IREX
EN

Engineered for ethics

IREX operates on a doctrine of Transparency by Design. Accountability is built into the architecture, the access controls, and the audit trail, rather than added afterwards as policy. This page describes the six pillars of that doctrine and, for each one, the platform mechanism that enforces it.

The distinction matters. A policy states what operators must not do. An architectural constraint decides what the platform will do at all. Where the two are available, IREX chooses the constraint, because a constraint survives staff turnover, deadline pressure, and a change of administration.

The six pillars

Magnifying glass with an eye icon representing full transparency Full transparency

Every action is recorded in a detailed, non-erasable log: adding a person to a database, running a search by photo, configuring an alert, exporting video, changing a permission. The log is fully searchable and auditable, which protects the agency and the public at the same time. Transparent policing is what builds trust between law enforcement and the community it serves.

Padlock icon representing cybersecurity and privacy protection Cybersecurity and privacy protection

The platform is security-hardened for public deployment. Controls include TLS 1.3 in transit, a multilayer web application firewall, intrusion detection and prevention, network filtering, signed API tokens with rotated keys, and multi-factor authentication. Penetration testing runs against every major release. Customer data is owned by the customer alone.


Icon of a group of people with one person highlighted, representing narrow constraints for law enforcement Narrow constraints for law enforcement

The platform detects pre-identified persons only: to prevent an attack, to find a missing child, or to address another public safety threat. The number of persons that can be enrolled for real-time biometric identification is limited by the platform itself. There is no recognition or tracking of random people, and the neural networks are trained only on scenes and subjects that bear on public safety.

Icon of three connected user profiles with one highlighted, representing permission-driven access Permission-driven user interface

Platform features, person databases, match alerts, and search results are visible only to authorized staff. A user who has not been added to a group with sufficient permissions does not see the data, and configuration changes are limited to top-level roles. Compartmentalization means no single account holds unchecked access to the whole system.


Balance scale icon representing bias awareness Bias awareness and mitigation

Every face recognition algorithm, and the human eye as well, performs differently across demographic groups of gender, age, and skin color. Natural variation in training data makes the effect hard to remove. IREX data scientists work continuously to reduce it by balancing datasets and tuning algorithms. IREX also participates in the NIST Face Recognition Vendor Test (FRVT), so the demographic effects are measured independently. Where a systematic bias remains, IREX policy is to disclose its existence and estimated size to the customer rather than to leave it unstated.

Checklist icon representing ethics best practices and checklists Ethics best practices and checklists

IREX works with privacy, civil-liberties, and law-enforcement experts to produce the governance material an agency needs: user guidelines for each capability, and audit processes for internal and external oversight. Compliance checklists cover data handling, from what is created to how, by whom, and for how long it is accessed, stored, and shared.

What enforces each pillar

Each pillar corresponds to a mechanism an auditor can inspect and an administrator can configure. The right-hand column points to the chapter that documents it.

Pillar Enforcement mechanism Documented in
Full transparency Mandatory Case ID on privacy-sensitive actions; append-only Logbook; daily signed export to an external archive. Case ID and accountability, View the Logbook
Cybersecurity and privacy Encryption in transit, defense in depth, signed API tokens, multi-factor authentication, per-release penetration testing. Access your workspace, Regulatory alignment
Narrow constraints Watchlist-only recognition; a platform-enforced cap on the real-time biometric database; models trained only on public safety subjects. Person and vehicle lists, Manage databases
Permission-driven interface Hierarchical role-based access control down to the individual camera, recording, media item, and module; configuration limited to top-level roles. Manage users and user groups
Bias awareness A 40-million-image CCTV-optimized training dataset, continuous dataset balancing, NIST FRVT participation, and disclosure of known bias. FaceTrack Pro
Ethics best practices User guidelines, audit processes, and compliance checklists, plus the audit tooling that makes review practical rather than theoretical. Regulatory alignment, View the Logbook

Ethics under agentic AI

Autonomous investigation agents raise an obvious question: if software can plan and run a multi-step search, who answers for it? The platform's answer is that an agent is never a separate actor.

  • Every agent action is logged and attributed to the operator who initiated it and to the Case ID that authorizes it.
  • Agents operate inside the initiating operator's permission boundary and cannot escalate their own access.
  • Agent-initiated searches, watchlist queries, and alert configurations are auditable through the same Logbook and the same daily export as manual actions.
  • No consequential decision is taken by an agent alone. The agent returns findings; a person decides.

See Ask IREX for the capability and its guardrails, and StreamVLM for how the same rules apply to detectors defined by a prompt.

Where responsibility sits

The platform's architecture supports lawful, ethical operation. It does not deliver legal compliance on its own. The lawful basis for processing biometric data, the retention policy, the watchlist enrollment criteria, and the oversight process are decisions for the operating agency and its counsel. Each rests on the law that applies to them. What IREX guarantees is that those decisions are enforceable in the system and verifiable afterwards. See Regulatory alignment.

Related topics