activity,[2] drought status,[3] and the location’s own fire history, and in August 2026 the enrichment was completed with hour-matched historical weather, two weeks of trailing rainfall, dead-fuel moisture, terrain, vegetation condition, and land cover for every confirmed fire plus a large population sample. The result is a research dataset that does not exist anywhere else.
17.1M
satellite heat detections swept from the NASA FIRMS archive
887,730
candidate heat clusters formed by the same rules as our live engine
3,605
clusters matched to officially named wildfires under the audited, burn-scar-verified match rule
5
full fire seasons covered, August 2021 through August 2026
8.27M
lightning grid cells sampled from the GOES Geostationary Lightning Mapper archive
260
weekly US Drought Monitor snapshots stamped onto clusters as they formed
Only 0.4% of these clusters ever earned an official fire name, a figure we re-verified in August 2026 by cross-checking claimed matches against satellite burn scars and tightening the match rule accordingly. That imbalance is the core of the problem. Almost everything hot a satellite sees (controlled agricultural burns, industrial heat, small fires that are out before anyone logs them) never becomes an incident, and a useful early-warning system has to tell the 0.4% apart from the rest.
What five years of outcomes taught us
With every cluster labeled, we can measure exactly how much each piece of evidence shifts the odds that a satellite heat cluster becomes a real, named wildfire. The figure below shows the strongest observed condition for each factor, expressed as a multiple of the 0.43% baseline conversion rate for US clusters (95% confidence intervals computed on every bin):
Figure 1. Conversion lift by factor, strongest bin vs the 0.43% base rate
Peak heat intensity (FRP)5.3x
150+ megawatts
Forecast wind, next 24 hoursNew4.8x
30+ mph
Terrain steepnessNew4.6x
15-30 degree slopes
Air dryness at detection (VPD)New3.9x
30+ hPa
Dead-fuel dryness (100-hr fuel moisture)New3.6x
below 8%
Fire history at that location3.5x
10+ fires/yr nearby
Persistence across satellite passes2.3x
5-8 overpasses
Drought severity2.0x
D4 exceptional drought
Population: 887,730 satellite heat clusters, Aug 2021 to Aug 2026, measured against the corrected, burn-scar-verified labels (August 2026 audit). Factors marked new come from the environmental enrichment and are measured on a population-weighted sample (every confirmed fire plus a 1-in-37 sample of the rest); their confidence intervals exclude 1.0x. Factors are not independent of one another; the model below learns them jointly.
Physics leads
The strongest signals are physical: how hot the detection burns,[1] how hard the wind will blow over it, and how steep the ground under it is. Detections above 150 megawatts convert at 5.3x base rate, the faintest at 0.4x; a 30+ mph forecast wind nearly matches that. These outrank every bureaucratic or geographic cue, which is exactly what an honest label should produce.
History still matters
A heat cluster in a cell that historically sees 10+ fires a year is 3.5x more likely to become a named fire, while a cluster somewhere with no fire history converts at a quarter of the base rate. Where fire has been tells you a lot about where it will be, just less than the physics of the moment.
Lightning surprised us
The literature on lightning-initiated fires[5] led us to expect recent strikes to raise the odds. On their own they do not: heat near strikes from the prior 24 hours converts slightly below heat with no strikes in two weeks, largely because thunderstorms bring rain. Rainfall context explains the inversion: strikes that arrived with wetting rain converted at 0.5x base rate, while strikes hitting ground that had stayed dry for two weeks ran near 3x, a six-fold spread (the dry-strike sample is still small, so we hold that number loosely). Findings like this are why we measure instead of assume.
Geography ties all of these findings together. In the map below, every dot is a place where satellite heat clustered over the five years. The size of the dot is how much heat the satellites saw there; the color is how often that heat went on to become a named wildfire. The same raw signal means almost opposite things depending on where it lands:
Figure 2. Five years of satellite heat, judged by outcomes
Share that became named wildfires:Never became a wildfireUnder 1% (near base rate)1-4%4-10%10%+ fire country
3,307 half-degree cells covering 773,110 US heat clusters, Aug 2021 to Aug 2026, judged by the corrected labels. Dot area scales with cluster count; cells with fewer than 5 clusters and Alaska/Hawaii are not shown. Nearly two-thirds of all cells (2,125) never produced a single named wildfire in five years. Notice that the biggest dots on the map are green: the places satellites see the most fire are the places it least often means a wildfire, and that broad green heart of the country is the subject of the next section.
Mapping the fires that are not wildfires
The biggest source of confusion in satellite fire detection is not technology, it is intent: enormous amounts of American landscape are burned on purpose, on schedule. Kansas Flint Hills ranchers burn pasture every April, Mississippi Delta rice country burns stubble every fall, and Idaho’s Camas Prairie burns after every August harvest. These are real fires that satellites correctly detect; they are simply never going to become wildfires with names and evacuation zones.
Because every cluster in our dataset carries a verified outcome, we could turn that burning calendar into a map. We identified the cells of the country where satellite heat returns in the same weeks across multiple years, yet five years of outcomes show it essentially never converting to a named wildfire. Burning rotates between fields from year to year, so the map is judged at neighborhood scale rather than field scale. The result held up better than we expected:
44,348
map cells flagged as recurring managed-burn areas, from the Flint Hills to sugarcane country
634,000+
historical heat clusters formed inside those flagged areas over five years
0.00%
of them became a named wildfire: not one, against a 0.43% national base rate
Figure 3. What the flagged cells are, by how much of the year they burn
55% burn 1-3 months a year: the classic agricultural and prescribed-burn signature, tied to harvests and spring green-up.
33% burn 4-9 months: longer burning seasons and mixed-use land.
12% are hot 10-12 months: year-round industrial heat, mostly oil-field flaring, caught here as a backstop.
This map is one layer of a deliberately layered defense. Known industrial heat (gas flares, steel plants, landfills) is identified separately and removed before candidates ever form, because a refinery can never be a new wildfire. Seasonal burn country is different: an escaped field burn is a real ignition source, so Ember treats the map as evidence to weigh rather than a reason to look away. A candidate in the Flint Hills in April starts heavily discounted, but if independent signals keep piling up, it can still surface.
Meet Ember, our early-detection model
The five-year dataset trains Ember, the statistical model at the center of the lab. It began with a single question, the probability that a brand-new heat cluster becomes an officially named wildfire, and has grown into a set of calibrated probabilities covering a fire’s whole opening arc. The name fits the job: an ember is the small, persistent heat that exists before anyone calls it a fire, which is exactly the moment this model lives in. Each Ember release is versioned, so you can always see which iteration produced a number and what changed between them.
The first two Ember versions, v0 and v1, were trained before the August 2026 label audit and were retired with it: models are only as honest as their labels, and theirs had been over-crediting distant coincidental matches. Ember v2 was the first version trained end to end on the corrected, burn-scar-verified labels, and it now serves as the baseline. Ember v3, its recalibrated successor trained under a frozen, pre-registered protocol, is the current model: it scores every live candidate, and every prediction is logged and tracked to a verified outcome.
One candidate, three questions
Ember is one system asking three questions about the same heat cluster, each answered by its own calibrated probability. All three ride the identical pipeline: the same candidates, the same full-picture context, the same discipline of logging every prediction and grading it against a verified outcome under a frozen protocol. They differ only in the question they answer.
P(concern)
Will this become a fire of concern?
The flagship number. Trained on official wildfire outcomes with negatives certified by Sentinel-2 burn scars, it estimates the chance that brand-new heat becomes a fire someone must act on. This is the probability our internal dashboard ranks by, and the one whose holdout evaluation is running now.
Ember v3 artifact · shadow-scoring every live candidate since 23 August
P(real fire)
Did the ground truly burn?
Plenty of real fire never earns a name: a landowner knocks it down, a local crew handles it, no record forms. This companion head predicts physical burning as judged by Sentinel-2 burn scars, an evidence source independent of naming bureaucracy. It keeps the concern head honest and gives real-but-unnamed fire a score of its own.
Same v3 artifact · shown beside P(concern) on the internal dashboard
P(growth)
If it is real, how big, how fast?
The newest heads (31 August 2026). Growth g0 is a separate model trained on reconstructed growth history (daily incident-size reports and perimeter archives back to 2021). From first detection it stamps two probabilities on every candidate: reaching 1,000 acres ever, and reaching 100 acres within 24 hours, with claims deliberately capped while the evidence is still weak.
Growth g0 artifact · shadow stamps, graded prospectively in January 2027
Together they cover a fire’s whole opening arc: is this heat worth attention, is it physically real, and how large could it get. Every head is versioned, every stamp is logged, and none of the numbers reach the public map until its registered evaluation passes.
Model card: Ember v3Updated: 23 August 2026
Version
Ember v3 (live scoring)
Status
Scores every eligible live US cluster; every prediction is logged and tracked to a verified outcome. Nothing it says affects the public map yet
Population
US satellite heat clusters
Labels
Fires of concern, not just officially named fires: real wildfire outcomes with negatives certified by Sentinel-2 burn scars, and ambiguous cases excluded rather than guessed. Size-scaled incident matching, re-audited automatically as new fires arrive
Design
Time-sliced hazard framing (the model re-answers "is this a fire of concern?" at fixed ages: 1h, 3h, 6h, 12h, 24h, 48h) behind a full-picture gate: no probability is printed until weather, terrain, land cover, and fire-history context have all arrived
Two heads
One head predicts a fire of concern, P(concern); a second predicts whether the ground truly burned (satellite burn-scar label), P(real fire), because plenty of real fire never earns a name. The growth probabilities are a separate artifact, growth g0, scored on the same candidates
Internal validation (May-Aug 2026)
ROC-AUC 0.945 at the 6-hour checkpoint, with precision concentrated 160x over base rate; calibration error 0.05 points. All pre-registered stress gates passed
Headline evaluation
Reserved. A pre-registered untouched holdout window (23 August to 21 November 2026) is scored once, in early 2027, after outcome labels mature. Negative controls must pass first
Why you don’t see Ember’s numbers on the map yet
Ember scores every eligible live cluster and every prediction is logged and tracked to a verified outcome, but nothing it says affects what the public map shows. Strong results on historical data are a promising start, not proof; a model earns its way onto the map only by proving itself on live fires it has never seen, in an evaluation whose rules were written down before training began. That bar applies to every head the same way: P(concern), P(real fire), and P(growth) all run in shadow until their registered evaluations pass. Those holdout windows are running now and are scored once, in early 2027.
Accountability, in public
A prediction system you can’t audit is just marketing. Everything on Fire Nearby is timestamped and sourced, and Ember will be held to the same standard: when it goes live, this page will carry a running public scoreboard of every call it makes, correct and incorrect, with lead times on the hits and an honest count of the false alarms. Until then, here is the live research pipeline at work:
6,622
early signals fused in the last 24 hours
72,219
live candidates tracked to a verified real-world outcome
Detection is becoming free. Judgment is not.
Wildfire detection is entering a golden age. New satellite constellations are being launched to scan every point on Earth for fires within minutes, and AI camera networks already spot many ignitions before the first 911 call. We think that is excellent news, and not just for the obvious reason: every new sensor is another raw signal our fusion engine can ingest.
But our five-year dataset shows why detection alone was never the hard part. Satellites already flag millions of heat anomalies, and only 0.4% ever become a named wildfire. As sensors multiply, that flood of raw detections grows, and the scarce resource shifts from seeing heat to knowing which heat matters. A response system cannot roll trucks to every warm pixel; it needs a trustworthy probability, and a reason to trust it.
Our position: as detection becomes abundant, the product is the judgment layer on top of it. A calibrated probability for every candidate, verified against official outcomes, with the track record published for anyone to audit.
That judgment cannot be launched on a rocket. It has to be learned from years of candidate-versus-outcome history, each candidate stamped with the conditions at the moment it formed, and that record only grows in real time. Every sensor that comes online makes the sensing side cheaper and the outcome record more valuable, which is exactly why the Signal Lab logs its own calls, right and wrong, every day.
Methods, briefly
For readers who want the rigor without the full paper:
Temporal holdout. Ember v3 trains only on data through early May 2026; its headline evaluation is a pre-registered, untouched holdout window (23 August to 21 November 2026) scored exactly once after outcome labels mature, so its headline numbers describe prediction, not memorization.
No peeking at the future. Every training example is featurized only from signals that had arrived by that moment, verified by a deterministic proof (deleting later signals must reproduce the identical features), and the fire history feature is built strictly from records that predate the evaluation window.
Labels are audited by machine. In August 2026 we cross-checked our claimed detections against Sentinel-2 burn scars, found the incident-matching radius over-crediting, corrected the rule, and re-derived every label. A standing automated audit now re-verifies claimed credit against the incident record, burn scars, and a displacement placebo on an ongoing basis, and a failed audit freezes our public claims until explained.
Uncertainty is reported. Every conversion rate in the audit carries a Wilson 95% confidence interval, and thin bins are treated as inconclusive rather than as findings.
The label is conservative. Ember v3 predicts fires of concern, and a claim only counts in its favor once an official record or a satellite burn scar proves the fire. Plenty of real fires are knocked down by landowners or local crews before any record exists, and every one of those counts against the model until physical evidence lands. Measured precision is therefore a floor, not a ceiling, on precision against real fires.
Known limitations. The historical dataset is satellite-driven; the live engine also sees dispatch traffic and other fast human signals that archives cannot reconstruct. Canadian clusters are currently excluded from training because official Canadian labels lag by years. We publish limitations here for the same reason we timestamp everything else.
Research log
New entries are added as the work progresses.
18 August 2026
Five-year dataset completed; Ember v0 trained and moved to shadow evaluation
Rebuilt five fire seasons of satellite history through the live engine’s rules (887,732 labeled clusters), audited every scoring assumption against outcomes, corrected the ones the data disagreed with (heat intensity now weighted in production), and trained Ember v0 (holdout AUC 0.878). The model now scores live clusters silently while we compare it against the current system.
18 August 2026
Managed-burn map built from outcomes; Ember v1 protocol pre-registered
Mapped 43,141 cells of recurring managed-burn country directly from five years of verified outcomes (0.00% wildfire conversion across 627,000+ clusters) and began stamping it onto live candidates. Also froze the Ember v1 evaluation protocol before any v1 training run: population, metrics, and the exact bars v1 must clear are now written down in advance, so its results will be predictions rather than selections. Collection of v1’s dry-lightning ingredient, precipitation history at every candidate, is underway.
19 August 2026
Ember v1 trained; both models now shadow-score every live candidate
With the environmental enrichment finished across all five years, we trained Ember v1 on the full feature stack. In rolling backtests (train on the past, test on a full unseen fire year, repeat for three years) it scored AUC 0.849 to 0.863, and its sharpest top-0.1% slice converted to named fires at 25 to 29 times the base rate. The single biggest gain came not from weather but from the managed-burn map: knowing which heat is routine agriculture removes false alarms that no forecast can. v1 now scores live clusters beside v0; its one-shot holdout evaluation, the promotion decision, runs in early September when outcome labels mature.
21 August 2026
Labels audited against physical ground truth; Ember rebuilt as v2
Cross-checking our claimed detections against Sentinel-2 burn scars, an evidence source independent of every human reporting system, showed the incident-matching radius had been over-crediting: a distant, coincidental incident could claim a cluster that never burned. We corrected the match rule (size-scaled radius, discovery time window, nearest incident wins), re-derived every historical and live label, and made the check permanent as an automated audit. Ember v0 and v1 were trained on the superseded labels, so both were retired without their planned promotions. Ember v2, trained on the corrected labels with a time-sliced design and a second physical-burn head, now shadow-scores every live candidate, and every number on this page was re-measured against the corrected labels.
23 August 2026
Ember v3 live; ground truth completed; the tracking record opens
Ember v3, a recalibrated successor trained under a frozen, pre-registered protocol, began scoring every live candidate with a full-picture gate: no probability is printed until weather, terrain, land cover, and fire-history context have all arrived. An audit prompted by two missed Florida fires found the national incident feed silently dropping active fires (1,326 of 1,678 recent wildfires were missing or deactivated), so we rebuilt ground truth: the year-to-date national record with our own activity rule, plus direct state forestry feeds (Florida, South Carolina, Kentucky) and Alaska’s interagency feed. Heat from the flanks of already-known fires is now mechanically attributed to those fires instead of being graded as a new claim. With the answer key complete, our internal tracking record started fresh on August 23; a retrain on five years of recovered state-fire labels was registered, trained, and honestly failed its ship bar (certification lag in the newest labels), so it waits for the labels to mature rather than shipping on a technicality.
26 August 2026
Population rules built on independent ground truth; a fourth state feed
In the days after v3 went live, engineering audits added a set of registered population mechanics to grading. Heat on the flanks of already-known fires is attributed to those fires using corrected multi-part perimeter geometry and a footprint buffer. Military live-fire ranges get a dedicated channel: satellite-only heat on a published DoD range with no human signal and no incident match is excluded rather than graded. Confirmations proven only by a burn scar leave the concern-graded numbers (the physical-burn head keeps them), including dispatch-center dispositions that arrive late. Texas A&M Forest Service incidents now flow in as a fourth direct state feed beside Florida, South Carolina, and Kentucky, and every dispatch entry is kept in a permanent log so feed coverage itself can be audited later. One standing rule governs all of it: a population filter may key only on independent ground truth (incident geometry, published military boundaries, Sentinel-2 burn scars, dispatch dispositions), never on the model’s own inputs, and no decided cluster is ever restamped.
29 August 2026
January grading pre-registered as two populations, side by side
The single-shot January evaluation will report every metric twice, on identical rows: a raw column with no filters, which remains the registered headline for the shared August-to-November window, and a filtered column with the population rules applied. The filters are symmetric by construction: an exclusion removes would-be wins exactly as it removes would-be false alarms, and every excluded count is published. Registering both columns now, in advance, means the choice of which population to quote can never be made after seeing the results.
30 August 2026
Ember v5 registered: the whole system, frozen before its window opens
v5 registers no new weights. It registers the system: the byte-identical v3 artifact operating inside the audited engine, with four population filters (flank attribution, military ranges, cross-border heat, scar-only confirmations) as registered rules rather than engine policy. Its evaluation window is fully prospective: 1 September to 21 November 2026, opening two days after the freeze, sitting entirely inside the untouched holdout, graded once in January 2027 in the same sitting as v3. Registered alongside it: a grading runner that refuses to run before 5 January 2027 and fails closed without authorization; ground-truth coverage tiers, so states where our label network is thin are reported as such instead of silently counting against the model; a named canonical database (the always-on collector, so overnight laptop sleep cannot bias the population); and an underpowered clause: if an unusually quiet fall produces fewer than 300 decided positives, the window extends once, automatically, rather than letting a thin sample masquerade as a verdict.
30 August 2026
History mining opens the v6 docket; the cross-border filter goes into v5
A mining pass over the pre-lock five-year history (nothing after 10 August was touched) asked where the model stamps meaningful probabilities on clusters that structurally cannot become named US fires. The clearest theme: heat just across the Canadian and Mexican borders carrying a US state label, 3.1% of all historical clusters, matching at 0.50% against 2.96% nationally, because US incident feeds cannot see across the border. That cross-border exclusion was adopted into v5 before its window opened, with its symmetric cost disclosed: 121 would-be wins are excluded along with the noise. Two further candidates (recurring agricultural practice burns keyed on burn-scar recurrence, and a second tier of intermittent industrial heat) are documented for v6 with the constraints that kept them out of v5. Also measured, on history alone: filtering by itself shifts per-band observed rates upward by 0.3 to 4.1 points, a denominator effect now on record so that in January it can never be mistaken for a calibration change.
31 August 2026
The growth program: from “is it real?” to “how big, how fast?”
Ember now asks a second question about every candidate: if this heat is a real fire, will it become a big one? An engineering audit found the day-by-day story of every fire’s growth was being thrown away (official feeds keep only the latest perimeter per fire), so the engine now archives every perimeter shape it ever sees, plus aerial infrared mission shapes from California’s intel flights, burned-area footprints from Sentinel-2, and five-minute power curves as outage insurance. On the historical side, daily incident-size reports and perimeter archives back to 2021 were joined to the five-year cluster history, producing growth labels for every pre-lock cluster. On those labels we trained growth g0, two calibrated probabilities stamped on live candidates from their first detection: reaching 1,000 acres ever, and reaching 100 acres within 24 hours. It shipped the same way every Ember model ships: negative controls first (its first artifact failed two stress gates and was mechanically blocked from stamping until the fix passed), probability claims capped by what weak evidence can support, and a frozen protocol registering a fully prospective evaluation window (1 September to 21 November 2026) graded once in January 2027. The stamps are shadow-only and dashboard-only: growth numbers reach the public map only after a passed registered evaluation, announced here.
1 September 2026
Planning for noise: Ember v5.1 registered before its window opens
A contingency question surfaced an arithmetic fact about January: v5’s band-honesty criterion is a strict point test, and a probability band can fail it on sample noise alone. An observed rate of 4% in the 1-to-3% band is statistically indistinguishable from a true 3% at any sample size a single fall can produce; separating them cleanly would take roughly 1,500 decided clusters in that one band. The discipline here leaves only one honest response: v5’s bar cannot be touched now that its window is open, so it stays exactly as frozen and will fail on noise if noise comes, with the confidence interval printed beside every band so the run log can say which case occurred. Instead, we registered a successor the same day. Ember v5.1 is the identical system, graded over a longer window (2 September 2026 to 28 February 2027) with the band criterion in its statistically correct form: a band fails only when its whole 95% confidence interval sits outside the stated range. The freeze predates the window by a day, so it is fully prospective; the January sitting is registered in advance as the single interim look, with every parameter fixed now so nothing learned there can steer the April 2027 grading. The failure plan itself is also pre-registered while still outcome-blind: if a band truly misses, the fix is recalibration validated on a fresh window, and the band edges never move to fit a graded result.
8 October 2026
Geostationary satellites for earlier detection: four models, one live test
Ember v3 can only score a fire after a polar-orbiting satellite passes over it, sometimes hours after ignition. GOES geostationary satellites scan the US every five minutes, so we built four models to use them: v6 and v6.1 added GOES evidence to Ember’s existing candidates, and v7 and v7.1 score fires GOES sees before any polar pass. Each had one pre-registered training run on historical data, and none cleared its ship bar. The final v7 candidate, frozen unchanged, now scores those fires live in shadow; nothing it says reaches the map or alerts. It is graded once, no earlier than 15 July 2027, on fires from 22 November 2026 to 31 May 2027, and earns a place only if it flags real fires before official discovery that Ember v3 would have missed, without overstating its confidence.
8 October 2026
Release rules written before any results: how Ember would reach the public
Passing an evaluation shows a model is accurate; it does not decide what the public sees. We wrote that part down in advance too. Each possible public use (a map layer, a push alert near a saved place, a daily watchlist, a “could grow large” label on a fire page) gets its own narrow, measurable claim with a minimum sample and a bar it must clear, and the draft registration covering them passed an independent review whose findings were all fixed. It will be frozen before its window opens on 22 November 2026, and each claim is graded once, no earlier than July 2027, after that month’s other registered sittings. The rules that decide what counts as shown and when are a single shared piece of code used by both the grader and the product, so the thing graded is the thing that would ship. Two release rules are decided as well: no public new-fire surface goes live while any registered window is still open or ungraded, and any surface that does go live is paused automatically if its measured accuracy falls below what it claimed or it fires too often. The product code is built and deployed switched off; nothing on the map, in alerts, or on this site has changed.
19 October 2026
A winter correction for Ember’s percentages, registered before its window
On five years of history, all from before any test window, Ember’s earliest claims on a candidate come true less often than its later claims at the same stated percentage, mostly outside summer. Rather than wait for a sealed test to fail on that, we registered the fix in advance: Ember v5.3 reads each printed score through a frozen correction table fit only on that history. The model itself does not change, only how its numbers are stated. The correction passed a pre-registered safety check on history before it was frozen, and it is graded on candidates first seen from 22 November 2026 to 31 May 2027, no earlier than 15 July 2027.
31 October 2026
The 2027 fire season plan: two updated models, graded on the whole summer
Summer is peak fire season, so Ember should not spend 2027 without learning from 2026. Two updated models will be frozen by 31 May 2027 and scored live from 1 June. The first keeps Ember v3 exactly as it is and refits only the translation from its score to an honest percentage for summer. The second, Ember v8, is a full retrain on more data, with a go or no-go decision on 15 May; if it is not ready, the first carries the season alone. Both are graded on candidates first seen from 1 June to 30 September 2027, no earlier than 15 November 2027. If the summer is too quiet to judge fairly, the rule is set now: each collects through October once and is graded no earlier than 15 December. Training data from this fall stays sealed until its own tests are graded, and is released day by day through a written ledger that the training code checks and refuses to run without. We also wrote down how passed results become features: a model whose percentages pass may back the map, alerts, or the watchlist, but only for the season it was tested on, and a passed public feature may later be made stricter, never looser. The public scoreboard will report early-signal calls shown from 1 November 2027 on. The 2027 summer then becomes training data for the 2028 model, frozen before 1 June 2028.
Planned: 2027
Sealed test sittings through 2027, and promotion decisions
5 January: the pre-registered single-shot evaluation of Ember v3 (against the v2 baseline) on the untouched 23 August to 21 November 2026 window, after outcome labels mature (burn-scar labels freeze 45 days after a fire), with pre-registered negative controls required to pass first. The same sitting grades the v5 system and the growth model on their 1 September to 21 November windows. 15 April: v5.1, the same system on its longer 2 September to 28 February window under the interval-form honesty criterion. 15 July: the winter registrations on 22 November 2026 to 31 May 2027: the prescribed-burn guard, the winter correction, the GOES-first model, and the release claims for each public feature. 15 October: second chances for release claims, on summer data. 15 November: the two 2027 summer models. Most registrations also fix in advance what happens if too few fires occur to judge them fairly: they keep collecting and are graded at a later sitting. If a model clears its bar, it becomes eligible to put its numbers on the public map; we flip that switch deliberately, not automatically, and announce it here.
[3]US Drought Monitor, National Drought Mitigation Center, University of Nebraska-Lincoln / USDA / NOAA. droughtmonitor.unl.edu/
[4]NIFC Wildland Fire Interagency Geospatial Services (WFIGS): official incident records and discovery times. data-nifc.opendata.arcgis.com/
[5]Schultz, C. J. et al. (2019). Spatial, Temporal and Electrical Characteristics of Lightning in Reported Lightning-Initiated Wildfire Events. Fire 2(2), 18. doi.org/10.3390/fire2020018
Dataset statistics and model metrics on this page are from the Fire Nearby historical replay and evaluation pipeline, August 2026. Questions about the methodology are welcome: firenearbysupport@gmail.com.
See the signals yourself
The early-signal layers this research is built on (satellite heat, 5-minute geostationary scans, live dispatch reports) are already on the live map, and the About page documents every source and its freshness in real time.