A single property attribute, say roof age, can arrive from four disagreeing sources: an aerial image, a floorplan, an exterior photo and a tax assessor record. US Patent 12,694,669, granted to Cape Analytics on July 28, 2026, does not resolve that disagreement by trusting the newest or most detailed source. It assigns each source its own confidence metric, combines them mathematically, and lets a rule-based module pick the final value, which hands pricing actuaries a source-level uncertainty number vendor property data has never carried.

Key Takeaways

  • US Patent 12,694,669 claims per-source confidence estimation and mechanical arbitration between disagreeing property-attribute values, not the computer vision that produces them.
  • A B1 grant with no pre-grant publication. The July 28 issuance is the first time the claim language has been public, three and a half years after the February 2, 2023 filing.
  • More than 100 carriers already use Cape's Roof Condition Rating, approved for ratemaking in more than 40 states across over 300 filed uses.
  • The combination step is explicit and computable, specified as averaging, weighting or probabilistic combination, feeding a rule-based final-value module rather than a black-box ensemble score.
  • The confidence metric does what a credibility factor does, one level below the rating variable: weighting an aerial roof-age read against a tax-record read rather than experience against a class mean.

Patent Details

  • Patent number: US 12,694,669 B1
  • Filed: February 2, 2023
  • Granted: July 28, 2026
  • Assignee: Cape Analytics, founded 2014, acquired by Moody's in January 2025
  • Classification: CPC G06V 20/176, recognition of urban or other man-made structures in imagery captured from aircraft or satellites
  • Publication history: issued directly as a B1 with no pre-grant A1 publication

What the Patent Actually Claims

The independent claim covers a method that determines raw property data comprising imagery, estimates an attribute value and a confidence metric for each of a set of property attributes using a pretrained neural network, retrieves third-party values for each attribute, determines a confidence metric for each of those, computes a combined metric, and automatically determines a final value before prefilling it into a property form (USPTO grant).

Three design choices in that language decide how a pricing actuary should read the output. Confidence is generated per source, not once for the finished attribute: an aerial-imagery roof-age estimate and a tax-record roof-age estimate each carry an independent score before anything merges. The combination step is explicit and computable, described as averaging, weighting or probabilistic combination rather than an opaque ensemble. And the final-value module is rule-based, using decision trees with heuristic selection rules, voting mechanisms and quantile-based selection mapping third-party values against confidence quantiles.

That is an auditable arbitration layer sitting on top of a neural network. The CPC placement at G06V 20/176 says the same thing from the classification side: Cape is not patenting generic scene classification, it is patenting the reconciliation logic applied after several separate recognitions of the same structure already exist.

The publication history is a strategy signal. Most US utility applications publish 18 months after filing as an A1 regardless of allowance; this one skipped that and issued as a B1, so the grant is the first public appearance of the claim language. Applicants can request nonpublication under 37 CFR 1.213 when representing that the invention will not be filed abroad. Competitors reconciling multi-source property data, including Verisk's Geomni unit, Nearmap's Betterview platform, Zesty.ai and HOVER, had no visibility into this scope and are now running freedom-to-operate analysis three and a half years after Cape locked a priority date.

Source Confidence Is a Credibility Problem

Classical credibility answers one question: how much weight belongs on a limited or noisy source relative to a stable prior. Buhlmann formalizes it as a weighted average between an observation and a collective mean, with weight rising as the observation's variance shrinks relative to between-group variance. The patented arbitration module solves a structurally identical problem one level down the pipeline, before the rating variable is assembled, weighting an aerial-imagery roof age against a tax-record roof age using a metric that plays the credibility factor's role.

The practical gap is what the carrier receives. A rating actuary handed a single "roof age: 14 years" field cannot tell whether it came from a high-confidence aerial read or a low-confidence compromise between two disagreeing sources. The site's work on computer vision gaps in commercial property COPE data found the same problem from the other direction: vendor attribute fields present one value with no visible uncertainty even where the imagery is occluded.

Expose the combined metric and the variable becomes genuinely credibility-weighted. Full weight where source confidence is high, partial credibility where it is not, and a flag for manual underwriting review below a set threshold. Consider a carrier applying one relativity to a roof age 10-15 years band regardless of source; with source-level confidence the same nominal band could carry two effective relativities, one where imagery, floorplan and assessor record agree within tolerance and a more conservative one where the final value depended on arbitration rather than consensus.

Scenario Source agreement Combined confidence Actuarial treatment
Aerial imagery, floorplan, and tax record agree High High Full weight; standard relativity applies
Two of three sources agree, one outlier Moderate Moderate Partial-credibility blend toward a class mean
Sources disagree; arbitration module resolves Low Low Conservative relativity or manual underwriting review flag

No filing currently discloses a rating plan built this way, and the table is illustrative rather than a claim about a live filing. The point is that the same nominal attribute value can carry a materially different effective relativity depending on which sources produced it, and a rate filing built on the arbitrated point estimate alone never discloses that to a regulator.

The same gap runs into catastrophe modelling. Property cat models consume exactly these attributes, and the vendors increasingly feed model inputs directly rather than through manual underwriting intake, the dynamic covered in the site's work on the aerial-imagery data moat. A model ingesting reconciled values without the confidence behind them produces an average annual loss built on mixed-confidence inputs, with no way to separate hazard uncertainty from exposure-data uncertainty.

What the Fence Encloses, and What It Leaves Out

The moat is not the vision model. Computer vision for property recognition is table stakes across Cape, Nearmap, EagleView, Zesty.ai and Geomni. What Moody's owns is the layer above: assigning independent confidence to each competing source and mechanically arbitrating between them. That is the more defensible claim, because vision architectures are widely published and hard to distinguish from prior art, while a fusion method tied to a specific commercial workflow is narrow and concrete, the shape that has fared better under post-Recentive Section 101 scrutiny.

What the patent does not do is oblige anyone to publish the number. The claim describes prefilling the arbitrated final value into a property form. The combined confidence metric is an internal input to that step, not a disclosed output, and whether Moody's makes it a standard deliverable, a premium add-on or keeps it inside the arbitration engine determines whether any of the credibility-weighted treatment above is available to a carrier at all.

That matters most where documentation is compulsory. If two policyholders share an identical roof-age value on the declarations page, one from high-confidence source agreement and one from low-confidence arbitration, a regulator reviewing for disparate impact needs to know the same nominal input carries two levels of reliability. As of mid-2026, 25 states and the District of Columbia had adopted the NAIC AI Model Bulletin in full or substantially similar form, and the vendor registry work pushes insurers toward documenting third-party data provenance rather than only the models built on it.

A source-confidence-aware variable is defensible in a way an unexplained point estimate is not, because the differential treatment ties to a measurable metric rather than an underwriting override.

The same model-drift filing gap applies to the arbitration logic itself. A retrained confidence module can shift which sources get weighted higher for the same property without the underlying imagery changing, so the filing has to specify the version of the reconciliation logic and not merely the values it produced. The carrier is documenting a vendor process it cannot see, described publicly for the first time by a patent that does not require the vendor to show it.

Further Reading

Sources