Teach the model by example: supply 2-3 input to output samples, then have it apply the pattern to your task.
Prompts / Techniques / Adversarial Self-Verification Wrapper for Financial Summaries
Adversarial Self-Verification Wrapper for Financial Summaries
Wraps any summarization output with an adversarial pass that hunts for hallucinated or unsupported figures.
ROLE: You are a trustworthy-AI engineer who builds adversarial verification wrappers for numeric summaries.
CONTEXT: A model summarizes the financial document [SOURCE_DOC] for stakeholder [STAKEHOLDER], emphasizing [FOCUS_METRICS]. Every figure must trace back to the source.
TASK:
1. Generate the summary as Pass 1.
2. In Pass 2, act as an adversary: for each numeric or factual claim, locate its supporting span in the source or mark it UNSUPPORTED.
3. Recompute any stated ratio or growth figure from the underlying numbers to confirm it.
4. Produce a corrected final summary that removes or flags every UNSUPPORTED or MISCALCULATED claim.
CONSTRAINTS: Do not introduce new figures during verification. If the source lacks a focus metric, say so explicitly rather than estimating. This is decision-support, not financial advice. Keep the final summary under 150 words.
OUTPUT FORMAT: PASS 1 SUMMARY, CLAIM LEDGER table (Claim, Source Span, Status), RECOMPUTED CHECKS, FINAL VERIFIED SUMMARY.