1. Vendor and Product Reality
Onfido, founded in 2012 in London and acquired by Entrust in 2024, is one of the most widely deployed identity-verification vendors in regulated digital onboarding. Its Real Identity Platform serves banks, fintechs, marketplaces, mobility platforms, and gaming operators across more than 195 countries, processing tens of millions of verifications annually. The product surface is anchored by two engines: a document-authentication engine that ingests photographs of government-issued credentials and runs them through computer-vision classifiers, and a biometric-matching engine that compares a live selfie or short video against the document portrait, with liveness detection layered on top to defeat photographs, masks, and screen replays.
The document engine is trained on millions of labeled fraud examples spanning passports, national identity cards, driver's licenses, and residency permits across hundreds of issuing authorities. It scores documents against template libraries, hunts for inconsistent fonts, examines security holograms and microprint, and flags pixel-level manipulation. The biometric engine uses face-embedding models that produce distance scores between the live capture and the document photo. The liveness layer mixes passive cues (texture, depth, illumination consistency) with active challenges (head turn, blink, randomized gestures) to confirm that a real person is in front of the camera.
The commercial wrapper is a cloud API that customers integrate into onboarding flows, with SDKs for iOS, Android, and the web that handle capture quality and user experience. Pricing is per verification, with tiered packages for repeat-customer flows, transaction-monitoring add-ons, and watchlist screening. The customer base is concentrated in jurisdictions with strong KYC and AML regimes, UK FCA, EU AMLD, US BSA, and Onfido's regulatory positioning emphasizes audit trails, ISO 27001 and SOC 2 compliance, and data-residency options for European customers. Within its scope the product is mature and well-engineered: the document and biometric pipelines are state of the art, the SDKs handle edge-case capture conditions gracefully, and the customer-services organization has internalized the operational rhythm of high-volume KYC.
Analyst placement consistently lists Onfido among the top three or four identity-verification specialists alongside Jumio, Veriff, and Persona. Its strengths are a deep document corpus, low friction in the consumer flow, and a mature anti-spoofing posture against the dominant attack patterns of the last decade: photograph swaps, manipulated PDFs, and synthetic deepfake injection through camera replay. Onfido is the reference implementation for what the industry calls "document-plus-selfie" verification, and the verification event is the unit of analysis around which the entire platform is organized.
2. The Architectural Gap
The structural property Onfido's architecture does not exhibit is biological-trajectory continuity across verification events. Each Onfido check is a standalone event: a document is ingested, a selfie is captured, classifiers fire, a decision is emitted, and the artifacts are archived. The product asks two questions, is this document genuine, and does the person in front of the camera match the document photograph, and answers them at a single point in time. There is no architectural place in the platform where a person's biological signals from prior verifications inform the evaluation of the current one. Even when a customer re-verifies the same individual, Onfido treats it as a fresh event and emits a fresh decision rather than as an observation that extends or contradicts an accumulated trajectory.
The gap matters because the dominant fraud pattern has shifted. A decade ago, document fraud, manipulated PDFs, swapped photos, counterfeit credentials, was the operative attack surface, and document classifiers were the right defense. Today, the operative attack surface is identity substitution with genuine documents: stolen credentials sold on dark markets, identities harvested through breaches, synthetic identities built around real attributes, and increasingly, real-time deepfake injection that defeats single-point biometric matching. In all of these patterns the document is authentic and the biometric capture matches the document, because the attacker has the document and is aiming the deepfake at it. The Onfido pipeline returns a clean verification because every check it knows how to run passes.
Onfido cannot patch this from within its current architecture because the platform was designed as a verification-event service, not as a substrate of accumulated biological observations. Adding a "repeat customer" flag to the API does not produce trajectory continuity in the structural sense; storing additional biometric templates does not produce stable sketching or trust-slope trajectory; layering a behavioral-analytics product on top does not produce biological identity in the primitive sense. The chain of reasoning the product runs is: document plausible, face matches, person is live, therefore admit. None of the three terms references prior interactions. The pipeline shape is a classifier stack, not a trajectory accumulator.
This shows up in concrete failure modes that customers already experience. Identity-recycling fraud, where a stolen credential is presented across many institutions, succeeds at every Onfido check because each check is local. Account-takeover fraud, where a legitimate user is replaced by an attacker who has the document, presents no continuity discrepancy that the platform can surface. Drift over years, a customer aging, gaining or losing weight, growing facial hair, is treated either as a false negative against the document photo or as a forced re-enrollment, neither of which captures that the change is consistent with the person's actual biological evolution. The structural shape Onfido lacks is the shape that would convert each verification into a contribution to a trajectory rather than a verdict in isolation.
3. What the Biological-Identity Layer Provides
The biological-identity layer disclosed in United States Patent Application 19/647,395 (Chapter 9) establishes human identity through continuity rather than through comparison against a stored template. A skilled implementer can build it from the pipeline the disclosure describes, and the sections below enumerate the load-bearing stages and the embodiment variations the disclosure contemplates.
The first stage is stable sketching. A biological capture is acquired from one or more modalities and passed through feature extraction and noise-tolerant normalization, then reduced by a stable sketching module to a noise-tolerant, non-invertible representation termed a stable sketch. In an embodiment the sketch is produced by dimensional reduction, projection, and quantization into band assignments, so a feature is represented by the partition region it falls within rather than by its raw value, and helper data generated alongside the sketch is updated at policy-governed intervals to accommodate drift. Non-invertibility is a design property: the sketch carries enough information for continuity validation but the mapping back to the underlying biological signal is infeasible. The sketch is not a matching template; it is the privacy-preserving bridge between the biological signal stream and the cryptographic layer above it.
The second stage is biological hash generation with domain separation. The stable sketch, a temporal binding value, a domain separation tag, and a rotating salt are combined and passed through a cryptographic hash function to produce a biological hash, a temporally bound, domain-scoped identifier for the individual's biological identity state at capture time. The domain separation tag makes hashes unlinkable across relying parties: two hashes derived from identical biological signals under different domain tags are computationally indistinguishable from hashes derived from different people, which prevents cross-context correlation into a single profile. Salt rotation lets the hash chain be refreshed without re-enrolling the individual. Because the input includes a temporal binding value, a replayed or stolen hash does not reproduce a later hash in the chain.
The third stage is trust-slope continuity validation. The trust-slope is not a score; it is a lineage, an ordered sequence of biological hashes and their underlying stable sketches that constitutes the identity record within a given domain. Each new resolution event generates a hash that is evaluated for continuity with the recent entries in the trust-slope: a new sketch consistent with the trajectory is appended at high confidence, a borderline sketch is appended with a reduced-confidence annotation, and a sketch inconsistent with the trajectory produces a detectable discontinuity. In an embodiment the disclosure builds a predictive identity trajectory with an acceptance envelope, so natural physiological drift falls inside the envelope while substitution falls outside it. Continuity, not similarity to an enrollment photo, is the anti-spoofing mechanism, which is why a stolen hash, a static template theft, and a deepfake aimed at a single point all fail to reconstruct a valid trajectory.
Around this core the disclosure specifies further structural properties that a deployment can adopt. Identity resolution modes are consent-gated: one-to-one verification against an asserted identity, one-to-many identification against a population, and a hybrid mode, with the mode selection structurally enforced so that a one-to-one interaction cannot silently perform one-to-many identification. Population-scale disambiguation uses coarse-band sketch assignments to narrow candidates and fine-band assignments to discriminate, with the temporal trajectory as an additional axis. Cross-modal biological hash fusion combines modalities to strengthen continuity. Biological state inference is offered as a first-class but explicitly non-diagnostic function: it classifies deviation relative to the individual's own continuity baseline (for example, to gate a safety-critical capability) and does not diagnose medical conditions, measure blood alcohol content, or output a clinical value. Quorum-based recovery restores identity continuity after a capture compromise or a broken chain without disclosing biometric contents, by accumulating credentialed attestations up to a policy-defined quorum. Throughout, no raw biological data and no reusable biometric template are stored; the record is the hash lineage and its helper data, and privacy-governed disclosure exposes the governance and decision context of a resolution rather than the signal or the sketch. The disclosure is technology-neutral as to the specific sketch construction, hash function, and slope algorithm, so these are implementation choices rather than limits on the inventive step.
4. Composition Pathway
In a composition of this kind, Onfido would act as a domain-specialized capture and document-authentication surface running over the biological-identity layer. What stays at Onfido: the document corpus and classifier stack, the SDKs and capture UX, the liveness engine, the watchlist and PEP screening modules, the regulatory-mapping work that translates KYC requirements into verification configurations, and the entire customer-services and account-management commercial relationship. Onfido's investment in document expertise, its broad multi-jurisdiction issuer-template library, and its anti-spoofing work against deepfake injection remain its differentiated layer and are not displaced.
What the layer adds: the biometric capture is reduced at the edge to a stable sketch, the sketch is hashed with a per-domain separation tag and a rotating salt into a biological hash, and that hash is appended to the individual's trust-slope lineage under the relying party's or coalition's authority taxonomy rather than scored against a single stored photo. The integration points are well-defined. Onfido's mobile and web SDKs emit stable sketches (and their derived hashes) alongside or instead of raw biometric captures, so no reusable template leaves the device; the Onfido decisioning engine consumes the trust-slope continuity state as an input to its current verdict logic, so a clean document plus a discontinuous trajectory produces a graduated outcome (step-up, defer, conditional admit) rather than a binary pass; and continuity updates persist as hash-lineage records under the customer's authority rather than as biometric templates in a vendor database.
The composition addresses the genuine-document attack surface directly on the continuity axis. An attacker presenting a stolen but genuine credential matches the document and can pass liveness, yet cannot reproduce the legitimate holder's biological hash lineage, so the new hash produces a continuity discontinuity and the layer returns a graduated step-up outcome rather than a clean admit. A legitimate holder re-verifying after years presents a sketch far from the enrollment photo but inside the predictive acceptance envelope of their own trajectory, so the drift is admitted as expected physiological evolution rather than failed as a mismatch. Cross-operator continuity follows from domain separation: a holder's trajectory accumulated at one relying party composes into a coalition trust-slope under a shared authority taxonomy without either operator holding the other's biometric data, because the hashes are unlinkable across domains except through governed composition.
The new commercial surface for Onfido is continuity-as-substrate for regulated customers that need verification capable of detecting genuine-document substitution and that need biometric-data minimization sufficient to navigate GDPR, BIPA, and emerging biometric-storage regimes. The trajectory belongs to the customer's authority taxonomy rather than to Onfido's database, which paradoxically makes Onfido stickier: the customer's continuity is portable, but Onfido's capture quality, document corpus, and liveness engine are what differentiate access to that substrate.
5. Commercial and Licensing Implication
One fitting arrangement is an embedded-layer license: Onfido, now under Entrust, embeds the continuity-based biological-identity layer into its verification platform and offers trust-slope participation to its enterprise customers as part of the verification subscription. Pricing on a per-credentialed-trajectory or per-continuity-update basis rather than purely per-verification would align with how regulated customers want to consume continuity, as an accumulating asset rather than a stack of disconnected events.
What Onfido gains: a structural answer to the genuine-document attack surface that document and liveness classifiers cannot close on their own; a defensible position against the wave of AI-driven verification entrants and against in-platform competition from cloud-hyperscaler identity products; biometric-data-minimization posture that converts a regulatory liability into a regulatory advantage under GDPR, the EU AI Act, BIPA, and the next generation of biometric-storage law; and a forward-compatible footing as deepfake injection and synthetic-identity attacks continue to erode single-point verification. What the customer gains: portable continuity across vendor and jurisdiction changes, detection capability that genuine-document substitution does not defeat, and a single trajectory spanning onboarding, transaction-time step-up, and periodic re-verification under one authority taxonomy. Honest framing: the disclosed layer does not replace identity verification; it gives document-plus-selfie verification a continuity substrate that a single-point match cannot provide on its own.
6. Disclosure Scope
The inventive subject matter described here, the continuity-based biological-identity layer built on stable sketching, biological hashing with domain separation and salt rotation, trust-slope continuity validation, consent-gated resolution modes, cross-modal fusion, explicitly non-diagnostic state inference, privacy-governed disclosure, and quorum-based recovery, is disclosed in United States Patent Application 19/647,395. The claims about what the disclosed layer does, including that no raw biometric data or reusable template is stored and that state inference is non-diagnostic, trace to that application and are the subject of this public disclosure.
References to Onfido, Entrust, and other named identity-verification vendors, and the descriptions of their products, markets, pricing models, and regulatory posture, are external context provided for comparison. They are based on generally available public information, describe those products at the architecture level, and are not claims of United States Patent Application 19/647,395. Named products are the property of their respective owners; nothing here asserts a specific false-match rate, storage practice, or regulatory finding against any named vendor. Where a competitor characteristic is stated, it is stated as a general architectural property of single-event verification, not as a proprietary defect.
