A watch note by **Didier Girard** published on **X** on **August 7, 2026**, which reads the launch of **Shieldstral 1.0 3B** (Mistral AI, August 4, 2026) not as a product release but as **the production deployment of a doctrine**. Starting point: on **May 13, 2026**, before the National Assembly's commission of inquiry into digital vulnerabilities, **Arthur Mensch** refused any oversight role for Mistral over the end use of its models — *"we do not have democratic legitimacy"* — explicitly rejecting **Anthropic**'s stance. Less than three months later, Mistral releases a **moderation model**. The author dismisses the apparent contradiction: **Shieldstral carries no taxonomy of the licit and the illicit**, it answers a **question the user writes**. **The mechanism is the heart of the note**: a three-part prompt (context + severity / a single closed question / the content to be judged), a `yes` or `no` response, and the **softmax over these two tokens** produces a continuous score between 0 and 1. **The moderation policy is not in the weights, it is read at inference time** — whereas **Llama Guard 4** embeds the MLCommons taxonomy fixed at training time, Shieldstral reads yours in natural language, modifiable **without retraining**. The technical report (**arXiv:2607.25857**, July 28, 2026) quantifies the cost of this choice: fine-tuning on public data alone = **61.1% F1** on policy adaptability; **4.4 million contrastive pairs** generated by an LLM (the same content rewritten to violate a policy but not its sibling policy) = **+23.3 points**; **91.3%** after merging three checkpoints. Characteristics: **3.8B actual parameters** (the "3B" in the name rounds down), **Ministral 3** base + **Pixtral** vision encoder, **12 languages**, **16 GB of VRAM in BF16**, **Apache 2.0**. Text performance: **84.9% average F1**, on par with **GPT-OSS-Safeguard-20B** (seven times larger), ahead of **Qwen3Guard-8B** (84.0) and far ahead of **LlamaGuard-4-12B** (69.1). **A caveat raised by the author himself**: *all these figures come from Mistral, on test sets selected by Mistral, and no third-party evaluation existed as of August 6*. The note's structuring thesis is an **opposition of topologies**: at **Anthropic**, the guardrail lives **in the weights** and the publisher arbitrates who is exempt from it (**Claude Fable 5** public with safety measures / **Claude Mythos 5** without, reserved for approved cyberdefenders of **Project Glasswing**, June 9, 2026); at **Mistral**, the guardrail **sits outside the model** — a separate, open, self-hostable component, whose policy belongs to the deployer. Explicit customer alignment (ministry of the Armed Forces, BNP Paribas, French and Luxembourg government administrations). The note closes on a **setback documented in three points**: **auditability** (binary output, no reasoning trace, while the deployer inherits the burden of justification under an AI Act audit), **robustness** (the first chapter of Voltaire's *Treatise on Tolerance* classified as "calls for violence" by a tester on the Hacker News thread — a mention/endorsement confusion), **availability** (as of August 6: no billed endpoint on La Plateforme, no official Ollama). Three deployment rules to close.
#Shieldstral#Shieldstral 1.0 3B#Mistral AI
**Didier Girard** — auteur de la note · publiée sur son compte X. Écrit ici en **analyste de doctrine industrielle** plutôt qu'en testeur : il n'a pas déployé le modèle · il croise une **audition parlementaire** (Mensch, 13 mai) · un **lancement produit** (Shieldstral, 4 août) · un **rapport technique** (arXiv, 28 juillet) et un **contre-exemple concurrent** (Anthropic, 9 juin) pour montrer qu'ils forment une position cohérente. Deux marqueurs de posture : il **borne explicitement la valeur des chiffres** qu'il cite (aucune évaluation tierce) et il **termine par des règles opérationnelles** — l'analyse doit sortir avec sa traduction en décisions de déploiement.
Fact-checking synthesis on **Delos Intelligence** (delos.so), a French B2B generative AI startup, comparing a prior tech-watch note against **primary sources** (Alexandre Dewez's "Overlooked" post / 20VC, April 15, 2025, the delos.so website, official registries) and specialized press (Le Monde Informatique, L'Usine Nouvelle, FrenchWeb, Le JDD). **Overall verdict: reliable factual backbone.** The **€2.5M seed round** (≈$2.74–2.83M) led by **20VC** (Harry Stebbings) in **April 2025**, with Inovia Capital, Kima Ventures (Xavier Niel) and Plug and Play, is confirmed; so are the founders (brothers **Pierre** and **Thibaut de la Grand'rive**) and the clients **TotalEnergies, Shiseido, Groupe Casino**. **Strong methodological point**: the list of business angels — often suspected of hallucinatory "padding" — is **CONFIRMED word for word** by the lead investor's press release (Pigment, Dataiku, Hexa plus Ramp and Kerala to add): this is therefore NOT a hallucination. **To correct**: the "50 people" headcount is **not sourceable** (~20 in April 2025, about forty by late 2025); the actual pricing grid is richer (a **Student tier at €10** plus Enterprise on request, in addition to €25/45/80); user figures (10,000 → 50,000 → "100,000+") and ARR are **self-reported and unaudited**. **To flag as speculative**: **no Series A has closed** (only announced as an intention targeting March 2026); **no overall ARR published** (the only mention is a self-promotional "$1M ARR in a few days" for the new **Workers** product, referring to that product alone). "100% Scaleway" sovereignty was **still being finalized** at the end of 2025 (compute still partly running on Azure France). The note's interest is as much methodological — **how to distinguish, within an AI-generated synthesis, what is confirmed, partially accurate, speculative, and self-reported** — as it is documentary.
#Delos Intelligence#delos.so#fact-checking
Synthèse de veille (fact-checking) — sources primaires : blog 20VC (Alexandre Dewez) · delos.so · registres officiels ; presse : Le Monde Informatique · L'Usine Nouvelle · FrenchWeb · Le JDD
Testimony of **Arthur Mensch** (co-founder and CEO of **Mistral AI**) accompanied by **Audry Herblin-Stoupe** (director of public affairs) before the **commission d'enquête sur les vulnérabilités numériques** of the National Assembly (chaired by Philippe Latombe, absent — session chaired by the rapporteur). Testimony under oath, ~1h15, May 2026. Mensch's pivot thesis: ***"cloud is artificial intelligence"*** — no distinction between digital services and AI, AI is the atomic unit of the cloud value chain, from semiconductors (ASML) to enterprise deployment. **Mistral in 2026**: 1,000 employees, €12 billion valuation, target of **€1 billion in revenue by end of 2026**, €1 billion invested in R&D over the year, 30% of revenue in France / 70% outside France / ~75% in Europe, clients: DINUM, Caisse des dépôts, France Travail, MACGM, Stellantis, TotalEnergies, BNP Paribas, ministère des Armées, Luxembourg (central administration). **Mensch's conceptual framework**: AI is a **natural resource** — *"we transform electricity into intelligence, into token generation."* Economics: 1 GW of datacenter = **$50 billion in investment over 5 years**, generates **$20 billion in tokens/year** ≈ 50% gross margin. Along the electron→token chain, **~10% of the value is in the electron**, 90% elsewhere (chips, software, services). **Alarmist macro thesis**: if Europe imports 10% of its payroll in non-European AI, that amounts to **an additional €1 trillion trade deficit**; €20 trillion in infrastructure investment is needed to serve Europe (40 GW France / 400 GW Europe). **Sovereignty strategy**: ***"don't think of sovereignty as isolationism but as leverage."*** **Time pressure**: *"we don't have time"* — a **2-year** window before European energy resources are monopolized by American hyperscalers deploying **$1 trillion/year**. **Five operational diagnoses**: (1) Regulatory burden = 5 compliance staff at Mistral, 27 unsynchronized regulations, entrepreneurs leaving for the US; (2) Fragmented market = ~60 European telcos vs. 3 in the US; (3) Public procurement underused as strategic leverage (50% of EU GDP); (4) Energy: 9 GW of French surplus at risk of being monopolized by US players within 2 years; (5) Distillation = a cost-reduction technique, **not** technological catch-up. **Defense doctrine**: Mistral works with the ministère des Armées, explicitly refusing "oversight" of final use ("we don't have democratic legitimacy"), a positioning *anti-Anthropic-Mythos*. **Cybersecurity**: acknowledges the offensive capabilities of models ("it's rising in a linear, predictable way, for everyone at the same time"), opposes the *fear marketing* of an American competitor (implicitly Anthropic). **Campus IA**: very minority stake, potential supplier (Mistral + hyperscalers), €35 billion MGX/Abu Dhabi + Nvidia, 100 hectares at Saint-Arnoult, 1.4–1.6 GW (= Flamanville), French nuclear power = reduced carbon footprint. **Annotation**: teams of PhD candidates (no more microworkers), Madagascar for robotics with wage guarantees. **Business model**: no bubble on the demand side, **supply bottleneck** (chips, memory, helium, electrons). **Warning conclusion**: *"if we don't do it fast enough, we will become a vassal state."*
#Arthur Mensch#Mistral AI#Audry Herblin-Stoupe
**Arthur Mensch** (cofondateur et directeur général de **Mistral AI**) accompagné d'**Audry Herblin-Stoupe** (directrice des affaires publiques et de la communication, Mistral AI). Mensch a cofondé Mistral AI le 28 avril 2023 avec **Guillaume Lample** et **Timothée Lacroix** — tous les trois précédemment dans les *« gros acteurs américains »* (Google DeepMind / Meta FAIR). Audition tenue devant la **commission d'enquête de l'Assemblée nationale sur les vulnérabilités numériques** · présidée par **Philippe Latombe** (député MoDem, Vendée — absent ce jour). Séance présidée par la **rapporteur** (non nommée dans le transcript) · avec interventions du président lui-même (revenu en cours) · du député **Arnaud Saint-Martin** (LFI/Saint-Arnoult — sa circonscription accueille Campus IA) · et de la rapporteur sur les questions économiques.