Cyphr

The removal test leaves nothing. The product is a financial language model fine tuned specifically for capital deployment, and everything the platform delivers is model output: a readiness score on a 0 to 300 scale, cash flow interpretation, document verification and a contextual financial profile assembled from incomplete records.

The company describes the model as trained on the nuanced, messy and incomplete data of underserved small business owners so it reads cash based operations, thin files and non linear growth correctly, and states that it interprets readiness and resilience rather than only predicting default. Strip that and there is no scoring engine, no readiness assessment and no product, only an application form.

The division of labour is stated at the right place: the platform produces a complete contextual financial profile and a readiness score for the lender, and the lender makes the credit decision. That keeps the machine on the analysis side of a regulated judgement. What is undescribed is everything inside it.

No confidence exposure on a score, no review path when the model reads an incomplete or unusual record wrongly, no escalation to a human analyst, and no account of what happens to an applicant the system judges not ready, which is the outcome the product exists to produce for the majority of a population it says is 80 percent denied.

No accuracy figure, validation evidence, error analysis or model documentation was located, and the framing makes the absence harder rather than easier to overlook. The company distinguishes its model from default prediction by saying it interprets readiness, potential and resilience, which are constructs with no external ground truth, so there is no obvious test against which the score can be scored.

A lender adopting it inherits an unvalidated component in a credit process, and a community institution with no model risk function of its own is the least equipped buyer in the market to supply the validation the vendor does not.

No lender is named anywhere, no loan volume or application count is published and no outcome figure exists for approval rates, time to decision or default performance, which is the measurement that would validate the central claim. What exists is external recognition and timeline: the product entered the market in April 2024, the company was a top twenty finalist in a major startup competition in 2025, and it carries regional incubator backing and a local business award.

The founding story is specific and relevant, with the chief executive having worked as a technology consultant to the federal small business agency and describing watching viable businesses fail inside legacy workflows. Funding figures conflict materially across sources, appearing as 220,000 dollars, 1 million and 2.2 million within roughly a year, so any citation needs a dated source.

The training arrangement is stated openly and never governed. The model's differentiator is that it was trained on real borrower data from underserved small business owners, which is the source of its claimed advantage over general purpose models, and nothing published states whether that corpus was consented, de identified or drawn from applications submitted through partner lenders.

Nor is there a boundary statement for the future: no indication whether applications processed for one lender improve models serving another, which matters in a market where community lenders in the same region see overlapping applicants.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The payload is small business owner financial data at its most personal: real time cash flow, uploaded documentation, and for the cash based and sole operator businesses the platform is designed to serve, the boundary between business and household finances is often nonexistent, so personal spending and circumstances enter the record by construction. The company also states the model was trained on real borrower data, and nothing describes the consent basis on which borrower records became training material or whether an applicant is told.

No attestation, certification, trust centre or dedicated security page was located, and no service organisation control report or international information security standard certificate is announced or offered on request. For a company at this stage that is unsurprising, and it is also the document a credit union or community bank vendor review will request before applicant financial records move, so it is the first commercial obstacle rather than a compliance detail.

No supervisor, statute, rule or programme framework is named. The gap is unusually visible for this vendor because its buyers are among the most heavily programme governed lenders in the market: community development financial institutions operate under federal certification with mandated impact reporting, government capital programmes carry their own eligibility and audit rules, and small business lending sits under the federal equal credit opportunity regime and a data collection rule aimed squarely at it. The platform offers impact reporting as a feature without naming the regime that requires it. Cyphr holds no lending licence and needs none, which is the correct posture for a technology supplier.

The company makes the strongest fairness claim in this index and publishes nothing that tests it. The platform is described as producing de biased lending decisions and as the financial backbone powering bias free capital deployment, and inclusive lending without credit scores is the headline proposition.

The underlying argument is coherent and better than most: traditional credit scores encode historical exclusion, so a model reading cash flow and operating reality rather than a score can reach businesses the incumbent system cannot see. But bias free is an absolute claim of the same family as hallucination free, and it is made about a model fine tuned on a corpus the company selected, scoring named businesses on a 0 to 300 scale.

No fairness testing, no outcome analysis across borrower groups, no adverse action or reason code handling, and no examination of whether readiness scoring reproduces the disadvantage it was built to correct through different variables. The claim raises the evidentiary bar rather than clearing it.

No guarantee or indemnity binds the vendor, but the product design puts the scored party inside the process rather than outside it, which is rare in this index and is what earns the grade. LoanReady guides the small business applicant through the application, tells them what lenders expect, and is explicitly built around the finding that most denials reflect unpreparedness rather than unsuitability, so the person being assessed receives the assessment as usable guidance rather than as a silent verdict.

That is the shape recorded for Worth AI, where the score is shown to the scored business, and for Persona's step up path. What is missing is the rest: no described route to challenge a readiness score, correct an input the model misread, or obtain the reasoning behind a low result.

This vendor does the thing almost nobody in the index does: it names its base model. The founders state that the current model is built on top of a commercial foundation model from a named provider and fine tuned for their use, having started by building a language model manually on their own training data before that path became available. Only OnFinance AI has previously identified its base model here, and this is the second instance. Two qualifications hold it at B rather than higher.

The disclosure reaches the reader through a press interview rather than the company's own material, and the data side is undisclosed, with no bank data aggregator, document processing service, bureau or subprocessor named despite cash flow analysis and document verification both requiring external sources.

A configurable modular architecture with microservices is described, and automated loan packet generation implies output lands in a lender's process, but no loan origination system, core banking platform, document repository, accounting integration or bank data aggregator is named anywhere, and no developer documentation was located.

That is a live question for the target buyer, since a small community lender cannot fund an integration project and needs to know how a readiness profile reaches the system where its loan officers actually work.

No hosting provider, region selection, residency commitment or private deployment option was located. The footprint is United States only so cross border transfer is not the live issue, but a community development lender handling federally programme funded capital still needs to know where applicant financial records and uploaded documents rest and under what controls, and nothing published answers it.

No pricing, packaging or basis of charge is published. The question carries particular weight for this buyer set, because community development financial institutions and government capital programmes operate on grant and programme funding rather than commercial technology budgets, so whether the platform is priced per application, per institution or per programme determines whether the intended customer can adopt it at all. Nothing addresses it.

Five buyer types are named and two of them are served by nothing else in this index. Community banks, credit unions and alternative lenders are familiar, but community development financial institutions and government capital deployment programmes are a distinct market with their own mandate, reporting obligations and mission constraints, and the platform addresses them directly with readiness checks, automated loan packets, grant intake and impact reporting.

That is a deliberate position at the underserved end of small business credit rather than a broad claim. The limit is depth rather than breadth: the footprint is United States only and no institution of any type is evidenced as live.