Add jakedgy codec: 6,402,499 bytes

[jakedgy]

Summary

This codec was developed as an experiment with Claude Code's Ralph Loop - an automated iteration system that explores solution spaces. The goal was to see how far automated exploration could push compression on this problem.

The Ralph Loop iteratively tested various compression techniques, keeping what worked and discarding what didn't. This submission represents the result of that exploration rather than a carefully hand-crafted solution.

Result: 6,402,499 bytes (~6.40 MB) on the 1M event dataset.

Techniques Discovered by the Loop

• 2-bit category encoding for ID deltas - Common zigzag values (0, 2, 4 representing deltas of 0, +1, +2) encoded inline; others as varint exceptions
• 2-bit category encoding for timestamp deltas - Common values (0, 1, 2 seconds) encoded inline
• Columnar layout with per-column zstd level 22 compression
• Timestamp sorting within row groups for optimal ID delta encoding
• 140K row group size (empirically tuned)

What Didn't Work

The loop also tried several approaches that made compression worse:

• RLE for sparse columns (zstd already handles this)
• Delta encoding for repo indices (no locality in alphabetical order)
• 3-bit category encoding (overhead > benefit)
• Sorting by (type, timestamp) (breaks ID delta locality)
• Skipping zstd for category-encoded columns

Validation

Tested on 5 different GitHub Archive datasets spanning March 2023 to January 2025 to verify the techniques generalize and aren't overfit to the training data.

🤖 Generated with Claude Code using the Ralph Loop

[agavra]

Thanks for the submission! I guess this proves that our AI overlords are pretty darn good 😨 now I'm extra hoping that someone comes in and improves on this with manual human intuition haha!

Just a format issue causing check failure. I'll merge and address when I update the leaderboard.

Confirmed with CI/CD:

┌────────────────────────┬────────────────┬────────────┐
│ Codec                  │           Size │ vs Naive   │
├────────────────────────┼────────────────┼────────────┤
│ Naive                  │    210,727,389 │   baseline │
│ jakedgy                │      6,402,499 │     -97.0% │
└────────────────────────┴────────────────┴────────────┘