fe2o3_austenite/tools/bench/corpus/parity.py
2 operations, oldest first
created by r1870400018:72510, which is this file's identity for as long as the history lasts, whatever it is later renamed to
read the file · who wrote it · compare two states
An operation carries no clock of its own. Each row is therefore dated by the first mark written after it by the same replica, which is one copy of the repository, and that mark is where the command that wrote the operation ended.
+ signed and verified, ? signed by a key this forge does not carry, - unsigned
| Operation | Author | What | When | |
|---|---|---|---|---|
| + | r1870400018:72510 | replica 1870400018 | created fe2o3_austenite/tools/bench/corpus/parity.py | 40 days ago under "fe2o3_austenite bench: render parity against typst over the whole corpus -- parity.py copies each document once and compiles that copy with Austenite's evaluator (--eval) and typst 0.15.1, each capped (120 s, 3G), and writes one JSONL row: both engines' exit codes and first error line, page counts, pdftotext -layout text (runs of spaces collapsed, Unicode Cf characters dropped, page breaks as line breaks) with the count of differing lines, every page to 20 at 50 dpi compared with compare -metric AE -fuzz 10% (worst page's fraction), a verdict (equivalent, near, differs, aus_fails, typst_fails) and for aus_fails a cause key naming the construct (unknown function or field, missing method, type mismatch, unsupported bibliography style, panic, timeout) with the package or document it sits in. Only a differs row keeps its worst page's PNG pair. report_parity.py prints counts and shares, the causes ranked by documents broken, and a sample of differs rows. Smoke, 20 Universe templates: 2 equivalent, 2 near, 4 differs, 11 aus_fails, 1 typst_fails; three rows checked by eye against their PNGs (tatras-ieee equivalent, mahou-cv near with its bars unfilled, homiework differs on justification, margins and highlighting)" |
| + | r1870400018:72511 | replica 1870400018 | inserted 14.2 KiB | 40 days ago under "fe2o3_austenite bench: render parity against typst over the whole corpus -- parity.py copies each document once and compiles that copy with Austenite's evaluator (--eval) and typst 0.15.1, each capped (120 s, 3G), and writes one JSONL row: both engines' exit codes and first error line, page counts, pdftotext -layout text (runs of spaces collapsed, Unicode Cf characters dropped, page breaks as line breaks) with the count of differing lines, every page to 20 at 50 dpi compared with compare -metric AE -fuzz 10% (worst page's fraction), a verdict (equivalent, near, differs, aus_fails, typst_fails) and for aus_fails a cause key naming the construct (unknown function or field, missing method, type mismatch, unsupported bibliography style, panic, timeout) with the package or document it sits in. Only a differs row keeps its worst page's PNG pair. report_parity.py prints counts and shares, the causes ranked by documents broken, and a sample of differs rows. Smoke, 20 Universe templates: 2 equivalent, 2 near, 4 differs, 11 aus_fails, 1 typst_fails; three rows checked by eye against their PNGs (tatras-ieee equivalent, mahou-cv near with its bars unfilled, homiework differs on justification, margins and highlighting)" |