T24d: final whole-card consistency pass

README.md

@@ -64,6 +64,17 @@ Built from arXiv's OAI-PMH `arXivRaw` interface (the only endpoint exposing
 full per-paper version history), cross-referenced against the GCS mirror, the
 S3 source tars, and direct fetches from arxiv.org.
 
+> **This repository is private, so every example on this page needs an access
+> token that has been granted access to it.** Anonymously — tested
+> 2026-09-15 — `load_dataset` raises `DatasetNotFoundError`,
+> `huggingface_hub` raises `RepositoryNotFoundError` and DuckDB's `hf://`
+> paths return **HTTP 401**; all three fail on the repository listing, before
+> a single byte of Parquet is touched, and none of them says "private" in so
+> many words. Run `hf auth login` (the old `huggingface-cli` is gone), or set
+> `HF_TOKEN`, or pass `token=` to the call. In DuckDB it is
+> `CREATE SECRET (TYPE huggingface, TOKEN '…')` once per session. Nothing
+> else about the examples changes.
+
 ## Snapshot and coverage
 
 This is a **one-off point-in-time snapshot**, not a live mirror.
@@ -556,9 +567,10 @@ replaced it with is exactly what is published here.
 own.** The withdrawal-marker caveat sets out the whole of it, corpus-wide:
 11,497 files whose entire content is that 12-byte marker, 1,872 more
 `paper_text` rows that add a reason under it, and 97 papers whose whole
-submission is such a note under a name `paper_text` never reads — of which the
-seven `removed.txt` and two `auto-ignore.txt` papers above are nine. (The other
-two notes here, `hep-th/9502028`'s and `gr-qc/9410009`'s, sit beside a couple of
+submission is one such note — 94 of them under a name `paper_text` never
+reads, the other three ending `.tex` — of which the seven `removed.txt` and
+two `auto-ignore.txt` papers above are nine. (The other two notes here,
+`hep-th/9502028`'s and `gr-qc/9410009`'s, sit beside a couple of
 leftover figures and so fall outside that sweep's single-file definition, which
 is why it is a floor.) What singles *these* out is that nothing else survived
 beside the note. The 11,369 bare stubs are named
@@ -971,10 +983,10 @@ census, not a sample:
 | the rest — OLE, AppleDouble, ELF, SQLite, FITS, zip, xz, TrueType, PCL, 154 part-text files, 25 empty ones | 717 | 211,672,079 |
 | **total** | **273,784** | **7,774,078,264** |
 
-So the bucket is overwhelmingly toolchain and notes: **88.9% of it is text or
-TeX fragments**, and only **1.68% is a whole document**. That 1.68% still has a
-consequence, because a document no extension test can see is text `paper_text`
-never collected.
+So the bucket is overwhelmingly toolchain and notes: the first two rows are
+**88.8% of it — 243,231 files of plain text or a TeX fragment** — and only
+**1.68% is a whole document**. That 1.68% still has a consequence, because a
+document no extension test can see is text `paper_text` never collected.
 
 **1,663 papers are in `paper_text`'s shortfall because of this bucket alone.**
 Take the extensionless files belonging to a paper with *no* `paper_text` row —
@@ -2118,7 +2130,7 @@ while what remains sits at 3.86:1. Changing basis changes the answer here
 exactly as it does for `.bib`. And note where deduplication lands beside the
 other filter measured for this config: dropping every `.bib` over 10 MB saves
 67,161,051,372 published bytes against deduplication's 65,950,601,244 — near
-enough the same number, because those over-10 MB `.bib` blobs are 71.0% of the
+enough the same number, because those over-10 MB `.bib` blobs are 70.4% of the
 config's duplication to begin with.
 
 `source` was not re-encoded; reading and rewriting 6.51 TB of Parquet to find
@@ -2406,7 +2418,7 @@ and two of its columns can be non-null and still empty.
 | `path` | relative, meaningful to a consumer — never a local filesystem path |
 | `path_encoding` | `utf8`, or `percent` for the few genuinely non-UTF-8 submission paths (recover with `urllib.parse.unquote_to_bytes`) |
 | `size` / `sha256` | of the **published** bytes |
-| `stored_as` / `stored_size` / `stored_sha256` | the original bucket object, where it differed — i.e. the `.ps.gz` a `.ps` was decompressed from. NULL otherwise |
+| `stored_as` / `stored_size` / `stored_sha256` | the original bucket object, where it differed — e.g. the `.ps.gz` a `.ps` was decompressed from, which is 2,070,188 of the 2,168,515 such objects and not all of them. NULL otherwise |
 
 **`paper_text`** — one row per paper, TeX resolved to readable text.
 `paper_id`, `text`, `main_file`, `resolution`, `n_tex_files`, `n_files_used`,
@@ -2495,11 +2507,12 @@ is counted once in every row it has a file in.
 particular paper — `tex`, `bbl`, `ltx`, `tikz` — are 2.2% to 14.9% duplicate.
 The other nine run 73.3% to 95.5%, because they are not the paper, they are
 *packages* shipped alongside it. `cls`, `sty`, `bst`, `clo`, `cfg`, `ins`,
-`dtx` and `def` together are 133,331,761,138 bytes of which **92.3% is a copy of something else
-in the config**, and the most-copied blob in `latex` is `IEEEtran.cls`, 281,957
-bytes appearing **60,967 times across 58,901 papers** — 17.19 GB of one class
-file. Across the whole config, 9,267,670 of the 12,348,082 rows (75.06%) are
-distinct blobs and **451,848,934,278 bytes (55.37%) are duplicates**.
+`dtx` and `def` together are 133,331,761,138 bytes of which **92.3% is a copy
+of something else in the config**, and the most-copied blob in `latex` is
+`IEEEtran.cls`, 281,957 bytes appearing **60,967 times across 58,901
+papers** — 17.19 GB of one class file. Across the whole config, 9,267,670 of
+the 12,348,082 rows (75.06%) are distinct blobs and **451,848,934,278 bytes
+(55.37%) are duplicates**.
 
 **One bibliography is 42.7% of this config.** 429.6 GB of
 BibTeX is not a million papers with long reference lists. 8,633 rows are over
@@ -2527,17 +2540,19 @@ number.
 
 **One blob is not the story.** The most-copied is the 43,619,593-byte copy the
 `paper_text` section describes, appearing 2,095 times across 2,077 papers under
-42 distinct path spellings corpus-wide — by itself 91,339,427,742 duplicate
-bytes, **11.2% of the config's 816,110,470,780 raw bytes** and 20.2% of its
-451,848,934,278 duplicate ones. Both denominators are used in this paragraph,
-so every figure in it says which. The twenty most-copied `.bib` blobs are *all*
-`anthology.bib`, and together they are 230,829,650,496 duplicate bytes:
-**51.1% of every duplicate byte in `latex`**. All 1,028 of the over-10 MB
-blobs reach 320,985,600,874 — 99.4% of `.bib`'s duplication and 71.0% of the
-config's. The largest row in `latex`, the 72.5 MB `anthology.bib` of
-`2508.04390` noted above, is one of them. Non-anthology giants exist and are a
-smaller version of the same thing: 316 rows over 10 MB, 9,804,072,995 bytes,
-192 blobs, led by `crypto.bib` at 112 copies and 3.57 GB.
+42 distinct path spellings **inside this config** — corpus-wide it is 2,099
+copies across 2,081 papers under 46 spellings — by itself 91,339,427,742
+duplicate bytes, **11.2% of the config's 816,110,470,780 raw bytes** and 20.2%
+of its 451,848,934,278 duplicate ones. Both denominators are used in this
+paragraph, so every figure in it says which. The twenty most-copied `.bib`
+blobs are *all* `anthology.bib`, and together they are 230,829,650,496
+duplicate bytes: **51.1% of every duplicate byte in `latex`**. All 1,028 of
+the over-10 MB blobs reach 318,277,378,364 — 98.5% of `.bib`'s duplication
+and 70.4% of the config's. The largest row in `latex`, the 72.5 MB
+`anthology.bib` of `2508.04390` noted above, is one of them. Non-anthology
+giants exist and are a smaller version of the same thing: 316 rows over
+10 MB, 9,804,072,995 bytes, 192 blobs, led by `crypto.bib` at 112 copies and
+3.57 GB.
 
 The concentration is extreme. **8,441 papers — 0.295% of the 2,860,286 in this
 config — ship a `.bib` over 10 MB, and those papers account for
@@ -3053,10 +3068,9 @@ and `files` carry no content of their own and are not candidates — but
 **`sample` does, and it sits outside that count**: one shard, 991 rows. Nothing
 there is unverified: each of its 991 rows matches a `paper_text` row on
 `paper_id` *and* `text_sha256` — 991 of 991, checked for this sentence — so the
-pass
-hashed those strings already, under their own `paper_id`. But "every blob in
-this dataset" is one config wider than the number beneath it, and this is the
-config it does not count.
+pass hashed those strings already, under their own `paper_id`. But "every blob
+in this dataset" is one config wider than the number beneath it, and this is
+the config it does not count.
 
 **Two raw-content totals appear in this section and they differ by exactly one
 config.** **23.63 TB** is 23,634,021,188,971 raw content bytes — `source`
@@ -3257,12 +3271,15 @@ content inside them.
   carry a PostScript name: `.ps` 125,966, `.eps` 90,413, `.pstex` 1,032,
   `.epsi` 344, `.epsf` 207, `.mps` 114, `.eps_col` 104, `.epsx` 52, `.ps_col`
   39, `.ps~` 38, `.eps2` 34, `.cps` 31, `.psc` 27, `.ps_tex` 26, and numbered
-  variants. The other 75,071 do not: `.png` 28,742, `.jpg` 22,992, `.gif`
-  15,733, `.jpeg` 1,659, `.tex` 860, 679 with no extension at all, `.prn` 229,
-  `.xcp` 210, `.pfb` 186, `.bb` 91, `.tif` 57, `.enc` 35, `.pro` 31, `.ai` 24,
-  `.pdf` 21, `.tikz` 19, `.tiff` 12, and a tail of 2,861 objects spread over
-  1,214 further spellings — mostly bare numbers like `.1` or `.31`, which are
-  simply the last dotted component of a figure's name.
+  variants — those fourteen spellings are 218,427 of the 218,546, so the
+  variants are the other 119. The other 75,071 do not: `.png` 28,742,
+  `.jpg` 22,992, `.gif` 15,733, `.jpeg` 1,659, `.tex` 860, 679 with no
+  extension at all, `.prn` 229, `.xcp` 210, `.pfb` 186, `.bb` 91, `.tif` 57,
+  `.enc` 35, `.pro` 31, `.ai` 24, `.pdf` 21, `.tikz` 19, `.tiff` 12, and a
+  tail of 3,491 objects spread over 1,214 further spellings — mostly bare
+  numbers like `.1` or `.31`, which are simply the last dotted component of a
+  figure's name. Those sixteen names and the extensionless rows are 71,580 of
+  the 75,071; the tail is the rest.
   **8,872 of those 1,953,094 (paper, version) pairs hold assets and no render
   at all** ��� the difference between the pair count and the 1,944,222 that do
   have a `<paper_id>v<n>.ps` — and between them they carry 29,023 of the
@@ -3601,15 +3618,16 @@ content inside them.
   of withdrawal**, though: 154 of the 13,302 run past ten thousand characters
   and are ordinary papers whose author put the marker at the top of a file.
 
-  **The same note under a name `paper_text` cannot see: 97 papers.** Sweeping
-  `files` for note-shaped basenames, and separately taking every paper whose
-  entire source package is a single file of 8,192 bytes or less that is *not*
-  named `<paper_id>.tex`, turns up 97 papers whose whole submission is one such
-  note — 96 of them hold nothing else at all, and `2603.12203` holds the note
-  beside arXiv's own 298-byte `00README.json`. All 97 were read back out of the
-  published Parquet and re-hashed against `files` — **0 size and 0 digest
-  mismatches** — and every one of them opens with `%auto-ignore`, four after a
-  little leading whitespace. They wear ten different names:
+  **The same note under a name `paper_text` mostly cannot see: 97 papers.**
+  Sweeping `files` for note-shaped basenames, and separately taking every
+  paper whose entire source package is a single file of 8,192 bytes or less
+  that is *not* named `<paper_id>.tex`, turns up 97 papers whose whole
+  submission is one such note — 96 of them hold nothing else at all, and
+  `2603.12203` holds the note beside arXiv's own 298-byte `00README.json`.
+  All 97 were read back out of the published Parquet and re-hashed against
+  `files` — **0 size and 0 digest mismatches** — and every one of them opens
+  with `%auto-ignore`, four after a little leading whitespace. They wear ten
+  different names:
 
   | the note is called | papers | |
   |---|---|---|