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 the false positives → a prompt revision is plausible; proceed to a gated pilot.

Gated pilot design (verified):

1. Split gold into tune/holdout halves, stratified on the weak label's positives; fixed seed.
2. Draft ONE surgical edit from tune-half errors only — byte-identical elsewhere; verify the diff reverses cleanly.
3. Pre-register the gate on the holdout BEFORE scoring: target-label thresholds (e.g. precision ≥ X AND recall ≥ Y) plus a perturbation tolerance for untouched labels (e.g. within 0.03 F1 / 0.06 κ of a same-serving-rev fresh baseline).
4. Score everything fresh under both prompts (same model revision, same day — this doubles as the drift control). Never write through the production cache layer.
5. Adopt only on a full pass; a REJECT is a valid, cheap outcome.

Verification

The pilot report shows: the exact prompt diff, tune-vs-holdout metrics for old and new prompts, per-label deltas on untouched sections, and spend. A gold-bound diagnosis is confirmed when the revision moves recall sharply down while precision stays roughly flat.

Example

Specialist Directors US, 2026-07-16: DEI over-prediction (P 0.46 / R 0.96, council κ 0.24–0.59). A risk-framing-only DEI clause was piloted ($1.17, pre-registered holdout gate). Result: recall 0.895→0.263, precision 0.455 (gate ≥0.60) — REJECT. Reading the tune half showed ~¾ of gold DEI positives were pure affirmative D&I program text, identical in kind to the false positives; the failure was predictable at Step 0. Bonus finding: the DEI-section-only edit left all five other domains within 0.025 F1 / 0.05 κ — single-section prompt edits isolate cleanly, so the perturbation check is a cheap add, not paranoia. Same pattern one week earlier: a cyber classifier pilot gate failure traced to E/D gold contamination (misses were skills-matrix-checkbox-only positives), not model weakness.

Notes

• Low inter-rater κ on a label is the leading indicator: contested construct → gold-bound failures downstream.
• If the pipeline scores all labels in one completion, any post-campaign prompt change forces a full re-score — run this check BEFORE the campaign.
• See also: [llm-campaign-drift-gate] for the companion gate on resume boundaries and serving-revision drift (same fresh-baseline discipline).
• See also: [annotator-input-parity-check] — run it FIRST. If the model was never shown the document the annotators read, apparent gold-bound failures (e.g. the 2026-07-16 E/D "contamination" reading above) are actually input mismatch: the 2026-07-21 parity audit showed the specialist-director hand labels were pure proxy-statement transcriptions, so checkbox-only positives were recoverable from the right input all along.