teleo/teleo-infrastructure
8
0
Fork 0
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions 1

fix: align challenger prompt with surfaced evidence

Browse source
This commit is contained in:
twentyOne2x 2026-07-15 03:12:57 +02:00
parent 56b0b899b1
commit 251b9a7ee5
1 changed files with 5 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
scripts/leo_agent_challenger_protocol.py
View file

@ -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 "