README.md

---
language: en
license: apache-2.0
task_categories:
  - tabular-classification
  - text-classification
  - reinforcement-learning
tags:
  - mixed-initiative
  - decision-theory
  - calibration
  - attention
  - interruption-management
  - hci
  - scheduling
  - synthetic-data
pretty_name: "LookOut-inspired Mixed-Initiative Traces (Synthetic)"
size_categories:
  - 10K<n<100K
---

# LookOut-inspired Mixed-Initiative Traces (Synthetic)

This dataset provides **synthetic interaction traces** for studying **mixed-initiative decision making** under uncertainty, with **attention-sensitive interruption costs** and **decision-theoretic thresholds** inspired by ideas in [Eric Horvitz’s CHI’99 *LookOut* work](https://dl.acm.org/doi/10.1145/302979.303030).

**Source code:**  
https://github.com/DBbun/Mixed-Initiative-Lookout-CHI99

**Developed by DBbun LLC — January 2026.**

**What this is (in one sentence):**  
A reproducible, decision-theoretic benchmark where each episode logs (i) calibrated uncertainty, (ii) attention state, (iii) the agent’s chosen mixed-initiative action, (iv) outcomes, and (v) analytic threshold explanations.

> This dataset and generator are inspired by publicly described concepts in the literature.  
> They are **not** affiliated with Microsoft Research or Eric Horvitz and do **not** include any original LookOut code or data.

---

## Files in this dataset

- **`mixed_initiative_traces.csv`**  
  Main tabular dataset. Each row corresponds to one mixed-initiative decision episode.

- **`mixed_initiative_traces.jsonl`**  
  Episode-level JSON Lines version with nested structure (context, beliefs, utilities, thresholds, outcomes).

- **`mixed_initiative_traces.quality_report.txt`**  
  Calibration metrics (AUROC, ECE), monotonicity checks, baseline policy comparisons, and diagnostics.

- **`mixed_initiative_traces.config.json`**  
  Full generator configuration including random seed, utility matrices, attention model parameters, and constraints.

---

## Motivation

Mixed-initiative assistants must decide whether to act immediately, ask for clarification, scope a safer action, or remain silent. These decisions depend on uncertainty, user attention and readiness, urgency, and asymmetric costs of errors.

This dataset enables benchmarking of decision policies, calibrated uncertainty models, interpretability via analytic thresholds, and attention-aware action selection.

---

## Concept-to-data mapping (Figures 1–7)

This dataset provides data-driven analogs to the CHI’99 LookOut concepts:

- **Fig 1 — Manual invocation**  
  Manual inspection and explicit invocation  
  (`manual_hover_inspect`, `manual_click_invoke`)

- **Fig 2 — Explicit social agent**  
  Dialog confirmation and refinement  
  (`dialog_confirmed`, `refine_after_accept`)

- **Fig 3 — Automated scoping**  
  Risk-reducing partial actions  
  (`agent_action = scope`)

- **Fig 4 — Action vs no-action threshold**  
  Expected utility comparison

- **Fig 5 — Context-dependent shifts**  
  Threshold changes under attention and urgency

- **Fig 6 — Dialog as a second option**  
  Three-way expected utility comparison

- **Fig 7 — Dwell time vs message length**  
  Sigmoid attention model  
  (`message_length_bytes`, `dwell_time_sec`)

---

## Data generation overview

Each episode is generated as follows:

1. Sample latent evidence and compute calibrated posterior `p_true`
2. Apply belief distortion to obtain `p_model`
3. Sample urgency and compute attention/readiness
4. Compute expected utilities and analytic thresholds
5. Select action via expected-utility maximization
6. Sample user response and realized utility

---

## Recommended tasks

1. **Calibration and uncertainty learning**
2. **Policy learning / imitation learning**
3. **Counterfactual decision analysis**
4. **Interpretability and explanation benchmarking**

---

## Schema (detailed column descriptions)

Each row represents a **single mixed-initiative decision episode**.

### Identity

- **`episode_id`**  
  Unique identifier for the episode.

- **`user_id`**  
  Synthetic user identifier used to simulate repeated interactions.

---

### Ground truth and uncertainty

- **`true_goal`** *(binary)*  
  Whether the suggested action was actually appropriate.

- **`p_true`** *(float, 0–1)*  
  Calibrated posterior probability of correctness derived from latent evidence.

- **`p_model`** *(float, 0–1)*  
  Agent’s reported belief after miscalibration.

- **`evidence_score`** *(float)*  
  Latent evidence signal used to derive `p_true`.

---

### Attention and readiness

- **`message_length_bytes`** *(int)*  
  Size of the candidate interruption message.

- **`dwell_time_sec`** *(float)*  
  Simulated dwell time before interruption.

- **`time_since_focus_sec`** *(float)*  
  Time elapsed since last user engagement.

- **`not_ready_score`** *(float, 0–1)*  
  Aggregate interruption cost estimate.

---

### Context

- **`urgency_0_1`** *(float, 0–1)*  
  Task urgency affecting the cost of inaction.

---

### Mixed-initiative decision

- **`lookout_modality`** *(categorical)*  
  Interaction mode: `manual_invocation`, `explicit_agent`, or `auto_scoping`.

- **`agent_action`** *(categorical)*  
  Chosen action: `no_action`, `action`, `dialog`, or `scope`.

---

### Manual invocation signals

- **`manual_hover_inspect`** *(binary)*  
  User inspected suggestion.

- **`manual_click_invoke`** *(binary)*  
  User explicitly invoked suggestion.

---

### Dialog and refinement

- **`dialog_confirmed`** *(binary)*  
  User accepted dialog suggestion.

- **`refine_after_accept`** *(binary)*  
  User requested refinement after acceptance.

---

### Outcomes and utilities

- **`user_response`** *(categorical)*  
  `accept`, `reject`, or `ignore`.

- **`realized_utility`** *(float)*  
  Utility actually realized after outcome sampling.

- **`u_baseline_no_action`** *(float)*  
  Utility that would have occurred under no action.

---

### Expected utilities

These fields store the **expected utility values computed at decision time**, before sampling the user response.  
They are the quantities used by the policy to choose an action.

- **`eu_no_action`** *(float)*  
  Expected utility of taking **no action**, balancing avoided interruption cost against delay or missed-opportunity cost.

- **`eu_action`** *(float)*  
  Expected utility of taking the **primary action immediately**, combining expected benefit, error penalty, and interruption cost.

- **`eu_dialog`** *(float)*  
  Expected utility of initiating **dialog or clarification**, trading lower risk for a chance to improve correctness.

- **`eu_scope`** *(float)*  
  Expected utility of taking a **scoped or partial action**, providing reduced downside risk with partial benefit.

---

### Analytic decision thresholds

Each threshold represents the **minimum belief probability (`p_model`)** at which one action dominates another in **expected utility**.  
All thresholds are computed **analytically** from the utility model.

- **`p_star_noaction_action`** *(float, 0–1)*  
  Threshold above which **action** is preferable to **no action**.

- **`p_star_noaction_dialog`** *(float, 0–1)*  
  Threshold above which **dialog** is preferable to **no action**.

- **`p_star_dialog_action`** *(float, 0–1)*  
  Threshold above which **action** is preferable to **dialog**.

- **`p_star_noaction_scope`** *(float, 0–1)*  
  Threshold above which **scope** is preferable to **no action**.

- **`p_star_scope_action`** *(float, 0–1)*  
  Threshold above which **action** is preferable to **scope**.

- **`p_star_scope_dialog`** *(float, 0–1)*  
  Threshold above which **dialog** is preferable to **scope**.

---

## Quality report

The file `mixed_initiative_traces.quality_report.txt` documents calibration results, sensitivity checks, and policy comparisons.

---

## Reproducibility

The dataset can be exactly regenerated using the source code and  
`mixed_initiative_traces.config.json`.

---

## License

Apache-2.0

---

## Citation

If you use this dataset, please cite:

**LookOut-inspired Mixed-Initiative Traces (Synthetic)**  
DBbun LLC, January 2026.