Order the layers you'd investigate when debugging a RAG faithfulness regression that just shipped to production, from most likely to least likely cause.

A **faithfulness regression** means the RAG system's answers stopped staying grounded in the retrieved context, it started adding claims the chunks don't support. Something changed, and your job is to find it fast. The trick is to order your hypotheses by how often each layer actually changes, not by which feels most interesting. The prompt template is touched constantly and ships to 100 percent of traffic the moment a pull request merges, so a bad prompt edit is the single most likely cause. Diff it against the last known good version first. Next check the model version in config. An endpoint or model name can change silently during a routine deploy, and a new model behaves differently. After that, run an isolated retrieval check: hold the model and prompt fixed, replay your eval set, and see whether context recall actually dropped. Only when those three come back clean do you investigate query drift, users slowly asking different questions, because it's the slowest to confirm and the rarest to cause a sudden regression.

The correct order is prompt diff, then model version, then isolated retrieval check, then query drift. The principle is **base rate triage**: investigate hypotheses in descending order of how frequently each layer changes, weighted by how fast you can confirm or rule it out. Prompt templates are first because they ship at 100 percent on merge with no canary in most stacks. A one line edit to chunk ordering or a dropped 'answer only from context' directive tanks faithfulness instantly. That single change affects every request. Diffing against the last known good version is also the cheapest check. Model version is second: it lives in config, can change during an unrelated deploy, and a silent swap to a newer or cheaper model shifts grounding behavior. It's tracked, so confirmation is quick. Retrieval is third because verifying it correctly requires a reproducible experiment, fix the model and prompt, replay a fixed eval set, and measure context recall. That's more work, so you only spend it after the cheap checks fail. Query drift is last. It's real but slow moving; it rarely causes a same day regression, and confirming it means comparing query distributions over time, the slowest hypothesis to verify.