How should a reader compare this quarter’s filing with the last eight filings from the same company? This note describes that comparison, using Amazon quarterly filings as the public example, and separates a real change in language from a change in the template.
Eight quarters is a window, not a magic span
A single filing is hard to read in isolation because companies reuse sentences. The interesting question is what changed. A practical window is the current document set beside the previous eight quarterly filings, long enough to see a habit and short enough to hold in one packet. Eight is a choice of scope for this kind of reading. It is not a claim that eight quarters is an optimal history, and it is not a performance horizon. It is the pile of documents on the desk.
The comparison has to be aligned. The same section in each filing, the same meaning of a figure, and the same treatment of restatements. A paragraph that moved because the company reordered its headings is not a new disclosure. A figure that changed because of a stock split or a reclassification is not a change in the business until the series is put on one basis. The first product of the comparison is a list of candidate changes. The list is not yet a list of facts that matter.
A filing comparison is a lineup of this quarter’s sections against a stated stack of earlier filings, used to find language that is new, missing, or in tension with the tables.
What an Amazon packet shows
Amazon’s quarterly filings are a useful public packet because the business is described in segments the company repeats. A reader can line up the current discussion with the eight before it and ask which sentences are new, which numbers moved, and which warnings were already there. A language model can help prepare that lineup by marking paragraphs that have no close neighbor in the prior eight. The marks are candidates. A person decides whether a candidate is a new risk, a new definition, or a sentence that was always present under a different heading.
Some changes are template changes. Lawyers and accountants settle on a phrase and then keep it. A sentence that appears in all eight prior filings and disappears in the ninth is more interesting than a sentence that rotates every quarter. A sentence that appears for the first time may be boilerplate the company imported from a different section. The comparison only works if the reader looks at the neighbors and not only at a similarity score. A score can rank paragraphs. It cannot know why a paragraph moved.
Claims that contradict the numbers are the point of reading the prose at all. If the discussion says a category grew and the table says it did not, the contradiction belongs on the list, with both locations cited. The eight-quarter window shows whether that category’s description has disagreed with its table before. One disagreement can be a wording slip. A repeated disagreement is a habit. Neither one is a recommendation. Both are items a person can verify in the public filing without trusting the lineup.
What the lineup must not become
The lineup is a preparation for reading. It is not a summary that replaces the filing, and it is not a score of management tone that can be averaged into a feature without the acceptance gate any other feature requires. If a candidate list is kept, it should store the filing dates, the section titles, and the model version that proposed the neighbors. A reader who disagrees with a candidate should be able to see the two paragraphs and reject the mark.
Amazon is the example because the filings are public and repetitive. The same method applies to any issuer with a stack of quarterly documents. The method does not say the current quarter is unusual in a way that should be traded. It says a person can see the current language against a stated stack, here the last eight, and can refuse the comparisons that are only movement inside a template.
This piece does not claim that eight quarters is a special horizon, that a wording change predicts anything, or that a model’s lineup is the filing. It describes a way to read one public document against the eight that came before it.
