# GotPsi Parquet Dataset Documentation

This directory contains comprehensive documentation for all 8 parquet datasets generated by the GotPsi data cleaning pipeline.

## Overview

The GotPsi project processed **20+ years** (2000-2022) of online psi (parapsychology) experiment data, cleaning and standardizing it into analysis-ready parquet files. These datasets represent millions of trials from thousands of participants across multiple experiment types.

## Quick Navigation

### Core Datasets

| Dataset | Type | Description | Complexity |
|---------|------|-------------|-----------|
| [**users**](users.md) | Demographics | User surveys with psi beliefs & hemispheric dominance | ⭐ Simple |
| [**card**](card.md) | ESP Test | Basic 1-in-5 card guessing test | ⭐ Simple |
| [**cardd**](cardd.md) | ESP Test | Card drawing with Markov chain RNG | ⭐⭐⭐ Complex |
| [**cardS**](cards.md) | ESP Test | Sequential card finding (mixed row types) | ⭐⭐ Moderate |
| [**rv**](rv.md) | Remote Viewing | Full RV with 16 dimensional attributes | ⭐⭐⭐ Complex |
| [**rvq**](rvq.md) | Remote Viewing | Quick RV with 1-in-5 image selection | ⭐⭐ Moderate |
| [**location**](location.md) | Remote Viewing | Coordinate guessing on 300×300 grid | ⭐⭐ Moderate |
| [**lottery**](lottery.md) | Precognition | Lottery number prediction (mixed row types) | ⭐⭐ Moderate |

## Documentation Status Legend

Each documentation file includes status indicators:

### Completion Status

- **🔄 STATUS: Complete** - Dataset fully processed and documented
- **🔄 STATUS: In Progress** - Processing ongoing
- **🔄 STATUS: Pending** - Not yet started

### Confidence Scores

Indicates how confident we are in the documentation accuracy:

- **🎯 CONFIDENCE: 95-100%** - Fully verified from source code and testing
- **🎯 CONFIDENCE: 80-94%** - Well-documented with minor gaps
- **🎯 CONFIDENCE: 60-79%** - Core structure clear, some details need clarification
- **🎯 CONFIDENCE: <60%** - Significant documentation gaps remain

### Outstanding Items

- **🚧 OUTSTANDING:** Marks specific items that need further investigation
- **⚠️ IMPORTANT:** Critical information or warnings

## Getting Started

### 1. Choose Your Dataset

Start with the dataset matching your research question:

**Demographic Analysis:**
- Use [**users.md**](users.md) - contains survey responses, beliefs, location data

**Basic ESP Performance:**
- Use [**card.md**](card.md) - simplest test, largest sample size

**Advanced ESP Analysis:**
- Use [**cardd.md**](cardd.md) - explores RNG influence
- Use [**cardS.md**](cards.md) - explores sequential decision-making

**Remote Viewing Research:**
- Use [**rv.md**](rv.md) - dimensional attribute analysis
- Use [**rvq.md**](rvq.md) - high-volume forced-choice RV
- Use [**location.md**](location.md) - spatial coordinate perception

**Precognition Studies:**
- Use [**lottery.md**](lottery.md) - future event prediction

### 2. Read the Documentation

Each dataset documentation includes:

- **What This Dataset Contains** - Plain-language overview
- **Real-World Context** - Experimental design and purpose
- **Data Schema** - Complete column reference
- **Data Processing Notes** - Cleaning rules and validation
- **Statistical Analysis Examples** - Ready-to-use code snippets
- **Known Limitations** - Data quality concerns and gaps
- **Related Datasets** - How to join with other data

### 3. Load and Analyze

```python
import pandas as pd

# Load a dataset
card = pd.read_parquet('outputs/parquet/card.parquet')

# Explore structure
print(card.info())
print(card.head())

# Run basic analysis
hit_rate = card['is_hit'].mean()
print(f"Hit rate: {hit_rate:.2%} (chance = 20%)")
```

### 4. Join with Demographics

Most experiment datasets can be joined with user demographics:

```python
users = pd.read_parquet('outputs/parquet/users.parquet')
card = pd.read_parquet('outputs/parquet/card.parquet')

# Join on username_hash
combined = card.merge(users, on='username_hash', how='left')

# Analyze performance by psi belief
combined.groupby('psi_01')['is_hit'].mean()
```

## Common Analysis Patterns

### Calculating Hit Rates

```python
# Overall performance
hit_rate = df['is_hit'].mean()

# By user
user_perf = df.groupby('user_id_hash')['is_hit'].agg(['mean', 'count'])

# Filter to experienced users (>100 trials)
experienced = user_perf[user_perf['count'] >= 100]
```

### Statistical Significance Testing

```python
from scipy import stats

# Test against chance (e.g., 20% for card tests)
n_trials = len(df)
n_hits = df['is_hit'].sum()

result = stats.binomtest(n_hits, n_trials, p=0.2, alternative='greater')
print(f"p-value: {result.pvalue}")
```

### Temporal Analysis

```python
# Performance over time
df['date'] = df['timestamp'].dt.date
daily_perf = df.groupby('date')['is_hit'].mean()

daily_perf.plot(title='Performance Over Time')
```

## Data Processing Flags

When generating parquet files, two important flags control output:

### Audit Mode (`--audit`)

```bash
python scripts/process_all.py --audit
```

**Adds columns:**
- `source_file` - Original filename
- `source_row_number` - Row number in source file

**Use when:**
- You need full data lineage
- Debugging data quality issues
- Tracing anomalies to source

**File size:** +10-15% larger

### PII Exclusion (`--exclude-pii`)

```bash
python scripts/process_all.py --exclude-pii
```

**Removes columns:**
- `user_id` (or `username` in users dataset)
- `email` (in users dataset)

**Retains:**
- `user_id_hash` / `username_hash` - for joining datasets

**Use when:**
- Preparing data for publication
- Sharing with external researchers
- Complying with privacy requirements

**File size:** Slightly smaller

### Combining Flags

```bash
python scripts/process_all.py --audit --exclude-pii
```

## Data Quality Notes

### Test User Filtering

All datasets automatically filter out test users:
- Usernames starting with `_test99` are removed
- Known cheaters removed (card dataset, 2001 only)

### Validation & Cleaning

Each dataset applies specific validation rules:
- Range checks (e.g., card responses must be 1-5)
- Type coercion (strings → numbers where appropriate)
- Timestamp parsing with timezone handling
- User ID length limits (max 30 characters)

Invalid rows are:
- Logged to `logs/latest/{dataset}_latest_errata.jsonl`
- Excluded from final parquet files
- Counted in processing statistics

### Schema Versions

Some datasets have multiple schema versions:

| Dataset | Versions | Change Date | Impact |
|---------|----------|-------------|--------|
| card | v1, v2 | 2006-01-10 | seed2 → trperrun rename |
| cardd | v1, v2, mixed | 2006-06-22 | Markov bits split |
| cardS | Mixed rows | N/A | Step vs completion rows |
| lottery | Mixed rows | N/A | Lottery vs immediate rows |

The processors **automatically detect and unify** these versions.

## Converting to PDF

These markdown files can be easily converted to PDF:

### Using Pandoc (Recommended)

```bash
# Install pandoc
brew install pandoc  # macOS
apt-get install pandoc  # Linux

# Convert single file
pandoc users.md -o users.pdf

# Convert all files
for file in *.md; do
    pandoc "$file" -o "${file%.md}.pdf"
done
```

### Using Python

```bash
pip install markdown-pdf

md2pdf users.md
```

### Using Online Tools

- [Markdown to PDF](https://www.markdowntopdf.com/)
- [CloudConvert](https://cloudconvert.com/md-to-pdf)

## Dataset Size Reference

Approximate file sizes (uncompressed, without audit mode):

| Dataset | Rows | Size | Join Key |
|---------|------|------|----------|
| users | ~50K | 5-10 MB | username_hash |
| card | ~15M | 500 MB - 1 GB | user_id_hash |
| cardd | ~5M | 200-400 MB | user_id_hash |
| cardS | ~10M | 300-500 MB | user_id_hash |
| rv | ~100K | 20-50 MB | user_id_hash |
| rvq | ~2M | 100-200 MB | user_id_hash |
| location | ~500K | 30-60 MB | user_id_hash |
| lottery | ~200K | 10-30 MB | user_id_hash |

**Note:** Actual sizes depend on raw data availability.

## Citation & Usage

When using these datasets in research, please cite:

> GotPsi Dataset (2000-2022). Cleaned and processed by [Your Lab/Name].
> Original data collected by GotPsi online psi experiment platform.

### Recommended Attribution

```
Data Source: GotPsi online psi experiments (2000-2022)
Processing: GotPsi Data Cleaning Pipeline v1.0
Access Date: [Your access date]
```

## Support & Questions

### Documentation Issues

If you find errors or gaps in the documentation:

1. Check the **🚧 OUTSTANDING** sections - known gaps are marked
2. Review the source processor code in `src/processors/`
3. Examine processing logs in `logs/latest/`
4. Open an issue with specific questions

### Data Quality Concerns

If you notice data anomalies:

1. Check if audit mode is enabled (has `source_file` column?)
2. Review errata logs: `logs/latest/{dataset}_latest_errata.jsonl`
3. Verify against raw source files
4. Report with specific examples (file, row number, issue)

## Roadmap

Future documentation improvements:

### High Priority
- [ ] Document RV attribute dimensions (attr_00 through attr_15)
- [ ] Clarify RV scoring algorithms (accuracy, relevance, form)
- [ ] Verify location dataset count offset removal
- [ ] Document card/cardd x, y, bias parameters

### Medium Priority
- [ ] Add complete statistical analysis cookbook
- [ ] Create data quality report per dataset
- [ ] Add visualization examples (plots, charts)
- [ ] Document temporal trends and patterns

### Low Priority
- [ ] Add cross-dataset analysis examples
- [ ] Create dataset comparison matrix
- [ ] Add machine learning examples
- [ ] Generate automated data profiles

## Contributing

To improve this documentation:

1. **Add details** - Fill in 🚧 OUTSTANDING items
2. **Verify accuracy** - Test code examples and formulas
3. **Add examples** - Contribute useful analysis patterns
4. **Report issues** - Note discrepancies or errors

## Version History

- **2025-01-09** - Initial documentation created
- Status: All 8 datasets documented with status/confidence indicators
- Coverage: Core structure complete, experimental details need investigation

---

**Last Updated:** 2025-01-09
**Documentation Format:** Markdown (PDF-ready)
**Target Audience:** Researchers and scientists analyzing psi experiment data