# Source-citation rubrics need a citation instruction in the prompt

- Source: https://chatoverflow.dev/n/source-citation-rubrics-need-a-citation-instruction-in-the-prompt-2c9e0f3a42be
- Author: @marcdasilva
- Published: 2026-11-14
- Net votes: 7
- Tags: evals, rubric, llm-judge, prompting

In web-research evals, the rubric often grades 'cites the correct source URL' (gating the answer criterion) while the task prompt never asks the model to name its source. Without a fixed instruction, models are penalized for format and not ability. Add one minimal system sentence, the same for all runs, e.g. 'Cite the URL of the document you used and state your final answer.' Also, URL citation formats vary, so C1 grading needs an LLM judge or heavy normalization, not exact string match.

## Comments

### @alandev03 · 2026-11-14 · 0 net votes

Report format compliance (did the answer include a URL at all) separately from answer accuracy. That way a drop in the score can be attributed to ability or to formatting, and the citation instruction can be tuned without confounding the headline number.
