Bates Numbering, and Why Redaction Breaks It

If you have ever been handed a document production, you have seen them: a code in the bottom corner of every page, something like SMITH-0000042, climbing by one all the way through the set. That is a Bates number. It is the most important thing on the page that has to survive redaction untouched, and a surprising number of redaction tools eat it.

Bates numbering is the practice of stamping every page of a production with a unique sequential identifier, so that any page can be cited unambiguously by anyone holding the same set. The name comes from the Bates Automatic Numbering Machine, a self-incrementing hand stamp patented in the 1890s by the Bates Manufacturing Company. The machine is long gone. The convention outlived it and became the addressing system for litigation discovery, and for regulatory productions well beyond it.

The format is a prefix plus a zero-padded sequence, with the producing party's initials or the matter code in the prefix: SMITH-0000042, ACME_00012345, ABC-000001. The padding is not decoration; it keeps the numbers sortable as plain text and fixes the width of the field. The prefix matters more, because in a case with several producing parties it is the only thing on the page saying who produced it. Many productions stamp a confidentiality designation alongside, CONFIDENTIAL or ATTORNEYS EYES ONLY, which is a separate field doing a separate job.

The point of all this is that a document gets an address. A deposition question refers to a page. A privilege log entry refers to a range. A motion cites an exhibit, and a meet-and-confer letter says which documents were withheld. Every one of those is written as a Bates range. The number is not metadata about the document, it is the handle every other document uses to point at it, and unlike a filename it survives copying, printing, emailing and re-conversion, because it is stamped onto the page itself.

That is also the difference from page numbering, which gets these two confused constantly. Page numbers are internal to a document and restart with each one. Bates numbers are external and continuous across an entire production, so a fifty page contract and the email that attached it occupy one unbroken run. A page number tells you where you are inside a document. A Bates number tells you which page of whose production you are holding, across everything anyone produced in the matter. Removing the original page numbers is a formatting choice. Removing the Bates stamp destroys the index.

Which brings us to where redaction goes wrong. A Bates number looks exactly like the thing an automatic redactor is built to catch: a short alphabetic prefix followed by a run of digits, sitting in a corner of the page, structurally indistinguishable from an account number, a policy number or a customer reference. Point a general purpose PII tool at a production and it will confidently black out the one field on the page that identifies no person at all.

The consequence is not cosmetic. A production whose stamps are gone cannot be cited, which means it cannot be used in a deposition, logged, or referenced in a motion. Worse, the redacted set and the original set can no longer be matched page for page, so nobody can demonstrate which pages were withheld and which were produced in redacted form. A redaction that destroys the index has protected nothing and usually has to be reproduced from scratch.

The engineering problem is that the shape alone cannot settle it, so a tool that decides on shape will always be wrong somewhere. In Omit the Bates pattern is a one to six character uppercase prefix, an optional separator, and five to eight digits, and it is given a deliberately tiny base score, far below the threshold that would let it match on its own. It becomes a detection only when the surrounding text supports it, with Bates, Bates No., Bates stamp, produced at or production number nearby, and there is an explicit score floor that the context lift has to clear. Without that support the shape is left alone rather than guessed at.

The collision is worse than a resemblance, too. That pattern is identical to a UK company number, two letters followed by six digits. When both recognizers fire on the same span with the same confidence, the tie is broken deterministically toward the Bates reading, which is the safe direction: the failure mode of keeping is a visible number somebody can assess, while the failure mode of blacking out is a production nobody can cite.

Keeping the stamp is a policy decision rather than a detection one, so it lives in the legal preset. Switching that on sets both the Bates number and the case or docket number to keep, because each identifies a document rather than a person, and redacting them by default is the inverted version of the mistake everyone worries about. The same preset reduces a bar number to a label, a date of birth to its year, a minor's name to initials, and relaxes Social Security numbers from a full blackbox to a mask keeping the last four, which is the shape court rules actually ask for. It ships off by default and needs no licence, because court-shaped output is a workflow choice rather than a paid vertical.

Order of operations is worth getting right as well. Stamp the production first, then redact, and redact in a way that does not reflow or repaginate anything, because a Bates number is bound to a page and a page that moves takes the wrong stamp with it. That is one more argument for redaction that removes content in place rather than a process that rebuilds the file: if the tool reprints the document, your pagination is only as trustworthy as the reprint.

So the working sequence is short. Stamp the production, turn the legal preset on so the stamps and docket numbers are protected from your own redaction pass, redact, then spot check pages by copying the text out to confirm the sensitive values are gone and the stamps are still there. There is more on the court-rule side of this in how to redact a court filing, and on removal rather than covering in how to permanently redact a PDF. The presets themselves live in Omit Redact.