Deconflict

Models do real work and the network is the moat. Pattern recognition across concurrent investigations is stated to surface structures invisible to single entity analysis, overlapping cases are surfaced autonomously across jurisdictions, and shared adversary infrastructure and overlapping victim wallets are identified across separate agencies, which is correlation at a scale no analyst reaches manually.

The engineering credential supports it, with a founding technologist who built on device machine learning and privacy architecture shipped to more than a billion consumer devices. But the removal test leaves a credentialed intelligence exchange with hashed identifier matching connecting 900 agencies to regulated institutions, and that network is the asset. Same position as Chainalysis, where models sharpen a data moat rather than constituting one.

The division is drawn correctly, with signals informing a decision the institution owns. Alerts enable policy based actions across three named tiers, block, review or escalate, so the institution defines in advance what a given signal triggers rather than the platform acting on its behalf. Each signal carries typology, source agency and a confidence level, which is calibrated uncertainty exposed to the user and something most vendors in this index omit entirely.

What is not described is any review point before a consequential action, and the product's central promise is pre transaction intervention that stops funds moving, which is a decision taken in seconds against an active investigation flag, so where a human sits in that sequence matters and is unstated.

Every signal is attributed and qualified rather than delivered as a bare flag, carrying typology, the originating agency and a confidence level, so an analyst can weigh a match rather than simply acting on it, and documentation cites independently sourced intelligence with full attribution and an audit trail.

Platform logging captures complete intelligence attribution and access records, which supports both oversight requirements and evidentiary standards in legal proceedings, and that last point matters because material of this kind may eventually be tested in court.

What is absent is measurement of the matching itself: no false positive rate, attribution accuracy or overlap precision is published, and for a system whose output can stop a payment those figures are what a model risk function would ask for.

One number carries this and it is a substantial one: signals are aggregated from more than 900 agencies, which is the scale that makes a deconfliction network work at all, since coverage below critical mass produces silence rather than intelligence.

The founding credentials are unusually specific and directly relevant, with a chief executive who spent more than 22 years in law enforcement, eleven as a municipal police officer and the remainder as a federal Secret Service agent specialising in cryptocurrency financial crime, departing in January 2026, alongside fusion centre detectives and a senior artificial intelligence engineer from a major consumer technology company.

Independent attention includes a long form interview hosted by another vendor indexed here. What is absent is the institution side: no bank, credit union or exchange is named as a customer, and testimonials are anonymous law enforcement quotations.

Here the cross party sharing question is not a risk attached to the product, it is the product, and it is governed rather than assumed. Participation requires verified law enforcement status or regulated financial institution classification, with credentialing processes validating organisational authority before any intelligence access is authorised, and sensitive investigative indicators remain restricted to appropriately credentialed entities under tiered controls keyed to operational necessity.

Hashed exchange, case compartmentalisation and complete attribution logging complete it. The chief executive frames the boundary explicitly, that this is not about sharing everything with everybody. Seventh benchmark answer to the cross party boundary problem after Mitigram, DiligenceVault, Saphyre, DwellFi, Omnisient and Tookitaki, and the first credentialed one.

The most sophisticated privacy architecture in this index, because it solves a genuinely hard two sided problem rather than restricting a one way flow. Institutions receive investigative context on wallets without personal information exposure, identifiers are exchanged in hashed form, and responses are formatted to carry case context without exposing source data.

On the other side, investigative files, tactics and operations remain fully siloed, with case compartmentalisation preserved for the agency and attribution preserved for the institution, so each party receives exactly what it needs to act and nothing that would compromise the other.

Tiered access controls distribute intelligence by participant category and operational necessity, encryption meets federal standards for sensitive law enforcement and financial data, and platform logging captures complete attribution and access records. That this was built by an engineer whose prior work was on device privacy architecture is evident in the design.

Encryption meeting federal security standards is asserted repeatedly and no framework, certification or attestation is named. That gap is more consequential here than for a conventional vendor, because carrying criminal justice information brings its own established federal security policy with specific personnel, audit and access requirements, and any agency participating will be bound by it. Naming the standard the platform is assessed against would be the single most useful addition to its published material.

Seven supervisors are named as the review standard the output is built to withstand without rework, covering the financial crimes network, the three federal banking agencies, the credit union regulator, the central bank and a major state regulator, alongside the statutory information sharing provision that permits institution to institution exchange.

Beyond naming, the company has engaged a regulator on the public record, filing a formal comment on a federal notice of proposed rulemaking concerning anti money laundering programme modernisation, which is the property that lifts Upstart on this axis.

That comment argues for pre transaction queries of law enforcement deconfliction databases to be recognised as constituting an effective programme, which is transparent advocacy for its own category and should be read as such, but the engagement itself is a matter of public record.

The due process exposure here is the sharpest of its kind in this index and follows directly from the product working as intended. A bank can query whether a counterparty is under active investigation and block a transaction before it settles, which means a person who has been investigated but never charged, and investigations frequently close without charge, can have payments stopped or an account closed on the basis of a signal they will never see, cannot contest and may never learn exists.

Every protective mechanism in the architecture serves the agency and the institution; none serves the subject. A second exposure sits beneath it: law enforcement attention is not evenly distributed across populations or geographies, so an active probe signal inherits whatever selection pattern exists in enforcement priorities, and a network aggregating 900 agencies aggregates that too. No analysis of outcomes, false attribution rates or subject remedies was located.

No guarantee, indemnity or falsifiable commitment was located. The institution is served better than almost anywhere in this index, receiving attributable, independently sourced intelligence with a full audit trail and documentation built to survive examination by seven named regulators without rework, which is exactly what a compliance officer needs when asked to justify closing an account.

The subject of a signal has nothing at all, and their position is worse than under an ordinary risk score because the basis of the decision is an active law enforcement matter that cannot be disclosed to them by design. No correction, notification or challenge route is described anywhere.

The intelligence source is disclosed in the way that matters for this product, both in aggregate and per signal: the network draws on more than 900 agencies spanning federal, state, local, tribal and international jurisdictions, and each response carries the source agency alongside typology and confidence, so an institution knows which body originated a finding rather than receiving an anonymous flag.

Attributed intelligence is unusual and it is the difference between an actionable signal and an unfalsifiable one. What is not disclosed is the technical layer, with no model provider, hosting arrangement or subprocessor named.

The design intent is to disappear into existing compliance operations rather than to add a destination, with an interface engineered to drop into transaction monitoring, case management and suspicious activity documentation systems already in use, webhook alerts on monitored wallets matching active investigations, and output formatted to flow directly into filing narratives, account closure memoranda and transaction decision records so an analyst is not retyping intelligence into a different template.

That output orientation is the practical detail most vendors overlook. What is not published is any named system on either side, so an institution cannot confirm its own monitoring or case platform is supported.

No hosting provider, region selection, residency commitment or private deployment option was located. Encryption is stated to meet federal standards for sensitive law enforcement and financial information, which speaks to protection in transit rather than to location, and with international agencies among the participants the question of where investigative signals rest and under whose jurisdiction is live and unaddressed.

No pricing, packaging or basis of charge was located for either side of the network. The two sided structure makes the question interesting and unanswered, since agencies and regulated institutions are unlikely to be charged on the same basis and a network of this kind conventionally subsidises one side to build the other. Nothing indicates whether institutions pay per query, per monitored counterparty or by subscription.

Two entirely different participant classes are served through one platform, with banks, credit unions, fintechs, exchanges and stablecoin issuers on the regulated side and federal, state, local, tribal and international agencies on the enforcement side, and the value to each depends on the presence of the other.

Fraud typologies addressed are named specifically and span money mule networks, investment grooming scams, romance fraud, business email compromise, account takeover and wire fraud. The limit is subject matter: the intelligence is centred on digital asset and cryptocurrency investigations rather than financial crime generally, so the coverage is deep in one lane.