Put access, config selection and limits first; consolidate repeated audit prose. Preserve measured tables and the crawl-rate disclosure. Correct overclaims, example predicates and licence-record scope. No dataset files or config metadata changed.
Every size and percentage in the card enumerated and recomputed against the published parquet.
Convention stated at the config table: `size` is compressed Parquet on disk; the other basis is
always named 'raw content bytes'.
Corrected:
- 'taking latex beside source costs 816.11 GB of transfer' -> 160.23 GB of Parquet to re-receive
816.11 GB of raw content (816.11 GB is not a transfer figure).
- 'you download the 22.57 TB either way' -> 15.84 TB of compressed Parquet (22.57 TB is raw).
- paper_text table: 247,014,274,151 UTF-8 bytes labelled 'what the download costs' -> it is what
memory costs; the download is 70 GB of Parquet, 3.5:1.
- '37% of the pdf config by bytes is superseded' -> 36.6% of raw content bytes, exactly
3,798,892,308,546 of 10,371,932,999,324 (37.03% by rows); both sites harmonised.
- '20,652 papers (0.659%)' had no denominator: it is 0.659% of the 3,131,959 papers holding a
PDF, 0.656% of all 3,148,796. Both sites.
- withdrawal caveat: 'the first two carriers…
Fixes statements that were false or that contradicted other statements on the card.
- 13,898 file-less papers: replace the 'withdrawn submissions and papers never mirrored'
gloss with the measured fact (all first posted 2026, earliest 2026-03-18, 12,965 on
2026-08-14+, 14,788 versions rows with no artifact, 16 metadata-only). Resolves a
four-way contradiction with the coverage section, the floor section, the withdrawal
stub proof and the marker account.
- withdrawal/removal now reads 'one event in five carriers' in both places that count them.
- ps caveat: separate the two scopes (159 pairs span 137 papers; 74 papers hold no PDF at
any version; 63+74=137) and state that the 99.99% redundancy is coverage, not bytes -
pdf and ps share zero blobs.
- truncated-.pdf headline: 752 rows, 750 distinct objects.
- 2401.18030v3's 267-byte HTML redirect page is now sourced: the reconstructed body
reproduces sha256 2e066757... and md5 db732060... (card and LICENSE).
- .git: 8,473 file…
Card T11b follow-up: harmonise the 1810.03044 sentence with the withdrawal caveat's own closing line, so 'four carriers plus one paper that left an empty file' and 'a fifth carrier' read as the same statement rather than two counts. Text only; no figure changed.
Card T11b: the degenerate small rows - a new Known-caveats section on the bottom of the size distribution. Zero-byte rows confirmed exactly: 242, every one in `source`, 33 papers, one digest (e3b0c442, the empty string); pdf/ps/html 0; latex 17 across 7 papers; paper_text and sample have no empty row. Counted twice - over all 54,597,326 rows of `files` and over each content config's own size/sha256 - and the passes agree to the row. PROVENANCE SETTLED, IT IS arXiv's: all 242 re-read out of the original unmodified arXiv_src_YYMM_NNN.tar archives - 241 already zero bytes as members of the paper's submission tar, and for 1810.03044 the payload member arXiv ships (1810/1810.03044.gz) is itself zero bytes. 0 differed, 0 missing. Classification: 122 are Python __init__.py package markers, 34 are one paper's macOS Photos-library sidecars, 86 the ordinary kind. Only ONE paper (1810.03044, an arXiv admin removal for copyright infringement) loses everything; the other 32 all hold their PDF in `pdf` and a paper_text row. Smallest row above zero per config read from the published Parquet and re-hashed: source 1 B, ps 3 B (= `%!` + newline), html 9 B (= arXiv's 00TOPLEVELFILE, `TOC.html`), pdf 322 B. The ps tail is entirely shipped assets - arXiv's 1,944,222 own renders bottom out at 9,241 B - and is 46 .bb bounding-box sidecars, 6 LaTeX files opening `%!TEX root`, and 10 bodyless PostScript headers. All 15 pdf rows under 4,096 B are structurally complete PDFs; quant-ph/0504128's 322 B declares zero pages against a `9 pages` comments field, confirmed independently. New finding: 60 files in `source` across 40 papers hold nothing but an HTTP error body (45 x `Not Found`), named .pdf/.png/.eps/.bbl/.jpg. Tail thresholds published per config. Boundary against the NUL population restated and cross-referenced both ways.
Card T7c: corpus-wide sweep for NUL-filled files - 913 files, 215,358,402 bytes, 189 papers, 719 blobs, every one in `source` and none in pdf/ps/html. Found exactly by comparing each row's sha256 against the digest of a zero-filled buffer of its own length (3,533,564 distinct sizes), then all 913 read back from the published Parquet across 178 shards - 0 length, 0 digest, 0 non-zero-byte mismatches. Provenance settled: 0 of 7,217,247 gcs objects and 0 of 6,475 web_scrape PDFs are NUL; all 913 were already entirely NUL inside arXiv's original S3 source tars (913/913, 189/189 papers, 0 differed). 71 are a valid EMPTY tar archive and 42 are Krita .defaultpixel records whose correct content is zero, leaving 800 genuinely missing. No power-of-two truncation signature. latex carries 86; paper_text 0, tested against the zero-digest of every length 0-46,000,000. Boundary against zero-byte rows (242 in source, 17 in latex) stated.
Card T3c: merge the three scattered withdrawal-marker families into one account (11,497 bare-marker files by digest, 1,872 marker+reason paper_text rows, 97 note papers under ten filenames invisible to paper_text, 12 notices shipped as PDFs, 1 zero-byte erasure; 83 arXiv-admin removals vs 14 author withdrawals), and add the .cry caveat: 7,387 XOR-encrypted TeX files in 4,393 papers, length-preserving, per-batch keystream, mechanism partly established and the rest left unstated
LICENSE T23b: complete the 54-defective-PDF arithmetic (42 identical + 3 repaired + 8 never-PDFs + 1 never-published = 54; damage is 45), and correct the gzip-wrapper claim (2,168,515 objects are not all named .ps.gz nor all PostScript)
T18: publish exact character and UTF-8 byte totals for paper_text, the per-paper character distribution, and how to rescale the token estimate to any tokenizer
Card T17: PDF provenance (6,475 web_scrape vs 4,967,481 gcs) and which gaps are permanent - repair-pass tally, era recovery, gap-by-era x attempted table, 144-URL live re-test
Card T14: dedup budget per config and corpus-wide, what the most-copied blobs are, the GROUP BY sha256 recipe, and the measurement that it saves storage but not download
Card T12 fix: one sampled .eps counted as a mismatch was headerless PostScript - .eps 0.10%->0.09%, overall 340->339, and note raw-operator EPS in the sniffing traps
Card T13: latex composition by extension — .bib is 52.6% of raw bytes and 42.7% of the config is the ACL Anthology; measured tex-only republish cost (63,724,684,922 B) and why WHERE ext='tex' prunes only 0.0037% of the download
Card T12: extensions are claims, not facts - complete .pdf census (13,035 non-PDFs, 13,013 in source), published-parquet verification, residual-22 resolution, unbiased cross-extension sample, and a magic-number remedy
Card T11: irreducible per-row memory floor - per-row size census for all five content configs, the six rows over 1 GB verified from published Parquet, NUL check on every row over 64 MiB, and measured peak RSS for the worst row
Card T7b: source composition table made mutually exclusive (.gif was in two rows); extension sets stated, remainder row added, rows now sum to the config
Card T7: state the definition behind four trivia counts (__pycache__ 1142->1141, video extension set, anthology.bib occurrences vs papers, nested-archive extension set + full magic-number census)
Card T6: ps config composition measured correctly - 1,944,222 of 2,237,839 objects are the version PostScript, 293,617 are shipped assets; full extension histogram plus 1,431-object magic-number sample
