Mixed-Initiative LookOut (CHI’99-inspired)

This repository contains a self-contained Python implementation of a synthetic mixed-initiative decision-making model inspired by the ideas in Eric Horvitz’s CHI’99 LookOut work on attention, interruption, and decision-theoretic control.

Hugging Face: https://huggingface.co/datasets/DBbun/Mixed-Initiative-Lookout-CHI99

The code generates synthetic interaction traces, analytic decision thresholds, figures, and dataset exports suitable for research in:

• mixed-initiative systems,
• decision theory under uncertainty,
• attention-aware assistants,
• calibration and interpretability,
• reinforcement learning and policy analysis.

Developed by DBbun LLC — January 2026.

This project is inspired by publicly described concepts in the literature.
It is not affiliated with Microsoft Research or Eric Horvitz and does not include any original LookOut code or data.

---

What this repository contains

• chi99horvitz v1.0.py
  A single, runnable Python script that:
  • generates synthetic mixed-initiative episodes,
  • computes expected utilities and analytic thresholds,
  • simulates attention and interruption costs,
  • produces figures (policy maps, threshold curves, attention dynamics),
  • exports datasets in CSV and JSONL formats,
  • writes a quality / validation report.

There are no subdirectories and no external assets required.

---

Conceptual overview

The model studies how an intelligent assistant should choose between:

• no action,
• dialog / clarification,
• scoped suggestions,
• direct action,

based on:

• uncertainty about user intent,
• user attention and readiness,
• urgency / time pressure,
• asymmetric utilities and costs.

Each decision is made by explicit expected-utility maximization, and the resulting decision thresholds are stored analytically, enabling interpretability and counterfactual analysis.

---

Outputs produced by the script

When run, the script produces:

• Synthetic datasets

  • mixed_initiative_traces.csv — tabular dataset
  • mixed_initiative_traces.jsonl — structured episode-level format

• Figures

  • Expected-utility threshold plots
  • Context-shifted policy curves
  • Attention (dwell time vs message length) dynamics
  • Mixed-initiative policy maps

• Quality report

  • Calibration metrics (AUROC, ECE)
  • Sensitivity and monotonicity checks
  • Baseline policy comparisons

• Configuration

  • mixed_initiative_traces.config.json capturing all generator parameters

---

How to run

python "chi99horvitz v1.0.py"