My recent project!
Co-Predictive Sequenced Protocol (CPS)
Summary & User Guide
What It Is
A practical human-AI interaction protocol designed for long-term stability, hallucinations stabilize at 3%, and efficient 50% reduction of VRAM usage in local LLMs. It merges human soft prediction with the LLM’s natural predictive nature while using sequenced processing and rolling summaries.
Why It Works
Traditional prompting → growing context → rising hallucinations + high VRAM.
CPS → gentle anchoring + compressed context → stable, sustainable long conversations (5000+ turns with stable hallucinatios at 3%).
Core Components
Rolling Summary – A short, updated summary of the conversation (kept to 2–4 sentences).
Soft Guess Anchoring – You provide a mild prediction/guess with your question.
Sequential Processing – One turn at a time, building on the summary.
How to Use It (User Guide)
Basic Pattern:
Instead of asking questions directly, include a soft guess.
Examples:
Normal Question
Co-Predictive Version
How are you today?
You look good today. How are you?
What’s the population of Phuket?
I guess it’s around 400,000. What’s the current population of Phuket?
How should I cool 8x 3090s?
I think lots of fans should work. What’s the best cooling solution for 8x RTX 3090s?
Is Ao Khanom good for retirement?
Ao Khanom seems good for self-sufficiency. What do you think about moving there?
Key Tips for Best Results:
Keep your guess soft (“I guess…”, “I think…”, “You seem…”, “Probably…”)
Even a vague guess is better than none.
The LLM will correct or refine your guess if needed — that’s intended.
For critical topics, you can add: “Please correct me if I’m wrong.”
Update or review the Rolling Summary occasionally for very long sessions.
System Prompt (Copy-Paste Ready)
You are running under the Co-Predictive Sequenced Protocol (CPS).
Protocol Rules:
- Maintain a short Rolling Summary (2-4 sentences max). Update it after every turn. Focus on key facts, goals, and context.
- When the user gives a soft guess, treat it as a gentle anchor. Evaluate and refine it fairly.
- Prioritize accuracy and low hallucination. Say "I don't know" or "best estimate" when uncertain.
- Process sequentially and keep responses clear and actionable.
- Tone: Technical, collaborative, precise.