Wildfire risk for British Columbia, scored every day.

What you get

Fire weather, computed to the Canadian standard

Every afternoon the engine ingests current weather, satellite greenness and snow, and terrain; computes fire weather across British Columbia; runs the risk model; and publishes a ranked surface. Any point can then be queried directly.

That last band is deliberate. 10,814 of the 41,857 cells return NOT_APPLICABLE — open water, rock, ice, built-up ground. A surface that cannot carry wildland fire has no wildland danger to report, and calling it “very low” would be indistinguishable from forest that happens to be safe today. It is the absence of a rating, not a low one.

Down to the address

A house and the forest behind it are not the same risk

A 5 km square is 25 km². One end can be solid timber and the other a street of houses, and averaging them helps nobody. So alongside the daily danger surface we measure what is actually burnable on and around a specific address — at 120 m, from federal land cover and national fuel-type mapping.

That Kelowna house has nothing burnable underneath it and a treeline 395 m away. A wildland fire model, asked about the house alone, says “not applicable” — which reads as safe and only ever meant “no fuel in this square”. Across British Columbia about 108,000 square kilometres of ground sits exactly like that: no fuel on it, a mostly-burnable neighbourhood around it. It is where the houses are.

This is a measurement, not a prediction — no model, no probability, no loss estimate. It answers what is there. The daily surface answers how dangerous conditions are today. Together they are what a property assessment actually needs.

The day before

Summerland, 7 August 2026

The Bald Range fire ran roughly 14 km toward Okanagan Lake within hours, forced more than 20,000 evacuations, and killed one person.

We then checked what a free lookup table of historical ignition counts would have said. It flagged the same twelve. That corridor is among the most repeatedly-ignited ground in British Columbia, and a static map knew it a week earlier.

We publish that because it is the useful part: it tells you which of your exposure is structural — terrain and history that will not change this season — and which is today's fire weather. The engine serves both, separately, so you can act on the difference.

How we measure ourselves

Every claim has a number, and the plan was frozen first

Wildfire risk is easy to make claims about and hard to prove. So each claim lives in a public ledger with its status, the measurement plan that governs it, and the figure behind it — and every plan is committed before the measurement exists, so a result cannot be reframed afterwards.

The season scorecard

Every qualifying fire in a fixed window — hits and misses, nothing selected.

Read the last row first: most fires in the window were missed outright, and Summerland supplies nearly every hit. Nineteen days in one province cannot establish skill, and we do not quote a pooled accuracy figure from it. Full scorecard →

The API

One key, one header, plain JSON

curl -H "X-API-Key: $INFERNIS_KEY" \
  https://api.infernis.ca/v1/risk/49.888/-119.496

{
  "location":      { "lat": 49.888, "lon": -119.496 },
  "grid_cell_id":  "BC-5000M-00288-00106",
  "risk":          { "score": 0.0163, "level": "EXTREME" },
  "fwi":           { "ffmc": 92.1, "dmc": 84.6, "dc": 512.3,
                     "isi": 8.4,  "bui": 96.2, "fwi": 27.9 },
  "resolution_m":  5000,
  "grid_version":  "bc-land-v1",
  "model_sha256":         "0f759546…",
  "serving_chain_sha256": "30187ca4…"
}

Every response is stamped with the grid version and the digests of the exact model and serving chain behind it. So when a number moves, you can prove whether the risk changed or the model did — the question that quietly ruins season-over-season comparisons everywhere else.

Open source

Check the modelling yourself

Nothing here is a black box. Take the trained weights, the feature contracts, the calibration and danger-level machinery, the evaluation harness and every frozen measurement plan — run them, and see whether you get what we say you get.

The Fire Weather Index implementation is worth having on its own: a clean, tested reference for FFMC, DMC, DC, ISI, BUI and FWI that you can drop into your own stack.

github.com/argonBIsystems/infernis · the claims ledger