Skip to content
Tyche Institute

Research programme

Tyche Attestable Agency Programme

Tyche studies verifiable agent-action evidence: how a consequential AI-agent action can leave a record that another party can inspect, recompute, contest, and retain.

The programme is broader than logging or transparency anchoring. It follows the full evidence path from authority and execution to portable records, carrier binding, independent appraisal, regulatory interpretation, and longitudinal observation.

Canonical relationship

EATF (the Agent Trust Framework) appraises Action Evidence Packages (AEPs).

Earlier Tyche artifacts may expand AEP as “Agent Evidence Package”. Historical titles and identifiers remain unchanged; “Action Evidence Package” is the canonical forward term.

The evidence stack

The anchor is the floor, not the house. A transparency receipt can support integrity, existence, inclusion, and time. It does not by itself prove authority, execution, semantic outcome, or legal compliance.

Layer 01

Mandate and authorisation

MachineMandate and EUDI-oriented delegation profiles bind actor, scope, action, and enforcement outcomes.

Zenodo 10.5281/zenodo.21233028 →

Layer 02

Execution evidence

Action evidence composes with platform Evidence under IETF RATS; execution support is kept distinct from semantic truth.

IETF AEP × RATS draft →

Layer 03

Portable action records

Action Evidence Packages carry typed claims, bindings, evidence references, limitations, and explicit claim ceilings.

Verifier artifact 10.5281/zenodo.21625781 →

Layer 04

Record binding and anchoring

Carrier bindings and AAC/SCITT-compatible services add integrity, inclusion, existence, and time evidence without inheriting stronger claims.

Joint SCITT payload-binding draft →

Layer 05

Independent appraisal

EATF profiles, independent verifier implementations, negative vectors, and disagreement analysis make policy-relative failure reproducible.

Corpus 10.5281/zenodo.21563876 →

Layer 06

Law and institutional observation

Clause-level crosswalks and longitudinal observatories preserve how rules, public evidence, trust infrastructure, and practice change over time.

Research corpus and observatories →

Research extensions

Current extensions apply the same claim discipline to machine unlearning, uncertainty-aware SAFE appraisal, human oversight, educational AI, X-Road carrier binding, and EUDI agent delegation. A domain profile may add claims and evidence types; it may not turn a valid container into proof that every statement inside it is true.

Open by design

Tyche publishes royalty-free specifications, software, datasets, technical notes, test vectors, and reproducibility bundles. The goal is a verification commons, not a central marketplace or trust service.

EATF framework → · Research corpus →