Reference
Glossary
Canonical definitions for terminology used in EPICORTEK's trusted synthetic intelligence infrastructure. Each term links to the relevant product page, protocol specification, or evidence register entry.
All terms
Synthetic intelligence
Software systems that exhibit goal-directed, adaptive behaviour — including AI agents, LLM-powered workflows, autonomous decision-making pipelines, and non-deterministic software. EPICORTEK uses this term to describe the class of systems its infrastructure is designed to govern.
How EPICORTEK uses this term
EPICORTEK's primary audience: any system that can act, decide, delegate, and evolve independently of real-time human instruction.
Non-deterministic systems
Systems whose outputs cannot be predicted purely from their inputs and code. LLM-powered agents are non-deterministic because model outputs vary by sampling strategy, context window, and temperature. Governance and audit for these systems requires evidence-based verification rather than deterministic code inspection.
How EPICORTEK uses this term
Used to describe the class of AI systems that require identity, policy, and observability infrastructure because their behaviour cannot be fully certified by static analysis.
Not to be confused with
Non-determinism in the cryptographic/algorithm sense. Here it refers to runtime output variability in ML-based systems.
Agent identity
A verifiable claim that a specific AI agent exists, holds a particular key, operates under defined capabilities, and has an auditable status. In AgIS, agent identity is expressed through an Agent Card, bound to a DNS record, and verified cryptographically.
Agent Card
A structured JSON document published at a stable URL by an agent operator. It contains the agent's identifier, public key (JWK), capabilities, delegation scope, and lifecycle status URL. AgIS evaluates the Agent Card as the primary identity evidence document.
Not to be confused with
DID Documents, Verifiable Credentials, or W3C identity schemes. Agent Cards are AgIS-specific.
Agent identity evidence
The set of verifiable artefacts that AgIS evaluates to produce a trust level decision: the Agent Card, DNS TXT binding, JWK thumbprint, Ed25519 signature, request signature (HTTP Message Signatures), Content-Digest, freshness window, nonce, and lifecycle status. Higher trust levels require more evidence types to be valid simultaneously.
DNS-backed verification
A verification method where the binding between an agent's identifier and its cryptographic key is published in a DNS TXT record. Verifiers query DNS to confirm the binding, then verify the card cryptographically. Requires no central registry. Works offline once the DNS record is cached.
Not to be confused with
DNSSEC. DNS provides discovery; cryptographic binding is via Ed25519 signatures, not DNS signatures.
Trust level
A numerical value (0–5) produced by AgIS reflecting the quality and completeness of identity evidence verified for an agent request. Trust levels are advisory; they inform but do not replace local authorization policy. Level 0 = no evidence. Level 5 = freshness, replay protection, and local policy all enforced.
How EPICORTEK uses this term
AgIS trust levels are the primary output of the AgIS evaluation engine. The offline reference implementation test vectors reach Level 4.
Delegation chain
A sequence of cryptographically signed tokens where each step grants a strictly narrower scope of capabilities from one agent to another. AgIS verifies that no delegate in a chain exceeds the authority of its principal, and that every token is valid and properly scoped.
Context governance
The discipline of specifying, enforcing, and auditing the conditions under which an AI agent may access, recall, or act upon information. In EPICORTEK's stack, context governance is addressed by ABDS/AUSF (intent and behavioural specification) and scarR (governed persistent memory with policy-bound recall).
Policy-bound recall
A memory retrieval model where an AI agent's access to past context, decisions, or records is constrained by active policy. Recall, retention, and retrieval are all governed — the agent cannot access memory outside its defined policy boundaries.
How EPICORTEK uses this term
The defining capability of scarR — EPICORTEK's Governed Agent Memory Layer.
Evidence-linked reasoning
An operational model where every agent decision or claim is linked to an auditable evidence artefact — a test vector, a verified identity record, a signed Agent Card, or a documented status. EPICORTEK's evidence-first approach requires that all maturity claims be supported by explicit, reviewable evidence.
Agent observability
The capability to collect, correlate, and analyse signals emitted by an AI agent during operation — including decisions made, actions taken, anomalies detected, and policy boundaries reached. Observability enables accountability and auditability for autonomous systems.
How EPICORTEK uses this term
Addressed by Sentinel — EPICORTEK's behavioral observability system for AI agents.
Autonomous workflow
A software workflow that executes without real-time human instruction — making decisions, calling APIs, processing data, and potentially delegating tasks to other agents or services. Autonomous workflows require the same identity, policy, and audit infrastructure as individual AI agents.