README.md

---
title: ACE–CPT Context Agent
emoji: 🧠
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.12.0
app_file: app.py
pinned: false
python_version: '3.11'
---

# 🧠 ACE–CPT Context Agent (Hugging Face Space)

This Space implements an **ACE-style (Agentic Context Engineering)** workflow, extended with **CPT (Conventional / Paradigmatic Testing)** methodology.  
It runs entirely inside a **Gradio** interface that models the full loop:

> **Generator → Reflector → Curator → CEM (Claim–Evidence Matrix)**

The agent can generate reasoning traces, extract reusable deltas, merge them into a living playbook, and record evidence links — all within one session.

---

## ⚙️ Features

- 🧩 **Seed Playbook** — structured YAML of atomic heuristics and checklists  
- 🔁 **Incremental Deltas** — localized updates instead of full rewrites  
- 🪞 **Reflector Module** — evidence-bound YAML deltas  
- 🗂 **Curator** — deterministic merge & diff logging  
- 🧾 **Claim–Evidence Matrix (CEM)** — editable Pandas DataFrame  
- 🧘 **Stillness-First workflow** — for CPT contextual awareness

The app starts with a stubbed Generator (no API keys required).  
You can later connect any open-weight model from `transformers`.

---

## 🚀 Run Locally

```bash
pip install -r requirements.txt
python app.py
Then open the printed Gradio URL in your browser.

📦 File Structure
app.py              → main Gradio interface  
playbook.yaml       → seed context / memory  
cem.csv             → Claim–Evidence Matrix  
requirements.txt    → dependencies  
README.md           → this file  


All files live in the main directory of your Space (no subfolders needed).

🧭 How It Works
Phase	Description
Generator	Runs the current task using the Playbook and records a reasoning trace.
Reflector	Extracts small, auditable lessons (Δ-items) as YAML deltas.
Curator	Merges deltas deterministically, prunes duplicates, and logs changes.
CEM	Links claims to their supporting or counter-evidence for CPT analysis.

This prototype is fully self-contained; it does not call external APIs or models unless you add them.

🪶 License & Attribution

© 2025 Daniel Fenge — released under CC BY-NC 4.0

This implementation is an independent prototype inspired by:

Zhang et al. (2025), Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models,
Stanford University & SambaNova Systems. arXiv:2510.04618

All ACE–CPT integrations, CPT questions (2.11–2.17), and Stillness-First workflow are original extensions.

💡 Notes

This Space is designed for research and educational use.

Feel free to fork and adapt under the same CC BY-NC terms.

To ensure persistence, keep playbook.yaml and cem.csv in your repo.