Primary source

The original document rather than anyone's summary of it: the filing, the transcript, the regulation, the price data.

The discipline matters more now than it did, because a summary generated by a language model is fluent, plausible and occasionally wrong in ways that are invisible without the original. Every number that enters a piece of analysis should be traceable to a document a reader could open.

The source stack for a company is the set of documents worth assembling before starting: several years of annual reports, the last eight quarters of results and transcripts, the shareholding patterns, the credit rating rationales, and whatever the regulator publishes about the sector.

Assemble it first. The analysis is faster afterwards.

The original document. Nothing between you and it.

The habit has become more valuable rather than less as summarisation has got cheaper. A generated summary of a filing is fluent, fast and confidently wrong in ways that are undetectable without the original open beside it: a figure from the wrong year, a segment misattributed, a caveat dropped. The rule that survives is that any number reaching published analysis must be traceable to a document a reader could open and check, and any number that cannot be traced does not go in, however plausible it looks.