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, a structure that hands pricing actuaries a source-level uncertainty number vendor property data has never carried before.

The patent, filed February 2, 2023 and issued as a B1 with no prior published application, describes a system that takes raw imagery of a property, estimates each attribute value with a pretrained neural network and an attached confidence score, retrieves competing third-party values for the same attribute, assigns those a confidence metric too, and then combines the confidence metrics before a final-value module resolves one answer (USPTO, granted July 28, 2026). Cape Analytics, founded in 2014 and acquired by Moody’s in January 2025 after raising $75 million from investors including State Farm Ventures, The Hartford, and Cincinnati Insurance, has built its underwriting business on exactly this kind of attribute: its Roof Condition Rating is already used by more than 100 carriers and approved for ratemaking in more than 40 states across over 300 filed rating and underwriting uses (Cape Analytics). What changed on July 28 is not the underlying computer vision. It is that the reconciliation logic behind those attribute values is now a matter of public patent record, and that record describes a data-fusion problem actuaries have priced around for years without a formal name.

What the patent actually claims

Coverage of Cape and Moody’s AI products tends to stop at "computer vision reads the roof." That undersells what US 12,694,669 protects. 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 a set of third-party values for each attribute, determines a respective confidence metric for each third-party value, computes a combined confidence metric by combining those respective metrics, and then automatically determines a final value for each attribute based on the third-party values, the estimated value, and the combined confidence metric before prefilling that value into a property form (USPTO grant, US 12,694,669 B1).

Three design choices in that claim language matter for how a pricing actuary should read the output. First, confidence is generated per source, not once for the finished attribute; an aerial-imagery estimate of roof age and a tax-record estimate of roof age each carry their own independent score before anything is merged. Second, the combination step is explicit and computable, described in the specification as averaging, weighting, or a probabilistic combination of the individual metrics, rather than a black-box ensemble score with no visible components. Third, the final-value module that picks among the reconciled candidates is rule-based: decision trees with heuristic selection rules, paired with the ML-based confidence estimates, alongside voting mechanisms and quantile-based selection where third-party values get mapped against confidence quantiles (USPTO grant, US 12,694,669 B1). That is an auditable arbitration layer sitting on top of a neural network, not a single opaque model producing one number.

The patent is classified under Cooperative Patent Classification code G06V 20/176, the subclass covering recognition of urban or other man-made structures in imagery captured from aircraft or satellites (USPTO CPC scheme, Class G06V). That placement is itself informative: Cape is not patenting a generic scene-classification method. It is patenting the specific reconciliation logic applied after several separate G06V-style recognitions of the same structure have already been produced.

The B1 grant and what no pre-grant publication signals

Most US utility applications publish 18 months after filing regardless of whether they have been allowed, appearing first as an A1 publication and only later, if granted, as a B2. US 12,694,669 skipped that step and issued directly as a B1, meaning the July 28 grant is the first time this specific claim language has been public. Applicants can request nonpublication under 37 CFR 1.213 when they represent the invention will not be filed abroad, and the roughly three-and-a-half-year gap between the February 2023 filing and the July 2026 grant is consistent with a domestic-only filing strategy rather than an unusually slow prosecution.

The practical effect is that competitors reconciling multi-source property data, Verisk’s Geomni aerial-imagery unit, Nearmap’s Betterview platform (acquired 2023) and its itel claims-data business (acquired May 2025), and modeling vendors like Zesty.ai and HOVER, had no visibility into this specific claim scope until the grant published (Reinsurance News). A no-publication B1 grant is a recognized signal of litigation readiness in patent strategy: the applicant forwent the deterrent value of an early published application in exchange for denying competitors advance notice of exactly what claim scope to design around. Any vendor whose current multi-source blending pipeline resembles the claimed structure, per-source confidence estimation followed by a combined-confidence arbitration step, now has a freedom-to-operate question to run, and it is running that analysis three and a half years after Cape locked in a priority date.

Why source-level confidence is a credibility problem

Classical credibility theory exists to answer one question: how much weight should an actuary put on a limited or noisy data source relative to a more stable prior. Bühlmann credibility formalizes that as a weighted average between an observation and a collective mean, with the weight increasing as the observation’s own variance shrinks relative to the variance between groups. Cape’s patented arbitration module is solving a structurally identical problem one level down in the data pipeline, before the rating variable is even assembled: instead of weighting a policyholder’s own loss experience against a class mean, it weights an aerial-imagery roof-age estimate against a tax-record roof-age estimate, using a confidence metric that plays the same role a credibility factor plays in classical theory.

That is the specific reframe most coverage of Cape and Moody’s AI stack misses. A rating actuary who receives a single "roof age: 14 years" field from a vendor has no way to know whether that value came from a high-confidence aerial read or a low-confidence, conflict-resolved compromise between two disagreeing sources. If the vendor exposed the combined confidence metric the patent describes, rather than only the arbitrated point estimate, the actuary could build a genuinely credibility-weighted rating variable: full weight on the attribute where source confidence is high, a blended or partial-credibility treatment where it is not, and a flag for manual underwriting review where confidence falls below a set threshold. actuary.info’s coverage of computer vision gaps in commercial property COPE data found the same underlying issue from the opposite direction: vendor attribute fields routinely present a single value with no visible uncertainty, even where the underlying imagery is ambiguous or occluded.

What a source-confidence relativity would look like

Consider a simplified version of what the patent’s combined-confidence metric could feed into a rating plan. A carrier currently applies one relativity to a "roof age 10-15 years" band regardless of source. If Cape exposed the underlying source-level confidence, the same nominal roof-age band could carry two different effective relativities: one for records where aerial imagery, floorplan data, and the tax assessor record agree within a narrow tolerance, and a second, more conservative relativity for records where sources disagree and the final value depended on the rule-based arbitration module rather than source 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 above is illustrative, not a claim about any live rate filing. But the mechanism the patent describes, a combined confidence metric attached to every reconciled attribute, is exactly the input a credibility-weighted variable needs, and it did not exist as a documented, patented process before this grant. The same nominal attribute value could carry a materially different effective relativity depending on which sources produced it, a nuance that a rate filing built on the arbitrated point estimate alone would never disclose to a regulator.

Where this flows into cat and property models

Property catastrophe models consume the same class of attributes this patent reconciles: roof geometry, roof age and condition, construction type, number of stories, presence of secondary structures. actuary.info’s prior coverage of the aerial-imagery data moat forming among EagleView, Cape, and Zesty.ai noted that these vendors increasingly feed cat model inputs directly rather than through a manual underwriting intake step. If a cat model ingests Cape’s reconciled attribute values without also ingesting the combined confidence metric behind them, it inherits the same problem a rating plan does: an average annual loss estimate built on a mix of high- and low-confidence inputs, with no way to separate model uncertainty attributable to hazard from model uncertainty attributable to the underlying exposure data. Calibrated confidence scores, if licensed alongside the attribute values rather than filtered out at the arbitration step, give a cat modeling team a genuine input for exposure-data uncertainty rather than a hand-waved data-quality haircut applied uniformly across a portfolio.

Documentation and unfair-discrimination testing

A rating variable whose effective relativity depends on which data source populated it raises a specific documentation question under state unfair-discrimination review. If two policyholders share an identical roof-age value on the declarations page, but one value came from high-confidence source agreement and the other from low-confidence arbitration between disagreeing sources, a regulator reviewing the filing for disparate impact needs to know that the same nominal input can carry two different levels of reliability. The NAIC’s ongoing third-party data and model governance work has already established that regulators expect documentation to follow data through the full vendor pipeline, not just the model that consumes the vendor’s output. As of mid-2026, 25 states and the District of Columbia had adopted the NAIC AI Model Bulletin in full or substantially similar form (NAIC bulletin adoption tracker, Q2 2026), and actuary.info’s reporting on the NAIC vendor registry found regulators pushing insurers toward documenting exactly this kind of third-party data provenance, not just the models built on top of it.

A source-confidence-aware rating variable is defensible in a way an unexplained point estimate is not: it gives the actuary a documented, quantitative basis for why two nominally identical inputs receive different treatment, tied to a measurable confidence metric rather than an ad hoc underwriting override. That documentation trail is also exactly what the same rate-filing governance gap around model drift requires when a vendor updates its confidence-scoring model: the filing needs to specify not just the attribute values used at filing time, but the version of the confidence-reconciliation logic that produced them, since a retrained arbitration module could shift which sources get weighted higher for the same property without changing the underlying imagery at all.

The fencing strategy: data fusion, not computer vision

"Patents offer a window into where a company believes its durable advantage sits," is a framing actuary.info has used before in covering carrier AI patent strategy, and it applies here with a specific twist: Cape is not fencing the neural network that reads a roof from a satellite image. Computer vision for property recognition is now table stakes across Cape, Nearmap, EagleView, Zesty.ai, and Verisk’s Geomni unit, all of which run some version of aerial or oblique imagery through convolutional or transformer-based recognition models. What Moody’s now owns exclusively, at least until it is challenged or designed around, is the layer above the vision models: the specific method for assigning independent confidence to each competing source and mechanically arbitrating between them.

That is a more defensible moat than a vision-model patent would be. Vision architectures are widely published, iterate quickly, and are difficult to distinguish from prior art given the volume of published computer vision research. A data-fusion and uncertainty-quantification method tied to a specific commercial workflow, reconciling property attributes across aerial imagery, floorplans, photos, and public records into one prefilled property form, is a narrower, more technically specific claim, and narrower claims tied to concrete implementations have fared better under the Federal Circuit’s post-Recentive Analytics Section 101 scrutiny than broad "apply machine learning to property data" claims would (actuary.info coverage of the Section 101 reset). If Cape enforces this patent, or simply holds it defensively while licensing the underlying confidence-scoring output to carriers, it has claimed the part of the property-data supply chain that carrier actuaries actually need: not a better roof photo, but a number that says how much to trust the roof photo.

What this means for the rest of the property data market

Verisk, Nearmap, Zesty.ai, HOVER, and any other vendor blending multiple property-attribute sources into a single delivered value now has two questions to answer, one legal and one product. The legal question is freedom-to-operate: does an existing or planned multi-source reconciliation pipeline resemble the claimed method closely enough to create infringement exposure, particularly if that pipeline also generates and combines per-source confidence metrics before a rule-based selection step. The product question is more interesting for actuaries: will competing vendors respond by exposing their own source-level confidence data to carriers, either to differentiate against Cape’s patented approach or because carriers, once they have seen what a combined confidence metric can do for a rating variable, start asking for it as a contractual deliverable rather than accepting a single point estimate.

Carriers that license Cape’s (now Moody’s) property attribute data have a near-term, low-friction ask worth making: request the combined confidence metric alongside the final attribute value, not just the arbitrated point estimate the patent describes prefilling into the property form. Whether Moody’s makes that metric a standard deliverable, a premium add-on, or keeps it internal to the arbitration engine will determine how quickly credibility-weighted rating variables built on this kind of data move from a theoretical exercise to something that shows up in an actual rate filing.

Further Reading

Sources