Ezra

The removal test leaves a document repository. Models convert unstructured data rooms into structured datasets, extract deal terms, identify risks and generate the diligence artifacts a credit team would otherwise write by hand, covering investment memos, research reports and due diligence question sets.

The company's entire framing is a technical argument about model behaviour, that general purpose systems are insufficiently reliable for credit analysis and that a purpose built grounded system is required, which is a proposition that exists only because of the models.

The design constrains what the system can assert rather than governing what a person does with it. Operating as a closed loop where every output is grounded in the underlying deal documents means the platform cannot introduce a fact the data room does not contain, which is a structural limit on the failure mode that matters in diligence.

Outputs are diligence artifacts, memos, research and question sets, produced for a credit team to work from rather than conclusions delivered to a committee, and the stated benefit is analysing more deals at the same headcount rather than deciding without people. What is not described is any confidence indication, escalation threshold, or treatment of documents the system cannot parse.

The published benchmark is the most substantive thing here and almost nobody in this index attempts it. Over a year of internal testing across common private credit tasks, the company measured general purpose models producing incorrect or unsupported answers roughly 30 percent of the time on credit deals, named the specific tools it tested, and built its architecture in response, concluding that for institutions deploying billions into asset backed transactions that error rate is unacceptable.

That is a falsifiable, quantified claim about the alternative, and grounding every output in source documents with full traceability is the engineering answer to it. What holds this below the top grade is the obvious gap the benchmark itself invites: the company publishes the error rate of the alternative and not its own.

The strongest usage claim is carefully worded and worth reading precisely: investment firms managing more than six billion dollars in assets have used the system during its development, which describes design partners rather than a customer base, and no firm is named. Transactions analysed span renewable energy, infrastructure, fintech and real estate.

The founding credential is substantial and directly relevant, since both founders came from a platform that financed more than 15 billion dollars in clean energy and home improvement lending, so they have originated the asset class the product serves. An eight million dollar seed closed in March 2026 led by a climate and infrastructure focused venture firm.

This is the ToltIQ problem with an additional dimension that makes it sharper. Competing private credit lenders diligence overlapping transactions on one platform, so a shared system holds confidential material from rival bidders as it did there.

What is new here is the network layer: because the platform also matches issuers to lenders, it observes not only the deals but which institutions are looking at them and how hard, which is information about a lender's appetite and pipeline that the lender would never disclose to a competitor. Nothing published describes tenant separation, contribution boundaries, or what the network layer may infer from analysis activity.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located, and the payload is among the most commercially sensitive material a company ever assembles. A data room contains contracts, financial models, customer concentrations and the weaknesses a borrower would least like disclosed, provided under confidentiality for one purpose. Nothing published states how long that material persists after a transaction closes or fails, or what happens to the structured dataset derived from it.

No attestation, certification, trust centre or enumerated framework was located, and for this product the absence is more consequential than for most. The platform ingests confidential data rooms belonging to third party borrowers, provided to lenders under non disclosure, so the security assessment is not merely the customer's own risk decision but a question of whether the customer can lawfully put that material into an external system at all.

No supervisor, statute or instrument is named. The capital network side raises the question most directly, because introducing companies raising capital to institutional lenders sits close to activities that securities regulation governs, and the boundary between providing analytical software and arranging finance is one a platform of this design will eventually need to state. Nothing addresses it, and no data confidentiality or diligence standard is identified either.

No consumer decision applies and the adapted exposure sits in the network rather than the analysis. A system matching companies raising capital to lenders actively deploying funds is making allocation decisions by another name, since a company not surfaced to the right lender does not get financed and never learns why, and the criteria determining that visibility are not published.

That is the Gain finding in a more direct form, because here matching is the stated purpose rather than a derived signal. A second exposure follows from document quality: a well resourced sponsor produces a well organised data room, and a system that structures what it is given will read a professional package more accurately than a thin one.

No guarantee or indemnity was located, and the architecture makes error findable rather than merely regrettable. Because every output is traceable to the source document that produced it, an analyst can verify any extracted term or flagged risk against the underlying contract, so a mistake surfaces during review rather than after a transaction closes, which is the property that matters when a diligence memo informs a nine figure commitment. What is missing is anything owed by the vendor when the grounding itself fails, and no correction or notification process is described.

There is an unusual asymmetry here: the company names the general purpose models it benchmarked against, identifying three major providers by name in its published argument, while saying nothing about which models power its own platform. Naming the competition and not the ingredients leaves a buyer unable to assess the fourth party exposure that a benchmarking exercise of this kind would otherwise make salient. No subprocessor list or hosting arrangement was located.

Data room ingestion is the primary connection and no provider is named, despite virtual data rooms being a concentrated market where a handful of platforms host most transactions. Nothing is published about connections into portfolio monitoring, loan administration or investment management systems on the lender side, no developer documentation was located, and the capital network's mechanics are described only as connecting parties directly.

No hosting provider, region selection, residency commitment or private deployment option was located. The question matters because institutional lenders reviewing confidential borrower material typically impose their own requirements on where that data is processed, and several would expect a dedicated environment rather than shared infrastructure for a data room derived dataset.

No pricing, packaging or basis of charge was located. The two sided structure makes the question more interesting than usual, since an analysis platform sold to credit teams and a capital network connecting issuers to lenders are conventionally monetised in entirely different ways, one by subscription and the other by transaction, and nothing indicates which applies to either side.

Both sides of the transaction are addressed, with credit teams at investors and lenders using the analysis platform and companies raising capital using it to structure their own deal information and prepare for diligence before connecting to capital providers. Sector coverage spans renewable energy, infrastructure, financial technology and real estate, which is genuine breadth within asset backed and project finance. The limit is that this is one corner of private credit rather than the wider institutional market, and no geographic footprint is evidenced beyond a United States base.