Barkr

The removal test leaves traditional appraisal, which the company identifies as failing on two counts, accuracy and absence of liability. A domain specific language model trained on proprietary datasets produces valuations for assets with thin comparable data and grey market movement, and the output is constructed for a specific purpose competitors do not serve: what an asset would realise in liquidation within a set time window rather than open market fair value. Assets are then marked to market monthly through the life of a loan.

Human review is stated as part of the architecture rather than an operational afterthought, with the company describing a domain specific language model built with human review in the loop, which matters for assets where a single unusual attribute can move value substantially. Monthly mark to market keeps a person engaged with the position over the loan's life rather than at origination only. Held at B because no threshold is described for which valuations attract review, and nothing states what a reviewer may override.

This is the strongest model risk position in the index and it works differently from every other one. Rather than publishing an accuracy figure and asking for trust, the company has a major reinsurer underwrite the accuracy of its output through a performance guarantee product, which means an independent institution with actuarial capability has priced the model's error distribution and taken the other side of it.

That is validation by someone with money at stake rather than by self assessment. Supporting controls are consistent: human review in the loop, monthly mark to market against realised conditions, and valuations purpose built for liquidation so the prediction is testable against an actual sale.

Around 2 billion dollars of valuations have been processed since early 2025, and the company states it is approved for use by large banks and private lenders who remain confidential given the nature of the business, which is a credible explanation rather than an evasion in collateral finance.

A major reinsurer entered a partnership in the first quarter of 2025 and its head of artificial intelligence insurance is quoted publicly, which is meaningful because that firm underwrites the accuracy of the output. A 3.5 million dollar seed round was led by a venture firm whose general partner is quoted. No customer is named.

No boundary statement was located. The model is trained on proprietary data and improves with the transaction and realisation outcomes it observes, which across a lender base means one institution's liquidation results inform valuations supplied to others. Nothing states whether that pooling occurs, whether asset level outcomes are retained, or what a lender contributes about its own book by obtaining a valuation.

No data protection agreement, retention schedule or subprocessor list was located. Exposure is materially lower than at most vendors here because the subject of analysis is an asset rather than a person, though valuations attach to identifiable borrowers and their collateral positions, and knowing which lender holds what against whom is commercially sensitive information that is not addressed.

No attestation, certification, trust centre or enumerated framework was located. Large banks have approved the service for use, which implies supplier assessment was completed, and none of the resulting control documentation is published for other institutions to rely on.

No regulator, statute or valuation standard is named. That is a notable omission for this business specifically, since professional appraisal operates under recognised valuation standards and lenders relying on collateral values for capital and provisioning purposes face supervisory expectations about how those values are derived. Insurance partners are named while the regulatory basis of the valuation itself is not addressed.

No individual is assessed and the adapted exposure runs through assets and the borrowers who own them. Valuation models trained on observed transactions inherit whatever the market already prefers, so unusual, regionally traded or culturally specific assets are likelier to be undervalued than mainstream equivalents, and an undervalued asset means a smaller loan or worse terms for its owner. The company itself identifies limited data, asset uniqueness and grey market movement as the core difficulty. No analysis of valuation error by asset class, provenance or geography is published.

This is the first grade of its kind in the index and it is the company's central proposition rather than a policy buried in terms. Every valuation carries a contractual warranty backed by a major reinsurer's performance guarantee insurance, so if an asset is liquidated for less than the model predicted, the shortfall is paid, stated by the company as simply that if they are wrong, they pay.

The founder's diagnosis of the alternative is exact: traditional appraisal firms hedge liability by design, page one carries the price and the remainder of the report is disclaimer. Converting a model output into an enforceable instrument, with an insurer's balance sheet behind it, is what every other vendor in this index leaves to the customer to absorb.

The model is proprietary and domain specific, trained on the company's own specialised datasets rather than wrapping a general purpose service, and the insurance chain is named with unusual candour including the progression across three carriers to the current reinsurance partner, which matters because the warranty is only as good as the balance sheet behind it.

Held at B because the valuation data itself has no named source, and for assets whose pricing depends on auction records, dealer networks and secondary market observations, those dependencies determine coverage and accuracy.

No named integration, interface documentation or connected system was located. Delivery appears to be as valuations and reports into a lender's own process, with the company emphasising speed to term sheet in days rather than weeks. For monthly mark to market across a loan book, some connection into portfolio or servicing systems would be expected, and none is described.

No hosting provider, region selection, residency commitment or private deployment option was located. Collateral positions and borrower identities are commercially sensitive to the lender, and nothing describes where that material is processed or held.

No pricing, packaging or basis of charge was located. The structure raises a question specific to this model: a warranted valuation bundles analysis with insurance, and nothing indicates how the premium relates to the fee, whether cover scales with asset value, or where the warranty limit sits, which is what a lender would need to price its own residual exposure.

Buyers span large banks, specialty lenders, asset-based finance firms and specialty credit funds, and asset coverage is where the breadth genuinely shows, running from fine art and vintage vehicles to private aircraft, industrial equipment and graphics processors, all of which price differently and none of which has a liquid reference market. The company describes building toward a broader class of alternative assets. Held at B because no geographic footprint is stated and no institution is named in any market.