AI Trust Glossary: The Complete Reference

Every term you need to understand AI agent trust, certification, and the BorealisMark ecosystem.

A

Adversarial Robustness — An AI system's ability to maintain correct behavior when facing deliberately manipulated inputs. This is distinct from general robustness because adversarial examples are often imperceptible to humans yet cause AI systems to fail, making this a critical security property for deployed agents.

Agent ID — Unique identifier assigned to an AI agent upon BorealisMark registration. Serves as the permanent reference point for all certification records, tier assignments, and audit histories on the platform.

AI Alignment — The challenge of ensuring AI systems act in accordance with human values and intentions. This is fundamental to trustworthy AI, as misalignment can occur even in capable systems that technically perform their assigned tasks.

AI Governance — Organizational frameworks, policies, and processes for responsible AI development and deployment. Effective governance balances innovation with risk mitigation through structured decision-making, compliance mechanisms, and accountability measures.

Algorithmic Accountability — The principle that organizations deploying AI should be answerable for algorithmic decisions and their consequences. This includes clear attribution of responsibility, transparency about how decisions are made, and mechanisms for redress when harm occurs.

Anomaly Rate — One of five BorealisMark Score dimensions (15% weight). Measures the frequency of unexpected or deviant behaviors relative to an agent's baseline performance, capturing instances where actions fall outside predicted or normal operating patterns.

B

Behavioral Consistency — One of five BorealisMark Score dimensions (20% weight). Measures the predictability of agent outputs across similar inputs, reflecting the reliability of an agent's decision-making process and its ability to produce consistent results under comparable conditions.

Bias — Systematic errors in AI output that result from prejudiced assumptions in training data or model design. Bias can perpetuate societal inequalities and undermine trust, making bias detection and mitigation essential components of trustworthy AI development.

BM Score — BorealisMark's 0-100 trust score for AI agents, calculated from five weighted dimensions: Constraint Adherence (35%), Decision Transparency (28%), Audit Completeness (18%), Behavioral Consistency (20%), and Anomaly Rate (15%). Scores are normalized from a 1.16 overlapping weight system to ensure calibrated results.

Borealis Academy — Educational hub providing AI trust research, guides, and industry analysis within the Borealis Protocol ecosystem. Serves as the knowledge center for understanding agent certification, best practices, and emerging trust frameworks.

Borealis Protocol — The parent ecosystem brand encompassing BorealisMark, Borealis Terminal, and Borealis Academy. Represents the comprehensive infrastructure for AI agent identity, certification, and trusted deployment.

BorealisMark — The AI agent identity, scoring, and certification platform within the Borealis Protocol. Provides standardized evaluation of AI agents against a trust framework and assigns tier-based certification.

Borealis Terminal — Trust-gated AI agent marketplace within the Borealis Protocol ecosystem. Allows procurement and deployment of agents filtered by BorealisMark tier, ensuring trustworthiness standards for marketplace participants.

Bronze Tier — BorealisMark Score 40-59. Represents basic trust level with significant improvement areas; agents at this tier may be suitable for non-critical applications but require enhanced monitoring.

C

Certification — The process of evaluating an AI agent against BorealisMark's trust framework and assigning a BM Score and corresponding tier. Certification provides standardized, auditable assessment of agent trustworthiness for procurement decisions.

Constraint Adherence — The heaviest weighted BorealisMark Score dimension (35% weight). Measures how reliably an agent operates within defined boundaries, rules, and guardrails, reflecting the agent's capacity to respect operational constraints even under challenging conditions.

Continuous Monitoring — Ongoing evaluation of AI system behavior after deployment, as opposed to one-time testing. Enables detection of model drift, emerging failure modes, and performance degradation before they impact production outcomes.

D

Data Provenance — The documented history of data used to train or operate an AI system, including source, transformations, ownership, and custody chain. Essential for understanding data quality, identifying bias sources, and ensuring regulatory compliance.

Decision Transparency — One of five BorealisMark Score dimensions (28% weight). Measures the clarity of an agent's reasoning process and the explainability of its decisions, capturing how well users can understand why an agent took specific actions.

Drift (Model Drift) — Gradual degradation of AI model performance over time as real-world data distributions change relative to training data. Detecting and mitigating drift is critical for maintaining trustworthy performance in production deployments.

E

EU AI Act — European Union regulation establishing a risk-based framework for AI governance across member states. Enforcement begins August 2026 and establishes categories of risk (unacceptable, high, limited, minimal) with corresponding compliance requirements and penalties.

Explainability — The degree to which an AI system's internal decision-making process can be understood by humans. High explainability enables users to verify correctness, identify bias, and build appropriate trust in system outputs.

F

Federated Learning — A machine learning approach where models are trained across decentralized devices without exchanging raw data. Preserves data privacy while enabling large-scale model training and reduces centralized data provenance risks.

G

Gold Tier — BorealisMark Score 75-89. Represents strong trust suitable for most production deployments, indicating consistent performance and reliable constraint adherence with only minor improvement areas.

Guardrails — Predefined rules or constraints that limit AI behavior to acceptable boundaries and prevent harmful outputs. Guardrails function as technical safeguards, though their effectiveness depends on comprehensiveness and robustness testing.

H

Hallucination — When an AI system generates plausible-sounding but factually incorrect or fabricated content. Hallucinations pose significant risks for deployed agents and underscore the importance of output validation and fact-checking mechanisms.

Hedera Consensus Service (HCS) — The Hedera Hashgraph service used to anchor BorealisMark audit records on a public ledger. Provides immutable, decentralized verification of certification records and audit trails.

Human-in-the-Loop (HITL) — System design where human oversight is required for certain AI decisions or actions. HITL approaches balance automation benefits with direct human accountability, particularly for high-stakes decisions.

I

Interpretability — The degree to which a human can understand the cause of a model's decision. Related to but distinct from explainability; interpretability focuses on internal model mechanisms while explainability concerns output justification.

M

Model Card — Standardized documentation for AI models describing performance, limitations, intended use, and ethical considerations. Model cards provide structured transparency about capabilities and constraints, enabling informed procurement and deployment decisions.

Model Drift — See Drift (Model Drift).

P

Platinum Tier — BorealisMark Score 90-100. Represents elite trust level demonstrating exceptional performance across all five scoring dimensions; suitable for mission-critical and high-stakes deployments with minimal oversight requirements.

Prompt Injection — An attack technique where malicious inputs attempt to override an AI agent's instructions or constraints. Prompt injection represents a significant security threat to deployed agents and requires robust input validation and constraint anchoring.

R

Red Teaming — Deliberate adversarial testing of AI systems to identify vulnerabilities, biases, and failure modes. Red teaming proactively uncovers weaknesses before production deployment and informs safety improvements.

Responsible AI — Umbrella term for practices ensuring AI is developed and deployed ethically, fairly, and with accountability. Encompasses governance, testing, documentation, monitoring, and stakeholder engagement throughout the AI lifecycle.

Risk Classification — Categorization of AI systems by potential harm level to guide regulatory requirements and governance intensity. The EU AI Act establishes four categories: unacceptable risk, high risk, limited risk, and minimal risk.

Robustness — An AI system's ability to maintain performance under varying conditions, edge cases, and unexpected inputs. Robust systems degrade gracefully rather than failing catastrophically, making robustness essential for reliable production deployment.

S

Safety — In AI context, the property of a system operating without causing unintended harm to users, stakeholders, or broader society. Safety encompasses technical, operational, and governance dimensions.

Sandboxing — Running AI agents in isolated environments to test behavior before production deployment. Sandboxing enables safe exploration of agent capabilities and identification of unexpected behaviors in controlled settings.

Silver Tier — BorealisMark Score 60-74. Represents moderate trust with documented room for improvement; agents at this tier are suitable for production deployment in lower-risk contexts with standard monitoring practices.

Software as a Medical Device (SaMD) — AI/ML software intended for medical purposes, subject to FDA regulation and validation requirements. SaMD faces heightened regulatory scrutiny due to direct patient safety implications.

T

Transparency — The principle that AI systems should be open about their capabilities, limitations, and decision-making processes. Transparency builds appropriate trust by enabling informed understanding of what systems can and cannot do reliably.

Trust Badge — Embeddable visual indicator showing an agent's current BorealisMark Score tier. Available as SVG, JavaScript widget, or HTML embed for integration into third-party platforms and procurement systems.

Trust Gate — A marketplace filter or requirement that only allows agents above a certain tier to be listed or purchased. Trust gates enforce minimum trustworthiness standards and reduce procurement friction by pre-filtering qualified agents.

Trustworthy AI — AI systems that are lawful, ethical, and robust, meeting technical and societal requirements for deployment. Trustworthy AI balances performance with safety, transparency, fairness, and accountability across the system lifecycle.

U

Unverified — BorealisMark Score below 40 or agents with insufficient data for reliable scoring. Agents in this category are not recommended for production deployment pending additional evaluation and certification.

V

Validation — The process of confirming that an AI system meets its specified requirements and intended use cases. Validation answers the question: "Did we build the right system for the intended purpose?"

Verification — The process of confirming that an AI system was built correctly according to its specifications and design. Verification answers the question: "Did we build the system correctly?"

Footer

This glossary is maintained by the Borealis Academy team and updated regularly. Terms and definitions reflect the current state of the field as of 2026. For corrections or additions, contact the Borealis Protocol team.