Compressing Documents for Legal Filing and HR

There is no federal number. PACER says so itself: “Each court sets a limit on the size of PDF files submitted for electronic filing.” Six real ones, quoted:

The Illinois row carries a trap the others do not. A 25 MB per-document cap sits inside a 50 MB envelope cap, so three documents that each clear the per-file limit can still bounce the submission.

Do not trust an old announcement. These sites keep superseded notices up indefinitely: the Western District of Washington’s live FAQ says 100 MB while a 2017 news item on the same domain still announces the raise “from 10 megabytes (MB) to 35 MB”. Check the court’s current FAQ, not whichever page a search engine hands you.

One question answered plainly, by the Bankruptcy Court for the District of Nevada: there is no page limit, and you should stop thinking in pages. “The number of pages that make up 35 megabytes is dependent on a variety of factors … It is recommended you use the number of megabytes rather than the number of pages to determine file size.” No federal e-filing guide we read publishes a page cap at all.

A practical target across portals: aim for 70% of the stated cap. An 11.8 MB file against a 12 MB limit fails for reasons the form will not explain, usually after you have waited out the upload. The wider set of portal caps and a measured preset table are in when a PDF is too big to upload.

Several systems constrain how the PDF is built, and those rules decide your compression settings more than the byte cap does.

Two of those change what you do at the scanner rather than the compressor. If a court requires bitonal 300 DPI, scanning in colour and compressing afterwards is the wrong order: you throw away sharpness to hit a byte target bitonal scanning would have hit for free.

Note the verbs. The Supreme Court says documents should be PDF/A and text-searchable; NYSCEF says it requires PDF/A and that non-conforming documents “may not be accepted”. PDF/A also constrains compression, since it requires embedded fonts and forbids encryption.

Not if the text layer already exists. An OCR pass writes an invisible second copy of the words over the picture of the words. That layer is text, not pixels, so image compression cannot touch it. We measured it on a three-page 300 PPI scan for compressing a scanned PDF without killing OCR:

All 1,405 words in every row, including the preset where the page is illegible to a human. Do it in the other order and there is nothing to rescue: without a prior text layer, re-OCR accuracy held at 96.8% down to 82 PPI then collapsed to 1.42% at 55 PPI. That is the stroke width of 10.5 pt type falling below one pixel, after which the letterforms stop existing.

One mechanical detail for filings that specify a resolution: Smol’s PDF presets are long-edge pixel caps (600, 900, 1,200, 1,800 and 2,400 px), not DPI targets, so effective PPI drifts with page geometry. The same preset measured 200 PPI on one document and 109 on another. Verify the output rather than trusting a preset name.

No. Not one preset, in any tool we tested. Our 162,058-byte test contract compressed at the Tiny preset came out at 57,377 bytes, a 64.6% reduction, and every Document Info key, every XMP tag, all three comment authors and the embedded attachment survived. Ghostscript behaves the same way: it copies Author, Title, Subject, Keywords and Creator through unchanged and replaces only the Producer string. These are compressors.

The courts know. The Northern District of California’s own redaction guidance tells filers to keep documents “completely free of any hidden metadata which may contain redacted data or other data you do not wish to make public”, because “Metadata can reveal anything that was contained in the file at any time, even text that was previously deleted or changed, and even if the file was re-saved.”

For lawyers it is an ethics question too. North Carolina State Bar 2009 Formal Ethics Opinion 1, adopted 15 January 2010, rules that “a lawyer must use reasonable care to prevent the disclosure of confidential client information hidden in metadata”. D.C. Bar Ethics Opinion 341 (September 2007) requires a sending lawyer to “employ reasonably available technical means to remove such metadata before sending the document”.

What a routine word-processor or scanner PDF carries into a filing:

The annotation row catches HR teams. A policy circulated for comment carries the name of everyone who commented, permanently, and no metadata tool removes it because it is not metadata. Delete or flatten the annotations; find attachments with qpdf --list-attachments and remove them with qpdf --remove-attachment. The cross-format version is in does compressing a file remove its metadata.

None of this is redaction. A black rectangle drawn over a name in Preview leaves the underlying text selectable and searchable. Verify a redaction by extracting text from the output, never by looking at it.

Two commands. Strip, then rewrite. The rewrite is not optional and it is the step almost every set of instructions omits.

brew install exiftool qpdf

exiftool -overwrite_original -all= filing.pdf
qpdf --linearize filing.pdf filing-clean.pdf

Why the second line exists, from the Writing Notes section of exiftool’s manual:

3) Changes to PDF files by ExifTool are reversible (by deleting the
update with "-PDF-update:all=") because the original information is
never actually deleted from the file.  So ExifTool alone may not be
used to securely edit metadata in PDF files.

It warns you at the time, too. Every exiftool -all= run on a PDF prints Warning: [minor] ExifTool PDF edits are reversible. Deleted tags may be recovered! Measured on our fixture, that is not a footnote:

One command, no forensics tooling, no hex editor. A byte search of the supposedly cleaned file still found the original author string twice.

After the qpdf --linearize pass, the same recovery command returned Error: File contains no previous ExifTool update, all 10 Doc Info keys and 13 XMP tags were gone, and a byte search for the author, company and authoring-tool strings returned nothing. Order matters: linearize after stripping.

Four ways to get this wrong that we measured, so you do not have to:

The full test, including the five other routes that failed, is in removing metadata from a PDF on Mac. For a folder rather than a document, see batch-stripping a whole folder, and for the command-line version across file types, removing metadata from the Terminal.

Different systems, same problem, and one useful number almost nobody knows.

Greenhouse accepts a 100 MB resume and cannot read a 3 MB one. Its documentation states that “Candidate uploads can be up to 100 MB, which will support most cover letters and resumes”. A separate article states that “Greenhouse Recruiting can’t parse resumes larger than 2.5MB”. The upload succeeds, the parse fails, the candidate profile arrives empty. A design-heavy PDF resume with embedded photography clears 2.5 MB easily, and that 40× gap between the two ceilings is the whole argument for compressing before submission.

The EEOC Public Portal publishes no size limit and bans encryption. Its FAQ says “The EEOC has not restricted the size limit for submissions” but adds that “Under no circumstances should you upload password-protected files to the EEOC Public Portal” (Public Portal FAQs). Compression is a courtesy there; removing password protection is not.

We looked for published upload caps for Workday, Lever and Workable and found none in official documentation. The figures circulating trace to third-party blogs, and at least in Workday’s case the limit appears configurable per employer. Ask the employer.

Four cases. The first is the one we most want you to read.

You are cleaning a single document and confidentiality is the point. exiftool -all= plus qpdf --linearize is free, takes about a second, and is strictly better than what we ship. Our strip-metadata output on the test PDF was byte-for-byte identical to exiftool -all=, SHA-256 and all, because exiftool is the bundled binary doing the work. So we inherit the reversibility, and we do not run the qpdf pass afterwards even though qpdf is bundled too. Add it yourself.

You need a defensible audit trail. Disclosure and privilege review want hashes at every step and a written procedure that survives cross-examination. That is a forensics workflow, not a drop target.

You need redaction, OCR, Bates numbering or form filling. We compress and convert; we do not edit PDFs. OCR belongs to OCRmyPDF or Acrobat.

Your PDFs are text-only. Our presets work on embedded images, and a word-processor export with no images gives them nothing to do. On a 21-page text-only contract Ghostscript reached 56,321 bytes against our 74,731. The measured comparison is in how to compress a PDF on Mac.

Where a local app earns the $29: sixty exhibits at once with the originals left untouched, mixed piles where PDFs sit next to photographs, a Finder right-click instead of a command someone has to type correctly, and colleagues who will never install Homebrew. Nothing leaves the machine, which for privileged material is the entire argument rather than a feature. Smol is $29 once.

The sibling vertical pages cover the same tools under different constraints: web developers, photographers, real estate listings and podcasters.