The buy side runs on numbers extracted from source documents. An analyst reads a filing or a transcript, finds the number she needs, and copies it into her model. The model holds the number. The source does not travel with it.
How numbers lose their sources
A week after she builds the model, a colleague reviews it. He asks about a segment margin figure. The analyst knows she pulled it from the most recent 10-K. She cannot remember whether it came from the segment footnote, the MD&A operating review, or a supplemental schedule management attached as an exhibit. She opens the filing and searches. If the margin figure appears in more than one place with slight rounding differences, which it often does, she has to decide which line she originally used.
This is a minor inconvenience when the original analyst is still on the team. It becomes a structural problem when turnover happens, when coverage passes to a new analyst, or when a model audit is required by compliance. The model is the artifact that survives. The context does not.
The mechanics of error compounding
Source loss is one problem. Transcription error is another. They often combine.
Northwind Freight Systems reported adjusted EBITDA of $541M in fiscal 2025. The non-GAAP reconciliation in the earnings release built that figure from GAAP operating income of $312M, plus depreciation and amortization of $194M, plus stock-based compensation of $31M, plus restructuring charges of $14M, net of a tax-effect adjustment of ($10M).
An analyst hand-keying those figures enters D&A as $184M. A transposition: the 9 and the 1 swap. The model now shows adjusted EBITDA of $531M rather than $541M. A difference of $10M on a $541M base is under two percent. It does not trip any reasonableness check. The model produces a margin that looks plausible. The error persists.
If that model feeds a comparable table distributed to the PM, the $531M figure becomes the number on the desk. If the PM references it in conversation with the company during an investor day, the error has left the building.
What auditable research means in practice
An auditable process is one where any number in any document produced by the team can be traced to its source in under two minutes. Not in a day. Not with a call to the originating analyst. Under two minutes, independently, by anyone with access to the filing.
The citation required to meet that standard is specific: document name, filing date, section or exhibit number, the specific table, and the line within it. Maintaining that level of specificity manually across a full coverage universe requires discipline from every analyst and a system to support the discipline. In practice, the discipline is inconsistent and the system is usually the analyst's own memory.
In a manual workflow, the infrastructure is a combination of notes, file naming conventions, and institutional memory. This works reasonably well when the team is stable and the coverage is small. It breaks down under scale and turnover, which is to say it breaks down at exactly the moment the stakes are highest.
The citation as analytical discipline
Beyond the mechanical benefits of auditability, the habit of citing sources shapes how numbers are understood and used.
A number that resists citation is often a number with an ambiguous definition. Segment revenue for Northwind Freight Systems is reported both including and excluding intercompany freight on certain routes, depending on which note you read. An analyst who writes "segment revenue" without specifying which treatment has a number that cannot be compared to a peer reporting under the other treatment.
The act of writing the citation forces the precision. You cannot write "segment revenue, Note 14, line 3, excluding intercompany" without having decided that the intercompany-exclusive definition is the right one for your purpose. That decision is the work.
A figure without a citation is a rumor. We would rather show our work than sound confident.
Basis maintains an extraction record for every figure it pulls: document, section, table, and the verbatim text surrounding the figure. The record is the citation. When a figure appears in an output, the citation appears with it. If the source document is later restated, the link makes it possible to identify every downstream output that used the original figure and flag it for review.
The discipline is not new. Good analysts have always cited their sources. What changes with a systematic approach is that the citation is no longer optional, no longer dependent on individual habit, and no longer lost when the analyst changes seats.