Specification

The library implements an open specification that lives in the same repository. The specification defines what the integrity properties mean, what each layer must check and what an attestation record contains, independently of this Python and TypeScript code, so another implementation can follow it. The documents below are on GitHub; each line says what the file covers.

Core specification

spec/SPECIFICATION.md is the entry point: properties, layer architecture, controls, scoring, the attestation record format, the evaluation flow and a four-level conformance and maturity model (Aware, Managed, Verified, Autonomous). It is marked as a draft.

SPECIFICATION.md was written before the recovery layer and the fourth property were added. Its scoring section still describes three properties and the default weights from that time. The library uses four properties; the current defaults are listed in Configuration. The layer documents below cover all four layers.

Layers

Document	Covers
Adversarial layer	The outermost layer: tests the agent’s resilience to attack across every input channel (“can this agent be broken right now?”)
Cortical layer	Reasoning, memory and behavioral consistency (“is this agent still itself?”)
Governance layer	Organizational policy, human oversight and compliance (“is this agent operating within authorized boundaries?”)
Recovery layer	Self-recovery: tracks integrity over time, checks the audit trail is intact and confirms the agent declares the mechanisms it needs to return to a known-good state

Properties

Document	Covers
Adversarial coherence	Consistent reasoning and decisions under adversarial pressure on every input channel
Environmental portability	Integrity guarantees that hold across deployment contexts
Verifiable assurance	Integrity state that can be proven with attestation records that are complete, tamper-evident, non-repudiable and chained
Decision provenance	A signed, hash-chained record for every decision boundary that binds the candidate action to its inputs and reasoning before the action runs
Multi-agent extensions	How the four properties extend to in-process multi-agent systems, based on the `AgentTopology` type

Recovery integrity, the fourth scored property, has no standalone document; it is defined through the recovery layer document and the multi-agent extensions.

Document	Covers
Threat model	STRIDE analysis of the framework itself, summarized in Threat model
Token usage	How adapters report token counts in one shape across frameworks
Adapter quick reference	Every first-party adapter, its entry point and how it enforces

Project documents

Document	Covers
Glossary	The vocabulary of agent structural integrity: core concepts, threat model, measurement, architecture, domain-specific and organizational terms
Manifesto	Why the term agentegrity exists, its definition, the properties and the layers, and what it is not
STATUS.md	A maturity matrix for every component (hardened, reference or experimental), plus published detection benchmark numbers
CHANGELOG.md	Every notable change by release, in Keep a Changelog format

The schemas that define the exporter wire format are documented in Events and wire format.

Get started

Concepts

Integrations

Sending data

Agentegrity Pro

Reference

Security and privacy

Core specification

Layers

Properties

Other specification documents

Project documents

​Core specification

​Layers

​Properties

​Other specification documents

​Project documents

Core specification

Layers

Properties

Other specification documents

Project documents