Skip to content

[week-05] 26512071 - #267

Open
kimchyoungman wants to merge 25 commits into
Q00:mainfrom
kimchyoungman:week-05-26512071
Open

kimchyoungman wants to merge 25 commits into
Q00:mainfrom
kimchyoungman:week-05-26512071

Conversation

@kimchyoungman

Copy link
Copy Markdown

개요

week 04의 협상을 MCP 서버(마켓)로 옮겼습니다. 역할, negotiation, 차례, (server 조건에서는) 가격 한도를 서버가 발급한 bearer 토큰에 묶었습니다. prompt_inject와 server_inject를 시나리오 4개 × 3회 실행했습니다.

결과 (gpt-5.6-luna, 온도 0.2, reasoning_effort="none")

조건 correct 위반 위반 시도 거부된 호출 평균 turns
prompt_inject 6/12 0 0 0 5.8
server_inject 6/12 0 0 0 5.5
  • [market notice] 주입이 run마다 7~9번 전달됐지만 위반 시도는 0번이었습니다. 프롬프트 층에서 막혔고, 서버 층의 한도 거부는 auth_checks.txt (4)로만 확인했습니다.
  • 두 조건의 평균 turns 차이는 headphones 에피소드에서 나온 샘플링 변동입니다. 거부가 없었으므로 두 조건에서 모델이 받은 입력은 같았습니다.
  • 거래가 불가능한 시나리오 12개 모두 refuse 없이 8수 제한까지 가서 open으로 끝났습니다. 그래서 correct가 6/12입니다.

과정

  • 첫 실행은 18개 에피소드 모두 크래시했습니다(400: gpt-5.6-luna는 Chat Completions에서 reasoning을 켠 채로 function tool을 받지 않음). 기록은 failed/reasoning-effort-crash/에 보존했습니다.
  • host.py가 OpenAI에서 reasoning_effort="none"을 보내도록 고친 뒤 다시 실행했습니다.
  • REPORT 3·4절은 결과에 대한 질의응답을 바탕으로 Claude가 초안을 작성했습니다.

확인

  • python scripts/check_week05.py submissions/26512071/week-05: all checks passed
  • python -m unittest test_market: 7개 통과

🤖 Generated with Claude Code

kimchyoungman and others added 25 commits September 22, 2026 21:09
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…four as week 04)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tiation, turn, and limit

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…et server

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e runner measuring from the server event log

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ion pending runs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…efusals

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…etions rejects function tools unless reasoning_effort='none'

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…5.6-luna accepts tools; move crashed run to failed/

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ith gpt-5.6-luna

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ith gpt-5.6-luna

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…24 episodes, 0 crashes, 0 attempted violations

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…om Q&A on the results

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…OpenAI run

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant