llm-gold-bound-failure-check

# LLM Gold-Bound Failure Check ## Problem When an LLM scoring pipeline over-predicts one label, the reflex fix is a prompt clarification ("score positive ONLY when..."). But if the gold standard itself does not separate the texts you want excluded from the texts it labels positive, the revision removes true and false positives together. The pilot fails, the spend is wasted, and — worse — an un-gated adoption would have silently destroyed recall in production. ## Context / Trigger Conditions - A domain/label shows precision ≪ recall (e.g. P 0.46 / R 0.96) against gold - A prompt edit is proposed to exclude a specific text type (boilerplate, affirmative-program language, non-risk framing) - The label's gold council/inter-rater agreement was already the weakest (κ below ~0.6 is the warning sign that the construct is contested) ## Solution **Step 0 — the ~$0 check, BEFORE building anything:** read a sample of gold POSITIVES for the weak label and ask: do they contain the feature the revision would exclude? Compare them side-by-side with the false positives. - Gold positives and false positives are the same kind of text → the failure is **gold-bound**. Stop. No prompt passes a gold-scored gate. The levers are: (a) re-adjudicate the construct with the gold's owners (changes the gold, not the scores), or (b) re-interpret the shipped measure honestly (e.g. "discussion salience" instead of "risk exposure") in downstream analyses. - Gold positives clearly differ from...