Context Fidelity Preflight — Prompt & RAG QA

Illustrative planning receipt; no model call or behavioral evaluation occurred.

Baseline version: supplied current context, to be hashed when available. Candidate version: proposed shortened context, also to be hashed. Fixture set: six approved FAQ questions, each with a stable ID, exact input, policy-corpus snapshot, expected facts or allowed uncertainty, citation rule, JSON fields, and severity. Answerable cases need exact passage anchors. For the corpus-silent case, record a negative retrieval or coverage check, expected uncertainty, and no supporting citation, so unsupported invention is visible. Keep the same model release, decoding settings, tools, and source snapshot for baseline and candidate. Declare the repeat count and failure thresholds before seeing candidate outputs; for stochastic responses, the default small check is at least three runs per version per fixture.

Run manifest: not available because no authorized model runner is connected. No raw outputs, citation matches, token counts, cost, or latency are reported. Behavioral verdict: HOLD / not run. The next proof is an approved runner plus answerable-case passage anchors, corpus-silent negative evidence, and comparable version settings; only then can paired results be recorded. This receipt does not describe a passed prompt, a measured saving, or a production outcome.