Hudson Labs

The models are the entire product and the removal test leaves nothing behind. There is no data terminal underneath, no workflow or portfolio system, no proprietary content library that would retain value on its own; what the customer buys is the interpretation of unstructured filing text, and that interpretation is model output end to end.

Forensic risk analysis reads narrative disclosure and judges what matters, guidance identification and multi document comparison are inference tasks rather than lookups, and the investment memo is generated. The company has built finance specific models rather than wrapping a general purpose one, and positions that specificity as the reason it can be relied on where a generalist assistant cannot, which makes the models the differentiator as well as the mechanism.

Provenance is the named control and it is applied as a default rather than on request. Output is described as source verified, with source selection and noise suppression presented as deliberate stages that decide what the model is allowed to draw on before it answers, and auditability is stated as a design requirement for institutional use. Constraining the retrieval set is a genuine mechanism positioned before generation rather than a caveat attached after it.

Held off the top grade because confidence never accompanies provenance, so a user is told where an answer came from and never how reliable it is, no abstention behaviour is described for questions the filings cannot support, and no threshold or review step is specified anywhere in the pipeline.

Strong comparative claims are made and none is accompanied by a published measurement. The company states it achieves state of the art results in financial workflows and outperforms large research laboratories, and describes its accuracy and reliability as industry leading, without naming a benchmark, publishing a score, describing an evaluation methodology or identifying who conducted it. An unverifiable comparative claim is weaker evidence than a modest published number.

No error rate for the forensic flags, no retraining or monitoring cadence, no artificial intelligence management system certification and no validation documentation was located, which is a notable gap for a vendor whose whole argument is that generalist systems are insufficiently reliable for this work.

A named individual is quoted with a specific result and the institutional references remain anonymous. Brett Caughran of Fundamental Edge describes a two to three hour process reduced to twenty minutes, which is an attributable claim from a person who can be asked about it. Alongside it sit unattributed figures: a beta cohort of hedge funds, asset managers and family offices representing over one trillion dollars of assets under management, and research time reduced by more than half.

Independent parties carry money at stake through Y Combinator, FoundersX Ventures, Born Capital and named angel investors, and two research firms are listed as partners. Held at this grade because no institutional client is named anywhere and the quantified outcomes attach to the anonymous cohort rather than to the named voice.

No safety practice, output screening, red teaming or data boundary is described. Noise suppression and source selection are quality mechanisms aimed at answer relevance rather than safety controls, and the distinction matters because the failure mode here is a confident, well sourced but wrong characterisation of a company's accounting.

The multi client question is weaker than usual since the corpus is public filings rather than client data, and it is not absent: several competing funds research overlapping universes on the same platform, and nothing states whether one institution's queries, flags of interest or generated memos inform anything served to another.

No privacy policy detail, processing agreement, subprocessor list or retention schedule was located. The underlying corpus is public disclosure, which keeps the personal data surface small, and the unaddressed exposure sits on the client side instead: a research platform records which companies an institution examines, which risk flags it pursues and when, and that pattern reveals investment attention that a fund would regard as confidential. Nothing states whether that activity is retained, for how long, or whether it is used to improve the product.

No certification, attestation, trust portal or security page was located, and the absence is recorded as unlocated rather than proven. The company is small enough that the omission is unsurprising rather than negligent, and the content it processes is public regulatory filing text rather than client portfolio data, so the confidentiality exposure is narrower than for most vendors in this index. What is not narrow is the buyer set, since hedge funds, insurers and law firms run vendor security reviews regardless of vendor size, and nothing published answers one.

The company is a software provider to investment institutions and holds no licence, registration or supervisory standing of its own, which is the ordinary posture for this shape and carries no penalty. One adjacency is worth recording without drawing a conclusion from it: output that characterises the accounting quality of named public companies sits close to territory that is regulated when produced by rating agencies or research analysts, and the company operates outside any of those regimes while serving customers inside them. No regulator engagement or accreditation was located.

No fairness evaluation, governance programme or model assessment material was located. The exposure specific to this product is not demographic but corporate, and it is real: forensic risk analysis assigns accounting and governance concern to named public companies, and if the models flag more readily on certain filing styles, sectors, company sizes or non native English disclosure, then some issuers are systematically more likely to be marked as risky than their conduct warrants. Nothing published examines how flags distribute or what proportion prove unfounded.

No liability position, error rate or remediation route is published, and this product has an unusual second party with something to lose. Beyond the investor who acts on a wrong summary, there is the company that gets flagged: a forensic risk signal asserting accounting or governance concern is reputationally consequential for the issuer, and that issuer is not a customer, has no relationship with the vendor and no way to know a flag exists. Nothing describes how a mistaken flag is identified or withdrawn, whether an issuer can contest one, or whether corrections reach the investors who already saw it.

No provider, model family, version or hosting arrangement is named. The material does indicate build rather than buy, describing custom models, finance specific training and in house pipelines, which is a partial answer about origin and more than pure silence, and the company's technical leadership is a plausible basis for it.

But nothing states whether any third party foundation model sits underneath the custom layer, where inference runs, or what the company depends on that it does not control, so an institution cannot enumerate the chain behind an answer it is relying on.

No integration was located in any source reviewed, and the absence is recorded as unlocated rather than proven. The platform is described consistently as web based and self contained: an analyst opens it, asks questions of filings and receives memos and summaries back.

Nothing identifies an order management, portfolio management, research management or customer relationship system it connects to, no application interface is documented publicly, and no partner directory or integration count appears. For a research product this is a real constraint rather than a formality, because output that cannot flow into the note or model an analyst actually maintains has to be moved by hand.

The platform is described as web based and nothing further about deployment is published. No hosting region, residency commitment, tenancy model or single tenant option appears, and no statement addresses where client queries and generated memos are stored. The company is headquartered in Canada and serves United States institutions, so the data crosses a border by default, and nothing addresses that even to say it is immaterial.

No price, tier or billing basis is published. The only commercial term visible anywhere is a recruitment offer of two free months in exchange for an introduction to a qualifying fund, which indicates a subscription model without indicating its cost.

Nothing states whether charging is per seat, per company covered or by usage, and nothing distinguishes the pricing of the forensic analysis product from the newer research platform, which are separately described capabilities that a buyer might want separately.

Three institution types are named and they are genuinely different readers of the same filings: asset managers and hedge funds researching positions, insurers assessing holdings, and securities law firms examining disclosure. Family offices appear alongside them in the beta cohort, which is stated to represent over one trillion dollars of assets under management, a figure that speaks to the calibre of institution engaged rather than the count.

Held off the top grade because the coverage is bounded to United States public equities by the company's own framing, no client count is published, an analyst database records only two known customers, and the breadth is described through categories rather than evidenced through named institutions across them.