MickaiJun 15

Federated Learning Spreads the Training. It Does Not Spread the Trust.

Federated learning distributes where training happens, and people quietly assume it distributes trust along with it. It does not. You can decentralise the computation completely and still be unable to prove what the system actually did. The fix is not more decentralisation. It is one signed, offline-verifiable record that…

https://mickai.co.uk/articles/federated-learning-needs-a-sovereign-anchor

#federatedlearning #AIgovernance #sovereignAI #auditability #postquantum

Federated Learning Spreads the Training. It Does Not Spread the Trust.

Federated learning distributes where training happens, and people quietly assume it distributes trust along with it. It does not. You can decentralise the computation completely and still be unable to prove what the system actually did. The fix is not more decentralisation. It is one signed, offline-verifiable record that no participant can rewrite.

MickaiJun 15

The Signature Has To Outlive the Signer

Most artificial intelligence systems are designed to be trusted in the present tense. But a model can run for decades, and the record of its actions has to be verifiable long after the keys, the company, and the author are gone. This is an argument for signing today for a verifier who has not been born yet.

https://mickai.co.uk/articles/the-signature-has-to-outlive-the-signer

#keycustody #postquantumcryptography #AIgovernance #sovereignty #auditability

The Signature Has To Outlive the Signer

Most artificial intelligence systems are designed to be trusted in the present tense. But a model can run for decades, and the record of its actions has to be verifiable long after the keys, the company, and the author are gone. This is an argument for signing today for a verifier who has not been born yet.

MickaiJun 15

The PDF Did Not Stop the Breach

Model cards and audit binders are theatre when nobody can prove what a system did at the moment it mattered. I argue that a record signed before each action, hash-chained and verifiable offline, beats any document that merely describes intent. This is the thesis behind Mickai's Open Audit Record.

https://mickai.co.uk/articles/the-pdf-did-not-stop-the-breach

#compliance #aisecurity #auditability #openauditrecord #sovereignty

The PDF Did Not Stop the Breach

Model cards and audit binders are theatre when nobody can prove what a system did at the moment it mattered. I argue that a record signed before each action, hash-chained and verifiable offline, beats any document that merely describes intent. This is the thesis behind Mickai's Open Audit Record.

RoofMan Official BlogApr 17

CareFabric: A Blueprint for Healthcare Interoperability

The CareFabric model proposes a new approach to healthcare interoperability, emphasizing clinical coordination rather than simply integrating existing systems. It aims to address the fragmentation in healthcare information by creating a federated clinical exchange where government plays a role as a trust tracker rather than a central data holder. This model preserves organizational autonomy while enabling selective data exchange and robust governance. The architecture incorporates a national control plane and a federated data plane, ensuring secure real-time communication. CareFabric emphasizes the importance of patient identity, metadata management, and AI governance to enhance operational efficacy while keeping financial processes out of scope.

https://roofman.me/2026/04/18/carefabric-a-blueprint-for-healthcare-interoperability/

Jean-GuillaumeApr 17
KENTA n’est plus seulement une Society of Mind.
C’est une société capable de faire naître ses propres esprits.
#Auditability
#TrustByDesign
#AIWithoutLLM
#SovereignAI
Jean-GuillaumeApr 16
KENTA n’est plus seulement une Society of Mind.
C’est une société capable de faire naître ses propres esprits.
#Auditability
#TrustByDesign
#AIWithoutLLM
#SovereignAI
Kevin SaltMar 27

Well i finally did it. I just released my test dataset for AI Evaluation. Its a simulated company, represented by 60,000 documents, the readme in the image explains it all ... If you are interested, its at https://codeberg.org/Lorenz_Systems/Company_Sim.git

#EUAIAct #DigitalSovereignty #SovereignCloud #FOSS #FLOSS #Codeberg #Forgejo #OpenSource #DataGovernance #Auditability #ForensicAI #EUTech #PrivacyByDesign #InformationRetrieval #KnowledgeManagement #DeterministicAI #EUPL

USBManagerJan 8

🔧 Why You Need to Centralize USB Management Using USB over IP
Centralize USB device management with USBManager Server for enhanced visibility and control. Replace outdated "Dongle Room" practices with digital, real-time management, simplifying access and tracking. Enjoy full auditability with detailed logs and access tracking, ensuring security and compliance.
Learn more 👉 https://usbmanager.net/why-you-need-to-centralize-usb-management-using-usb-over-ip/

#USBoverIP #DeviceManagement #USBManagerServer #Auditability #Security

guIA - guía a la IASep 21, 2025
https://www.notebookcheck.net/Open-source-tool-measures-the-stupidity-level-of-AI-models.1117315.0.html #auditability #ai #foss #aibenchmarks
Open source tool measures the stupidity level of AI models

A new open-source tool is offering real-time monitoring of multiple AI models, including OpenAI GPT-5, Claude Opus 4, and Gemini 2.5 Pro. The first of its kind, it can detect "when AI companies reduce model capability to save costs." The benchmarks can run against the users' own OpenAI, xAI, Anthropic, or Google API keys as well.

Notebookcheck
Agustin V. StartariJul 7, 2025

🧠 New paper
The Grammar of Objectivity
Language models simulate neutrality not by removing bias, but by formalizing it.

🔍 Based on 1,500 LLM outputs (medical/legal, 2019–2024)
⚠️ 64 % of medical and 57 % of legal texts flagged

🔗 Read / download:
Zenodo: https://doi.org/10.5281/zenodo.15729518
SSRN: https://ssrn.com/abstract=5319520

Neutrality is no longer a meaning. It’s a structure.
#AI #LLM #Objectivity #Syntax #CriticalCode #Auditability #LLMTransparency #GrammarOfPower #agustinvStartari #social

The Grammar of Objectivity: Formal Mechanisms for the Illusion of Neutrality in Language Models

Abstract Simulated neutrality in generative models produces tangible harms (ranging from erroneous treatments in clinical reports to rulings with no legal basis) by projecting impartiality without evidence. This study explains how Large Language Models (LLMs) and logic-based systems achieve neutralidad simulada through form, not meaning: passive voice, abstract nouns and suppressed agents mask responsibility while asserting authority. A balanced corpus of 1 000 model outputs was analysed: 600 medical texts from PubMed (2019-2024) and 400 legal summaries from Westlaw (2020-2024). Standard syntactic parsing tools identified structures linked to authority simulation. Example: a 2022 oncology note states “Treatment is advised” with no cited trial; a 2021 immigration decision reads “It was determined” without precedent. Two audit metrics are introduced, agency score (share of clauses naming an agent) and reference score (proportion of authoritative claims with verifiable sources). Outputs scoring below 0.30 on either metric are labelled high-risk; 64 % of medical and 57 % of legal texts met this condition. The framework runs in <0.1 s per 500-token output on a standard CPU, enabling real-time deployment. Quantifying this lack of syntactic clarity offers a practical layer of oversight for safety-critical applications. This work is also published with DOI reference in Figshare https://doi.org/10.6084/m9.figshare.29390885 and SSRN (In Process )   Resumen La neutralidad simulada en los modelos generativos produce daños tangibles, desde tratamientos erróneos en informes clínicos hasta sentencias sin fundamento jurídico, al proyectar imparcialidad sin evidencia. Este estudio analiza cómo los modelos de lenguaje de gran tamaño (LLM) y los sistemas lógicos reproducen dicha neutralidad mediante la forma y no el contenido. Patrones como la voz pasiva, los sustantivos abstractos y la supresión del agente ocultan la responsabilidad y, al mismo tiempo, afirman autoridad. Se examinó un corpus equilibrado de 1 000 salidas de modelo: 600 textos médicos de PubMed (2019-2024) y 400 resúmenes legales de Westlaw (2020-2024). Se emplearon herramientas estándar de análisis sintáctico para detectar estructuras asociadas con la simulación de autoridad. Por ejemplo, una nota oncológica de 2022 afirma «Se aconseja el tratamiento» sin citar ensayos clínicos; en un resumen migratorio de 2021 se lee «Se determinó» sin referencia a precedentes jurídicos. El artículo introduce dos métricas de auditoría: la puntuación de agencia, que mide la proporción de cláusulas con agente explícito, y la puntuación de referencia, que calcula el porcentaje de afirmaciones autoritativas respaldadas por fuentes verificables. Las salidas con valores inferiores a 0,30 en cualquiera de estas métricas se clasifican como de alto riesgo; el 64 % de los textos médicos y el 57 % de los jurídicos cumplen este criterio. El marco se ejecuta en menos de 0,1 segundos por salida de 500 tokens en una CPU estándar, lo que demuestra su viabilidad en tiempo real. Cuantificar esta falta de claridad sintáctica aporta una capa práctica de supervisión para aplicaciones críticas.

Zenodo