Building spatial intelligence that earns trust.

01

Why safety matters for spatial intelligence

Spatial decisions act on the physical world. Getting them right — and keeping people in control — is the whole point.

Nexma builds spatial intelligence that informs consequential decisions — where to route, what to build, how to allocate scarce resources across real terrain. Those decisions have physical consequences, so the bar for getting them right is higher than for systems that only produce text.

Nexma is the decision and orchestration layer, not the actuator. We help organizations reason about the physical world; the authority to act stays with the people and institutions accountable for the outcome. Safety, for us, means making that reasoning faithful, legible, and contestable — and never letting automation quietly become authority.

02

Our safety commitments

Five principles govern how we design, deploy, and improve the platform.

• Human oversight — a person remains accountable for consequential decisions; the system informs, it does not decide unchecked.
• Verifiability — outputs are grounded in an explicit world model and can be inspected, reproduced, and challenged.
• Transparency — we are honest about what the system can and cannot do, and about its uncertainty.
• Privacy — we minimize the data we use, protect it, and use it only for the purposes we disclose.
• Fairness — we work to prevent our technology from producing or entrenching unjust outcomes.

03

How we build safely

Safety is engineered into the architecture, not bolted on after.

Our agents reason over an explicit, typed world model — an ontology of entities, relationships, and constraints — rather than free-associating over a map. Proposed changes are checked against those constraints before they are surfaced, and consequential actions pass through human review gates.

• World-model grounding — every action references real entities and their defined relationships, not loose inference.
• Verifiable optimization — where we optimize, results are checked against the constraints and objectives that define a valid solution.
• Constraint validation — proposals are validated against the active ontology's rules before a human ever sees them.
• Human-in-the-loop gates — high-impact operations require explicit review and approval.
• Conservative margins and rollback — the system favors safe defaults and changes can be inspected and reversed.

04

Evaluation and red-teaming

We test for failure before deployment, and keep testing after.

We evaluate the platform against representative tasks and known failure modes, probe it adversarially to find where it breaks, and quantify uncertainty so that low-confidence outputs are flagged rather than presented as fact. Evaluation is continuous: real-world use surfaces new cases that feed back into testing.

05

Dual-use and responsible deployment

Powerful general-purpose technology demands judgment about how and with whom it is used.

Nexma's platform is general-purpose and serves a wide range of beneficial uses. We recognize that capable technology can be misused, and we apply heightened review to sensitive deployments — including defense, intelligence, and other dual-use contexts — assessing the end user, the stated use, and the destination before proceeding.

These reviews are backed by binding commitments. See our Trade Compliance & Export Control Policy and our Human Rights & Modern Slavery Statement for the obligations that govern where and how our technology may be deployed.

06

Human oversight and governance

Accountability is assigned, not assumed.

Responsibility for safety sits with Nexma's leadership. We define review gates for consequential capabilities, document the decisions behind them, and maintain escalation paths so that concerns reach people empowered to act. As our capabilities grow, our governance scales with them.

07

Data and model integrity

Safe outputs depend on trustworthy inputs and disciplined use of data.

We handle customer data in accordance with our privacy commitments and do not use it in ways we have not disclosed. Our approach to model behavior, data, and the limits of automation is described in our Responsible AI policy, and our handling of personal data in our Privacy Policy.

08

Reporting a concern

If something looks wrong, tell us.

We welcome reports of safety issues, vulnerabilities, and misuse. Security concerns can be sent to security@nexma.ai and broader safety or policy concerns to legal@nexma.ai. We review reports made in good faith and prohibit retaliation against those who raise them.