The argument, the specification, and how to test both.

Version

v0.5.4

Status

Open Review

Issued

26 August 2026

Extent

37 pages

File

PDF · 991 KB

AI no longer only produces text — it reads mailboxes, joins records across systems, prepares transactions and triggers actions, which makes a model's output a business event. Most organisations now control which model may be used. Far fewer can demonstrate, for one specific action, who authorised it, on whose behalf, under which mandate and policy version, and whether the thing approved is the thing that executed. This paper argues that access is not authority, and sets out what an executable control layer would have to do.

• Shadow Execution: how the risk outgrew the Shadow AI framing
• Impact tiers and provenance assurance — the two control variables the other documents reuse
• What this architecture does not solve, stated before the case for it
• Where the EU AI Act, NIS2 and the BSI Act, DORA, the GDPR and the Cyber Resilience Act converge on the same technical objectives

Version

0.4

Status

Candidate — Open Review

Issued

26 August 2026

Extent

67 pages

File

PDF · 3.0 MB

The normative counterpart to the position paper: a wire profile for a governed action, written so that two systems that have never met can exchange one and independently verify it. It defines the Governed Action Envelope, the control invariants an implementation must hold, the processing procedure and state machine, and the tests an implementation has to fail correctly. It deliberately defines no conformance scheme and no conformance levels.

• The Governed Action Envelope: principal, intent, parameters, mandate, preconditions, approval record
• Digest definitions, domain separation and the signer matrix
• Canonicalisation and cryptographic profiles for JSON and CBOR
• Target-adapter assurance modes, thirteen test vectors and the required negative tests

Version

0.1.4

Status

Open Review

Issued

26 August 2026

Extent

22 pages

File

PDF · 584 KB

The practical layer. It walks one worked example end to end — AI-assisted invoice processing, including what still fails in it — then gives an adoption path that starts by making AI actions visible rather than by migrating anything. The measurement protocol is written so that an organisation can run it, and publish the result, without using Vianordis, including a result that goes against us. The guide is non-normative and contains no field measurements.

• A worked example, with the failure cases left in
• A six-phase adoption path and the exit criteria for scaling it
• Brownfield routing: enforcement modes M0–M4 and the integration patterns
• A maturity model for leadership, and a pilot scorecard with a publication template