ctoon, measured honestly

Two separate numbers, because they answer different questions. Success is the share of corpus files a library converted without an error. Correct is, of the outputs it actually produced, the share that were right, so “98% success, 0% correct” reads as what it is: a library that never complains and never gets it right. Correct means: TOON that a library produced is decoded with the official reference decoder (@toon-format/toon), not with the library's own, because a lenient decoder hides encoder bugs from its own round trip; and JSON that a library produced is parsed back. Either way the result has to equal the original document. Equality is the spec's JSON-model equality: numbers compare as IEEE-754 doubles (the spec lets an encoder fall back to the host's numeric approximation), objects are unordered maps, arrays are ordered.

MB/s and docs/s count only correct conversions, so a wrong answer never looks like a fast one; the time is the time of all attempts, as before. TOON → JSON decodes the TOON that ctoon produced for each file; if that shared input is itself missing or wrong, the file is left out of every library's TOON → JSON row (the encoder is charged for it under JSON → TOON), so no decoder is blamed for someone else's output.

Every "Key order preservation" table below round-trips this exact JSON through that library and checks whether the keys come back in this same order. The sample is deliberately non-alphabetical, with a nested object and an array of objects, so sorting or hash-based reordering would show up.

{"zebra": 1, "apple": 2, "mango": 3, "nested": {"beta": true, "alpha": false}, "list": [{"z": 1, "a": 2}, {"z": 3, "a": 4}]}

Average across the whole corpus

JSON → TOON

Throughput vs. input size: one line per library, marker per size bucket.

TOON → JSON

Throughput vs. input size: one line per library, marker per size bucket.

JSON → TOON → JSON

Throughput vs. input size: one line per library, marker per size bucket.

TOON → struct (load)

Throughput vs. input size: one line per library, marker per size bucket.

struct → TOON (dump)

Throughput vs. input size: one line per library, marker per size bucket.

Key order preservation

Round-trips the fixed sample above through each library and checks whether keys come back in the same order they went in.

Average across the whole corpus

Key order preservation

Round-trips the fixed sample above through each library and checks whether keys come back in the same order they went in.

Average across the whole corpus

Average across the whole corpus

Average across the whole corpus

Average across the whole corpus

Average across the whole corpus

Average across the whole corpus

Average across the whole corpus

Recorded automatically by system_info.py at run time, not filled in by hand. Numbers above are only meaningful relative to each other on this specific machine; they won't transfer directly to yours.