Revisions

Every revision, what changed, and what it did to the published figures.

Rev	Prepared	Published	What changed
19	2026-09-21	2026-09-21	A correction to revision 16, and the respondent's own account in Twigg. AMC-FED-0144's ai_tool is RESTORED to 'Implied', the source compilation's value, after revision 16 overwrote it. The field carries the compiler's vocabulary and is not this register's to author, and the reasoning was backwards besides: 'Implied' means a machine was inferred and no system named, which is exactly a matter where no court mentioned AI at all. Nothing caught it for three revisions because both values are non-attributions inside the same unattributed total, so the edit moved no figure and tripped no rule. Logged as C-2026-013; guard G11 now asserts the field against the source. Twigg's ECF 236, the respondent's letter brief, is recorded as a party filing: he admits an unnamed 'AI software platform' and describes quotations spliced from two separate sentences of one opinion. No figure moves.
18	2026-09-21	2026-09-21	THE TWIGG SANCTION IS IMPOSED. M.D. Pa. ECF 239, 18 June 2026: Attorney Nicholas W. Mattiacci, Sr. to pay $1,500 personally to the Clerk within fourteen days, that sum expressly NOT counted against the litigation costs taxed by the Clerk, and suspended from practicing in the district for six months with automatic reinstatement from 22 June 2026. Suspension from practice is a remedy form this register had not carried and is heavier than the pro hac vice revocation in Wadsworth; a reader given only the dollar figure has the smaller half. This docket carries six merits dollar figures against one sanction and none is in any penalty total. No confirmation figure moves; amount_recorded 102 to 103.
17	2026-09-21	2026-09-21	THE MINLEON SHOW CAUSE IS DISCHARGED, and this is the first row carried from finding to disposition entirely on free primary documents. M.D. Pa. ECF 53, 14 January 2026: on the attorneys' response the court held their apology and 'the payment already tendered to Plaintiff's counsel' a sufficient reprimand and ordered that no further sanctions issue. The court imposed nothing, so penalty_recorded moves to none_imposed. Money moved anyway and the court did not order it: the payment was voluntary and the order recites it without an amount. The zero sits in the penalty fields, the payment in the respondent rows, and it is in no total.
16	2026-09-21	2026-09-21	AMC-FED-0144 completed from the companion order. M.D. Pa. ECF 47, 17 December 2025, names what the memorandum left unnamed: Attorneys Joshua L. Schwartz and Justin A. Tomevi, counsel for the DEFENDANTS, ordered to show cause by 2 January 2026 why non-existent quotations to two real Federal Circuit decisions appeared in their brief. Invented-quotation subtype, not invented-case: both cases are real at the reporter pages given and only the quoted language does not exist. No figure moves.
15	2026-09-21	2026-09-21	AMC-FED-0287 read at source from two free USCOURTS documents, M.D. Pa. ECF 231 and ECF 232, both filed 21 May 2026 — the decision date the row already carried. THE DISPOSITION IS AN ORDER TO SHOW CAUSE, NOT AN IMPOSITION, so the row does not move to none_imposed: the court did not decline to sanction, it continued the proceeding. The only dollar figure in the order, $384,697.45, is the molded final judgment and not a sanction. The row leaves the paid-retrieval queue at no cost. No figure moves.
14	2026-09-21	2026-09-21	AMC-FED-0139 struck from the confirmed set. It had been confirmed since revision 3 at a match score of 0.961 with an empty docket number, and the docket it matched was filed in January 2012 against a decision dated December 2025. The age-gap guard that should have caught it reads the year out of the docket number, so an empty field silently disabled it. Confirmed falls 322 to 321 and the seam 323 to 322, which withdrew the seam collision revision 13 had created. Nothing about the September retraction was reconsidered: a row that should never have been in the confirmed set left it. A guard now fails any row confirmed to a docket without a docket number, and the claim that revision 13 closed the last entry with no docket number at all is corrected — forty-three rows carry none.
13	2026-09-21	2026-09-21	United States v. Hayes, the best-specified of the unresolved entries, is closed. It resolved to E.D. Cal. No. 2:24-cr-00280-DJC, ECF 62, 17 January 2025, from the court's own order: $1,500 personally on an Assistant Federal Defender under a local rule and the court's inherent authority — not Rule 11, because it is a criminal case — with the order served on two state bars and on every district and magistrate judge in the district. The search was narrowed by a footnote in an unrelated sanctions order read the same day. For one revision the seam figure stood at 88.3%, the value retracted on 20 September, reached from a disjoint set of rows; revision 14 withdrew that collision.
12	2026-09-21	2026-09-21	Wadsworth v. Walmart carried $5,000, which is $3,000 plus $1,000 plus $1,000 across three named attorneys with both law firms expressly not sanctioned. No respondent owed $5,000. Found by the cross-citation check — two independent courts describing the sanction differently from this register — and then decided by reading the order. The first defect here found by a tool rather than by hand.
11	2026-09-21	2026-09-21	THE MATCHED RATE MOVES FOR THE FIRST TIME SINCE REVISION 5: 320 of 366 to 321 of 366. One entry, Mattox v. Product Innovations Research, had no docket number at all; the court's own order supplies one. That same entry carried $29,495 — a figure that is not a sanction anyone was ordered to pay, but three individual fines, a fourth attorney sanctioned without money, and a fee award split between two firms, added together. A second entry, Whiting, carries $30,000 that is $15,000 on each of two people and an understatement besides. Three columns and a per-respondent companion file are added so that an amount can say whose it is.
10	2026-09-21	2026-09-21	Two further amounts read from the orders ($12,997.60, $250) and two more orders recorded as imposing nothing. Also the first correction to an amount this register had already published: one entry carried $4,997 where the opinion says $5,000, caught because another court described that sanction and the description did not match. An amount recorded here is not an amount verified.
9	2026-09-21	2026-09-21	A fifth amount: the $10,000 Rule 11 sanction in Virgil v. Experian, which this register had recorded on 19 September as recommended with adoption not established. The adopting order had issued on 11 September, eight days earlier. The recommendation was read; it was all there was to read unless someone followed the docket forward. No docket changed.
8	2026-09-21	2026-09-21	Four sanction amounts read from the courts' own orders and recorded — $8,151, $45,763, $2,500 and $1,000 — three of them closing the gap between the order that imposes a sanction and the later order that fixes the figure. No entry's docket changed and the matched rate is unchanged. This is the first revision to make the remedy data stronger rather than weaker, and it did so by reading orders.
7	2026-09-21	2026-09-21	One sentence corrected in one entry's audit trail. The Sixth Circuit's sanctions opinion in a consolidated matter is indexed by CourtListener under a docket record belonging to an unrelated Social Security appeal; the entry had recorded that as an adjacent case rather than as the misfiling it is. The entry's docket is unchanged and no published figure moved.
6	2026-09-21	2026-09-21	No entry's docket changed. The penalty field gained a third value for an order that was read and imposed no monetary sanction, and twelve entries that had carried a one-dollar sentinel as a recorded amount were cleared. The share carrying no dollar amount rose from 71.0% to 71.9% because the sentinel had been counted as an amount.
5	2026-09-21	2026-09-21	Four federal appellate entries confirmed to the issuing court's own docket from the published opinions, at no cost; two of them left the paid-retrieval queue. Matched rate 86.3% → 87.4% (316 → 320). No unresolved federal appellate entry remains. The resulting figure equals the superseded revision 3 figure and is not a return to it: revision 3's four struck entries remain struck.
4	2026-09-21	2026-09-21	Four entries struck on caption collision — a caption informative about a person that resolves to many dockets in the same court, which is why the existing caption guard does not fire on it. One strike reversed a prior adjudication. Matched rate 87.4% → 86.3% (320 → 316). Guard added.
3	2026-09-18	2026-09-18	Two corrections. Label-variant court groups collapsed. Anchored; not overwritten.

Anchoring

Each published revision is hashed and the hash is timestamped. A superseded revision keeps its original hash and stays downloadable. Revision 3 is anchored and is not overwritten by revision 19; revision 19 is a new record, not an edit of the old one.

Revision 19 is confirmed. Its manifest, SHA-256 7c0726fa2b3f843bcb972de3fb107897ebabd8804ab16a25c2270eaf1f2ce7f0, is committed by Bitcoin block 968,034. Because those bytes are now anchored, the manifest is not edited again for any reason, including to remove the build-host path it carries in its source_revision field. That defect is disclosed in correction C-2026-016 and is fixed forward at the next revision rather than backward in this one.

Which way the rate moves

This section used to say that every revision had lowered the matched rate. That stopped being true at revision 5, and the sentence is replaced rather than quietly dropped.

Four revisions lowered the rate, by striking entries that could not be verified. Revision 5 raised it, by reading four published opinions and confirming four entries to the docket each order issued from. Those are not the same kind of move, and the difference is the one that matters: a rate falls here when a claim is withdrawn, and rises only when a primary document is read. It has never risen because a number was recomputed, and the one figure this project published that was reached that way — a retracted 88.3% — was withdrawn precisely because no build could produce it.

A register whose published figure only ever improved would be a register that was not looking. One whose figure could only fall would be a register that could never finish verifying anything.