Plumbline Risk scores commercial property submissions at intake and ranks them worst to best. Every score names the factors that moved it and by how much.
A submission arrives as a postal prefix and a handful of attributes. The engine attaches wildfire, flood, hail and earthquake exposure, a fire protection grade, distance to hall, construction, and a modelled reconstruction cost. What is measured and what is a regional estimate is set out below, layer by layer.
The model returns a score from 0 to 100, a band, and the drivers behind it in order of how much each one moved the score. The cutoffs are fixed in the engine and published here rather than tuned to make a demo look tidy.
Under-insurance against modelled rebuild cost, catastrophe exposure, aged roofs on combustible construction, and high fire-load occupancies in unsprinklered frame buildings. All of it raised before the submission reaches the rating engine.
| Risk | Score | Action | Leading drivers |
|---|---|---|---|
| Restaurant, frame, V1Y Kelowna BC | 88.0 · RED | Refer | Wildfire exposure (high); hazardous occupancy (high) |
| Contractor, frame, V2S Abbotsford BC | 48.4 · AMBER | Underwriter review | Flood exposure (high); combustible construction (high) |
| Office, fire resistive, M5V Toronto ON | 5.4 · GREEN | Straight through | Flood exposure (high); roof age (moderate) |
These three rows are lifted from the sample submission ranking, so the numbers on the two pages match. The risks are invented. The wildfire and earthquake numbers attached to them are real published data; flood, hail and fire protection are regional estimates, as set out below.
Take the worst risk in that sample: a frame restaurant in V1Y Kelowna, unsprinklered, roof last replaced in 1999. It scores 88.0, red.
If it turns out to be sprinklered, the same risk scores 77.8, down 10.2, and stays red.
If the roof had been replaced in the last five years it scores 79.3, down 8.7, still red.
Both figures ship with the score. When the broker comes back with a correction, you can see what the correction is worth before you re-rate, and you can see when it does not move the risk out of the band it is in. A score you can argue with is worth more than a score you have to take on faith.
Two hazard layers are measured. Wildfire exposure comes from the BC Wildfire Service Provincial Strategic Threat Analysis fire threat rating, 2021, joined at the forward sortation area. It covers 191 FSAs, all of them in British Columbia. A risk outside BC gets nothing from that source, and the report says how many of your risks the layer actually reached, in its first paragraph.
Earthquake exposure comes from the Natural Resources Canada 6th generation seismic hazard model, Geological Survey of Canada Open File 8950: the pre-calculated grid behind the National Building Code seismic hazard tool, which NRCan states serves both the 2020 and 2025 editions of the Code. It uses short-period spectral acceleration, Sa(0.2), at 2% probability of exceedance in 50 years, on site class C, and it covers the whole country.
Two things about that layer you should get from me rather than discover later. The 0 to 100 scale is a choice: ground motion arrives in units of g and has no natural mapping onto a 0 to 100 score, so it is anchored deliberately, at the point where Vancouver reaches a catastrophe referral. And site class C is firm ground, which means the layer understates soft-soil sites such as the Fraser delta, where liquefaction rather than shaking is the exposure. Both sentences travel with the data and print in every report.
Everything else is a regional estimate. Flood, hail and the fire protection grade are derived from province and city averages, not from surveyed property-level data. They are good enough to rank risks against each other. They are not good enough to price one, and both sample reports say that in the first paragraph.
Flood is on a regional estimate because four datasets were tried and rejected. The national NRCan flood susceptibility layer models where terrain lets water accumulate, so it cannot see dikes, river hydrology or sea level. Aggregated three different ways it rates Richmond, a diked delta at sea level, below downtown Kelowna, which is backwards for underwriting. The BC designated floodplain mapping is real regulatory data, but it covers riverine studies in the interior and the Fraser Valley rather than the diked delta, and it never reaches a level that would raise a catastrophe flag anywhere in the province. Each rejection is written down. Flood stays on the regional baseline until a flood hazard product with return periods and defences is licensed, because a flood number I cannot defend is worth less to you than an estimate that says what it is.
How a layer earns its place. A spatial join does not fail loudly. It returns a table, and the table looks correct until somebody who knows the geography reads it. So every hazard layer is checked against a dozen Canadian cities before it is allowed near a report, and the check runs again before every client run. That is how four flood builds were rejected, and how the two layers that did get through earned it: the seismic layer put Victoria above Richmond above Vancouver above Montreal above Toronto, which is what a Canadian underwriter would expect, and the flood layers did not.
Property-level enrichment is not licensed yet. In production it would come from vendors such as Opta or iClarify, Ecopia building footprints and provincial hazard layers. None of those are licensed today and nothing on this site depends on them. Data licensing is what a first paid licence funds, not a claim to make in advance of it.
On the synthetic 500-risk book in the sample portfolio review, red is 13% of the risks and 23% of the losses, running at a 67% loss ratio against a 62% book average. It also carries 16% of the premium.
That last number is the one a headline usually leaves out. Take the band out and its premium goes with it, which is why blanket declinature is rarely the right answer and rarely available in practice. What remains is price, deductible, a sub-limit on the flagged peril, and selective declinature on the worst of it. Knowing which risks warrant that attention before binding is the work. Declining an eighth of your submissions is not.
Those figures come from invented data, so read them as the shape of the arithmetic rather than a finding about your book. The same arithmetic run on your book is step two.
Most underwriting-automation tooling is built for the US market and adapted afterwards. This was built the other way round, on Canadian hazard geography, public fire protection grading, and Canadian construction and valuation assumptions.
That matters more than it used to. CatIQ put Canadian insured catastrophe losses at $2.4 billion in 2025, and reported it as a record year for fire catastrophes declared, with every one of them occurring in a province that had never had an industry fire catastrophe before. An appetite map drawn from where wildfire losses used to happen would have missed all of them. CatIQ, January 2026.
A scoring model is only as good as the book it is calibrated on. The weights this one ships with are priors about Canadian commercial property, and step two replaces them with your own loss experience. That is the whole point of the first cohort: the partners who go through it get a model fitted to their book rather than to an industry average, and they get it at no cost, because their data is what does the fitting.
It runs in two steps. The first one costs you about ten minutes.
Step one. Send 25 to 50 risks from your book. The first three characters of the postal code, the occupancy and the construction type. No street addresses, no insured names, no premiums, no loss data.
Back within 48 hours: those risks ranked worst to best with the reason behind each one, and the three worth looking at hardest. You will know quickly whether that ranking matches your own read of them. If it does not, that is a useful answer and we stop there.
Step two, if it does match. A full portfolio review on roughly 500 historical submissions with their loss outcomes. That quantifies how much of your loss experience concentrates in the risks the model flags, and it is where the weights get fitted to your book instead of to a prior. Also at no cost.
Two things are worth saying plainly before you ask:
Step one turns around in 48 hours. The full review is about a week from receiving the file. There is no obligation attached and nothing to sign beyond your own NDA.
Why this is easy to approve. Step one is a test, not a vendor relationship. Three columns, no insured names, no street addresses, no policy numbers. Nothing that leaves your building identifies anyone, so there is usually nothing for a privacy review to weigh in the first place.
If it does reach compliance, the answers are unusually short. Files are processed on one machine in Canada. The engine makes no network calls, so scoring a book sends nothing anywhere. There are no subprocessors, no cloud provider and no AI service in the path, because there is nobody else in it. That is a shorter vendor assessment than most firms can offer, and it is a deliberate design choice rather than an accident of size.
Your NDA gets signed before anything moves, everything is deleted on request and within thirty days regardless, and if you would rather send no file at all it can be run on a screen share while you watch.
Where this goes. If the review is useful and you want the engine scoring live submissions, that is a paid annual licence priced on your submission volume. Founding partners keep preferential terms. The review is free either way, and the point is to prove it on your book rather than describe it.
Your risks ranked worst to best, the drivers behind each one, and the three worth looking at hardest.
Run on a synthetic 500-risk book, since no real one has been through it yet. This is the format and the depth.
Step one asks for the first three characters of the postal code, plus occupancy and construction type. Nothing else. The hazard data is keyed to the forward sortation area, so V1Y is as much location as the engine can use and a street address would add nothing to the score. Nothing that identifies an insured ever leaves your building.
So: no insured names, no street addresses, no policy numbers, no premiums. At step two, add year built, declared TIV, floor area and a loss flag. Floor area is what makes the under-insurance check possible; without it the rebuild figure is a proxy and the check does not run at all. Still nothing that names anyone.
Your NDA gets signed before anything moves. If you would rather not send a file at all, we can run it on a screen share while you watch, and I keep nothing.
Postal prefix, occupancy and construction is enough. Download a three-column template if it is easier than exporting. They come back ranked, with the reasons, inside 48 hours.
Start a portfolio review or email ryan@plumblinerisk.caWritten out because an MGA under a carrier agreement has to be able to answer these questions, and because the honest answers are short.
It stays in Canada, on one machine. Files you send are processed on a single computer in Vancouver. The engine runs entirely offline: it makes no network calls, and scoring a book does not send anything anywhere.
No third parties, and no AI services. Your file is not uploaded to a cloud provider, an analytics tool, or any large language model service. There are no subprocessors, because there is no one else.
Deleted when you say, and within thirty days regardless. Say the word and the file and everything derived from it goes. If you say nothing, it goes within thirty days of the review being delivered.
Nothing is published, ever. No case study, no logo, no anonymised example, without you asking for it in writing first. The samples on this site are invented data and will stay that way until a partner offers otherwise.
Step one carries nothing identifying. The first three characters of a postal code, an occupancy and a construction type describe a building type in a region. They do not name an insured, an address, or a policy.
Your NDA gets signed before anything moves. If you would rather not send a file at all, we can run it on a screen share while you watch, and I keep nothing.
Plumbline Risk is underwriting decision-support for Canadian commercial property, built in Vancouver by Ryan Teymourian, who has spent his career building triage systems and quantitative applications: ranking a queue by severity, showing the reasoning behind each position, and designing for the fact that the cost of missing a serious case is nothing like the cost of a false alarm.
Commercial property intake is that problem in a different domain. A ranking function, an explainability requirement, and asymmetric costs, which is why the amber band exists and why a catastrophe exposure holds a risk that a frequency score would have waved through. The Canadian hazard geography and construction classes are built from published sources and checked against places an underwriter already knows. Public fire protection is still a modelled proxy rather than a looked-up FUS grade, and the report says so on the risk instead of in a footnote.
The scope is deliberately narrow. It does not replace underwriting judgment, and it does not ask you to move off your existing systems. It does one thing: turn an incoming submission into a scored, explained, ranked risk in seconds.
A plumb line is the reference a builder uses to check whether something is true. Same job here. Establish what a risk actually is before it is priced.
Plumbline Risk is pre-revenue and the founding partner work is free, which is the trade: you get the analysis, I get the loss experience that turns priors into a fitted model.