How the HOMAI property risk score works

A HOMAI analysis starts with one Norwegian address and ends with a report you can read on screen or export as an A4 PDF. It gets there in four steps.

1. You choose the property. Search for an address, click on the map or draw an area in the dashboard.
2. HOMAI collects official data. From the Matrikkel, Norway's cadastre, it reads the building's byggeår (build year), usable floor area (BRA), number of floors and building type. It adds the plot boundary, elevation, planning and hazard zones, and terrain data from LiDAR.
3. The analysis pipeline runs. Four stages (perception, synthesis, scoring and narrative) run as a background job. A full analysis can take minutes. Results that are already cached come back faster; vision and narrative results are kept for 90 days.
4. You read the report. It shows the scores, a hazards table with cited sources, registry and terrain data, an imagery gallery and a methodology section with the formula filled in with the property's own numbers.

The same result feeds every view in the HOMAI platform: the single-address report, the PDF and the portfolio overview. The rest of this page explains each step, so you can judge the property risk score on its method rather than on trust.

Every official value in a HOMAI report comes from a Norwegian public source. The table lists each source, what HOMAI uses it for, and links to where you can look at the data yourself.

AI models handle perception, synthesis and narrative. They work on the data above; they are not data sources in themselves.

Licences and attribution

Open data from Kartverket, Geonorge, NVE, MET Norway and NIBIO is published under licence terms that require attribution, and some imagery, such as Norge i bilder, can have stricter terms. HOMAI names the provider of every derived value in the report. These agencies publish the data; they are not partners of CUATRO and do not endorse HOMAI.

Two registry facts are worth knowing in detail. The SEFRAK register is shown as an information flag, not scored. Planning data is shown with the fields exactly as the municipality published them; HOMAI does not interpret a zoning plan for you.

Perception is the only stage that looks at pictures. The vision model gets up to seven images of the building: a map capture, the Kartverket orthophoto, aerial images from Google or Esri, street-level and multi-angle views, and a shaded-relief render of the national LiDAR height model. The images are enhanced on the server before the model sees them.

What it reports

The model answers in strict, structured data, not free text:

• roof type (gable, hip, flat, mansard, shed or complex) and the number of roof planes
• pitch in degrees and as a class, later cross-checked against LiDAR
• roofing material and overall condition
• visible defects
• the share of the roof that is flat
• counts of chimneys, roof windows and vents
• visible drainage, the surroundings and the quality of each image

The rules it works under

• Temperature 0. The model runs with its randomness setting at zero.
• Count, don't guess. It counts features quadrant by quadrant and prefers a slight under-count to an invented chimney or vent. Uncertain counts are shown as ~N.
• No age, footprint or area. It is forbidden to estimate the building's age, footprint or area from images. Those values come from Kartverket's registers.
• No imagery, no roof score. If no usable image exists, the report becomes a climate and registry report with a clear "Roof not yet assessed" banner.

The observations feed the 10 roof risk factors. The next section shows how those factors are weighted and turned into a number.

The score is calculated by code, step by step, from values that are all visible in the report. There are three layers: a roof risk index from ten factors, a climate score from four hazards, and an overall property score that blends the two. Together they make up HOMAI's property risk rating.

The 10 roof risk factors

Each factor is scored from 0 to 100, where higher means riskier, and carries its source, provider, confidence and basis. These are the default weights:

The Age factor uses the building's byggeår as a proxy for roof age. It dates the building, not the roof, so a roof replaced later does not show up in it. The historical-change factor looks for visible change between older and newer images, but it does not date the roof either.

From factors to the risk index

Risk index = Σ (factor score × normalised weight) + (1 − confidence) × 8 + 10 if fewer than three factors were scored

• Normalised weights: the weights are scaled so that they add up to one.
• Confidence drag: (1 − confidence) × 8 adds up to eight points when the evidence is uncertain, so doubt raises the index instead of lowering it.
• Coverage penalty: ten points are added when fewer than three factors could be scored.
• Unknown factors default to a neutral 50 with low confidence instead of being dropped.

The health index is the mirror image of the risk index: 100 minus the risk index.

The overall property score

Overall score = climate score (normalised) × 0.6 + roof risk index × 0.4

The result falls into one of three bands: Low below 40, Medium from 40 to 69, and High from 70 upwards. The portfolio view uses the same bands, so you see risk bands across all your properties.

Defaults, not fixed values

All weights and constants above are defaults. An administrator can tune them for each deployment, and client PDFs can hide the weight columns. The report's methodology section shows the formula with the property's own numbers filled in, so a reviewer can follow the arithmetic.

The climate score rates the climate exposure of an address from 0 to 10. It is HOMAI's physical climate risk assessment methodology in compact form: four hazards, each banded from 1 to 5 with public Norwegian data.

The climate score appears in two places. By default it carries 60 % of the overall property score, and it is also the Climate factor inside the roof risk index. NVE maps each flood zone for a set return period, and HOMAI uses the return period of the zone a building lies in when it bands the flood exposure.

Any climate risk assessment methodology rests on choices, and two of ours matter most. First, the score describes exposure today, from current hazard maps and ten years of observed weather; it is not a forward-looking climate scenario model. Second, it covers these four hazards only: snow load, storm surge and fire are not scored. For each hazard in detail, see the hazards we assess.

Every derived value in a HOMAI report carries its provenance, in four parts:

• a label that says how the value was obtained
• the provider it came from
• a confidence between 0 and 1
• a basis: a short, plain-language explanation

The label stays with the value wherever it appears in the report, so a reader can tell at a glance which numbers come from a public register and which come from the model.

HOMAI never fills a gap with a plausible guess. Missing data stays visible:

• A value that cannot be found is shown as Not assessed.
• A report field that cannot be derived from the data is shown as Not derivable, with low confidence.
• A risk factor that cannot be scored defaults to a neutral 50 with low confidence, which increases the confidence drag. When fewer than three factors are scored, the coverage penalty applies as well.
• Without usable imagery, the roof is not scored and the report says Roof not yet assessed. The climate and registry parts still work.
• If the narrative model fails, the deterministic text is kept.

The result is a report that tells you what it does not know, instead of one that looks complete.

Knowing the limits of a method is part of using it well. HOMAI:

• Does not replace an inspection. It is remote screening, not a physical inspection and not a condition report (tilstandsrapport) with condition grades.
• Does not let the AI set the score. Code does.
• Does not estimate age, footprint or area from images. These come from Kartverket's registers.
• Does not train its own machine-learning models. The only AI is a multimodal model used for perception and for writing text.
• Does not put a price on risk. The report contains no repair-cost, loss-cost or property-value estimates.
• Does not monitor properties continuously. Analyses run on demand per address, and there are no automated alerts. The portfolio view summarises the latest analyses.
• Is not instant. A full analysis can take minutes.
• Covers Norway only. Every official data source is Norwegian.
• Is not a physics-based or climate-scenario model. It combines registers, hazard maps, observed weather statistics and image perception with weighted scoring.
• Does not use live satellite passes. Imagery comes from aerial and satellite orthophotos and LiDAR, and is only as recent as the provider's latest capture.
• Does not publish accuracy percentages. Instead, every value carries a confidence and a basis you can check.

See the industries we work with for how insurers, lenders, investors and property managers use HOMAI within these limits.