I measure the distance between what an organisation declares about one AI workflow and what its own material establishes. Independent, Brussels.
Founder of Sprinkling Act. The method is Human in the Map: in the loop, on the loop and in command each describe one control point at a time, and none of them answers the question a director actually asks, which is where the humans are across everything now running and what the whole set costs. That is a question about the map. I delimit, I don't recommend.
- Role debt. The distance between what a job description says and what a person does once a model sits in their workflow. Not role drift, which identity management already uses for the creeping accumulation of access rights. Like any debt it has a principal, a rate and a debtor, which is why it can be counted.
- The evidence layer. Four statuses on every line, and a gap established only after a second source of a different kind. The hard part is not the finding, it is deciding in advance what a finding is allowed to rest on.
- The deployer-side cascade of the AI Act (Art. 4, 26, 27, 50): the obligations that land on the organisation using an AI system, not only on the one that built it. It belongs downstream of the reconstruction, never upstream of it, because a regulatory reading applied to a workflow nobody has reconstructed reads the system as designed rather than as run.
- Agent reliability as AI Act evidence: mapping how an agent is tested to what Articles 9, 10 and 15 require. In progress, and unreleased until someone outside this project has read it.
Until July 2026 I sold a regulatory position: where an AI system falls under the AI Act, and which obligations follow. I stopped for four reasons, all of them checkable. My own April screening of 50 European companies found 96% with no published AI Act position, which reads as latent demand only if you assume those companies are looking for one. The Digital Omnibus provisional agreement of 7 May 2026 moved standalone Annex III obligations to 2 December 2027, and an offer whose argument is a date does not survive the date receding sixteen months. Classifying a system article by article assumes you know how the work runs, which is the part nobody knows. And the 0-100 band I published aggregated exactly what the analysis exists to pull apart.
The gate framework was moved downstream rather than dropped, and the papers from April and May are not being edited to match what I think now. They are timestamped, so editing them would destroy the only thing a timestamp guarantees. Full account: methodology/TRANSITION.md.
Open access, CC BY 4.0, DOI-indexed, cross-filed on SSRN, cryptographically timestamped.
- EU AI Act Readiness: A Structured Screening of 50 European AI Companies (April 2026). DOI 10.5281/zenodo.19671328.
- Annex A: The Deployer Multiplier (May 2026). DOI 10.5281/zenodo.20042174.
- The AI Act as a Third Structural Pole (May 2026). DOI 10.5281/zenodo.20343243.
- The Deployer Multiplier (Eliva Press, June 2026, 82 pages, ISBN 978-99993-4-609-2). A print consolidation of the three above. The Zenodo versions are the ones to cite.
sprinkling-act/methodology· the method, documented openly, with the AI Act gate framework it sits above and the record of why it moved there.emeth-agent· signed, timestamped boundary attestations for the EU AI Act. Rule-based, no LLM.lean-lab· a self-taught Lean 4 formalisation log. Public because the failures are the point.trajectory-probe· how much hallucination can be removed from a language model without supplying any outside knowledge, and what each exit costs. Thirteen stages, every protocol declared before it runs, every control logged whether it passes or fails. It is the empirical half of the same questionlean-labproves: a system does not decide its own truth from the inside.
Every conclusion states what it rests on, every figure carries its reserve every time it is reused, and a figure I cannot support is withdrawn rather than defended.