Methodology

How an entry is made, what counts as a match, and what stops a build.

What qualifies

An entry records an instance in which a court made a finding about a citation or assertion produced by a machine system and filed by a person. The finding must appear in the court's own document. A report of a finding is not a finding, and a news account of an order is not the order.

Matching

An entry is matched when the caption, court and date resolve to a single docket in that court's file. 321 of 366 entries — 87.7% — are matched. The remainder are recorded as unmatched rather than approximated, and an unmatched entry is never given the nearest plausible docket.

Guards

11 rules run against every build. A FAIL stops the build; a WARN queues the entry for adjudication. An adjudication recorded against a filed order downgrades a FAIL to a WARN — the guard is not removed and the reasoning is kept. Current state: 0 FAIL, 15 WARN.

Two of the nine exist because the register got something wrong. One fires on a repeated sentinel value in a monetary field, after twelve entries across eleven courts were found carrying a penalty amount of exactly 1 — a field default, not eleven one-dollar sanctions. The other fires on a caption that is informative about a person but collides with many dockets in the same court.

Attribution

283 of 366 findings do not identify the system involved. The register does not infer one. A product is named only by quoting a public order.

The attribution rate is falling as volume rises — 40.0% in 2024, 28.1% in 2025, 18.2% in 2026. Two explanations fit and the public data cannot separate them: courts may name systems less often as the phenomenon becomes routine, or the growing share of brief orders may carry less detail than early written opinions did.

Remedy

9 entries — 2.5% — establish a remedy.

No amount distribution is published. Any amount statistic from this corpus describes amounts STATED IN A FIRST ORDER, a biased subsample that excludes the dominant remedy — fee-shifting whose figure is fixed later by invoice or accounting. From revision 6 the penalty field distinguishes an order that was read and imposed nothing from an amount that is simply unknown; before revision 6 it could not, and twelve entries carrying a one-dollar sentinel were counted as recorded amounts. Revision 8 followed four dockets forward to the order that fixed the amount and recorded it. One quantifying order — Coronavirus Reporter Corp. v. Apple, N.D. California — is FILED UNDER SEAL, so for that entry the amount is not available to anyone outside the case, by any means. A corpus-wide distribution of sanction amounts is therefore not obtainable even in principle.

Non-monetary remedies recorded, per respondent

Generated from the published respondents file, not carried by hand. A remedy cannot appear here that is not in the data a reader can download. None of these is expressible as an amount, and several were imposed instead of a fee award — a register denominated in dollars records them as no remedy at all.