AI Risk and Accountability Assessment

System Identification

Core identifying information for procurement and regulatory review.

Risk Classification

Assessment against Canada's Directive on Automated Decision-Making and equivalent institutional AI governance frameworks.

The Level I classification is supported by four independent findings, each of which independently places the system at the lowest impact level:

• No administrative decisions are made or supported COMPAiSS is an information navigation tool. It directs users to verified institutional sources. No eligibility determination, ranking, scoring, prioritization, or decision affecting rights, benefits, or entitlements is made by the system at any point. All decision authority remains with the institution and the user.
• No personal information is collected, stored, or processed Queries are stateless. The system receives text input only. No names, identification numbers, account information, or any personally identifiable information is collected or retained. Nothing persists after the transaction completes. A Privacy Impact Assessment is not required.
• All impacts are immediately and fully reversible The system provides information only. No action is taken on the user's behalf. A user who receives an unhelpful or incorrect response retains full ability to contact institutional staff directly through existing channels at any time. No rights, benefits, or entitlements are affected by the system's output.
• The pre-inference execution gate eliminates the primary unauthorized institutional hallucination risk category. Source validation and scope enforcement occur before the AI generates any response. COMPAiSS does not generate answers beyond authorized institutional sources, cannot cite unauthorized sources, and cannot produce content outside the authorized institutional source set.

Data Governance

What the system collects, processes, and retains - and what it does not.

Aggregated, anonymized usage patterns - for example, topic areas generating high query volume - can be made available to the institution for service planning. This analytics function operates exclusively on de-identified aggregate data and cannot be traced back to any individual user.

Architecture Accountability

How the system enforces scope, maintains audit trails, and supports institutional oversight.

COMPAiSS is built on an execution-gated inference architecture. Authorization and scope validation occur before any AI generation takes place. If no authorized institutional source exists for a query, the AI model does not execute. This is the core architectural accountability property - and it is not a configuration setting that can be overridden at the user level.

• Pre-inference execution gate Every query is validated against the institution's authorized source list before any AI generation occurs. Out-of-scope queries are rejected at the gate. The AI model only runs when authorization conditions are satisfied. This constraint is architectural - it cannot be prompted, reframed, or manipulated away.
• Authorized source list (greenlist) governance The institution retains full authority to review, approve, and modify the authorized source list at any time. COMPAiSS cannot access any page, document, or source that is not on the approved list. Every response is traceable to specific authorized sources. See Greenlist Control, Authorization Governance, and Ongoing Source Auditing for how this scales across departments and functions as institutional documentary evidence for procurement and audit review.
• Audit trail and logging System logs capture every query received, the gate authorization result, and the sources parsed for each transaction. Each response is traceable to specific authorized source pages. Logs identify the platform version, governance configuration version, and AI model version used for each transaction.
• Version control and change management All changes to the governance configuration, authorized source list, prompt architecture, and platform infrastructure are committed to a version-controlled repository with timestamps and change records. The institution maintains a complete history of all governance configuration changes.
• Greenlist integrity auditing A purpose-built audit server checks every URL in an institution's authorized source list against the live web on a regular maintenance cycle, classifying each entry as active, redirected, broken, or unreachable. Broken or moved URLs are remediated and the corrected list is deployed automatically through the version-controlled repository. See Greenlist Integrity and Maintenance for the full audit process.
• No continuous learning from user interactions The system does not learn from individual user queries. No interaction data is fed back into any model. The AI model improves only as Microsoft updates the Azure OpenAI models - a process entirely external to institutional data and interactions, with no institutional data used for model training.

Risk Register

Identified risks by category with architectural and operational mitigations.

No ethical, financial, privacy, or legal risks have been identified that would impede institutional deployment. The following risks have been identified and mitigated:

Technical Risks

• AI inference API availability Mitigation: Standard error handling and fallback messaging direct users to institutional staff when the inference layer is unavailable. The system degrades gracefully - it does not produce unreliable outputs when the API is degraded.
• Authorized source content becoming outdated Mitigation: Active greenlist maintenance on a regular audit cycle. The institution flags when source content changes. Outdated or removed pages are removed from the authorized list, preventing the system from citing content that no longer reflects current policy.
• Source URL unavailable at query time Mitigation: If a source is unreachable at the moment of a query, the system draws from other authorized sources that did respond, or returns a safe failure response directing the user to institutional staff. It cannot fabricate content from a source it cannot reach.

Governance Risks

• Scope creep beyond authorized content Mitigation: Architectural - the pre-inference execution gate cannot be overridden at the user level. Out-of-scope queries are rejected on the first attempt and every subsequent attempt regardless of how they are phrased or reframed.
• Adversarial prompt manipulation Mitigation: Input filtering blocks excessively long inputs, garbage text, and adversarial prompt attempts before they reach the gate. The gate itself enforces scope boundaries deterministically. Attempts to manipulate the AI's instructions are blocked before they reach the model.

Equity Risks

• Digital accessibility barriers Mitigation: Web-embedded design accessible from any browser-enabled device without specialized software or installation. Users who cannot use or prefer not to use the system retain full access to all the same information through existing institutional channels.
• Language barriers Mitigation: Multilingual query support - users can submit queries in any language. Translation is treated as a preprocessing step outside the authorization framework, ensuring that the language of the query does not affect what institutional content is considered authoritative.

Reputational Risks

• Public or institutional concerns about AI deployment Mitigation: The governance-first architecture, full auditability, stateless privacy-by-design approach, and pre-inference execution gate that structurally eliminates all standard hallucination sources directly address the concerns that generate institutional scrutiny. The system is designed to be auditable and defensible from the concept stage.

Human Oversight and Recourse

How institutional authority is preserved and how users escalate when the system does not meet their needs.

Human oversight and intervention operate at multiple levels within every COMPAiSS deployment. The system is designed from the concept stage to preserve institutional authority and user recourse at every point.

• Immediate user recourse Every system response includes links to official institutional pages and directs users to contact institutional staff directly for further assistance. A user who receives an incorrect or unhelpful response can contact staff through existing channels at any time. No action is taken on their behalf and no rights or entitlements are affected.
• Institutional override authority The deploying institution retains full authority to modify the authorized source list, update governance configuration, or disable the system at any time. All such changes take effect immediately and are logged with timestamps in the version-controlled repository.
• Crisis query handling When a query contains language associated with personal distress or crisis, the system returns a pre-approved institutional response directing the user to appropriate support resources - counselling services, campus security, or emergency contacts - as configured by the institution in alignment with its own duty-of-care policies. The AI model does not engage with crisis content.
• Developer accountability Frank Harvey holds direct technical accountability for all design, development, maintenance, and governance configuration. COMPAiSS Inc. is the incorporated entity that holds the intellectual property and platform operating rights. Service levels and accountability obligations are defined in each institutional licensing agreement.
• Non-automated alternative always available COMPAiSS is a supplementary channel. All existing institutional information channels, staff resources, and service options remain available at all times. The system's unavailability does not affect access to institutional services or information.

Procurement Readiness

Documentation and compliance status for institutional procurement and governance review processes.

Canada's National AI Strategy and Regulated Institutions

How Canada's evolving AI governance framework aligns with execution-gated institutional AI.

Canada's National Artificial Intelligence Strategy recognizes both the opportunities and risks associated with widespread AI adoption. The strategy emphasizes responsible deployment, public trust, transparency, accountability, and governance as essential conditions for the successful use of AI across public institutions and regulated sectors. It reflects a broader international trend toward trustworthy AI, accountable AI systems, and governance frameworks for regulated institutional environments.

As AI systems become more capable, the governance challenge becomes increasingly important. Institutions are expected not only to demonstrate that AI systems are useful, but also that they operate within clearly defined accountability frameworks, respect institutional authority, and maintain public confidence.

What This Means for Regulated Institutions

• Governance must be demonstrable Institutions increasingly require evidence that AI systems operate within documented accountability frameworks rather than relying solely on user disclaimers or post-generation monitoring.
• Authorized information boundaries matter Institutions must be able to explain where information originates, which sources are considered authoritative, and how unauthorized content is prevented.
• Auditability becomes essential Governance frameworks increasingly emphasize traceability, transparency, and institutional oversight of AI-mediated interactions.
• Public trust depends on accountability The ability to demonstrate how AI systems are governed may become as important as the performance of the models themselves.

For institutions evaluating AI under emerging Canadian governance frameworks, the distinction between generation-first architectures and governance-first architectures such as execution-gated inference may become increasingly significant as expectations surrounding accountability, transparency, and responsible AI continue to evolve.

AI Governance Framework | Why AI Developed This Way | Governance Stress Tests and Benchmarks

Policy and Regulatory Alignment Report

How execution-gated inference maps to Canadian and international governance frameworks.

COMPAiSS has published a structured policy and regulatory alignment report evaluating the execution-gated inference architecture against Canada's Voluntary Code of Conduct on Responsible Generative AI, the National AI Strategy AI for All (June 2026), the NIST AI Risk Management Framework, and Zero Trust Architecture standards.

Request Documentation

How to obtain formal governance documentation for procurement or regulatory review.

The following documentation is available on request for any institution conducting a formal procurement or governance review:

• Complete Algorithmic Impact Assessment Full AIA responses across all sections, including risk classification rationale, data handling assessment, procedural fairness analysis, and system accountability documentation.
• Technical architecture documentation Detailed description of the pre-inference execution gate, greenlist governance model, audit trail architecture, and defense-in-depth security controls.
• Pilot deployment results Governance configuration documentation, query volume and gate pass/fail rate analysis, and institutional feedback from active deployments.
• Institution-specific governance package A tailored governance documentation package scoped to your institution's specific procurement framework, AUP requirements, and review process - prepared in advance of formal submission.