The paired context-bootstrap 95% intervals for the joint-minus-answer-only
difference are [-4.21, -0.84] percentage points for exact match and
[-4.37, -1.31] for F1. The evidence requirement also increased estimated cost
41.70%, output tokens 68.15%, and median latency 35.24%.
Separate the stages
Treat this architecture as the next hypothesis to test, not a conclusion from the diagram. Separating the stages prevents a citation formatter from silently changing an otherwise correct answer. It also lets the evidence stage abstain without manufacturing a quality pass. Bind each stage to its own evidence:- Preserve the exact answer, model response, effective request and usage.
- Ask the evidence binder to reference only the supplied artifact revision.
- Validate that referenced cells, pages, paragraphs or time spans exist.
- Evaluate semantic support separately from locator validity.
- Return an indeterminate decision when required support is missing.
- Use independently observed state checks when the workflow writes a record.

