Candor Technology

The automated decision is the entire company, and the technique is less learned than the marketing suggests. On centrality there is no argument: strip the underwriting engine and nothing remains, since every product in the estate either feeds the decision or extends it. That places this well above the payments and core banking vendors in this index, where large deterministic businesses survive the removal test intact. What holds it out of the top band is what the engine actually is.

The company describes expert systems and aerospace heritage alongside artificial intelligence, and markets its income module explicitly as deterministic. An expert system encodes human knowledge as rules and inference; it is a recognised branch of artificial intelligence but it is not machine learning, and a deterministic system by definition does not infer probabilistically.

The company's own positioning against a finite rules engine indicates something more sophisticated than a decision table, but nothing published describes learned components, training or model updating. The determinism is not incidental to the business either, since it is what makes the warranty underwritable.

High autonomy by design, bounded in ways the company publishes rather than merely asserts. The product fully automates decisions that previously required a human underwriter, and does so in as little as 90 seconds, so this is not decision support. Three things nonetheless keep it out of the lower band.

The system fires lender conditions where its analysis stops, with the published example of routing recurring debt review back to the underwriting team, so the boundary of automation produces a human task rather than a silent gap. Lenders can implement their own overlays on top.

And the company addresses the displacement question directly rather than avoiding it, arguing that underwriters are among the most capable people in a lending business and are typically redeployed into risk management and distressed servicing when volumes fall, which is an unusually candid position for a vendor selling automation.

What remains unpublished is a stated floor: nothing describes what the system will not decide autonomously, what review a lender must retain, or how an automated decision is escalated when the file is marginal.

Decision logic documented at a level of operational specificity that almost nothing else in this index approaches. The published questions page tells a lender exactly how the system treats particular income and asset situations, including the threshold at which a deposit is examined, which account types are in scope and which are not, and where a condition is generated instead of a determination. A model risk function can read that and understand the behaviour rather than inferring it.

Determinism compounds the advantage, because a system that returns identical output for identical input is reproducible and therefore testable in a way a probabilistic model is not, and repurchase provides an externally defined failure measure the vendor cannot redefine.

What is missing is the formal apparatus: no validation methodology, no exception or condition rate, no accuracy or error measurement beyond the repurchase count, no version or change disclosure for guideline updates, and no technical documentation behind the patents. The shifting basis of the headline figure, from 600,000 underwritten loans to three million underwrites within six months, is the specific item to raise in diligence.

One outcome metric that is genuinely well chosen, and a figure that has moved in a way a buyer should ask about. The good metric is zero repurchases. Repurchase is the objective failure event in mortgage origination, defined by the government sponsored enterprises rather than by the vendor, triggered when a loan is found defective after sale, and a vendor cannot quietly redefine it to flatter itself.

Reporting zero across a large population is therefore a stronger claim than a self defined accuracy percentage. Scale is stated at more than 100 banks, credit unions and independent mortgage banks, with productivity gains of two to three times, a ten day cycle time reduction and a 90 second underwrite. The question concerns the denominator.

In October 2025 the company and a named partner described the engine as powering over 600,000 underwritten loans with zero repurchases; by April 2026 the figure was more than three million underwrites with zero repurchases. That is a fivefold movement in six months and the unit changed from underwritten loans to underwrites, which are not the same thing. Named customers appear as video testimonials rather than attributed quotes.

The published questions page is the strongest safety artefact located in this sweep, and its virtue is that it documents what the system will not do. It states that the system does not gross up social security income, that it analyses checking and savings statements for large deposits above half of qualifying income but does not analyse retirement or stock accounts for the same purpose, describing that explicitly as a future enhancement, and that it does not analyse asset statements for recurring debt but can fire a general lender condition directing the underwriting team to review that itself.

A vendor publishing the boundary of its own automation, naming one gap as not yet built and routing another to a human by design, is doing something almost no one else in this index does. Determinism supports the same end, since a system that returns the same answer for the same file cannot drift quietly between decisions.

Against that, across two passes no model card, evaluation methodology, red team result or incident disclosure was located, and an executive has publicly described the technology as producing bias free decisions without any published testing behind the claim.

A thin public surface over an unusually sensitive data set. Mortgage underwriting requires the most intimate financial documentation a consumer ever assembles: tax returns, pay statements, bank and asset statements, social security award letters and employment records, and this system ingests, classifies, extracts and validates all of it.

Across two passes no privacy programme description, data processing disclosure, retention schedule or subprocessor list was located, and the Gramm Leach Bliley Act appears nowhere despite the customer base consisting of institutions the statute governs and despite the vendor handling exactly the non public personal information it protects.

The service organisation control attestation announcement lists personal information privacy among the trust categories examined, which is the closest thing to a privacy commitment located, though the quoted criteria in the same announcement centre on security. Nothing published describes how long borrower documents are retained after a decision, whether they are used to improve the system, or what happens to the file of an applicant who is declined.

A properly evidenced attestation, with a currency question a buyer should raise first. The company announced a service organisation control attestation of the second type covering the underwriting system, and the announcement is unusually substantive: it names the chief technology officer, states the examination was completed by an independent service auditor, describes the auditor's opinion as providing reasonable assurance that controls are appropriately designed and that service commitments are achieved, and cites the specific trust services criteria framework by reference.

That is materially more than a badge. Three reservations. The announcement dates from February 2022, and an attestation of this type covers a defined observation period rather than conferring lasting status, so a four year old announcement does not establish current coverage and no more recent report was located across two passes, while product pages continue to assert compliance in the present tense. The announcement repeatedly describes the attestation as a certification, which this framework does not produce. And no trust centre exists and no report is obtainable without contact.

An unregulated technology supplier whose product is nonetheless built tightly to named public programme rules. The engine cross references Federal Housing Administration guidelines, produces income analysis described as aligned to government sponsored enterprise standards, handles the standard residential loan application form including reconciling documents against it, and covers Veterans Affairs streamline refinances, so the regulatory content of the product is specific and verifiable rather than generic.

The insurer backing the warranty provides a further indirect discipline, since a top rated carrier underwriting decision defects will have examined the compliance of the process it is standing behind. Neither is regulatory standing.

Across two passes no financial services licence, no published supervisory examination outcome, and no position on the equal credit opportunity rules or their implementing regulation was located, which is a conspicuous absence for a system making automated mortgage credit decisions, where fair lending supervision has the longest and most active enforcement history in United States consumer finance.

An affirmative fairness claim made in public with nothing published to support it. The company's chief operating officer has stated in trade press that the technology helps lenders make more accurate and bias free decisions. Bias free is an absolute, and in mortgage it is the most heavily litigated and examined claim a vendor can make, since automated underwriting sits directly under the equal credit rules, home mortgage disclosure reporting and decades of fair lending enforcement.

Across two passes no fair lending testing, disparate impact analysis, model card, explainability documentation or adverse action reasoning methodology was located. One structural argument in the company's favour deserves recording because it is genuine: a deterministic system applying published programme guidelines produces the same answer for the same file, which eliminates the variance between individual human underwriters that is itself a documented source of unequal treatment. Consistency is not fairness, however. A uniformly applied rule can still fall unevenly if the rule correlates with a protected characteristic, and nothing published tests for that.

The only vendor in this index located to transfer the consequential loss rather than refund the fee. Candor warrants its own underwriting decisions, specifically income calculations and cleared conditions on funded loans, backed by a warranty from a top rated insurer for up to 60 months after closing, and provides assistance with repurchase claims. The distinction from everything else graded here is categorical.

Other vendors offer a refund on the report you were unhappy with, a marketplace credit within 30 days, or a complaints procedure. This is an insurer standing behind the specific downstream event that ruins a mortgage originator, namely a loan found defective after sale and demanded back, for five years after the decision was made. It also aligns incentives in a way disclosure never can, since a vendor carrying insured exposure to its own errors has a direct financial reason to be right.

Three reservations belong on the record rather than in the grade: the located scope covers income and conditions rather than the whole decision, with collateral explicitly outside the agency automation; the insurer is unnamed; and the warranty terms, exclusions, caps and claims process are not published.

Provenance is clear and composition is not. The engine is named, stated to be patented, and presented as the company's own, with a described lineage in expert systems and aerospace decision technology traced to a founder who worked as a scientist at the national space agency. A buyer therefore knows the decisioning is built in house rather than licensed, which is the primary question this axis asks. Below that the record is empty.

Across two passes no description was located of the technique or architecture, no third party model provider is named for the document intelligence pipeline that classifies and extracts data from borrower documents, and no data or verification suppliers are identified, though a named automation partner supplies document preparation in one joint offering.

The patents are cited repeatedly as proof of differentiation without any accompanying technical documentation, which means the strongest public evidence about how the system works is a legal filing the marketing references rather than reproduces.

Integration into the origination stack at more than one depth, with the documentation kept private. The engine is stated to integrate with lenders' loan origination software so that it calculates income and assets and clears conditions inside the system of record rather than beside it, and the company positions the platform as built for existing origination systems.

Around the engine sit components deliberately made separable: the income analysis module is sold as a standalone service callable by interface and embeddable in a point of sale system, an origination system or a customer relationship system, and the document intelligence pipeline exposes indexing, classification, extraction and validation through an interface so a lender can embed document handling anywhere in its stack.

A partnership with an automation vendor demonstrates the engine being consumed by another platform rather than only sold directly. Holding this below the top band: across two passes no public developer documentation, sandbox or named connector catalogue was located, so a lender cannot confirm support for its own origination system without asking.

Not addressed publicly. Across two passes nothing was located stating the hosting arrangement, cloud provider, available regions, tenancy model or any data residency commitment, and no customer hosted or in perimeter option is described. The company states that it is United States based and serves lenders nationwide, which speaks to where the business operates rather than where borrower files are processed and stored, and the two should not be conflated.

The gap carries weight in proportion to what moves through the system, since a complete mortgage file contains the borrower's tax returns, pay statements, bank and asset statements and identity documents, and a lender's own third party risk assessment will require placement, segregation and retention answers before onboarding. Those answers evidently exist, because more than 100 regulated institutions have completed that assessment, but none of it is establishable from outside. The service organisation control attestation covers availability among its criteria without describing the architecture that delivers it.

No price, unit or tier is published, and two passes across the company's site, its product and specification pages, its published questions, its news coverage and industry directory listings produced nothing on how any part of the estate is charged. The published entry route is a demonstration booking. The omission is more consequential here than at a typical enterprise vendor because of how the estate is structured.

Six distinct offerings are sold, spanning the underwriting engine, prequalification, document intelligence, a standalone income module callable by interface, a quality control service and a full fulfilment service that is effectively outsourced processing, and those are very different commercial shapes. A per loan fee, a subscription and an outsourced fulfilment engagement would each price differently, and a lender cannot tell from anything published which applies to which.

The warranty raises a further unanswered commercial question, since an insurer backed guarantee of decisions carries a cost that must sit somewhere in the pricing, and nothing indicates whether it is bundled, optional or separately charged.

Genuine breadth within one asset class, and no breadth beyond it. Within mortgage the coverage is real and specific: full automated decisioning on conventional, Federal Housing Administration and Veterans Affairs streamline refinance loans, with the agency programme automation described as an industry first when it launched.

Institution types span banks, credit unions and independent mortgage banks at more than 100 in total, and directory listings show the product positioned to capital and secondary markets desks, deposit banks, wholesale lenders, brokers and regional branch operations, which is most of the origination market. Borrower types are addressed at a useful level of detail, covering salaried, self employed, gig and mixed income earners, which is where mortgage underwriting is actually hard.

What is absent is everything outside mortgage: no consumer, auto, card or commercial lending, and no geography beyond the United States, which follows necessarily from a product built around domestic agency guidelines. No large national lender is named.