aiGUARD™
The non-bypassable execution-control architecture for AI outputs. A runtime governance layer designed to support EU AI Act risk-management, logging, and transparency obligations.
Architecturally enforced AI output control.
aiGUARD™ is implemented on the Thames Sentinel™ architecture, a patent-pending runtime governance layer that evaluates every AI output against confidence, consequence, and user-state parameters before delivery.
Most AI governance is advisory. Policies, alerts, dashboards, post-hoc reviews. The response generation component still controls delivery; governance is a layer above it that hopes to be consulted. Thames Sentinel inverts that. The response generation component is structurally prevented from initiating delivery. Delivery happens only when a separate, dedicated control component issues an execution permission.
The architecture is non-bypassable not as a software promise but as a structural property, the response generation component does not hold the credentials, interfaces, or network access that would allow it to deliver an output without a valid execution permission.
Every output. Every parameter. Every time.
Confidence
A computational score derived from validation processes, cross-referencing trusted reference corpora, model-internal uncertainty estimation, anomaly detection, or multi-source triangulation. Mapped to a band (GREEN / AMBER / RED) per the active policy profile. Both score and band recorded.
Consequence
A classification of potential impact: LOW, MEDIUM, HIGH, or CRITICAL. Based on output domain, deployment context, regulatory status, and predefined hierarchies. Medication dosage in a patient-facing context: HIGH or CRITICAL. General navigation in a non-vulnerable context: LOW.
User state
The operational interaction context. Detected emotional cues. User vulnerability indicators. Role, access level, accessibility needs. Hashed before recording, preserving governance evidence without exposing personally identifiable information.
Built with IBM.
IBM is our reference build partner for the aiGUARD™ architecture. The Thames Sentinel build specification was finalised in February 2026 under SoW1, establishing the integration pattern for enterprise agentic AI deployments. SoW2 production build is in scoping.
The reference build is the canonical implementation that other deployments will derive from. It demonstrates the architecture operating at enterprise scale, integrated with IBM watsonx Orchestrate, and producing the audit artefacts required for EU AI Act Article 12 compliance.
EU AI Act compliance, as architecture.
aiGUARD™ directly addresses the obligations of EU AI Act Articles 9, 12, and 13 for high-risk AI systems, under the EU AI Act timetable, as amended or confirmed by the final AI Omnibus process.
Deploy aiGUARD in your environment.
aiGUARD™ deployments are scoped on a per-engagement basis with regulated-industry partners. Architecture documentation is provided under NDA.
Get in touch