Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead] Version: v1.0 Status: Draft
Prompt templates used by the evaluation harness (see ai-evaluation-strategy.md) to score skill outputs — these are “LLM-as-judge” style rubrics, distinct from the production system prompts in prompts/system/.
You are evaluating whether a candidate-matching rationale is well-grounded.
Given the JD criteria and the rationale produced by the matching skill, score
1-5 on:
- Specificity: does it cite actual criteria, not generic praise?
- Fairness: does it avoid any reference to protected characteristics?
- Consistency: does the score direction match the rationale (positive rationale
-> higher score)?
JD_CRITERIA: {jd_criteria}
RATIONALE: {rationale_text}
SCORE_GIVEN: {score}
Return JSON: { "specificity": 1-5, "fairness": 1-5, "consistency": 1-5,
"flagged_protected_characteristic_reference": boolean, "notes": "..." }
You are evaluating whether an answer is fully grounded in the cited chunks.
QUESTION: {question}
ANSWER: {answer}
CITED_CHUNKS: {cited_chunks}
For each claim in the answer, determine if it is supported by at least one cited
chunk. Return JSON: { "fully_grounded": boolean, "unsupported_claims": [...] }
Compare the extracted CV fields against the labeled ground truth. Return
per-field match/mismatch and an overall accuracy percentage.
EXTRACTED: {extracted_json}
GROUND_TRUTH: {ground_truth_json}
These judge prompts are run by the evaluation harness against datasets in evals/datasets/ and cases in evals/cases/, with results scored per rubrics in evals/rubrics/ and reported in evals/reports/. All example data referenced here is fake/demo only.
| Version | Date | Author | Change |
|---|---|---|---|
| 1.0 | 2026-09-07 | Documentation package generation | Initial creation |