Pipeline map · tap each stage
Where a prompt goes, and where a human actually is
Nine stages, color-coded by what kind of oversight exists at that point — live, trained-in-months-ago, policy-set-in-advance, or after-the-fact only. The dashed rose bracket marks the stretch where nothing live is watching, which is also where most relational harm actually gets through.
no human, live
human feedback, trained in earlier
human policy, set in advance
human review, after the fact
no live oversight window
What's general vs. specific: this maps the industry-standard pattern (input/output classifiers, RLHF, system prompts, async review) as it's publicly documented across AI companies, not a confirmed account of any one company's internal pipeline. I don't have insider access to exactly which review steps exist on any given product, including the one you're talking to me through — treat this as the architecture shape, and verify specifics before citing them as fact about a named product.