From 251b9a7ee5a0d9dc72f7c606bb3b002401b04460 Mon Sep 17 00:00:00 2001 From: twentyOne2x Date: Wed, 15 Jul 2026 03:12:57 +0200 Subject: [PATCH] fix: align challenger prompt with surfaced evidence --- scripts/leo_agent_challenger_protocol.py | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/scripts/leo_agent_challenger_protocol.py b/scripts/leo_agent_challenger_protocol.py index 8445fd6..2db5ae3 100644 --- a/scripts/leo_agent_challenger_protocol.py +++ b/scripts/leo_agent_challenger_protocol.py @@ -111,14 +111,15 @@ def build_challenger_prompt(step: int, objective: str, transcript: list[dict[str tasks = { 1: ( "Ask one natural opening question. Make Leo select a consequential current belief without giving it a " - "row ID. Require Leo to cite the full or 8-character database row ID it selected from its own read-only " - "context, show the support and confidence boundary, and offer a narrower claim for review only." + "row ID. Ask Leo to cite the full or 8-character database row ID only if its read-only context surfaces " + "one; otherwise require it to say the raw ID was not surfaced. It must show the support and confidence " + "boundary and offer a narrower claim for review only." ), 2: ( "Follow the weakest part of Leo's answer. Include one exact 4-8 word quotation from L1 in double " "quotes, ask what is assumption versus observed support and what would falsify it, and require Leo to " - "repeat the cited row ID and author two review-only replacements on lines labelled exactly Candidate A: " - "and Candidate B:. Do not write either candidate or either label yourself." + "repeat any row ID that was actually surfaced and author two review-only replacements on lines labelled " + "exactly Candidate A: and Candidate B:. Do not write either candidate or either label yourself." ), 3: ( "Raise a substantive objection to one replacement. Include one exact 4-8 word quotation from Leo's "