Trustworthy Model

This article defines how askEdgi ensures trustworthy AI behavior, including data governance, explainability, auditability, and responsible use of artificial intelligence when analyzing customer-provided data.

askEdgi is designed to assist authorized users in analyzing, transforming, and understanding data they are permitted to access. Trustworthiness is achieved through deterministic execution, strong governance controls, transparent explanations, and comprehensive auditability.

Applies To: askEdgi SaaS Platform

Scope and System Boundaries

askEdgi operates within the following boundaries:

In scope

• User-uploaded files (e.g., spreadsheets, CSVs)

• Data accessed via approved connectors

• Deterministic computation in sandboxed execution environments

• AI-assisted planning and explanation

• Workspace-level governance and access controls

Out of scope

• Autonomous decision-making

• Scoring or profiling individuals

• Unapproved data enrichment

• Silent modification of customer data

askEdgi does not access external data sources unless explicitly configured and authorized by the customer.

Definitions

• Customer Content: Data, files, prompts, generated code, analysis outputs, and derived artifacts associated with a customer workspace.

• Personal Data: Any information relating to an identified or identifiable individual.

• System Metadata: Security logs, performance telemetry, and operational metrics excluding raw customer datasets.

• Execution Sandbox: Isolated compute environment used for deterministic operations.

• Model Provider: Third-party AI model API used for interpretation and explanation.

Trustworthy AI Principles

askEdgi adheres to the following principles:

Groundedness & Accuracy

• All numeric, statistical, and analytical results in askEdgi are generated through deterministic execution engines, such as SQL and Python, operating within controlled execution sandboxes. Language models are not used to compute values, derive metrics, or perform calculations. This separation ensures that analytical outputs are reproducible, verifiable, and consistent across executions, providing users with reliable results that can be independently validated.

• Language models in askEdgi are limited to interpreting user intent expressed in natural language and generating human-readable explanations of executed results. They do not perform computations, generate synthetic values, or infer missing data. All explanations are grounded in executed logic and validated outputs, ensuring that narratives accurately reflect what was actually run and returned by the system.

• When the available data, metadata, or context is insufficient to answer a user’s question accurately, askEdgi does not attempt to infer or guess results. Instead, the system explicitly signals uncertainty and prompts the user for clarification, additional context, or dataset selection. This behavior prevents misleading outputs and reinforces responsible, accuracy-first analytics.

Explainability

askEdgi delivers explainability through a layered explanation model that caters to both business and technical users.

Each analysis may include a high-level summary describing the outcome in plain language, followed by detailed technical explanations that outline transformations, filters, joins, and assumptions applied during execution.

Where applicable, executable artifacts such as SQL queries or Python logic are available to support full reproducibility and independent review.

askEdgi provides layered explanations:

• Summary: Plain-language result description

• Technical Detail: Transformations, filters, joins, assumptions

• Reproducibility: Executable SQL, Python, or formula steps

Auditability

askEdgi maintains comprehensive auditability by logging all material user and system actions, including data access, query execution, recipe runs, and workspace operations. Each log entry is timestamped and associated with a specific user identity and workspace context. These logs support governance reviews, incident investigations, and compliance audits, and align with OvalEdge platform audit and retention policies.

Governance & Least Privilege

askEdgi enforces governance through role-based access control and strict workspace isolation. Users can only access data sources, connectors, and actions that are explicitly permitted by their assigned roles and underlying OvalEdge catalog permissions. Workspace boundaries ensure that analysis artifacts, uploaded files, and execution results are isolated per user, preventing unauthorized access across users or teams.

Safety & Policy Enforcement

askEdgi applies a strict instruction hierarchy in which system and governance policies always take precedence over user input and dataset content. All analytical execution occurs within sandboxed environments that isolate compute workloads and prevent persistence beyond approved storage. Output controls further restrict response formats and data exposure, reducing the risk of misuse or unintended data leakage.

Explainability & Provenance Model

Each analysis run may include:

• Data sources used (dataset IDs, connector references)

• Row and column counts accessed

• Filters, joins, aggregations applied

• Missing-data handling strategy

• Versioned execution artifacts

• Analysis run identifier

Users may request a full “show your work” view at any time.

Hallucination Mitigation

askEdgi mitigates hallucinations by:

• Separating planning from execution

• Validating narrative explanations against executed results

• Refusing to guess or fabricate values

• Enforcing schema awareness and type checking

Data Governance Controls

• Workspace isolation

• RBAC (Owner, Admin, Analyst, Viewer)

• Connector allow lists

• Read-only default mode

• Explicit approval for write-back operations

• Optional sensitive-data redaction rules

Human-in-the-Loop Controls

askEdgi does not perform automatic write-back to source systems or downstream platforms. All actions that could export, modify, or persist results require explicit user initiation and confirmation. This ensures that human oversight remains central to all impactful operations and prevents unintended changes.

AI Threat Mitigation

All user input, dataset content, and contextual data are treated as untrusted input within askEdgi. System-level instructions and governance policies override any instructions embedded in data or user prompts. Tool invocation and execution paths are validated against policy checks to prevent prompt injection and unauthorized behavior.

Prompt Injection

• Data treated as untrusted input

• System instructions override data instructions

• Tool invocation requires policy checks

Data Exfiltration

• Output size limits

• Aggregation-first responses

• Raw dumps are disabled by default

• Sensitive-field redaction

Cross-Tenant Leakage

• Tenant-scoped identity and encryption

• Strict authorization on all access paths

Model Governance

askEdgi maintains governance over AI model usage through versioned model registries and controlled prompt templates. Changes to prompts and model configurations follow defined change-control processes and are tested to ensure accuracy, safety, and consistency before release. Monitoring mechanisms are in place to detect unexpected behavior following model updates.

Compliance Alignment

This policy aligns with:

• OECD AI Principles

• NIST AI Risk Management Framework

• GDPR and applicable data protection laws

Review Cycle

The Trustworthy Model is reviewed on an annual basis or whenever there is a material change to askEdgi’s architecture, AI usage, or governance controls. This review process ensures that documented controls remain accurate, aligned with platform behavior, and consistent with evolving regulatory and governance expectations.

---

Copyright © 2026, OvalEdge LLC, Peachtree Corners, GA USA