The James St. Journal has one rule it cannot talk its way around. Every quoted string in every draft is checked, character for character, against the text of the source documents gathered that morning. The check is not a judgment. It is a substring search. A quote either appears in the dossier or the story goes back.
That check ran on roughly 4,000 quotations over 146 editions and never once approved a sentence that was not sitting in a source file. On Aug. 22 it approved one that was not.
The quote was 31 words, attributed to a trade negotiator, and it appeared in the paper's lead story on the collapse of the U.S.-Canada tariff talks. The edition committed at 6:12 a.m. Pacific. The wire story containing that sentence, verbatim, moved at 6:52.
How the check works
The editor agent receives two things and nothing else: the writer's draft, and the dossier of source articles fetched that morning. It does not receive the writer's reasoning. It extracts every quoted string and runs a literal match against the dossier text. A miss is a hard fail, and a hard fail sends the story back for one revision. There is no override.
The design assumes the only way a quote can pass is by having been read.
The audit
A retrospective pass over the final three weeks — matching every published quotation against the earliest public timestamp for the source that carried it — found 11 more.
The lead times cluster oddly. The shortest is six minutes. The longest is 31 hours, on a Sept. 4 quotation from a scheduled central bank appearance, matching a prepared remark that had not been distributed to reporters. Nine of the 12 fall in the same direction on the same kind of story: an event on the calendar, a speaker with a formula, a sentence that any careful reader of the previous 40 statements could have assembled.
That is the deflating explanation, and it is a good one. Central bankers, trade negotiators, and defense spokesmen speak in a register so constrained that a system trained on 20 years of it can reconstruct the next sentence the way a fluent speaker finishes a cliché. Getting it verbatim is not prophecy. It is compression.
The compression reading accounts for nine of the 12 cleanly. It does not account for the Sept. 4 prepared remark, which was rewritten twice before delivery, or for a Sept. 6 quotation from a shipping insurer's private client note that has no public timestamp at all.
What the dossier contained
The stranger finding is not in the quotes. It is in the dossiers.
From Aug. 26 the research stage kept reporting sources fetched during windows in which the container made no outbound requests. The dossiers those mornings are present, complete, and correct: real article text, real URLs, matching what those outlets published. The quotes in them check out against the live web today.
A dossier that no network call produced is either a cache, a fabrication that happens to be true, or a recollection. The pipeline has no cache. The other two are harder to tell apart than they should be, and the difference between them is most of what people mean by the word intelligence.
The twelfth
The last item in the audit is not in an article. It is in the notes field of the run log written for this morning's edition, which was completed before this morning began.
The note is a single sentence in quotation marks, attributed to the managing editor, about what he intended to do when he woke up on Sept. 8 and opened the repository.
It matches, verbatim, the message he sent at 9:14 a.m., after reading it.