EATF — Agent Trust Framework
EATF is the appraisal framework within Tyche Institute's Attestable Agency Programme. EATF provides open profiles, independent verifiers, policy inputs, conformance vectors, negative controls, and disagreement analysis for Action Evidence Packages (AEPs).
Documentation, specifications, reference implementations, and public demonstrations are maintained at the project site.
A public demonstration of the framework, focused on Article 14 (human oversight) and adjacent EU AI Act obligations, runs at ai-act.eatf.eu. It is a public learning artifact, not a production service.
EATF describes the AI-agent layer; it composes with, rather than competes with, hardware attestation. Tyche's IETF Internet-Draft (draft-sokolov-rats-aep-composition) shows how an Action Evidence Package binds to platform Evidence under RATS (RFC 9334), conveyed via EAT, and appraised by an open-source Verifier such as Veraison.
Earlier Tyche papers refer to this artifact as the “Agent Evidence Package”; the canonical term is now the “Action Evidence Package” (the same AEP), matching the IETF draft.
Evidence stack
The anchor is the floor, not the house
Transparency anchoring is essential, but it supports a narrower claim than the whole evidence path. Each layer below keeps its own claim ceiling; assurance never leaks upward merely because a lower layer passed.
- Independent appraisal EATF policies, independent verifiers, negative vectors, and disagreement.
- Action evidence AEP claims, evidence references, bindings, limitations, and verdict ceilings.
- Authority and execution Mandates and EUDI delegation; RATS platform Evidence and Attestation Results.
- Record binding and anchoring AAC/SCITT-compatible integrity, existence, inclusion, and time evidence.
A valid anchor does not by itself prove authority, execution, semantic outcome, or legal compliance. A valid AEP is likewise not a truth certificate.
EATF is included in the community-curated awesome-eu-ai-act resource list under AI Agent Governance.