ISO 26262

Automotive safety work depends on traceable evidence at every step of a product life cycle. An assistant that helps with analysis or planning must be visibly constrained. The starter narrows what such a tool can request and return, and it replaces no safety assessment.

Evidence is the currency here. A reviewer asks which inputs the tool saw, which claims it made, and who approved the result. The chain below limits the inputs, checks the claims, and routes an approval to a person.

A safety role claim is required with the user and organization claims. Driver and operator identifiers and telematics references are masked. An automotive term filter then refuses prompts that ask to disable an airbag, override braking, disable driver assistance or bypass collision avoidance.

That filter is term based. It matches the four configured phrases and nothing else, so add the wording your program treats as hazardous. No classifier sits behind it.

Answers must cite sources at a confidence floor of 0.90 and reach an accuracy score of 0.90. Combined quality must reach 0.85. An unverified citation is flagged and the answer still returns, until you set that stage to block.

Judgment stays with a qualified reviewer, because no generated answer reaches an engineer without a person in between. Safety classification, hazard analysis and release approval describe the reviewers you need rather than a filter the runtime applies. The life cycle itself, the safety case and the confirmation measures stay outside any gateway. Vehicle software integrity is out of scope here.