On 22 May 2026, ICE finalised a $25.1 million sole-source contract with Bi2 Technologies, a Massachusetts-based biometric firm, for iris recognition hardware and database access to be deployed in field operations by late June. The award, first reported by Project Salt Box and subsequently confirmed via SAM.gov, represents a fivefold increase over ICE's initial $4.6 million engagement with the same vendor just eight months prior. The device count scales even more dramatically: 1,570 units compared to the original 200, an increase of nearly 685 percent. This is not an incremental procurement. It is a rapid, large-scale deployment of a biometric identification system with, at present, no documented independent audit trail, no congressional notification requirement, and no FedRAMP compliance mandate before operational use.
The technical and governance implications deserve careful unpacking. Iris recognition is among the most accurate biometric modalities available, with well-studied false match rates in the range of one in a million or better under controlled conditions. However, field deployment introduces a set of confounding variables that laboratory benchmarks rarely capture: variable illumination, subject non-cooperation, occlusion from glasses or contact lenses, and the physiological effects of stress or injury on iris texture. The gap between benchmark accuracy and operational accuracy in adversarial or uncontrolled field conditions is a persistent and underappreciated problem across all deployed biometric systems.
What the Biometric Information System Actually Contains
The contract grants ICE's Enforcement and Removal Operations division continuous access to Bi2's database of over five million booking records. Booking records, in this context, are biometric templates captured at the point of arrest or detention, typically in a law enforcement or corrections setting. This is a meaningful distinction. The database does not represent a random sample of the population; it is drawn almost entirely from individuals who have had prior contact with the criminal justice or immigration detention system.
Several structural properties of this dataset warrant scrutiny:
- Selection bias: A database constructed from booking records will systematically over-represent communities that face disproportionate policing. Any model trained or calibrated against this data inherits those distributional properties.
- Template provenance: Biometric templates captured in custody settings may not have been collected with informed consent for secondary use in field identification. The legal basis for repurposing these records across jurisdictions and agencies is not addressed in the publicly available contract language.
- Staleness: Iris biometrics are generally considered stable across a lifetime, but template quality degrades with image resolution, sensor variation across devices, and age-related changes in the periocular region. A database of five million records of uncertain vintage and capture quality introduces non-trivial false non-match risk.
- Scale of false positives: Even at a false match rate of one in a million, deploying 1,570 devices across a large operational population means statistically inevitable misidentifications at scale. The contract describes no adjudication protocol for disputed matches.
The FedRAMP Gap and What It Signals
FedRAMP, the Federal Risk and Authorization Management Program, exists precisely for situations like this. It provides a standardised security assessment framework for cloud systems that handle sensitive federal data. The fact that this procurement explicitly does not require FedRAMP authorisation before deployment is not a bureaucratic footnote; it is a substantive decision with real security consequences.
Biometric templates are, by definition, non-revocable credentials. A compromised password can be reset. A compromised iris template cannot. The storage, transmission, and access control architecture for a database of five million iris templates is a high-value target for both criminal actors and foreign intelligence services. Without FedRAMP or an equivalent independent security review, there is no public assurance that the system meets baseline controls around encryption at rest and in transit, access logging, incident response, or data retention limits.
This is particularly notable given the trajectory of recent high-profile government data breaches. The OPM breach of 2015 exposed fingerprint records for approximately 5.6 million federal employees. The lesson drawn by the security community was that biometric databases require exceptionally rigorous protection precisely because the damage from exposure is permanent. Deploying a comparable system without mandatory pre-deployment security review suggests either that this lesson has not been institutionally absorbed, or that operational speed was prioritised over security assurance.
No-Bid Procurement and the Accountability Deficit
The sole-source, no-bid nature of this contract deserves separate analysis from its technical dimensions. Competitive procurement exists to ensure that the government obtains best value, that vendors are evaluated against defined technical requirements, and that the process is open to challenge and review. Sole-source awards are permissible under the Federal Acquisition Regulation when there is only one responsible source capable of satisfying the requirement, but that determination should be documented and defensible.
The rapid scaling from $4.6 million to $25.1 million within eight months, awarded without competition, suggests that the initial contract functioned as a de facto vendor lock-in mechanism. Once an agency has deployed 200 devices, trained agents on a specific interface, and integrated with a proprietary database schema, the switching cost to a competing vendor becomes substantial. This is a well-documented procurement pathology in government technology acquisition, and it is especially pronounced in biometric systems where database interoperability is technically complex.
The absence of congressional notification or independent oversight review compounds this. Congress has historically exercised oversight over large-scale biometric programmes through appropriations riders, committee hearings, and inspector general referrals. A $25 million deployment of a novel field biometric system, scaled to nearly 1,600 devices within weeks, would ordinarily attract that scrutiny. The contract structure, as described, appears designed to move faster than that oversight mechanism can engage.
Broader Context: Field Biometrics and the Erosion of Operational Constraints
This contract sits within a broader trend that researchers in AI policy and civil liberties have been tracking for several years. The cost of biometric hardware has fallen sharply. Iris recognition, once confined to high-security fixed installations, is now available in handheld form factors affordable enough for mass field deployment. The combination of cheaper hardware, larger databases, and faster network connectivity has collapsed the practical barriers that previously limited biometric identification to controlled environments.
What has not kept pace is the governance infrastructure. The academic literature on algorithmic accountability, from Raji and Buolamwini's work on commercial facial recognition to more recent studies on iris recognition accuracy across demographic groups, consistently identifies a gap between deployment speed and evaluation rigour. Field biometric systems used in enforcement contexts carry particularly high stakes for false positives: a misidentification during an immigration enforcement operation can result in wrongful detention, family separation, or deportation of a person with legal status.
The Bi2 contract does not, based on available public documentation, describe any independent accuracy evaluation on the specific population against which the system will be deployed, any demographic disaggregation of error rates, or any mechanism for subjects to contest a biometric identification. These are not exotic requirements. They are the baseline that responsible deployment of any consequential automated decision system should meet.
What Rigorous Oversight Would Look Like
For researchers and policy practitioners thinking about what adequate governance of a system like this would require, the minimum viable framework would include:
- Pre-deployment FedRAMP High authorisation, given the sensitivity of the data and the scale of the deployment.
- Independent accuracy evaluation on a demographically representative sample of the operational population, with published false match and false non-match rates disaggregated by age, sex, and ethnicity.
- A documented adjudication protocol specifying what happens when a field match is disputed, who has authority to override the system, and what human review is required before any enforcement action is taken based solely on a biometric match.
- A data retention and deletion schedule specifying how long templates and match logs are stored, and under what conditions they can be shared with other agencies.
- Congressional notification and an inspector general review mechanism with a defined timeline.
None of these are technically infeasible. Several are standard practice in comparable deployments in allied jurisdictions. Their absence from this contract is a policy choice, not a technical constraint.
The trajectory here is clear. Biometric field identification is becoming a standard tool of immigration enforcement, scaled rapidly through sole-source procurement, without the oversight architecture that the sensitivity of the technology and the stakes of its application require. The technical community has a responsibility to be precise about what that means: not just in terms of civil liberties abstractions, but in concrete terms of false positive rates, data security exposure, and the systematic absence of accountability mechanisms. The 1,570 devices due at ICE locations by late June will be operational long before any of those questions are publicly answered.