Changelog

[Unreleased]

Every format page and testing guide has a Markdown copy at its URL plus /index.md, for example /pdf/index.md, linked from the page and described in llms.txt. The pages' structured data now gives each file's size and SHA-256.
Twelve format pages (mp3, mp4, parquet, docx, pptx, mkv, png, jpg, pdf, csv, zip, wav) have titles and descriptions that name what sets their files apart, for example Sample MP3 test files: L/R channels, ID3v2 tags.
The sqlite, tiff, avi and eml pages are selection guides too, and their titles and descriptions name what sets their files apart, for example Sample SQLite database files for testing: foreign key. The status page now says which agent the site's own checks send.
Every page's meta description is now 120 to 155 characters. The other 65 format pages describe their file count and size range (a single file reads 1 SQL test file), and edge reads 21 edge-case test files.
The privacy notice names the authority responsible for loremfile.dev beside the reader's own, and states that the pages load no third-party resource.

mkv/vp9-opus-720p-5s.mkv: VP9 video with Opus audio in general Matroska, the codecs of webm/720p-5s-vp9.webm in the container that also holds H.264. Published from the bytes CI built, because VP9 and Opus encoders do not reproduce across processors.

Four new files each on three pages that held two or three, and three on a fourth. wav/: mono, 8-bit (which WAV stores unsigned), 32-bit float behind format tag 3, and a LIST/INFO chunk that moves the samples to byte 94. mkv/: two audio tracks (440 Hz and 880 Hz), an SRT subtitle track, and three chapters. parquet/: zstd, gzip, ten row groups, and nested types (structs and a map). pptx/: 4:3 slides, a native table, a native chart, and a hidden slide.
New measured props, existing bytes unchanged: info_tags for WAV; track_tone_hz, scodec and chapters for video, only where present; column_types for every Parquet file; tables, charts and hidden_slides for every PPTX file.

Every file with audio now records the tone it carries, measured per channel from the decoded audio as tone_hz: 440 Hz on 45 files, 440 Hz left and 880 Hz right on mp3/stereo-lr-5s.mp3, and nothing on the silent mp3/silence-5s.mp3. The chained Ogg file records each chain's tone. The files are unchanged.

Four new files each on three pages that held one. ogg/: FLAC in Ogg, mono 22.05 kHz Vorbis, Vorbis with four comments and a cover that spans several header pages, and two chained streams. tiff/: LZW, three pages, 16-bit big-endian grayscale, CMYK. arrow/: zstd-compressed buffers, the IPC stream format (people-1000.arrows, served as application/vnd.apache.arrow.stream), ten record batches, and nested and logical types.
New measured props, existing bytes unchanged: TIFF compression, byte order and bits per sample; Ogg chains, header pages, comments and cover; Arrow IPC format, buffer compression, column types and null counts.

Format titles without em dashes: all 79 that carried one, each of which reaches its page's <title>, og:title and Dataset.name. 14 format descriptions and 17 fixture descriptions follow, reworded with the same meaning. The files are unchanged.
jpg/cmyk-640x640.jpg replaces jpg/cmyk-640x480.jpg, which has measured 640x640 since it was published under a 640x480 name. The old URL keeps serving the same bytes, and its entry now carries supersededBy and deprecated.

m4a/aac-64k-44k-30s.m4a: AAC at 64 kbps and 44.1 kHz, beside the 128 kbps, 48 kHz file.
opus/cbr-32k-30s.opus: Opus at a constant 32 kbps.
wav/sine-440hz-3s-48k-24bit-stereo.wav: 24-bit PCM behind the 68-byte extensible header.
md/long-with-toc-and-footnotes.md: a long document whose contents links and footnotes all resolve.
docx/anchored-image-with-alt-text.docx: one JPEG floating in an anchor, with alt text.
hls/multi-bitrate/master.m3u8: a master playlist over a new 360p rendition in hls/360p-10s/ and the existing 720p one, its BANDWIDTH, RESOLUTION and CODECS measured from the segments.
New measured props, existing bytes unchanged: bitrate_kbps from the audio packets and bit_depth for PCM on every audio file, the WAV header's format and size, the video profile and H.264 level, inline and anchored images and alt text on every DOCX, and headings, anchor links and footnotes on Markdown.
Four audio descriptions now say what was measured: opus/30s.opus averages 114 kbps, not the 96 it was encoded for; aac/adts-30s.aac is 130.6 kbps with its frame headers; ogg/vorbis-30s.ogg no longer names an encoder quality; and the 3-second MP3's extra 24 ms is one 1152-sample frame.

contains_text in the manifest: three to five strings each text-bearing PDF, DOCX, PPTX, XLSX, EPUB, HTML and Markdown file is known to contain, so a parser test can check what it read, not only that it read something. Compare by containment. The files are unchanged.

jpg/exif-orientation-8-640x480.jpg: an EXIF rotation case for viewers that ignore orientation tags.
mp4/non-faststart-720p-5s.mp4: metadata at the end, so sequential playback waits for the download.
zip/zip64-70000-empty-files.zip: 70,000 empty members force ZIP64 central-directory handling.
pdf/scanned-1page.pdf: readable raster text with no text layer, so extraction returns an empty string.
docx/with-tracked-changes.docx: an insertion, a deletion and one comment for revision-aware extraction.
txt/utf16be-no-bom.txt: UTF-16 big-endian text without a byte-order marker for encoding detection tests.

heic/640x480.heic: the colour-bar test card as HEIC, the format iPhones capture photos in, for testing upload forms and converters. A new format.
csv/people-10-crlf.csv: the ten people rows with every record ending in CR LF.
pdf/form-fields-1page.pdf: a fillable form of four text fields and two checkboxes, half of them filled.
xlsx/3sheets-with-formulas.xlsx: three sheets, with summary formulas that read the other two through a quoted sheet name, results cached.
New manifest props on these entries: the HEIC container's brands and primary item, form_fields and filled_fields for a PDF form, and cross_sheet_formulas for workbooks.

Two polyglots: edge/pdf-zip-polyglot.pdf and edge/gif-zip-polyglot.gif are each valid as two formats at once, for testing what your type detection does when the bytes do not decide. Scanners and proxies sometimes refuse polyglots — the /edge page says so.

Three of the broken files describe their damage in terms of their own bytes rather than one parser's error message. The files are unchanged.

csv/people-10-quoted-commas.csv: ten people rows with a comma inside a quoted field, for parsers that split on commas.
The 19 broken files now say what is wrong with them, in manifest.json and on the site: one sentence about the damage, a valid file of the same type to compare against, and whether a reader must fail, may recover, or may do either.

5 more media files: H.264 MP4 at 1080p for 10 seconds, at 10 MB and at 50 MB, a 30-second Opus file and a 720p VP9 WebM.

First public release.

223 sample files in 77 formats — documents, images, audio, video, office files, data, archives, fonts and text — all CC0.
19 files broken on purpose (truncated, empty, malformed) for testing error handling.
A SHA-256 for every file in manifest.json and sha256sums.txt, and a full archive on the GitHub release for mirroring.
A website with a page per format, search, and llms.txt for AI agents.