data/README.md

# GotPsi Experiment Data

Raw data files for the GotPsi psi-testing experiments. Each subdirectory
contains `.dat` files for one experiment, plus a `manifest.json` that
lists every file with its md5 checksum.

## Quick Start

### 1. Download archives from Hugging Face

Use `snapshot_download` to pull all `.tar.gz` archives into `data/archives/`:

```bash
python3 -c "
from huggingface_hub import snapshot_download
import os

snapshot_download(
    repo_id='instNoeticSciences/gotpsi_preprocess',
    repo_type='dataset',
    token=os.environ['HUGGING_FACE_TOKEN'],
    local_dir='data/archives/',
    allow_patterns=['*.tar.gz'],
)
"
```

This places `card.tar.gz`, `cardd.tar.gz`, etc. in `data/archives/`.

### 2. Extract archives

```bash
# Extract all experiments
python data/unpack.py

# Extract specific experiments only
python data/unpack.py card lottery

# Verify existing data integrity (no extraction)
python data/unpack.py --verify
```

Archives remain in `data/archives/` after extraction — rerun `unpack.py`
any time without re-downloading.

### Alternate: download and extract in one step

If you prefer to stream directly from the URL without saving archives locally:

```bash
python data/unpack.py --source https://huggingface.co/datasets/instNoeticSciences/gotpsi_preprocess/resolve/main
```

## Experiments

| Experiment | Files | Description |
|------------|------:|-------------|
| card       | 8,341 | Basic ESP card test -- standard 1-in-5 Zener card guessing with forced-choice trials. Schema changed in 2006 (seed2 to trperrun). |
| cardd      | 6,934 | Card Draw test -- Markov-influenced card selection where the RNG uses transition probabilities. Two format versions (pre/post 2006-06-22). |
| cards      | 7,627 | Sequential Card test -- participants guess card positions in a sequence. Mixed row format: step rows (4 cols) and completion rows (11 cols). |
| location   | 7,563 | Location (Remote Viewing Coordinates) -- participants guess geographic coordinates. Variable format: 13-14 columns (older files lack seed column). |
| lottery    | 2,450 | Lottery number prediction -- participants predict lottery draws. Mixed row format: lottery rows (9 cols) and immediate-draw rows (16 cols). |
| rv         | 6,468 | Full Remote Viewing -- dimensional attribute scoring (0-100) across 16 image attributes. 24-29 columns with multiple scoring methods. |
| rvq        | 6,248 | Quick Remote Viewing -- 5-choice image selection task, similar to Card but with photographs. 14-15 columns. |
| users      |     2 | User survey responses -- Psi Quotient and Hemispheric Dominance questionnaires plus demographics. Two files: users14.dat (2000-2015) and questions.dat (2014-2022). |

## Directory Layout

```
data/
├── archives/           # .tar.gz archives from HF (gitignored)
│   ├── card.tar.gz
│   ├── cardd.tar.gz
│   └── ...
├── card/               # extracted .dat files (gitignored) + manifest.json (tracked)
├── cardd/
├── cards/
├── rv/ rvq/ location/ lottery/ users/ lost_and_found/
├── pack.py             # archive experiments for distribution
├── unpack.py           # extract archives into experiment dirs
└── README.md           # this file
```

The `.dat` files and archives are gitignored. Only `manifest.json` files are tracked in version control.

### lost_and_found/

Holds orphan files that could not be assigned to a specific experiment during
data migration. Local only — not distributed or published.

## Manifests

Each `manifest.json` contains:

- **filename**: The `.dat` file name
- **md5_hash**: MD5 checksum for per-file integrity verification
- **archive_md5** (after packing): Checksum of the `.tar.gz` archive

Processors use manifests to discover files (no recursive glob needed).

## For Maintainers

Pack data for distribution:

```bash
# Create archives for all experiments (output to data/archives/)
python data/pack.py --output-dir data/archives/

# Pack specific experiments
python data/pack.py card lottery --output-dir data/archives/
```

After packing, upload the `.tar.gz` files to the HF dataset repo, then update
`download_url` in each manifest and commit.

## Dataset Nuances for Parquet Users

The processing pipeline (`scripts/process_all.py`) produces cleaned parquet
files in `output/parquet/`. A separate README ships with those files, but
the key nuances are summarized here for reference.

### card: `is_hit` and schema versions

The card dataset has two schema versions (`schema_version` column):

- **v2** (post-2006, ~63.7M rows): `is_hit` equals `target2 == response`.
  Chance hit rate: 20% (1-in-5).
- **v1** (pre-2006, ~25.8M rows): `is_hit` was precomputed by gotpsi using
  a harder criterion. Observed hit rate is ~4% (1/25). Do not recompute
  `is_hit` from `target2 == response` for v1 rows.

The `response` distribution is non-uniform (center bias toward card 3).
This is participant behavior, not a data quality issue. The `target2`
distribution is uniform, confirming RNG integrity.

### rv: Uses `start_time` / `end_time` instead of `timestamp`

### rv: Sentinel values in score columns

The `accuracy`, `relevance`, and `form` columns use -1 as a sentinel for
missing data (~2.1M rows for accuracy/relevance, ~24K for form). Filter
these before computing score statistics. `total_score` is clean (0-100).

### cardS: Null patterns by row type

The cardS dataset has two row types: step rows (4 columns of data) and
completion rows (11 columns). Step rows have NULLs in `response`, `steps`,
`response_array`, and `target_image` (~112M of 150M rows). This is
structural, not missing data.

### lottery: Null patterns by row type

Lottery rows (`row_type = lottery`) have NULLs in target/match columns
(~1,575 rows). Only immediate-draw rows carry the full 16-column schema.

### cardD: Null patterns by schema

`image_filename` is NULL for ~18.8M rows and `target_image` for ~41M rows,
corresponding to older schema versions that did not include these fields.

### users: `email` column sparsity

Only ~22K of ~284K users have an `email` value (from `questions.dat`,
2014-2022). The remaining rows have NULL emails (from `users14.dat`,
2000-2015). All non-null emails are HMAC-hashed.

### All datasets: Timestamps are America/Los_Angeles

All timezone-aware timestamp columns across every dataset use
`America/Los_Angeles`. Convert before comparing across external sources.

### users: PII considerations

The `username` and `email` columns are HMAC-hashed, but `city`, `state`,
`coordinates`, and `country` contain raw self-reported location data.
Public releases should evaluate whether to include these columns.