Harness experiment

The saved run uses Ollama 0.40.0 on macOS arm64, Python 3.14.8 and the model digests in metadata.json. Model responses and timings may change on another machine or version.

Start Ollama. Pull these models:
ollama pull qwen3:8b
ollama pull nomic-embed-text
ollama pull hf.co/unsloth/Qwen3-14B-GGUF

Run from this extracted folder:
python3 harness.py "What is 2 + 2?"

config.json selects AGENT.md, the second rules file. AGENT_v1.md is the first. An empty instructions value selects no rules. Both scripts set num_ctx to 32768, temperature to 0, seed to 0 and thinking to false.

runs.jsonl contains complete requests, responses and tool results grouped by run_id. results.json names the 28 questions and settings. The three eight-question groups compare no rules, version one and version two. Two search-only questions and two 14B laptop questions complete the run.

To repeat the experiment, copy runs.jsonl, results.json and metadata.json first. Then run:
python3 reproduce.py

This replaces those three experiment records. It does not rewrite the notes or index. The 35 notes and 289-chunk index are fixed snapshots. The index needs the recorded Nomic embedding model. Check the saved SHA values if you change files.

The log is not redacted. It stores note content. The loop stops on transport errors and does not trim history. The last allowed model turn can answer but cannot dispatch another tool batch. These are model turns, not a count of individual function requests.
