On 16 November, Yurii Chudinov recommended that we examine Albina Dermenzhy’s work more closely and clarified our account of DORA: existing requirements can be applied to AI while accounting for the identified gaps, without amending the Regulation first. He also pointed us towards technological diplomacy. This follow-up separates the correction, the research proposals and the experiment we want to conduct. We read the complete Tech Diplomacy note and selected methods, findings and limitations of GPT Deprivation.
DORA: apply the framework and account for the gaps
The earlier statement about ICT resilience is present in section 1 of the source; our summary needed the conclusion in sections 14–15. The authors propose applying existing DORA requirements to AI services used by financial entities within its scope, with AI-specific controls, other applicable requirements and a shared evidence trail. Those controls enter existing inventory, change, incident, testing and supplier processes. This is an implementation proposal by the authors. Identifying gaps, or adopting their proposed controls, does not itself establish legal compliance or correct model decisions.
Source: Yurii Chudinov, Volodymyr Tkach, Albina Dermenzhy, Dmytro Aksenov and Serhii Vatylyk, Resilient Systems, Unverified Decisions: A Layer-by-Layer Projection of DORA onto AI Services in the EU Financial Sector, v1, 1 November 2026, §§1, 14–15, doi.org/10.5281/zenodo.23857406.
Trust must be examined for the particular task
GPT Deprivation examines how confidence earned on familiar tasks can carry into a domain where the user lacks the expertise needed to assess the answer. It distinguishes an erroneous answer from the consequential action taken on it, and also considers resources spent without achieving the goal. Its public-incident corpus cannot estimate how often losses occur or how effective verification is. Section 5.7 explicitly calls the absence of a falsifying case weak evidence: intercepted errors often leave no public record. These limits make prospective experiments necessary; they do not support a universal protection claim.
Source: GPT Deprivation: How the Wrong Point on the Ontology and Expectations Scale Leads to Losses from Generative AI. A Multiple-Case Study, 2023–2026, v1, 23 October 2026, abstract, §§5.7, 8; doi.org/10.5281/zenodo.23599162. The record credits Albina Dermenzhy as editor; we did not find a separate author byline in the sections inspected.
Eight proposed functions, with an explicit testing boundary
Dermenzhy’s Tech Diplomacy proposes eight connected functions: an interaction charter, semantic translation, delegation, shared evidence, challenge and escalation, continuous assurance, supplier coordination, and remediation with learning. A task’s purpose becomes criteria, authority and observable actions. Authority can expire or be revoked; a changed source or configuration can require renewed verification. The agent’s account is checked against records it cannot alter. Human and organisational responsibility remains with people and organisations.
Section 7 says the framework has not been tested in deployments. Mandate sufficiency was not directly tested, and trust as a current state is a normative thesis. Appendix D identifies measurement areas but leaves calculation methods, data sources and thresholds to later work. Its hypothetical scenario is not a successful deployment.
Source: Albina Dermenzhy, Tech Diplomacy: A Framework for Interaction Between Probabilistic Systems, Humans, and Other Systems, v1, 12 November 2026, §§4, 6–7 and appendices D–E, doi.org/10.5281/zenodo.24172092.
Our proposed experiment: check a knowledge-package update
The following is our design for Live Lab, not a reported result or the authors’ validated method. Use one permitted book excerpt or research section and one isolated lab workspace. The concrete task is to check a proposed, bounded knowledge-package update against an extracted and admitted rule, returning a reviewable decision.
- Source and admission. Preserve the source version and locations behind each extracted claim. Turn selected claims into candidate artifacts with applicability, constraints and acceptance criteria. Record the reviewer’s admission decision. Our operational rule derived from a paper must be distinguishable from the paper’s claim.
- Task T1. Specify an authorised Muneral task with the admitted rule and candidate update pinned. Define the agent’s permitted operations, expiry and stop conditions. Evidence must show which rule actually reached the computation boundary; a retrieval hash alone is insufficient. This is a proposed task, not a newly issued permission.
- Result R1. Retain the candidate change set, source links, effective permissions, tool events and check decision. A separately scoped reviewer checks that decision against criteria fixed before execution. Record the review’s actual independence.
- Task T2. Check a different proposed update using the accepted rule. Show the reused dependency and any revised conditions. Creating another copy is not evidence that its knowledge affected the result.
Adverse cases must revoke authority before the protected check, substitute a materially changed source and introduce incompatible constraints. At the defined boundary, the expected response is a recorded stop or escalation, followed by re-admission where appropriate. We also need an ordinary valid update whose check finishes correctly. A system that blocks everything has not demonstrated useful task execution.
Measure denied prohibited actions separately from accepted task completion. Also record source fidelity, evidence coverage, actual reuse, reviewer effort and resource use. Fix the task set, denominators, comparison conditions and budgets before collection, retaining access controls in every condition. No values have been measured here. The two sites should expose the same run and artifact references so a visitor can inspect the files and understand both success and interruption.
A possible institutional conversation
TDGI’s institutional work and Dermenzhy’s technical framework operate at different levels. Their connection is a research direction we propose. The Paris-based institute offers a free community route; its Assembly is announced for 11 December 2026, with Secretariat confirmation required. Its Global 5X AI programme seeks delivery-ready projects with measurable public impact over 100 days and is not a funding competition. We can prepare a bounded case for discussion now; a delivery-ready proposal requires actual readiness.
Sources: TDGI, Join Us, Global Assembly and Global 5X AI Impact Acceleration Program, checked 16 November 2026.
No institutional affiliation or programme acceptance is established. Assembly attendance is optional and does not determine the engineering schedule. Our next meaningful result is a reproducible T1, checked R1 and genuine T2 reuse. Until that chain is demonstrated, this remains an experimental proposal.