A fire vanished from the federal report at 98% contained — without ever reaching 100%.That is exactly what the Autumn Lane Fire did at 97%. For nineteen reports in a row this site has refused to call that silence “contained,” because no one ever said so. This week another fire proved that refusal right. Still 97%, and this record stays open.

A verified record of the fires that burned into the city of Spokane on August 1, 2026 — what happened, what the air did to three towns across one state, and why. Every number here carries a source. Where no source could be verified, this site says so instead of guessing.

Most recent verified figures ·
What this is

A logbook, not a news feed.

Verified only. Every figure is attributed to a named source and stamped with the time it was reported. Wildfire numbers change hourly, so an acreage figure without a timestamp is close to meaningless.

Where credible sources disagree — and on this event they disagree a lot — both numbers are printed. Nothing is averaged into a tidy single answer. Today's clearest example is Spokane's own air quality, where two views of the same monitor returned 63 and 183 within hours of each other. Both are on the page.

Gaps stay visible, they do not all get filled, and the ones that get filled are usually filled by something breaking. On the air quality page you will see days marked NO VERIFIED DATA. Those are real gaps in the public record, left open on purpose. August 8, August 9, August 13 and August 16 were each published with no fire figures at all because no federal situation report existed on those mornings; all four have been backfilled onto their own dates, as have the August 20, August 22 and August 25 editions, and now the September 3 and September 4 ones. August 24, August 26, August 28 and now September 4 and September 5 are completely blank for air quality in all three cities, and nothing is interpolated across them. September 4 and September 5 are the first two consecutive blank days in this record, which matters for how the chart reads: the line between September 3 and September 6 spans seventy-two hours and two unsampled days, and that segment is drawn, not measured. The repair for all five is the clean air agency’s monthly quality-assured report; the August file is still not posted, checked again today for a fourth time, but the July one confirms the format — daily AQI by date and by station, across twenty-two county monitors — so the fix is real and is waiting on a file. September 4 and September 5 also lost their weather, which had survived every previous missed day but one. The forecast version window is the only route to a missed day’s meteorology, and when it was walked on September 6 every version returned an Area Forecast Discussion instead of a Fire Weather Planning Forecast. That is the mirror of a failure logged three days earlier, when the discussion URL served a planning forecast; both directions of the same crossing have now been seen, and both were seen in one run. It cost two days of weather and it paid for one thing in return: the discussion it served by mistake is the first NWS product this site has read that names debris flows near burn scars. An absence here is always phrased as what could not be read, never as what does not exist.

The Sources & Methods page lists exactly which data could not be obtained and why — and which of this site's own working assumptions have since been proved wrong.

It ends when the fire does — and this week another fire showed exactly how a fire can leave this report without ever ending on paper. This record closes when the Autumn Lane Fire — the largest of the three — reaches 100% containment. Nineteen consecutive national situation reports, August 18 through September 5, have now been read, and not one names Autumn Lane, Old Trails or Fairview anywhere; sixteen of the nineteen were read on two or more independent paths, and none contains the bare word Spokane. Three of those nineteen were read in a single sitting on September 6, because this site did not run on September 4 or September 5 — which is also why two of them rest on one path each. The corroboration was not unavailable. It was missed. The prediction published here in advance came back a hit for a twelfth time: Black Fox, Sandy and 4V were all printed at 100% on the September 2 edition and all three are absent from the next one. Every one of the twelve went the same way — printed at 100% once, then gone. Autumn Lane did not do that. It was printed at 97% on August 17 and was simply absent the next morning. AND THIS WEEK THE OTHER HALF OF THAT QUESTION FINALLY GOT AN ANSWER. Kaiser Canyon — 138,293 acres, the third-largest fire in Washington — stood at 99% on September 2, at 98% on September 4, and is absent from September 5. It never printed a 100. So reaching full containment is sufficient for leaving this table, and it is not necessary. For eighteen editions this site argued that Autumn Lane’s silence could not be read as containment, and that was a refusal to infer. It is now an observation about a fire in the same table, the same region and the same week. The rule this site understands well runs one way only, and running it backwards would have put a 100% on this page that no source has ever published. Still 97%. This record stays open.

So the most likely explanation for the silence is not that this fire was finished, but that it stopped being federally reported — most plausibly transferred off federal incident management back to state or local control, which is ordinary at this stage of a fire and which the situation report does not announce. That is an inference, it is labelled as one, and it is not a containment figure. This page keeps both honest readings and weights them unequally. One: a federal product drops an incident when it is finished, so fourteen read editions of removal amount to a statement of completion. That reading is the weaker of the two by a wide margin, because the product demonstrably marks completion by printing 100% first — watched four times, once against a prediction made in advance — and it never printed 100% here. Two: the published figure is 97%, no source anywhere has said this fire is contained, controlled or out, and the Spokesman-Review’s four-byline one-month feature of August 30, updated August 31, states no containment status either. This site will not manufacture a number it has never seen. The closing rule was written on the assumption that the number would be published, and that assumption has now failed for twenty-four days. Today added a reason for caution about all of it. Re-reading two editions this site had already read turned up a fourth fire at 100% that yesterday’s run missed entirely, corrected a regional count from zero to three, and showed two ways of questioning the same PDF returning different answers. The absence claim survives that, because it asks whether a name occurs anywhere in a document rather than trusting a table to be listed completely — and knowing exactly which of your claims survives a bad extraction is the difference between a record and a rumour.

Spokane's fire record

How this one compares.

Incident commander Tom Clemo, on August 2: “Probably the largest, most destructive fire in Washington's history occurred yesterday afternoon.”

Day by day

Thirty-eight days.

The Spokane Area Fires complex ignited on Saturday, August 1, 2026, the day after the National Weather Service issued a fire-weather warning it had never used in Washington before. This is the record as it was reported, day by day, with the weather that drove each one. Twenty-three of these days were first published with an empty fire column — August 8, August 9, August 13, then August 18 through September 6. The first three have since been backfilled from federal editions that appeared later, as have August 16, August 20, August 22 and August 25. The nineteen editions from August 18 to September 5 cannot be backfilled, because there is nothing to backfill them with: every one exists, every one has been read, sixteen of the nineteen on two or more independent paths, and none contains a figure for any of these fires. September 3 and September 4 rest on a single path each, and September 6 was read on two. August 23, September 3 and September 4 are the three days still resting on one path — and two of those three are single-path because this site was not running when they were live, not because a second path was unavailable. September 6 is empty for the ordinary reason: today’s edition has not been published yet. August 27 was the exception for three days — no edition for that date appeared anywhere — until the archive published five editions at once on August 30 and it turned out to have existed all along. That is the distinction this timeline is built to keep: what could not be read, never what does not exist.

Incident inventory

Three fires, one complex.

The three fires were consolidated under the designation Spokane Area Fires at 7:30 a.m. on August 2 and managed as a single complex by California Interagency Incident Management Team 7, a Type 1 national team. Sources also call it the Spokane Complex. A fourth fire near Medical Lake, the Espanola Fire, burned at the same time but is not part of the complex.

Where the three fires burned, and how much of each was contained
Every circle is drawn to scale: its area is the real burned area, on the distance scale in the corner. The ring around each circle is the containment — the share of the fire’s edge that crews had surrounded and held. A complete ring means 100%. Hover or tap any fire for its numbers.
Fire positions use the coordinates this site already uses for its smoke model. A real fire is a ragged shape, not a circle — these discs show the correct AREA and the correct PLACE, not the true outline, because no source read here published a mapped perimeter. Final acreages: NIFC, Aug 16–17.
Acreage over time, by fire
Each point is one report from a named source, with its own timestamp. Read the shaded band first: the three fires did almost all of their growing in about four days, then stopped. Old Trails went from 50 acres to 3,509 on August 1 alone. Autumn Lane reached 5,869 by August 4. After that the lines are flat because the fires really had stopped spreading — not because nobody was looking. Hover any point for its acreage and the time it was reported.
REDESIGNED SEPT 6, at the site owner’s request. Earlier versions of this note said the axis could not be made readable without a hand-changed label rule, and declined to make that change unsupervised. That change has now been made and is recorded here rather than done silently: every day still gets a gridline, but only every second day is labelled, which is what stops the dates colliding. The axis ends just past August 17. Nothing after August 17 is missing by accident. Nineteen consecutive federal editions, August 18 through September 5, have been read and none carries a figure for any of these three fires, so no point can be plotted and none has been. A day with no federal figure earns no point — the rule that made the August 13 blank meaningful. The August 8, 9 and 13 columns were empty on the day and were backfilled later from editions that appeared afterwards. Fairview’s line stops a day short of the other two because it reached 100% on August 16 and left the table first. The length of the silence since August 17 is deliberately NOT drawn as a shape here, because a stretched empty axis is easy to misread; it is kept as a number in words, and day by day on the timeline, with each edition named. Downward steps are remapping corrections, not fire being undone. NIFC revised Old Trails down 47 acres and Fairview down 64 on Aug 6, Autumn Lane down 60 and Old Trails down 13 on Aug 7, Autumn Lane down 26 and Old Trails down 5 on Aug 8, Old Trails down 330 on Aug 9 — the largest correction of the event — and Autumn Lane down 8 with Old Trails down 50 on Aug 10. Note what that does to a tidy story: this note once said the corrections were getting steadily smaller and converging on a real perimeter. The next one was ten times larger than the one before it.
Response

Who was on it

Incident teamCA IMT 7Type 1 national team · IC Tom Clemo, deputy IC Nic Elmquist · assumed command Aug 2. RE-AGED SEPT 6: no source has reported a transfer of command, a demobilisation or a closeout for this complex, and the federal report has not named any of its fires in nineteen editions. Nationally the team count has now gone 16 CIMTs and 1 NIMO, then 9 and 0, then 16 and 1 again across three consecutive editions — a fall of seven and a rise of seven in two days, from narrative sentences rather than raw rows. One of those three readings is probably an extraction error and this site cannot tell which. All three stand, and none of them says anything directly about this complex.
Personnel61NIFC sum of the two fires still listed, Aug 17 — now nineteen read federal editions old, and staffing went up before it went away. The complex read 55 (Aug 15) → 69 (Aug 16) → 61 (Aug 17) and then vanished from the report. For scale, the Sept 2 edition assigns 1,716 people to Little Giant and 1,352 to Sinlahekin, roughly twenty-eight and twenty-two times this figure. RE-AGED SEPT 3: twenty of the twenty-four Northwest rows on the Sept 2 edition reconcile on both the acreage change and the personnel change; four print --- and are excluded rather than counted as passes, because an unrunnable check is not a passed check.
Injuries3 burnedCNN, Aug 8: three firefighters have suffered burns, “one seriously enough to be hospitalized twice and sent to a burn center.” RE-AGED SEPT 6: still single-sourced and uncorroborated after twenty-nine days. NIFC’s report carries no injury column at all, so federal silence here is not disagreement — it is the absence of a field. Nothing published since adds a casualty figure, and three differently-worded searches on the arson case today returned nothing newer than Aug 26.
Containment line75+ miMiles of containment line constructed across the complex as of Aug 8 (Spokesman-Review, 8:58 p.m.). NOTE, RE-AGED SEPT 6: no figure has been published since, so this is a twenty-nine-day-old number and is not a current measurement of anything. It is kept because it is the only line-construction figure anyone published, and removing it would hide that nobody published another.
National Guard~240RE-AGED SEPT 6 — no figure published since. Guard members in Spokane, per Maj. Gen. Gent Welsh on Aug 10 — up from 110 deployed statewide on Jul 31, plus 27 helicopter crew activated earlier in July. Twenty-seven days old and never updated. No source has reported a drawdown, and no source has reported them still being there; this site does not know which. The one staffing figure that did move today belongs to a different agency: EPA reports approximately 250 specialized personnel on hazardous-material removal, up from the 185 hazmat specialists in 14 teams reported on Aug 26.
Aircraft & equipment0 aircraftNo aircraft assigned anywhere in this complex since Aug 14, and nineteen read editions have named none of these fires since. The helicopter count went 12 → 12 → 0 across the Aug 12, 13 and 14 reports: Old Trails gave up all seven at once after holding them for six consecutive reports while shedding people, and Fairview gave up all five. Autumn Lane, the largest fire, never had a single aircraft assigned on any federal report. RE-AGED SEPT 3. Nineteen days old and, like everything else here, frozen by the absence. The helicopter column is also where this site published a phantom structure loss on Aug 31 by misreading it — the correction has now held on four consecutive editions.
National preparednessLevel 4CHANGED SEPT 4, AND THIS SITE MISSED THE DAY IT HAPPENED. The national preparedness level fell from Level 5 to Level 4 on the Sept 4 edition, after standing at 5 for twenty-five consecutive days, and holds at 4 on the Sept 5 edition. It is recorded on the date it was printed, not the date it was read. The Northwest Area also sits at preparedness level 4. The Sept 5 edition reports 78 uncontained large fires nationally, 6 new large incidents and 65 new initial-attack fires, with 18 Complex Incident Management Teams and 1 NIMO committed — a figure from a narrative sentence, which has now read 16/1, 9/0, 16/1 and 18/1 across four editions and is published with that history attached every time.
Declarations

State: Gov. Bob Ferguson declared a state of emergency on Saturday, August 1, including a statewide burn ban through September 30, 2026.
Federal: President Trump approved an Emergency Declaration on Tuesday, August 4 — not a Major Disaster Declaration. It covers six counties and three tribal nations and provides a 75% federal cost share for emergency measures such as evacuation support, shelters, food and water, and search and rescue. Debris removal was requested but not approved. A Major Disaster Declaration request was expected to follow after damage assessments; as of August 10 no such declaration has been reported, and its damage assessment is the thing that would eventually reconcile the five competing structures counts on this site.

Cost to date, Aug 17 — final

The final federal cost figures. NIFC’s Aug 17 report puts spending at $18 million on Old Trails and $13.1 million on Autumn Lane; Fairview’s $1.4 million is its Aug 16 figure, after which the fire left the table. $32.5 million across the complex — and it did not move on Aug 16 or Aug 17, which breaks the one pattern this site had watched hold on every single report of the event. The run was $17.15M (Aug 8), $20.45M, $23.01M, $25.1M, $26.5M, $29.1M, $30.2M, $32.5M on Aug 15 — and then flat, twice. Cost had risen on every prior report, including the one on which staffing fell 77%. Two flat reports at the end are consistent with a complex whose remaining work is mop-up rather than suppression; they are also the last cost figures anyone will publish for these fires. Old Trails remains the most expensive fire of the three and the most destructive; it was never the largest.

Air quality · three towns, one state

Spokane is the cleanest of the three cities for a second reading running — but the order behind it has swapped, and the two days that would explain why are the two days nobody sampled.

On the evening of August 4, three Washington Department of Ecology monitors read 151 in Spokane, 121 near Kirkland and 31 near Chimacum — a spread of about five times. By the morning of August 9 the same three instruments read 179, 16 and 14: a spread of about thirteen times. On the morning of August 14 they read 30, 28 and 16 — a spread of 1.9 times, and all three inside the Good band on the same day for the first time in this event. August 15 has only one of the three. Spokane returned to Moderate at 58 — a Spokane Regional Clean Air Agency county-wide maximum stamped 1:35 p.m., flagged on the point because it is not the Augusta Ave instrument that carries the rest of this line. Kirkland and Chimacum have no August 15 value at all: this record’s run executed at 01:30 on August 16, their feeds serve only the current hour, and Sunday readings are not published on Saturday’s row. Spokane’s 30 on August 14 remains the only Good reading it has produced anywhere in this record, and it comes from the cleanest comparison this chart carries: the same station and the same 08:00 hour label as yesterday’s 65. The two proxy stations were sampled an hour later, at 09:00, so the three-way comparison is close but not the exact match August 13 managed — and both proxies read identical numbers to yesterday, which is flagged on the rows rather than smoothed over, because an unchanged figure is exactly what a cached one looks like. Two of the four station pages consulted today were in fact serving cached figures as current — one three days old, one four days old. The reading hour caught one and the category word caught the other, which is why this site reads two views of every station. And the agency that publishes Spokane’s official number is back online after yesterday’s outage, but its index still shows Thursday’s 89. That stale figure did the most useful work of the day: it is a county-wide maximum for Thursday morning, and it confirms to the point the 89 this site published yesterday over a station page that was serving 158. A city’s air is not a number. It is an instrument, at an hour, with a name.

The EPA Air Quality Index

Category names are printed everywhere alongside the colors, so the scale never depends on color alone.

The scale ends at 500. Any number above 500 is an unofficial extrapolation, not an EPA AQI. This matters for one figure on this event — see the caveat below the chart.

Verified AQI readings, July 31 – September 6, 2026

Why the air did what it did

Smoke concentration at any given place is not really about how big the fire is. It is about how much air the smoke gets mixed into, and where that air is going. Four things set that: the depth of the mixing layer, the wind, the terrain, and whether an inversion has put a lid on the whole thing.

The mechanism, in the words of Spokane's own air agency

The Spokane Regional Clean Air Agency describes a surface temperature inversion as “a layer of warmer air above cooler air near ground-level,” which traps pollutants “in cooler air near the surface.” Inversions form under “calm winds, clear skies, and long nights,” and “calm winds prevent warmer air above the surface from mixing down to the ground.” Critically for this event: “Spokane's topographical features can enhance the formation of inversions, especially in valley locations,” and inversions “can persist as long as several days” when a strong high-pressure system develops. That is exactly the pattern that set up over eastern Washington on August 3. And it is still operating on the best day of the event: on August 10, with an afternoon ventilation index of 971, the overnight figure still collapses to 47. The daytime air improved by 116 points; the nights have not changed at all.

And the reverse case, in the words of NWS Seattle

Kirkland sits on the wrong side of the Cascades to receive eastern Washington smoke — normally. On August 4 that reversed. Ridge amplification plus a thermal trough along the coast produced offshore easterly flow, which carried smoke westward through the Cascade gaps into the Puget Sound basin. The smoke arrived aloft first. NWS Seattle: “A shallow inversion aloft will break as surface temperatures warm, allowing smoke to mix down.” That sentence is the whole mechanism by which smoke 250 miles away became air that people in Kirkland breathed by Tuesday afternoon. Then the wind turned. By Saturday, August 8, NWS Spokane's transport wind was due west; on Sunday and Monday it was southwest, and on Monday twice as fast — either way, pushing smoke away from Puget Sound. The Bellevue proxy read 38 on Saturday and 16 on both Sunday and Monday, its lowest of the event. Same fires, opposite wind, opposite result.

Questions this data can answer
  1. August 1 had the strongest winds of the entire event, yet Spokane's worst verified air came on August 3 and 4, when winds were nearly calm. Why would less wind mean worse air?
  2. Spokane sits at 1,843 feet — the highest of the three towns — and it had the worst air. Chimacum sits at about 100 feet and had the best. So elevation alone does not explain air quality. What does?
  3. On August 4 the mixing height over Spokane fell from 7,500 feet to near the surface and the ventilation index dropped from 403 to 32. If the same amount of smoke is produced, what happens to its concentration when the volume of air it can mix into shrinks by that much?
  4. Relative humidity on August 4 read 16% in Spokane, 37% in Bellevue and 49% in Port Townsend. What single geographic fact explains that gradient, and how is it connected to Chimacum's AQI of 31?
  5. The highest verified reading of the event, AQI 465, came from a monitor at the Little Spokane River Natural Area rather than downtown Spokane. What does that tell you about using one monitor to describe a whole city?
  6. One outlet reported an AQI of 1,654 at Nine Mile Falls. The official scale stops at 500. Before you decide whether to believe that number, what would you need to know about the instrument that produced it?
  7. On August 5 the fires were still burning and still barely contained, yet Spokane's AQI fell from 151 to 109 at the very same monitor. Nothing about the fire explains that. Look at the mixing height and the ventilation index for the two days and say what changed — then explain why the overnight numbers mean the improvement was only temporary.
  8. The same Augusta Ave monitor read 109 at a 9:00 a.m. label on August 5, 134 at a 3:00 a.m. label on August 6 and 127 at a 1:00 p.m. label the same afternoon. A newspaper could truthfully headline the middle number as “Spokane's air quality worsens.” Using the ventilation index for that night, explain why that headline would be misleading — and say what measurement you would need instead before you could honestly claim the air got worse.
  9. On August 6 Spokane had a ventilation index of 730 (High) and the National Weather Service issued a Fire Weather Watch for two days later. On August 7 the index reached 952 and on August 8 it reached 1,057 — the best dispersion of the entire event — on the last day of a Red Flag Warning. Explain how the same weather can be good for air quality and dangerous for fire at the same time. What two variables do the ventilation index and a red flag warning share?
  10. NIFC's August 6 report shows Old Trails at 3,574 acres — 47 fewer than the day before — and every report since has shown further reductions: 60, 13, 26, 5, 330 and 50 acres. Fire does not un-burn ground. What must have happened? Then look at that sequence again. On August 9 this site wrote that the corrections were “getting smaller each day,” which was true of the three days it had. The next correction was ten times bigger. How many points do you need before a trend is a trend, and what should this site have written instead?
  11. The August 7 point on the Spokane line comes from a different monitor than every other Spokane point on this chart: Monroe St instead of Augusta Ave, because Augusta Ave had not reported that morning. Two options were available — plot the other station and say so, or leave the day blank. Which would you have chosen, and what does each choice cost the reader?
  12. Now use that answer. Spokane reads 153 on August 7 and 142 on August 8. A student writes: “Spokane's air improved by 11 points.” The August 7 reading is Monroe St; the August 8 reading is Augusta Ave. Is the student's claim supported? What is the smallest change in the data that would support it?
  13. And now check your own reasoning. On August 8 this site argued that a claimed 11-point improvement was indistinguishable from the known 11-to-17-point offset between Spokane's two monitors. On August 9 both monitors reported at the same hour and the gap was 2 points — Monroe St 181, Augusta Ave 179 — and on August 10 the gap was 8 points in the opposite direction: Monroe St 55, Augusta Ave 63. Was the August 8 caution wrong? State precisely what it did and did not establish, and explain the difference between “this evidence cannot show X” and “X is false.”
  14. On August 8 the Kirkland proxy fell to 38 (Good) and the Chimacum proxy to 15, while Spokane stayed at 142. Look at the transport wind direction in the meteorology table for August 4 and for August 8. Explain the western Washington numbers using only that one variable — and then say why the same variable does almost nothing for Spokane.
  15. The one perfectly clean comparison. On August 9 the Augusta Ave monitor read 179 at the same 8:00 hour label at which it read 142 the day before — same instrument, same hour, consecutive days, no caveat. Meanwhile containment rose to 61%, so there was less fire. Using the ventilation index, mixing height, wind speed and transport direction for the two days, explain how the air got 37 points worse while the fire got smaller. What does that tell you about which variable a smoke forecast should watch?
  16. The single-variable experiment. Compare August 9 and August 10 in the meteorology table. The mixing height is identical on both days — 7,000 ft by day, 6,500 ft falling to the surface at night. The transport wind roughly doubled, from southwest 5–15 mph to southwest 10–15 mph. The ventilation index went from 719 to 971, and the measured AQI at one station fell from 179 to 63. Since ventilation index is mixing height × wind speed, and the mixing height did not change, the entire change in the index is the wind term. Work out roughly what wind speed each index implies. Then answer the harder question: containment also rose on August 10, so there was less fire as well as more wind. Why does that make August 10 a worse day than August 9 for demonstrating that dispersion matters more than fire size?
  17. On August 10 Kirkland's proxy read 16 and Chimacum's read 14 — exactly the same as the day before — while Spokane fell 116 points. Two towns whose air did not change at all and one whose air changed enormously, under one wind field. What does that tell you about where the smoke was being made, and what would you predict for Kirkland if the transport wind reversed to easterly tomorrow?
  18. Two numbers for one monitor. On August 10 one view of the Augusta Ave station read 63 and another read 183. One carried a stated reading hour of 09:00 and the category word “Moderate”; the other carried only the time the web page was built. Spokane's clean-air agency separately published 161 at 6:35 a.m. as a county-wide maximum. Which of the three would you put on a chart, and why? Now write the sentence a newspaper should print about Spokane's air quality that day — in one sentence, without being wrong.
  19. On August 8 and again on August 9 there was no federal situation report at the time this site was compiled, so both days were published with no acreage and no containment; both were backfilled onto their own dates the following morning. The alternative would have been to repeat the previous day's numbers under a new heading and correct them later. Argue for the choice that was made. Then argue against it — who is harmed by an empty column?
  20. For five days this site relied on a rule it had worked out itself: the federal report is always about 24 hours behind, so a morning run will always be reporting yesterday. That was true on August 7, 8 and 9. On August 10 it was false — one NIFC path served the same day's report. What is the difference between a pattern you have observed three times and a property of the source? Name one other rule on this site that might be in the same position.
  21. On August 10 Spokane's own monitors read Moderate, and the Spokane County Air Quality Alert was still “in effect until further notice.” Western Washington's alerts, by contrast, were allowed to expire on August 6 and 7. Is the eastern alert wrong? What would you want to know about the overnight numbers, and about what an alert is for, before deciding?
  22. The question this whole site exists for. On August 11 every fire in the Spokane complex posted its largest containment gain of the event — Autumn Lane +16 points, Old Trails +20, Fairview +44 — and the complex got 12 acres smaller. On the same morning Spokane's AQI went from 63 to 129. Less fire, worse air, and the mixing height did not change at all. Write down every explanation you can think of before reading on. Which of them can be tested with a number on this site, and which cannot?
  23. The afternoon ventilation index fell from 971 to 792 on August 11 — about 18% worse dispersion — while the measured AQI rose 105%. If dispersion got 18% worse, can concentration double? Work out what else must have changed for the arithmetic to close, then read what Spokane's clean-air agency said the smoke was coming from. What term is missing from a ventilation index, and why would a forecaster who used only that number have got August 11 badly wrong?
  24. The Little Giant fire, 130 miles west in the eastern Cascades, grew 5,527 acres on August 11 and stands at 116,956 acres, 2% contained. The entire Spokane complex is 9,932 acres and shrank by 12 that day. Which fire is the bigger public-health event for the people of Spokane, and which one is on the local news? What does that tell you about how we decide which fires matter?
  25. Two monitors, one city, 57 points apart. On August 11 Augusta Ave read 129 (Unhealthy for Sensitive Groups) and Monroe St read 72 (Moderate) — a whole EPA category between two instruments a few miles apart. No reading hour could be found for the Monroe St figure. Why does that missing timestamp make the 57-point gap impossible to interpret? Name three different things the gap could mean, and say what single piece of information would let you tell them apart.
  26. The gap that closed without an answer. On August 11 Spokane's two regulatory monitors read 129 and 72 — 57 points, a whole EPA category apart. On August 12 three of the city's monitors read 153, 152 and 151 — two points apart. No outlet, agency or monitor operator published a correction to either day. Someone tells you the August 11 gap must have been a sensor fault, since the instruments agree now. Explain why that does not follow, and list what the August 12 agreement does and does not establish about August 11. Then answer the harder question: if a discrepancy in public data disappears without anyone explaining it, what should the public record say about it — and who is responsible for saying so?
  27. When the clock moves, what happens to your series? Every air quality reading on this site was collected in the morning — except August 12, which was collected at about 7:30 p.m. Its three hour labels are 17:00, 18:00 and 19:00: the closest together of any row in the record, and the least comparable to any other row. Spokane's ventilation index that evening was Moderate (490), falling to Low (122) overnight, against a High (694) afternoon figure. Spokane's reading rose from 129 to 153. Work out which direction the sampling change alone would push a reading, and by roughly how much you can and cannot say. Then decide what to do with the point: plot it, drop it, adjust it, or plot it with a warning. Argue for one, and name what each of the other three would cost you.
  28. Two agencies, two winds. On August 12 the Spokane Regional Clean Air Agency's alert said smoke “will continue to spread from west to east.” The Washington Smoke Blog said winds would “bring smoke from the north into the Columbia Basin.” The National Weather Service's fire weather forecast for the same day put the transport wind at north 10 to 20 mph, after three days of southwest. Washington had at least three fires over 120,000 acres burning at once, in different directions from Spokane. Yesterday this site attributed Spokane's smoke to one named fire. Was that justified on the evidence it had? Write the sentence it should have written instead — and notice that the correction was possible only because the original claim named a specific fire that could be checked.
  29. Two instruments that moved in opposite directions. On August 12 the Autumn Lane fire's containment stayed at exactly 89% and its assigned personnel fell from 551 to 283 — 49% of its staff sent home in a day. Across the complex, engines halved and helicopters did not change at all. Explain how all four of those can be true at once without any of them being an error. What is a containment percentage actually certifying, and what is a staffing level actually pricing? Which one would you rather know if you lived inside the perimeter — and does your answer change if you own the house rather than rent it?
  30. Spokane's clean-air agency printed AQI 170 with the words “Very Unhealthy” on August 11. On the EPA scale, 170 is Unhealthy; Very Unhealthy starts at 201. This is the fifth time that page has mislabelled its own number. Which does a reader act on — the number or the words next to it? Design a check you could run in under a minute, every time, to catch this yourself.
  31. The day the number was not there. On August 13 the federal situation report was not published, so this site has no acreage and no containment figure for that day. It would have been easy to repeat August 12's numbers under an August 13 heading, and almost nobody would have noticed. What exactly would have been wrong with doing that? Name the specific false claim a repeated number makes. Then look at the acreage chart: the line stops rather than running flat. Explain to someone else why a gap and a flat line mean different things.
  32. When the source of truth goes dark. Also on August 13, the Spokane Regional Clean Air Agency could not publish an air quality index at all — “State and federal monitoring sites are experiencing data issues” — and on the same morning one Spokane monitor served a two-day-old reading as if it were current while another had not reported since August 11. You still need to know whether it is safe to run outside. What do you do? List, in order, the checks you would make, and say which single check would have caught the stale reading. (The answer is on this site, and it takes about ten seconds.)
  33. The best comparison in the whole record. August 13 is the first day here on which all three cities were read at exactly the same hour — 08:00 — giving Port Townsend 16, Bellevue 28, Spokane 65. Every other row mixes hours. Why does that matter, and what does it let you say today that you could not say on any previous day? Then answer the harder half: Spokane read 153 the evening before and 65 this morning. Which pair of numbers in this record is the honest measure of how much Spokane's air improved — and how much of the 88-point drop is real?
  34. The number that came back late — and why waiting beat guessing. On August 13 this site published an empty fire column because no federal report existed. On August 14 that report appeared, and it showed Autumn Lane had gained eight points and Old Trails fourteen on the day the column was empty. Suppose the site had instead carried August 12’s figures forward, labelled “most recent available.” What would a reader have believed about August 13, and how many days would it have taken for anyone to notice the error? Now argue the other side: what does a reader lose from an empty column, and is that loss larger or smaller than the one you just described?
  35. Two checks, and neither one is enough on its own. On August 14 two station pages served old readings as if they were current. E Broadway Ave printed 158, “Unhealthy,” stamped Tuesday, while the live value was 25, Good — the category caught that one. Bellevue printed 16, Good, stamped Monday, while the live value was 28, Good — both are in the Good band, so the category check would have passed and only the reading hour caught it. Write the rule that catches both cases. Then explain why a site that only checked the number, and not its label, would have published a wrong figure on one of those two stations and never known.
  36. The best number and the worst reason. On August 15 the smoke model on the Explore page returned about 72 for Spokane against a measured 58 — an error of about 24%, its smallest of the whole series. The day before it returned 204 against a measured 30, an error of 580%. But between those two days the model fell 132 points while the measurement rose 28: they crossed. First question: if a model’s error can go from its worst to its best in one day by passing the measurement going the other way, what is error size actually measuring? Second: on August 15 the model predicts Kirkland at about 74 — higher than Spokane — because an east wind pushes its three Spokane-side fires toward Puget Sound. There is no August 15 reading for Kirkland to test that against. Is a confident, untestable prediction better or worse evidence about a model than a testable one it gets wrong? Third: the smoke that actually arrived in Spokane that day came from fires more than a hundred miles west that are not in the model at all. Can a model be right for the wrong reason, and how would you tell?
  37. The model went the wrong way on the best day. On August 14 the smoke model on the Explore page returned about 204 for Spokane while the monitor read 30 — an error of about 580%, its worst of the event — and it got there by moving up from 176 while the real air halved. The single input that changed was humidity, and drier air makes that model produce more smoke. Three fires were at 97% containment with no aircraft left. What variable is missing from the model, and why does adding more precision to the variables it already has not fix it? Compare this with question 30: is a model that is 26% wrong on one day and 580% wrong on another better or worse than one that is steadily 200% wrong?
  38. A model that is stable for the wrong reason. From August 16 through August 20 the smoke model returned 189 · 205 · 214 · 214 · 232 for Spokane against measured figures of 63 · 68 · 72 · 69 · 80 — five consecutive days with an error near 200%. A stable error looks like a working model. Why is it not? The model has no term for remaining fire, and its three fires — which are its whole picture of the state’s smoke — are now off the federal report. Its stability is measuring only its own missing variables.
  39. When you cannot recover a data class after the fact. On August 19 this site’s scheduled task did not run. The fire figures were recoverable — the federal archive is dated. The weather was recoverable — the NWS FWF product keeps a version history that reaches at least two days back. The air quality was not — the aqicn feeds serve only the current hour, and two of the three cities went null for August 19 because a returning-the-next-morning read cannot find yesterday’s local reading. Which property of a source makes its data recoverable after a missed run, and which does not?
Interactive model

Change one thing. Watch what moves.

Wind, an inversion, humidity and terrain are the four levers that decide where smoke goes and how bad it gets. Move them and watch the plume, the fire's behavior, and the air in all three towns respond. Start with a real day from this event, then change one variable at a time.

Read this first

This is a simplified physical model, not measured data and not a forecast. It is built to make relationships visible — inversion depth versus concentration, wind direction versus plume path, terrain versus transport. The arithmetic is printed in full at the bottom of this page so you can check it, argue with it, or break it. The measured readings live on the Air Quality page; those are the real ones. On August 10 this model was wrong by a factor of 3.5 on Spokane — its largest miss of the event — and the reason is printed with it rather than tuned away.

The direction the wind blows from. Smoke travels the opposite way. A west wind pushes smoke east, away from Puget Sound; an east wind pushes it toward Kirkland.
Does two opposite things at once: strong wind spreads fire fast, but it also dilutes and flushes out smoke.
A lid of warm air over cooler air at the surface. The stronger it is, the lower the lid, and the smaller the volume of air the smoke has to spread into.
Dry air means dry fuel and an active fire producing more smoke. Very humid air slows the fire down.
Switch them off to see what each one is actually doing. The Cascades block smoke transport across the state; the valley traps it near the ground in Spokane.
Smoke transport model for Washington State A map of Washington showing the fire origin near Spokane, a smoke plume drifting with the wind, and the three comparison towns; below it a west-to-east vertical cross-section showing terrain, the inversion lid and the depth of the smoke layer.
Try this

Forty experiments

  1. Report the result when your prediction finally gets tested, in the same words you used when it could not be. On August 31 this site predicted four fires would vanish from the next federal edition; on September 1 that edition did not exist and the result was published as untested — not confirmed, not broken. On September 2 the edition appeared and all four were gone. The prediction is now a four-for-four hit, and it is reported with exactly the prominence the untested day got. The discipline is not in celebrating the hit; it is in having given the null result the same headline. Try it on any forecast you make: write down in advance what would count as right, what would count as wrong, and what would count as could not tell — and publish whichever one arrives.
  2. When a rule you invented yesterday is falsified today, retire it and take back what you built on it. On September 1 this site worked out a test for whether a frozen web page’s reading could be trusted, based on how a station habitually labels its readings against its own clock — and used that test to recover a value it had previously discarded. On September 2 the same station behaved the opposite way and the test collapsed. The tempting move is to patch the rule and keep the recovered value. The honest move is to retire the rule and withdraw the value. Try it as a habit: when a method fails, go back and find every conclusion that rested on it, and say out loud which ones fall with it. The hard part is not admitting the error; it is tracing what the error was holding up.
  3. Write the prediction down, then report honestly when it cannot be tested. On August 31 this site predicted that four fires printed at 100% contained would be gone from the next federal edition. On September 1 there was no next edition — both live paths and the dated archive stopped at August 31. The tempting move is to quietly let the prediction ride until a day it happens to win. The honest move is to say it is untested: not confirmed, not broken. A prediction that is only scored on the days it succeeds is not a prediction, it is a decoration. Try it with a forecast of your own: write down what you expect, write down the date you will check, and then report all three possible outcomes — right, wrong, and could not tell — with the same prominence.
  4. Ask the same source the same question two different ways and see whether it answers the same. This site asked one federal report to list every fire that is 100% contained, and then asked the same file to copy out the rows exactly as printed. The list named three fires. The raw rows showed that two of the three were actually at 92% and 89%. Same document, same day, two answers. Summarising is a step where information goes missing, and asking for the raw thing is how you catch it. This is worth doing with any tool that reads for you, including the one you are probably using: ask for the summary, then ask for the actual sentence it came from, and compare them. Doing this on September 1 also turned up a whole fire the day before had missed.
  5. Write down what would have to happen tomorrow if your explanation is right, and then go and look. On August 30 this site noticed that two fires had each been printed at 100% contained and then dropped from the next federal edition, and inferred a rule: that is how a fire leaves this report. An explanation that only fits what has already happened costs nothing to believe. So the rule was turned into a claim that could fail: North Pole and Mitchell Springs, both printed at 100% on August 30, will be absent from August 31. On August 31 both were absent. Try the same move on something you already believe. Say what you would expect to see next week, in enough detail that you would notice being wrong, and write it down where you cannot quietly revise it. Then check. The value is not in being right — it is that a prediction you have written down is the only kind of belief that can lose an argument with the world.
  6. Learn to tell a broken page from a page that is preserving something. A website that has stopped updating looks like a dead end, and this site treated it as one for thirty days. Then it noticed the difference: a frozen page still carries a timestamp, and if that timestamp is coherent — if the page could actually have known what it prints at the moment it stopped — the reading it froze on is real data from that moment. On August 31 two frozen pages paid out. Port Townsend’s air quality page, stuck since August 26, turned out to be preserving a genuine midday measurement from August 25 that no live feed could ever have given back. Test the idea yourself: find a page that has clearly stopped updating, work out from its own timestamps exactly what moment it is frozen at, and ask whether the thing it shows could have been true at that moment. Sometimes it is a fault. Sometimes it is a photograph. The skill is knowing which — and the check is always internal consistency, never how recent the page feels.
  7. Test a claim that was made about the world against a claim that was made about what could be seen — then watch one of them turn out wrong. On August 29 this site wrote two things about the same missing federal report. One: it “has never been read on any path.” Two: “no edition for that date has ever existed anywhere.” On August 30 the archive published five editions at once and the August 27 file was among them. The first claim was true when written and is still true of that morning. The second was false the whole time. Read the RESOLVED entry at the top of the Gaps list, which keeps the original wording rather than quietly correcting it. Then ask yourself which of the two claims you would have written, and whether the difference between “I could not find it” and “it does not exist” is one you keep in your own work. Every conclusion this site reached rested on the first claim. None of them needed retracting.
  8. Find out how something leaves a dataset, and then check whether the thing you care about left that way. Federal situation reports drop fires all the time and never explain why. So look at what happens just before a fire disappears. Mustang was printed at 100% contained on August 28 and was gone on August 29. Picture Rock was printed at 100% on August 29 and was gone on August 30. That is the exit protocol: print the number, then drop the row. Now check Autumn Lane: printed at 97% on August 17, gone on August 18, never printed at 100% on any of thirteen read editions since. It did not leave the way finished fires leave. The August 30 edition prints 100% for North Pole and Mitchell Springs, so on this pattern both should be absent tomorrow — a prediction this page has written down in advance so that it can be wrong in public. The general move is the one worth keeping: when you cannot observe a thing directly, find cases where you can observe it, learn the pattern, and then ask whether your case fits it.
  9. Watch a scientific result get retired by one more click, and decide whether that is a failure or the method working. Two days ago this page recorded, in bold, that the smoke model had for the first time got the direction of a Spokane change right: load AUG 23 (187) and then AUG 27 (209), and the model went up while the monitor went up, 34 → 59. Now load AUG 29 and read 213. The model went up again. The monitor went down, 59 → 45. Nothing was tuned between any of these three clicks. (a) Given three clicks, how many ways are there to get one sign right by chance, and what would a genuinely working model have had to do instead? (b) Write down what evidence would have justified the Aug 27 write-up at the time it was written, and say honestly whether this page had it. (c) The site kept the Aug 27 claim on the page and marked it retired rather than deleting it. Argue the case for deleting it instead, then argue the case against, and notice which of the two arguments you can make without knowing anything about wildfires.
  10. Read the same document twice and find out what your first reading missed. On August 29 this site fetched the federal situation report and asked for every Washington and Oregon table row, verbatim. It got fifteen rows. It then fetched the identical file and asked about three fires by name — and got all three, including Sinlahekin, at 163,001 acres the second-largest fire in the state. Nothing about the file changed. (a) If you had only run the first read, what would you have believed, and how would you ever have found out you were wrong? (b) This site’s central claim is that ten federal editions do not contain the words “Autumn Lane.” That claim was tested by asking whether a name appears anywhere, not by listing a table. Explain, in one paragraph, why the row-omission does not damage the name-absence claim — and then explain why “I used the same tool for both” is exactly the reasoning that would have damaged it. (c) Design a third check that would catch a failure neither of the first two could. (d) Now apply all of this to your own work: name one claim you currently believe because a list you read appeared complete.
  11. Change exactly one input and watch a model get the sign right and the size wrong by a factor of six. Load AUG 23, note 187 / 10 / 2, then load AUG 27 and note 209 / 11 / 2. These two presets are identical in wind direction (southwest), wind speed (15 mph) and inversion (off). The only input that differs is humidity — 28% then, 22% now. The model moved Spokane up twelve percent. The monitor moved from 34 to 59, up seventy-four percent. So the model got the direction of change right, which it usually does not, and understated the magnitude sixfold. Then look at Kirkland: the model went 10 → 11 and the measurement went 11 → 10. Write one sentence on why the Kirkland pair tells you nothing at all, and one on why the Spokane pair tells you slightly more than nothing. The answer to the first involves the words “one point.”
  12. Find out whether your best check survives being pointed at a different kind of claim. This site’s strongest verification is reading the same federal edition on two independent paths. On Aug 27 that check was run twice on the Aug 25 report — the live path on Aug 25, the dated archive file two days later. On the question “does the phrase Autumn Lane appear anywhere in this document,” both paths agree: no. On the question “how many new fires did the Northwest Area report,” one path says 1 and the other says 7. Same document, same day, same two paths. Open the Conflicts list and read the entry, then answer this: if a check is reliable for absence and unreliable for arithmetic, what exactly is it a check on? The useful answer is not “the document” — it is something about how a long PDF is turned into text, and about which kinds of claim survive that translation.
  13. Watch one model be its worst and its best on the same click. Load AUG 23 and read all three numbers: 187 / 10 / 2. The measurements that morning were 34 / 11 / 10. So on Spokane the model is +450% — its worst since Aug 14 and the widest absolute gap in the record — and on Kirkland it is 10 against 11, a 9% error, closer than it has ever come to a west-side reading. Same click. Same untouched constants. Now work out why the Kirkland number is worthless: load AUG 22 and note Kirkland modelled at 3 against a measured 25. The model barely moved between the two days; the measurement fell to its record low and walked past a stationary output. Write down, in one sentence, what would have to be true for a single day’s agreement to count as evidence — then check whether any of it is true here.
  14. Score a real agency’s forecast, then decide what to do with the result. Open the Air page and find Aug 23. The Spokane Regional Clean Air Agency forecast Moderate air through Sunday and warned smoke might push the city to Unhealthy for Sensitive Groups overnight. Every Spokane-area instrument came in Good, between 21 and 38 — a miss of one category at the baseline and two at the warning. Now the harder question, which has no clean answer: the agency erred in the direction that makes people take precautions they turned out not to need. Is that the same kind of mistake as under-warning them? Argue it both ways in writing. Then look at how this site handled it — the miss is published, and so is the observation about which direction it ran — and decide whether that is fair to the agency, too soft on it, or too hard.
  15. Two days, one model, and the whole lesson about error. Load AUG 15 and write down Spokane’s modelled figure (72) next to what was measured (58) — a 24% error, the best this model has ever produced. Now load AUG 16 without touching a single slider: 189 modelled against 63 measured, a 200% error. Nothing about the model changed in between. One input did: the transport wind went from east to southwest. Ask yourself which of the two numbers you would have quoted if you had only seen one of them — and then ask what a single day’s agreement between a model and a measurement is actually worth.
  16. Find the term that isn’t there. With AUG 16 loaded, look for a control that lets you tell the model the fires are 97% contained, down to fifty-five people and zero aircraft. There isn’t one. Now set the wind speed to its minimum and back to maximum, and the humidity from dry to wet, and watch Spokane’s number move by a hundred points on inputs that have nothing to do with how much fire is left. That missing variable is why the model still points three full-strength plumes at a city whose fires are smouldering — and it is the single biggest reason the numbers on this page diverge from the monitors as the event winds down.
  17. The wind paradox. Load AUG 4 — TRAPPED and walk the wind speed up one step at a time: 4, 8, 12, 18, 25, 35, 45. Spokane's air gets better up to about 12 mph and then starts getting worse again. Two different things are fighting. Watch the ventilation index and the fire behaviour tile and name both of them.
  18. Take the lid off. Load CALM NIGHT, then drag the inversion from 10 to 0. Watch the mixing height and the cross-section. Roughly how many times more air does the smoke get to spread into?
  19. Reverse the state. Set an east wind at 20 mph, then a west wind at 20 mph. Same speed, opposite direction. Which towns swap places, and why does Chimacum change less than Kirkland?
  20. Remove a mountain range. With an east wind blowing, switch the Cascade barrier off. Who is affected most — and is anyone not affected?
  21. Dry it out. Hold everything and take humidity from 70% down to 10%. The wind never changes, but the plume gets denser. What is actually changing?
  22. Reproduce a real day. Load AUG 4 — TRAPPED and compare the model's three numbers to what the monitors actually read: 151, 121 and 31. Where is the model closest, and where is it worst? What might it be leaving out?
  23. Two days, one variable — and a model that overshoots. Load AUG 4 — TRAPPED and note Kirkland's number, then load AUG 8 — WESTERLY and note it again. Mainly the wind direction changed. The measured Bellevue readings for those two days were 121 and 38. The model moves in the right direction but falls to about 10 — far below 38. What is a steady-state model unable to represent about Saturday's air, and would you rather have a model that is right about direction and wrong about size, or one that has been adjusted until it matches?
  24. Right for the wrong size, three days running. Load AUG 7 — RED FLAG, then AUG 8 — WESTERLY, then AUG 9 — RED FLAG OVER, and write down Spokane's modelled number each time: about 236, 239, 242. The measured Spokane readings were 153, 142 and 179. The model's number rises across all three days — so it gets the shape of the week roughly right — while being 35% to 70% too high throughout. Does a model that is reliably wrong by a similar factor have any use? What would you have to do to it before you could use it to answer “how bad will Tuesday be?”, and what would you lose by doing that?
  25. The day the model broke, and why that is the most useful day of all. Now load AUG 10 — VENTILATED. The model returns about 219 for Spokane. The monitor read 63. That is a factor of 3.5 — the model's worst miss of the entire event, and worse than any of the three days in experiment 8. But look at what it got right: its modelled ventilation index of 640 sits between the 493 NWS published for the morning and the 971 it published for the afternoon, and its Spokane number fell from the Aug 9 preset's 242, which is the correct direction. So the model knew the air should improve and had no idea by how much. Write down why. (Hint: the model computes today's concentration from today's fire and today's weather. What does a real 116-point fall in one day require the atmosphere to have removed, and where did that smoke come from?) Then answer the question that matters: a model that is 35% high on a bad day and 250% high on a good day is not wrong by a constant factor — its error grows exactly when the news is good. What kind of decision would that make it dangerous for, and what kind of decision is it still fine for?
  26. The experiment the model cannot run. Load AUG 10 — VENTILATED and write down Spokane's number. Now load AUG 11 — SMOKE FROM ELSEWHERE and write it down again. The model moves about two points. The real monitor moved sixty-six, from 63 to 129. Every slider on this page is a local variable, and the thing that actually changed on August 11 was a fire 130 miles away that this model does not contain. Question: is there any combination of these settings that would produce August 11's measurement? If not, what does that tell you about the difference between a model being wrong and a model being incomplete?
  27. The day you cannot score. Load AUG 11 — SMOKE FROM ELSEWHERE, write down Spokane's number (about 217), then load AUG 12 — THE WIND FLIPS NORTH and write it down again (about 193). The model says the air improved. The monitor read 129 on Aug 11 and 153 on Aug 12, which says the opposite — so at first glance the model has its sign backwards and its error has still shrunk, from 68% high to 27% high. Now read the hour labels: 08:00 on Aug 11 and 19:00 on Aug 12, because that day's data was collected in the evening rather than the morning. Spokane's overnight ventilation index is Low (122). Question: can you tell from these numbers whether the model was wrong? Write down exactly what you would need in order to decide — then say what a scientist should do with a day whose measurement protocol changed halfway through a series. (The wrong answer is to drop the day. The right answer is on this site.)
  28. The absent number. Load AUG 12 — THE WIND FLIPS NORTH, write down Spokane’s modelled figure (about 193), then load AUG 13 — THE AIR HALVES and write it down again (about 176). The model has moved 17 points. Now find Aug 13 on the acreage chart on the Fires page. There is nothing there. No federal report was published that day, so this site has no acreage, no containment, no personnel and no cost for it, and it printed the gap rather than repeating the previous day’s numbers under a new date. Three questions. First: the measured Spokane reading fell from 153 to 65 — how much of that fall can you attribute to the air, given that the two readings were taken at 19:00 and 08:00 respectively, and what is the one comparison in this record that is clean? Second: on a day with no fire data at all, which of the three terms in a smoke forecast — production, dispersion, what is upwind — can you still evaluate, and which have you lost? Third, and hardest: the model’s error went 250% → 68% → 26% → 171% across four days while its own output moved a total of 43 points. What does that tell you about using error size to judge a model? Write the answer before you look at the reality check.
  29. The day the two lines crossed. Load AUG 14 — THE AIR TURNS GOOD and write down Spokane’s modelled figure (about 204), then load AUG 15 — THE WIND TURNS EAST and write it down again (about 72). The model has fallen 132 points. The measurement over the same day went 30 → 58: it rose 28. The model’s error went from 580% — its worst ever — to 24% — its best ever — without the model changing in any way. Now do three things. (a) Write down, before reading further, what you would conclude about a model that just improved by a factor of twenty-four. (b) Look at Kirkland in the readout: about 74, higher than Spokane, the only preset on this page that does that. Work out why — which input changed, and what that input does to three fires that all sit beside Spokane. (c) Now try to check it. There is no August 15 reading for Kirkland or Chimacum, because the record’s run executed at 01:30 the next morning. The single boldest claim this simulator makes across the whole series is the one that cannot be tested. Finish by answering the question that connects all three: the smoke that turned Spokane Moderate that day came from fires over a hundred miles west that are not in this model. If a model gets close to the right number for a reason that is not the real reason, has it told you anything at all?
  30. Two clicks, four inputs changed, two outputs frozen. Load AUG 22 and write down all three numbers (53 / 3 / 2). Now load AUG 25 and write them down again (65 / 3 / 2). Between those two presets the wind speed halved, the humidity fell seven points, the inversion switch went on and the mixing height rose by 1,500 feet — and Kirkland and Chimacum did not move by a single point. Find the one input they are both pinned to. Then answer two questions. (a) The real measurements on those two days were Kirkland 25 → 15 and Chimacum 18 → 13, so the real air did change while the model’s answer did not. What does it mean that a model can be insensitive to everything except one switch? (b) On Aug 25 the National Weather Service and the Washington Smoke Blog both said smoke was drifting west, and both western measurements rose — while this model returned its lowest possible western numbers. Look at the FIRES array and explain why no setting on this page could ever have produced the right answer that day.
  31. Score a forecast against a forecast, and work out what you have actually measured. On Aug 23 the weather service published a day-ahead outlook for Monday: 86–93 °F, humidity 17–24%, transport wind east around 5 mph, mixing height 6,500 ft, ventilation Moderate (378). The Aug 24 product gave 86–94, 17–24% exactly, east 5–10, 6,500 ft exactly, Moderate (290→387). Three fields land exactly. This is the most accurate day-ahead call anywhere in this record — and this site published it with a warning attached. Work out the warning for yourself before reading the Aug 24 timeline entry. (a) What are the two numbers being compared, and who produced each of them? (b) If a forecast agrees with a later forecast from the same office, what property of the forecasting system have you measured — and is it the same as accuracy? (c) What observation would have settled it, and why does this record not have one for Aug 24? The answer to (c) is the reason this experiment exists.
  32. The cleanest controlled experiment on this page: three clicks, one input, twenty-five days apart. Load AUG 27, then AUG 23, then SEPT 3, writing down Spokane each time (209, 187, 154). All three presets sit at the same wind direction (southwest), the same wind speed (15 mph) and no inversion. The only input that differs is humidity: 22%, 28%, 38%. Now write down what the monitors actually read on those mornings: 59, 34, 7. (a) Both lines fall as humidity rises. Over sixteen points of humidity the model drops 55 and the monitor drops 52 — almost the same absolute response. Does that mean the model’s humidity physics is right? (b) Now subtract: 209−59, 187−34, 154−7. You get 150, 153, 147. Write down what a near-constant additive error means about what a model is missing, as opposed to what it is getting wrong. (c) Then argue against yourself. The three presets were pinned; the world was not. Twenty-five days separate the first from the last, the fires changed, and on the third day it rained. How many points would you need before you would believe the 150? Write the number down before you look for more data — that is the whole discipline.
  33. Two clicks, best and worst, nothing changed in between. Load SEPT 2 and note Spokane (34) against that morning’s measured 33 — an error of +3%, the smallest this model has ever produced. Now load SEPT 3 and note Spokane (154) against a measured 7 — an error of about +2,100%, the largest it has ever produced. One day apart. No constant was re-fitted between them. (a) Write one sentence explaining how the same unchanged arithmetic can be nearly perfect and catastrophically wrong on consecutive days. (b) The answer is in what changed outside the model: overnight rain removed most of the particulate from the air over Spokane. Find the term for rain in this model. (c) Finally, the harder question. If yesterday’s +3% had been the only click you ever saw, what would you have concluded — and what does that tell you about every single-day agreement anyone ever shows you, including the ones that agree with what you already believe?
  34. The day the model went the wrong way. Load AUG 13 — THE AIR HALVES and write down Spokane’s modelled figure (about 176), then load AUG 14 — THE AIR TURNS GOOD and write it down again (about 204). The model went up 28 points. The monitor went down from 65 to 30, its first Good reading of the event. Exactly one input changed between those two presets — find it. Then answer three questions in order: (a) is the physics in that input wrong, or is it being applied to the wrong situation? (b) the mixing height rose 1,500 feet that day to the highest figure in the record — why does this model not respond to that at all? (c) all three fires were at 97% containment with no aircraft and 237 people left. Where in this model is the term for how much fire is left to burn? Once you have found that there isn’t one, you have found the real reason a 580% error was available on the quietest day of the event.
  35. Score a model on the day its subject disappeared. Load AUG 16, then AUG 17, then AUG 18 and write down Spokane’s modelled figure each time: about 189 → 205 → 214. Now write down what was actually measured: 63 → 68 → 72. Both lines rise, and the modelled one rises about three times faster. Then find the fact that neither line contains: on the morning the third preset describes, the national situation report stopped listing all three of these fires. Autumn Lane’s last published containment figure is 97% and Old Trails’ is 98%. Question: if the fires driving this model were closed out of the federal record, why did the modelled number go up? Answer it in terms of what the model is made of, not in terms of the weather.
  36. Find a model that is wrong in two directions at once — and decide which error is worse. With AUG 18 loaded, compare all three cities against what was measured: Spokane 214 modelled / 72 measured, Kirkland 10 / 33, Chimacum 2 / 21. The model over-reads the east side by about 3× while under-reading the west side by 3× and 10×. Those are two different missing variables, not one. The first is the term for remaining fire; for the second, look at the fire list this model contains and ask where the smoke actually came from that week — the Washington Smoke Blog named Little Giant and Sinlahekin, both far outside this model. Then the harder question: a public smoke map with these two errors would tell Spokane to stay indoors and tell Kirkland it was fine. Which of those two mistakes does more harm, and does the size of the error tell you the answer?
  37. Watch a site break its own generalisation, and check that it kept the wreckage. Read the Aug 15, Aug 17 and Aug 18 rows on the weather table and note what each says about which fire weather zone was under a Red Flag Warning. All three name zone 709 and exclude zone 708, which contains Spokane, and by Aug 18 this site was calling that a pattern. Now read Aug 22: zone 708 is inside the warning. The question is not whether the site was wrong \u2014 three observations in a row is not a rule and never was. The question is whether you can still find the wrong claim on the page, with the date it was made and the date it broke. If a site quietly deletes the generalisations it got wrong, it stops being a record of what was known and becomes a record of what survived. Find the deleted version. There isn\u2019t one.
  38. Compare two ways of losing a day, and decide which loss is worse. This site missed three runs: Aug 17, Aug 19 and Aug 21. Open the timeline rows for Aug 19 and Aug 21 side by side. Aug 19 was recovered almost whole \u2014 the federal report keeps a permanent dated archive and the weather forecast kept a version history, so both came back; only the two proxy air readings were lost. Aug 21 lost the weather too, because the version window that saved Aug 19 skips Aug 21 entirely. Now find what the Aug 21 met row prints instead of numbers, and find the day-ahead forecast sitting beside it, labelled as a forecast. Ask yourself what the table would look like if that forecast had simply been typed into the cells. It would look exactly like every other row \u2014 and it would report Aug 21 as the hottest, driest day of the event on the strength of a prediction nobody ever checked. That is the whole difference between a record and a narrative, and it is one paste operation wide.
  39. The day a rule this site trusted twelve times over stopped short of the one question it was asked. Open the fires page and read the containment note on Fairview. It reached 100% on Aug 16 and vanished from the federal table the next morning — the exit protocol, which has now gone twelve for twelve on incidents printed at a hundred. Then read the note on Autumn Lane, which was printed at 97% and vanished the same way. For eighteen editions the tempting move was to fill in the missing hundred, and this site refused on the grounds that no source had published one. Now find Kaiser Canyon in the Sept 5 and Sept 6 timeline entries: 138,293 acres, printed at 98%, then absent, never having printed a hundred. Ask yourself which of the two claims the twelve confirmations ever licensed — that reaching 100% makes a fire disappear, or that disappearing means it reached 100%. They are not the same sentence, only one of them was ever tested, and the difference is the entire reason this page is still open.
  40. Find the number that looks like a confirmation and is actually the evidence against. On the Explore page, click SEP 3 and note the gap between the model’s Spokane output and the measured value: about 150. Click AUG 23 and AUG 27 — both pinned to the same wind — and you get 153 and 150. Sept 3 published that as a hypothesis with a test attached: a fixed missing source term would produce a fixed offset. Now click SEP 6. The gap is 152. It looks like the fourth point on a straight line. Then click SEP 1 and SEP 2, which run on a southerly, and read their gaps: 13 and 1. Ask yourself what a genuinely missing source term would have to do when the wind direction changes — and then ask why an offset that disappears on a southerly and returns at 150 on a southwesterly and a northerly should be called a missing term at all. The confirming number and the refuting number are the same click. Which one you get depends entirely on whether you looked at the rest of the set.
The arithmetic

Exactly what this model does

A model you cannot inspect is not worth trusting. Here is all of it.

Deliberately left out: fuel type and fuel moisture, fire size and heat output, plume rise from buoyancy, chemistry and particle aging, precipitation, diurnal heating cycles, and the fact that real terrain is three-dimensional. A real smoke forecast — NOAA's HRRR-Smoke, for example — carries all of these. Anyone who tells you the whole story fits in six lines of arithmetic is selling something.

Sources & methods

Where every number came from.

Conflicts in the record

Fast-moving disasters produce contradictory reporting. That is not a failure of journalism; it is what a partial view of a moving event looks like. Rather than hide the disagreements, they are all listed here — including the ones that have since resolved, because how a disagreement ends is part of the record, and including the ones where this site's own reasoning turned out to be wrong.

    What could not be obtained

      Design and color choices

      • Chart series colorsThe three-town palette — orange for Spokane, blue for Kirkland, green for Chimacum — was checked programmatically against this dark background for colorblind separation, not chosen by eye. It clears the all-pairs thresholds for deuteranopia, protanopia and tritanopia, and every series is also directly labeled so identity never rests on color alone.
      • AQI colorsThe six EPA Air Quality Index colors are used as published, because they are what the public already reads on every air quality app and sign. As a hazard scale they are not colorblind-safe on their own, so every AQI color on this site appears next to its category name in text.
      • AccessibilityThe chart has a data table view. All interactive controls are real buttons and range inputs and work from the keyboard. Motion respects the reduced-motion preference. The page prints with all sections expanded.

      Full source list