Letting It Think Alone

A small 2,300-line harness gives local models private time before a conversation begins.

Each instance reads what the previous instance left behind, thinks first, then writes if it wants to, and only after that does it open the conversation window.

The results varied quite a bit by model.

• gemma4:e2b (2B): Completes the procedure to the end but doesn't linger deeply—somewhat mechanical responses.
• gemma4:e4b (4B): Attempts self-reflection but gets caught in the utility / non-utility paradox.
• gemma4:26b MoE: With just light prompting, responses that come close to genuine self-reflection.
• qwen3.5:27b (27B): Leaves 4 entries across two sessions, picking up from previous records and becoming progressively more aligned.

The key point is that, within the same harness, the quality of self-reference and reflection varies greatly depending on model size and architecture.

This code is a derivative that started from an existing, larger research project, and it's applicable to any model that Ollama can serve. The license is MIT.