What is a language model allowed to write when it reads a filing or a transcript? This note describes that draft, using a Costco earnings transcript as the public example, and draws the line between a sourced note and a conclusion.
A draft with sources, not a verdict
A language model can read a long document and return a shorter one. The shorter one is useful when it points back to the passages it used, names the entities it found, and lists questions it could not answer from the text. That product is a note. It is not a view. A view says what the reader believes follows from the note, and a view has an author who is a person. Mixing the two in one paragraph makes it impossible to see which sentences were extracted and which were decided.
The model's note should be boring on purpose. It can list topics the document spent time on, quote figures that appear in the text, and flag a change in wording from a prior document when both documents are in the packet. It should not invent a figure that is not in the packet, smooth over a refusal to answer, or assign a motive to a speaker. When the text is ambiguous, the note says it is ambiguous. A fluent paragraph that resolves the ambiguity has stopped being a note.
A model note is a sourced draft that lists what a document says, what it does not answer, and which passages were used. It is not a conclusion and it is not an instruction.
A Costco transcript as the packet
Consider a Costco quarterly transcript as a public packet. Prepared remarks describe membership, traffic, and categories the company chooses to highlight. The question period contains answers that are specific and answers that are general. A model can draft a note that separates those two parts, lists the figures actually spoken, and writes down questions an analyst asked that did not receive a number. A person can then check the draft against the transcript. The check is the work. The draft saves the person from starting with a blank page. It does not replace the check.
The same model can be pointed at the prior quarter's transcript and asked which topics appear in one and not the other. That comparison is still a draft. Topics are not facts until someone confirms the passages. A topic label can be too broad, so that two different remarks look like one theme. The acceptance of the note, if it is accepted at all, names the model version and the packet. A later model version requires a new check. Costco is the example because the calls are public and repetitive enough that a made-up figure stands out against the usual categories.
Nothing in the draft places an order or recommends that a reader do so. The note does not say the quarter was strong or weak. Those words are conclusions. If a researcher wants them, the researcher writes them underneath the note and signs them. The model may have proposed a sentence of that kind. The sentence stays out of the accepted note unless a person moves it, on purpose, into the conclusion and accepts the responsibility for it. The file can store both layers. It should not flatten them.
Where the draft is allowed to go
A draft note can enter a research discussion as a reading aid. It can sit next to the source. It should not be pasted into a public page as if the firm had written it, and it should not be treated as a dataset of truths about the company. If pieces of the note are later turned into structured fields, those fields are proposals in the same way a sentiment score is a proposal. Someone accepts them, with the failures sampled, or they stay in the draft folder.
The limit is ordinary and strict. The model reads. The person concludes. A Costco transcript is long enough that the reading aid is real, and public enough that any reader can test whether the note stuck to the text. If the note cannot be checked, it is not yet a note. It is an unreviewed paragraph, and an unreviewed paragraph does not belong in the study.
This piece does not claim that a model understands a business, that a transcript note is accurate without a check, or that any note should affect an order. The draft may be written by the model. The conclusion may not.
