Methodology & limitations

How we measure Overwrld, what each number means, and what we will not claim. We would rather lose a sale than publish a number that does not survive scrutiny.

The enwik9 result, precisely

The public board is the authority. On August 30, 2026, program altxs 1.0.0 reached #1 by official total-size ranking.

• 93,434,410 is the payload.
• 13,490,401 is the decompressor.
• 106,924,811 is the total the board ranks on.
• 91.7 MB is a later model. Later model. Not submitted. Not listed on any public table.

#2 on the same table is nncp v3.2 at 107,261,318. Next row on the same official total-size ranking. Hutter Prize scoring is different and is not claimed here.

What we measure

We report total bytes: the compressed output plus the decompressor needed to read it back, plus every dictionary, table and piece of side information it needs. Counting the decompressor is the convention used by the standard compression benchmarks, and it is the only honest way to score a method that ships its own decoder. We always name the baseline configuration we compare against, and we report on held-out data.

What every published result includes

• total counted size: payload plus decompressor plus dictionaries plus side information, in bytes.
• bits per byte on the named corpus.
• encode throughput and decode throughput, with the hardware stated.
• peak decode memory.
• verification: SHA-256 equality of decoded output against the full original bytes.
• baselines: tuned configurations, never defaults.

What we do not claim

• We reached #1 by official total size on the LTCB on August 30, 2026. Boards move.
• We do not claim the unpublished later-model figure has been independently verified.
• We do not claim a video or AI-workload result. Those have not been measured under a protocol we would defend.
• We do not claim to beat the information-theoretic limits of lossless compression.
• We do not claim a fixed ratio that holds across arbitrary files or sizes.
• The listed altxs 1.0.0 decode is not the commercial product. Decode cost on the board is a diligence point, not a feature we hide.

Limitations

• Encoding spends significant compute, so the method fits encode-once, decode-cheap data first.
• A result counts only when the decoded output matches the input byte-for-byte.
• A record on a text corpus is evidence about that corpus. Generalization to your data is a measurement we run with you, not an assumption.

Reproduce

The public row is on Matt Mahoney's LTCB. The benchmarks page restates that row with payload and total kept separate.

Benchmark: Large Text Compression Benchmark (enwik9), lossless, decompressor counted. Program altxs 1.0.0. We reached #1 on August 30, 2026 by official total-size ranking. Payload 93,434,410 + decompressor 13,490,401 = 106,924,811 total. The 91.7 MB figure is a later model, not on the board, and is not the rank.