LIVE
Mistral Makes Large 4 Available as Open Weights29/10/26 · Mistral AI|Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v529/10/26 · Google|Mistral Large 4 now available on Hugging Face29/10/26 · Mistral AI|Google Introduces "Gemini 4 Carbon" as Next Step in Its Model Lineup28/10/26 · Google|DeepMind unveils Genie 4, a new generation of world model27/10/26 · Google DeepMind|Anthropic files IPO prospectus, with risk disclosures taking center stage26/10/26 · Anthropic|Robot policy trained only in simulation matches top real-world systems26/10/26|OpenAI unveils GPT-6.1 Terra, a model built for computer use26/10/26 · OpenAI|OpenAI revises its Model Spec with five guiding values for AI behavior29/10/26 · OpenAI|OpenAI outlines its monitoring system for catching misaligned model behavior29/10/26 · OpenAI|Hugging Face releases open models for speech recognition and decision-making on Apple devices29/10/26 · Hugging Face|Hugging Face launches Project ProtoStar to make publicly funded science AI-ready29/10/26 · Hugging Face|Mistral Makes Large 4 Available as Open Weights29/10/26 · Mistral AI|Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v529/10/26 · Google|Mistral Large 4 now available on Hugging Face29/10/26 · Mistral AI|Google Introduces "Gemini 4 Carbon" as Next Step in Its Model Lineup28/10/26 · Google|DeepMind unveils Genie 4, a new generation of world model27/10/26 · Google DeepMind|Anthropic files IPO prospectus, with risk disclosures taking center stage26/10/26 · Anthropic|Robot policy trained only in simulation matches top real-world systems26/10/26|OpenAI unveils GPT-6.1 Terra, a model built for computer use26/10/26 · OpenAI|OpenAI revises its Model Spec with five guiding values for AI behavior29/10/26 · OpenAI|OpenAI outlines its monitoring system for catching misaligned model behavior29/10/26 · OpenAI|Hugging Face releases open models for speech recognition and decision-making on Apple devices29/10/26 · Hugging Face|Hugging Face launches Project ProtoStar to make publicly funded science AI-ready29/10/26 · Hugging Face|
+ nAIvigateA SHEET FOR PEOPLE WITHOUT 2 HOURS
+ nAIvigate · fiches / stack-ia-six-briques-garde-fous-evaluations
▣ fiche-009.md ×
□ index.md
☰ ContentsVerdictOne question per blockSigns of a demo stack

AI stack: six building blocks, only one says no.

An enterprise AI stack always assembles the same blocks. The gap between a demo and a reliable system comes down to two of them: guardrails, which block an unauthorised action, and evaluations, which measure before every release.

Verdict

the essentials in 10 seconds3 rows
BLOCKSEvery platform lines up the same blocks. The difference lies in how they are governed.EVERYWHERE
STOPOnly one block can stop a bad action: the guardrails. Ask where they sit.DECISIVE
MEASUREWithout evaluation on your real cases, a successful demo proves nothing. Measure before every release.OVERLOOKED

The chain, from document to action

where each block steps in6 blocks

The agent plans, the model reasons, data comes in through RAG. Before any real action, the request goes through the guardrails, then through MCP to your applications. Evaluations measure the whole chain at every version.

DATA DocumentsPDFs, wiki, tickets Vector databaseindexes by meaning RAGfinds the passage REASONING Requestuser Agentplans, chains steps Modelgenerates, reasons ACTION Guardrailsallow or block MCPconnects the apps Tools and dataCRM, email, IT Real actionsend, write blocked here CONTROL Evaluations: quality, cost and latency measured on your real cases no new version goes to production without passing → Ask your vendor where the blocking block and the measuring block sit

One question per block

to ask your vendor or your team7 rows
model ↗Generates and reasons. The most replaceable block: choose it on your cases, not on a leaderboard.FINDER
data ↗Documents split, indexed in a vector database, retrieved by meaning. Read permissions must carry through to the index.QDRANT
agent ↗Plans and chains the steps. Every tool it can call is a permission to justify.GUIDE
MCP ↗The protocol that plugs the agent into your applications. An unvetted server opens a door into your systems.MCP SERVERS
guardrails ↗Filter inputs and outputs, block unauthorised actions. The only block that can say no.DECISIVE
evaluationsMeasure quality, cost and latency at every version. Without them, nobody knows whether an update broke something.OVERLOOKED
vendor ↗Where the data goes, who reuses it, which contract. Check before the first line of code.TRUST INDEX

Signs of a demo stack

one is enough to dig deeper7 rows
demoThe demo runs on public data, never on yours.COMMON
permissions ↗The agent acts with the permissions of the user who launched it, or even more.BLOCKER
guardrailsThe guardrails amount to an instruction in the prompt.BLOCKER
MCPMCP servers installed from a public repository, with no code review.
measureModels or versions are swapped without measuring anything first.
logNo way to trace who asked for what, and what the agent actually did.
check ↗Your Claude Code configuration checked in the browser, without uploading the file.TOOL
Does your AI stack have a block that says no? We help you map your blocks, place the guardrails and set up evaluations.
TALK TO nAIvigate STUDIO →
⎇ mainfiche #009 · securitesources: Synthèse nAIvigate, octobre 2026 · Briques reliées aux fiches, outils et guides du site
Follow nAIvigate | naivigate.com/en/fiches