Run a skeptic against your own answer.
A second LLM tasked with refuting your first one catches what no eval set can.
Why it matters
You can build all the evaluation suites in the world. They’ll catch the failure modes you anticipated. They won’t catch the failure mode that ships next quarter. A second‑pass adversarial checker — another LLM whose only job is to attack your first LLM’s answer — catches things eval sets miss because it sees the same context the answerer saw, with fresh eyes and an explicit mandate to disagree.
The cost of not doing it
Your AI gives a confident, well‑cited answer based on a 2017 policy document. The policy was superseded in 2022 — but the 2017 doc was the most similar to the query (more explicit, more verbose, more readable). No eval flagged it because no eval anticipated the specific phrasing. Your customer reads the answer and acts on superseded policy.
How BRAG does it
After the Verifier passes the answer, BRAG’s Refutation Agent (a different LLM) reads the draft, picks the most attackable claim, generates a counter‑hypothesis, and searches the corpus for evidence that contradicts it. It runs seven strategies — restated_value, later_reversal, guidance_vs_actual, alternative_cause, materialization, policy_change, revised_value — each tuned to a specific way the answer could be wrong.
In the demo D2 query “Did Netflix ever say it had no plans to add ads?”, the Refutation Agent generated the hypothesis “Netflix’s Q4 2022 shareholder letter announced an ad‑supported tier, reversing its earlier no‑ads position,” searched for evidence, found it, and triggered a re‑retrieval loop. The final answer includes both the original position and the reversal.
Apply it in your own system
The pattern is robust. The cost is real.
- Pick a strong second LLM (different model family from your generator, if possible).
- Give it a sharp, single‑purpose prompt: “Read this answer and the supporting evidence. Find a reason to believe the answer is wrong.”
- Categorise the kinds of refutation it should try — the categories depend on your domain.
- Loop back to retrieval when refutation succeeds.
Cost: roughly 2× your per‑query LLM spend. Worth it for any application where being wrong is expensive.
This is one of six principles abstracted from BRAG — a retrieval‑augmented generation system with two LLM skeptics gating every confident answer. Tested on ten years of Netflix’s public financial reporting.