Measured notes from the workshop and the rack.
Put the agent inside the project's existing authority structure: bound its task and inputs, classify the workflow, retain provenance, keep acceptance criteria and safety decisions outside it, and route every influenced safety work product through named verification.
functional-safety iso-26262 agentic-ai engineering-process verification
AI can generate verification artefacts and useful findings, but a second model does not establish independence. The argument depends on independent sources, oracles, responsibility, retained evidence and acceptance authority around the AI activity.
functional-safety iso-26262 verification agentic-ai tool-qualification
Treat an AI engineering system as a software tool and classify its intended use and complete workflow. Tool Impact becomes TI2 when its output can introduce or fail to detect an error in a safety work product; the independent verification chain then determines Tool error Detection.
Measured LLM serving on a consumer RTX 5060 Ti 16 GB in a Dell PowerEdge R740: engine comparison, prefill against decode, the concurrency knee, what a full attention cache costs, and a hardware thermal slowdown that default chassis cooling never reported.