Manitoba's 2025 fire season, re-solved week by week.
In 2025, 354 wildfires burned 1.93 million hectares of Manitoba and drove more than 31,000 people from their homes. We rebuilt the season from public satellite and fire-perimeter data, then asked one question: if the same initial-attack crews had been re-positioned every week by an optimization model, using only information available at the time, how much closer would help have been?
Two waves, a thousand kilometres apart.
The fires came in two acts: a May wave in the west and centre, around Lac du Bonnet, Flin Flon, Lynn Lake, then a second, larger July wave in the far north and east, around Island Lake and Thompson. A fixed footprint can be right for one act. It cannot be right for both.
Watch the crews move before the fires do.
Each Monday the model re-solves: where should 20 crews stand this week, given only the trailing two weeks of satellite hotspots and ignitions? Bronze rings are optimized crew positions; blue rings are the fixed footprint; embers are the fires that actually came. Press play above to run the season.
Same crews. Same fires. A third more of them reachable.
Every decision was made with data a duty officer could have had that Monday morning. No hindsight, no foresight, just a weekly re-solve. The gap compounds all season.
The honest wrinkles.On brand-new starts, ignitions with no satellite hotspot within 20 km in the prior two weeks, the optimized plan still reached 66% within 100 km against 44% for the fixed footprint. And it wasn't right every time: the single largest fire of the season ignited 17 km from a fixed base while the optimized plan stood 86 km away. Optimization improves the odds. It doesn't grant prophecy, which is exactly why the plan is re-solved every week.
Same crews. Same budget. One change: where they stood on Monday.
Strip away the math and here is the whole study in one comparison. Nothing else differs between these two columns: not the fires, not the weather, not a single extra crew, aircraft, or base.
117 more fires had help within initial-attack range on day one, the day a fire is still hectares, not townships, for zero additional dollars of crews, aircraft, or infrastructure.
Initial attack is the cheapest moment in a fire's life. Every fire that escapes it becomes evacuations, air quality emergencies, and insurance losses. We don't claim each of those 117 fires would have been contained; we claim each one would have been met properly, and the whole economics of a fire season lives in that difference.
2025 was the audit. The product is next Monday.
Nothing in this study used information from the future, which means the loop runs identically as a live, forward-looking system. Pointed at the current season, it looks like this:
Latest hotspots, ignitions, and fire-weather feeds land. The week's risk picture assembles itself.
An exact solver, not an AI guessing, re-solves staging for the fleet you actually have: crews, caps, relocation budget. Provably optimal, in minutes.
Agentic simulation throws 500 versions of the coming week at the plan. If it only works in one future, it doesn't ship.
A staging recommendation is on the duty officer's screen, with the reasoning, not just the answer. The human decides.
Optimized by math. Pressure-tested by simulation.
487,769 satellite hotspot detections and 354 fire records become one weekly question: where should scarce crews stand?
A mixed-integer program allocates 20 crews across 27 candidate communities, maximizing risk-weighted coverage under real capacity limits. Solved to proven optimality by an exact solver, every week. No model guessing; the same inputs give the same answer, always.
Agentic simulation takes the solver's plan and attacks it: 500 alternate seasons, with fires resampled and locations perturbed. The advantage never fell below 31 points. A plan that only works in one version of reality isn't a plan.
Monday arrives, the world has moved, the model re-solves. 30 consecutive weekly decisions, the same loop Ourus runs on fleets, plants, and grids.
Where the AI is, and where it isn't. The staging plan is never generated by a language model. AI does two jobs at the edges: it builds the optimization model, reading the operation and translating it into an objective and constraints you can read in plain language, and it attacks the result, running agentic simulations of the weeks that could come. The decision in the middle is made by an exact mathematical solver: deterministic, auditable, and provably optimal. AI frames the problem and stress-tests the answer. The math is just math.
Built to be checked.
Everything here comes from public data, and every assumption is on the table. If you'd like the model pointed at your operation's version of this problem, the loop is the same.
National Burned Area Composite 2025 (Natural Resources Canada): 354 fires under Manitoba jurisdiction, 1.93M ha adjusted burned area.
CWFIS satellite hotspots (487,769 Manitoba detections, 2025). The model at week t sees only data before week t; cold-start weeks default to the fixed footprint.
20 crews. The fixed footprint spreads them across 8 established attack bases; optimized staging may use 27 candidate communities with airstrips or existing infrastructure.
A crew stationed within 100 km of the ignition point, a proxy for initial-attack reach, not a claim any given fire would have been contained.
Mixed-integer linear program (facility location with capacities), HiGHS solver. Roughly 7 crew moves per week; a relocation-budget constraint is one line, not a rewrite.
Your operation has a version of this problem.
Somewhere in your data there's a decision being made on habit that math could make on evidence, and simulation could make robust. The loop is the same.
Reconstruction on public data. NRCAN National Burned Area Composite & CWFIS hotspots, 2025. A counterfactual model, not a critique of the crews who fought this season.
Have a decision like this?
We take on a small number of design partners. Bring the plan you are afraid to run.