Use cases differ by an order of magnitude in latency envelope, acceptable false-match rate, and legal authority, and cannot share a single matching pipeline without compromise. Checkpoint screening and forensic latent matching both use the same enrolled database but have entirely different operator-in-the-loop expectations — one has two seconds before queue congestion becomes a vulnerability, the other has hours. The most common architectural mistake is one-sizing a checkpoint flow onto forensic exploitation and inheriting the wrong thresholds.

The sub-second checkpoint window is not a soft target — it is enforced by physical queue dynamics, and a 5-second matcher creates the queue that becomes the security vulnerability. Every millisecond spent on uplink to an ABIS query is a millisecond unavailable for template extraction, fusion, and operator confirmation. Decisions made after the SLA is set have 10× less leverage than decisions that set the SLA correctly the first time.

NIST FRVT 2023 consistently reports top-performer accuracy rates exceeding 99.7% on mugshot-quality imagery at FMR 1e-6, but operational imagery degrades this substantially — pose, lighting, partial occlusion, and standoff distance each reduce measured accuracy. A single uniform threshold almost always misprices the cost curve in at least one mission, and the cost of a false match at HVI-targeting tier is categorically different from the cost at base-access tier.

E.O. 13526 governs classification of national security information and ICD 710 governs classification marking — both are binding on any biometric database that draws from intelligence sources. Routing classified enrolled templates through a commercial cloud inference endpoint is a classification spill at worst and a policy violation at best. Classification boundary is a first-order architectural decision, not an afterthought, and cross-domain solutions (CDS) must be scoped before any data plane design.

The Privacy Act of 1974 restricts federal agency collection, retention, and sharing of records on US persons, and DoD 5400.11 operationalizes this inside the Department. A biometric system that may incidentally encounter US persons — refugee processing, domestic installation access, joint operations with US law-enforcement — must have a published SORN or a documented exclusion, and the architecture must enforce the filter in the data plane rather than relying on operator discipline. Privacy Act violations carry criminal penalties and are a reliable source of Inspector General findings.

CJCSI 5221.01D governs the delegation of authority to commanders to disclose classified military information to foreign governments and international organizations, and it is the authority document most often cited in coalition biometric sharing. Bilateral biometric sharing agreements — including Five Eyes — specify exactly which modalities, which metadata, and which retention periods apply, and a matching pipeline that aggregates across partner sources must enforce these distinctions at the data plane. Getting this wrong creates diplomatic exposure alongside operator risk.