Reading a Federal Civil Docket as Data

Nearly every mistake made with this material starts from a reasonable assumption: that something so detailed must be structured underneath. It is not. What sits under the detail is a chronological list of events in one case, each carrying a number, a date, and a short description somebody typed. That is the whole object, and expecting more from it is where the trouble begins.

What each entry actually commits to

An entry asserts that a document was filed or an event occurred on a date. It is highly reliable for that, because the case cannot proceed correctly if it is wrong, and errors get noticed by parties with an incentive to notice them.

It asserts very little else. The descriptive text is not drawn from a controlled vocabulary. Two districts will describe the same procedural event differently, and so occasionally will two deputies in one district. Anyone matching on entry text across courts is measuring local drafting habits as much as case behaviour, and the effect is not random noise. It clusters by court, which means it survives averaging and shows up as apparent regional variation in whatever is being studied.

The fields that hold up

A small number of things can be taken more or less at face value.

• Filing date and the date of each entry.
• The sequence of entries, which is the case's actual procedural history.
• Party names as captioned, with the caveat that captions change.
• The presiding judge, though reassignment happens and is itself a docket event.
• The existence of a document, separately from what the document says.

That last distinction is the one that trips up automated work. A docket tells you a motion for summary judgment was filed and that an order followed. Whether the order granted it, denied it, granted it in part, or disposed of something else entirely is in the order, not in the docket text, and the docket text will often gesture at the answer in a way that is right most of the time. Most of the time is not a research standard.

What the docket cannot tell you

Three questions come up constantly and none of them is answerable from the docket alone.

Who won. Termination is an administrative state. A case closes when the court has nothing further to do, and that covers judgments, settlements, voluntary dismissals, transfers, and consolidations without distinguishing them in a way that maps to winning. A very large share of civil cases end in a way that has no winner in any sense a client would recognise.

What it was worth. Amounts appear in a docket only when a document containing them was filed publicly. Settlement terms usually are not, and where they are, they are frequently partial. Any distribution of case values built from dockets is a distribution of the subset that became public, which is not a random subset.

Why anything happened. The docket records that a stay was entered. The reason lives in a motion, an order, or nowhere at all.

Getting at the records

Federal dockets are accessible through the courts' own electronic access system, per document and per page. That pricing shapes research practice more than any methodological consideration: it makes exhaustive collection expensive and therefore rare, and it makes convenience samples common.

The RECAP Archive, maintained by the Free Law Project, holds a large collection of dockets and documents contributed by people who already paid for them. It is genuinely useful and its coverage is not a random sample of federal litigation. It is a sample of what people were interested enough to fetch, which skews toward larger, more contested, more newsworthy matters. Using it to characterise the typical case will overstate how contested the typical case is.

This is not a criticism of the archive, which is explicit about what it is. It is a warning about a specific inference that gets made from it silently.

Practical notes for anyone parsing at volume

Parse dates and sequence. Treat entry text as a weak signal requiring validation against the underlying document, and quantify how weak by hand-checking a sample rather than assuming.

Expect docket numbering to be non-contiguous. Entries get sealed, stricken, and corrected, and a gap is a fact about the case rather than a fetch failure. Code them as distinct.

Watch for cases that enter and leave. Transfers and multidistrict consolidation mean the same dispute can appear as several dockets in several courts, and naive counting will treat one matter as three.

Above all, decide before starting whether the question actually needs dockets. If it is a question about case counts, dispositions, or the composition of a court's civil caseload, the coded case-level records exist precisely so that nobody has to reconstruct them from event logs. The next article in this section covers what those records contain and where their coding compresses away something that mattered.

For the tooling side of the same problem, which is how systems reading these documents should account for themselves, the full index of research notes connects the two sections.