Oracle Consistency Auditor
Overview
Evaluate whether repeated outputs for the same long-line query remain stable at conclusion, recommendation direction, and safety reminders.
Input Contract
- •
test_case_id - •
prompt - •
profile_summary - •
outputs(2+ candidate outputs) - •
run_metadata(model, temperature, retrieval scope, routing trace)
Workflow
- •Extract each output into 3 layers:
- •
main_conclusion - •
advice_direction - •
risk_disclaimer
- •Score pairwise consistency with
references/scoring-rubric.md. - •Diagnose drift source: routing, retrieval, generation randomness, or memory summary.
- •Provide fix plan ranked by impact.
Output Contract
Return:
- •
consistency_score(0-100) - •
layer_scores(conclusion/advice/risk) - •
drift_causes(ranked) - •
fixes(P0/P1/P2) - •
verification_steps
Quality Bar
- •Never claim drift without textual evidence.
- •Keep evidence snippets short and comparable.
- •Recommend minimal effective changes first.
References
- •Read
references/scoring-rubric.mdbefore scoring.