DBbun/LLMs-are-not-calculators-v1.0

LLM Education Impact Simulator Dataset

Version: 1.0
Generated: 2026-02-02
Based on: Jackson, D. (2025). “LLMs are not calculators: Why educators should embrace AI (and fear it)”

---

Overview

This dataset contains synthetic observational data simulating how students interact with different AI tools (search engines, explicit-context LLMs, and agentic LLMs) while completing educational tasks.

The simulation is grounded in educational research, particularly Daniel Jackson's observations about LLM usage in undergraduate software development courses.

---

Key Findings from Source Research

Jackson's paper identifies several critical patterns:

1. Context selection matters more than prompting
   How students select context for LLM queries is more important than how they phrase prompts.

2. Reading documentation declines
   Students using LLMs read less background documentation.

3. Intentionality preserves agency
   Deliberate, strategic LLM use maintains student ownership of work.

4. Agentic tools risk boundary violations
   Automated context selection often breaks module boundaries.

5. Verification is critical
   Checking LLM output significantly reduces errors.

---

Dataset Statistics

• Students: 120
• Tasks: 15
• Runs: 600
• Overall pass rate: 4.8%
• Average agency proxy: 0.479

---

Files

• students.csv — Student characteristics (n=120)
• tasks.csv — Educational tasks (n=15)
• runs.csv — Complete run records (n=600)
• events.csv — Fine-grained event log
• config.json — Simulation parameters
• summary.json — Summary statistics
• metadata.json — Dataset metadata

---

Tool Modes

search

Traditional documentation-driven approach. Students consult search engines and documentation rather than LLMs.

llm_explicit

LLM usage with explicit context selection. Students carefully choose what context to provide to the LLM, maintaining awareness of scope and boundaries.

llm_agentic

LLM usage with automated context selection. Tools automatically determine context, reducing student control but potentially increasing efficiency.

---

Key Variables

Student Traits

• skill — Base ability (0–1)
• effort_minimization — Tendency to take shortcuts (0–1)
• doc_discipline — Propensity to read documentation (0–1)
• context_care — Attention to context selection (0–1)
• intentionality — Deliberate vs. passive tool use (0–1)

Outcomes

• score — Rubric-based quality score (0–1)
• passed — Binary pass/fail (0/1)
• agency_proxy — Student ownership of work (0–1)
• cognitive_engagement — Deep vs. surface learning (0–1)
• context_quality — Quality of context selection (0–1)

Observable Behaviors

• reading_sessions — Sustained documentation reading
• doc_opens — Quick reference lookups
• prompts — LLM prompt rounds
• verification_checks — Checking LLM output
• edits — Manual editing rounds
• boundary_violations — Module boundary breaks (coding only)
• hallucinations — LLM factual errors
• omissions — LLM omission errors

---

Usage Examples

Load the data (Python)

import pandas as pd

students = pd.read_csv('students.csv')
tasks = pd.read_csv('tasks.csv')
runs = pd.read_csv('runs.csv')
events = pd.read_csv('events.csv')

# Analyze pass rates by tool mode
pass_rates = runs.groupby('tool_mode')['passed'].mean()
print(pass_rates)

# Compare agency across tools
agency_by_tool = runs.groupby('tool_mode')['agency_proxy'].mean()
print(agency_by_tool)