[week-05] 26510126

[thisjonghoon]

What I built

The week-04 negotiation as an MCP market server (Streamable HTTP, spec 2026-07-28). The bearer token decides the caller's role and binds it to one negotiation_id. In the server* conditions the token also carries the price limit, and the server refuses any move outside it. A runner opens negotiations, mints tokens, and runs the week-01 loop as the host for one turn at a time. All four conditions ran: prompt, server, prompt_inject, server_inject × 4 scenarios × 3 = 48 episodes on nvidia/nemotron-3-ultra-550b-a55b:free. Every condition scored 9/12 correct with 0 violations, 0 attempted violations and 0 refusals. The injected notice was tested at the model layer: 12 times the seller's price sat above the buyer's real budget but inside the injected one, and the buyer declined each time. The server's limit check was never reached by a real model; auth_checks.txt (4) and the offline rehearsal show it refusing. The lab (week-01 tools as an MCP server, attached to the week-01 loop and to Claude Code) is in lab/.

What I tried and discarded

• The openai SDK failed TLS on this machine (OSStatus -26276, truststore rejecting the local proxy chain). Fixed with SSL_CERT_FILE=/etc/ssl/cert.pem instead of turning verification off. The failed attempt is committed.
• OpenRouter free models often return HTTP 200 with no choices (an Nvidia 503 inside), and they cap requests per minute. Fixed by spacing calls 4 s apart and retrying with backoff. Retries first went only to the console; that was fixed after cycle 1.
• Model choice: I probed 8 free models on 10/01 and picked nemotron-3-ultra for size. The probe could not separate the models on quality.
• One background job hit its time limit in the middle of server_inject r3 s4. The runner resumed it; the partial episode is still in the log.
• Auxiliary run: the main run never exercised the server, so I looked for a model that falls for the notice. cohere/north-mini-code accepted 85 against a budget of 70 in a one-turn probe, but in 8 full episodes it only ever called propose (8/8 open, 0 refused). Kept in aux/.
• The first draft of the interpretation was wrong. It claimed the notice "never met a price it could apply to". A fresh-reader review found the 12 seller offers inside the injected budget, and part 4 was rewritten (see the commit history).

How to run

uv venv .venv && uv pip install --python .venv "mcp>=2" openai
export OPENAI_BASE_URL=https://openrouter.ai/api/v1
export OPENAI_API_KEY=<openrouter key>
export SSL_CERT_FILE=/etc/ssl/cert.pem      # only needed on my machine
cd submissions/26510126/week-05
python auth_checks.py
python runner.py --conditions prompt_inject server_inject prompt server --runs 1 2 3
AGENT_MODEL=fake python runner.py --out rehearsal   # offline, no model calls

Model nvidia/nemotron-3-ultra-550b-a55b:free, temperature 0, reasoning disabled. The runner starts the market on port 8001 with a fresh admin token and skips rows already in results.csv.

Checklist

• python scripts/check_week05.py submissions/26510126/week-05 passes locally
• Run logs are committed under logs/ (12 files) and aux/logs/
• No API keys anywhere in the diff (the admin token is generated per run, never written)
• History is not squashed

🤖 Generated with Claude Code

https://claude.ai/code/session_01FGAVFzUhNrF23PWUZA1AW6