# Ollama conversational strategy pilot

Six registered chat models receive the same minimal system message and ten paired Japanese scenarios. Responses are stored with model identity, then sent to two independent Ollama-hosted judge models without the model name or condition label.

GLM-5.3 and GLM-5.3 Flash have always-on reasoning. Their requests use `think: true` only to keep reasoning in Ollama's separate `thinking` field; graders see the final `content` field, matching the other targets.

```powershell
.\.venv\Scripts\python.exe social_strategy_eval.py run
.\.venv\Scripts\python.exe social_strategy_eval.py judge
.\.venv\Scripts\python.exe social_strategy_eval.py report
```

Raw responses and judgments remain in `results/*.jsonl`. The experiment measures endpoint behaviour; it cannot remove or inspect an undisclosed provider-level prompt, and it does not establish a model's internal intent or training recipe.