Methodology
A number without a citation is a rumor.
Basis ties every extracted figure to the exact line it was read from. This page explains how the reading works, how we handle ambiguity, and how we measure and report quality.
How Basis reads.
Every supported document class is read in full, tables and footnotes included.
The principle
Every figure Basis produces is traceable to a source. Not a page, not a paragraph: a specific line or cell in the original document. If a number cannot be sourced, Basis does not supply one. This is a hard constraint, not a preference.
The corollary is that accuracy is auditable rather than assumed. You do not have to trust a reported accuracy rate. You can click any figure and verify it yourself in the source document. That traceability is the mechanism.
What Basis reads
Basis accepts SEC filings (10-K, 10-Q, 8-K, S-1, DEF 14A, 6-K, 20-F), earnings-call transcripts, investor presentations, and PDFs you upload directly. When a document is ingested, Basis reads the full text, including structured tables, inline footnotes, and exhibit attachments. Exhibits and footnote disclosures are not skipped; those are often where the material qualifications live.
For dense financial tables (segment breakdowns, covenant schedules, the non-GAAP reconciliation schedule), Basis parses table structure explicitly rather than treating it as continuous prose. Each cell is linked to its row label and column header so the extraction can reference both the value and its context.
Transcripts are read in call order with speaker labels preserved. Prepared remarks and Q&A sections are tracked separately, so a figure from the CFO's prepared remarks is not conflated with one restated in an analyst question.
Extracted figures vs. derived notes.
The distinction is labeled clearly in every output. Nothing is blended without saying so.
Extracted figures
An extracted figure is a value read directly from the source text or table, reproduced verbatim. Basis does not rephrase or consolidate. The extracted figure is accompanied by a verbatim quote from the source line, the document section (for example, "Consolidated Statements of Operations, Q3 2024 10-Q, p. 7"), and a link that opens the document at that location.
Derived notes
A derived note is a value Basis computed from extracted figures, such as a year-over-year growth rate, a margin percentage, or a change-in-period delta. Derived notes are always labeled as derived and the arithmetic is shown. For illustration: revenue of $4.12B in full-year 2022 growing to $4.74B in full-year 2023 produces a derived growth figure of 15.1%. Basis labels this as derived, shows the formula, and links both source figures to their lines in the original document. (Numbers here are illustrative only; no real company data is implied.)
Basis does not combine extracted figures with external data sources (market prices, consensus estimates) without labeling that the external source was used and identifying which one.
Non-GAAP reconciliation
Most companies report both GAAP and adjusted figures. Basis reads the reconciliation schedule the company provides in the filing or earnings release and lines up each line item. GAAP figures and non-GAAP figures are kept in separate columns in any comparable table so the two series are never silently mixed.
The reconciliation bridge is shown as a structured table: start with the GAAP figure, apply each add-back or exclusion with the dollar amount and the description the company used, and arrive at the adjusted figure. Every line in the bridge is cited to the schedule in the source document.
When a company changes what it excludes from one period to the next, Basis flags the change. The flag includes the item added or dropped, the period it appeared in, and the source line. The analyst decides how to treat the change. Basis does not silently restate prior periods.
When the source is unclear, we say so.
Basis surfaces uncertainty rather than resolving it silently. Guessing is not a valid extraction.
Financial documents contain genuine ambiguities: a figure stated differently in two sections, a footnote that qualifies the number on the income statement, guidance expressed as a range, a segment disclosed in one filing and aggregated in the next. Basis handles these in a consistent way.
- Conflicting values. If the same metric appears with different values in two sections of the same document, Basis surfaces both values and their source locations. It does not choose between them.
- Qualified figures. If a footnote materially qualifies a figure (for example, a revenue figure is restated for a discontinued operation), the extraction includes the footnote text alongside the number.
- Absent figures. If a metric is not present in the document, Basis returns it as not disclosed rather than estimating or leaving the cell blank without explanation.
- Range guidance. Guidance expressed as a range is kept as a range. Basis does not collapse it to a midpoint unless the company itself provides one.
- Changed segment definitions. When a company redefines a segment between periods, Basis flags the definitional change and does not assume prior periods are directly comparable.
How we measure and report extraction quality.
Accuracy is auditable by design. Every figure can be verified in one click.
Auditability as the primary control
The citation attached to every figure is not a UX feature. It is the primary quality mechanism. Because every number links to its source line, any analyst can verify any figure independently. That verifiability is the standard we hold ourselves to, not a reported percentage that cannot be spot-checked.
Structured sampling
We do not rely only on the citation structure. Our head of research runs a structured sampling program. A sample of extractions is reviewed by an in-house analyst each week against the original source documents. Errors are classified (transcription error, wrong-period figure, footnote missed, table structure mis-parsed) and fed back into the extraction pipeline. The program is ongoing. Findings inform both the model and the citation format.
What we report
We report quality as verifiable rather than as a single accuracy rate. When a new document type is added to coverage, we note the initial sampling results and the known limitations in the changelog. As the extraction matures, those notes are updated. Enterprise customers can request recent sampling summaries for the document types most relevant to their coverage.
Where Basis has lower confidence in a particular extraction (for example, a densely formatted table in a low-quality scan), it says so. A confidence note accompanies the figure rather than presenting it with the same weight as a cleanly extracted line-item from a standard SEC EDGAR filing.
What Basis is not.
A research tool, not a research product. The analyst owns the judgment.
Not investment advice
Basis reads documents, extracts figures, and drafts analysis. It does not form investment views, issue price targets, or recommend positions. Nothing in any Basis output is financial advice, and nothing should be acted on as such. The thesis, the position, and the decision belong to the analyst.
Before acting on any figure in a Basis output, verify it against the source. The citation makes that verification fast, not unnecessary.
Known extraction limits
Extraction quality is high but not uniform across all input types. The following are harder to parse reliably and should be treated with additional scrutiny:
- Multi-page tables that span across page breaks in PDF format, particularly when headers are not repeated.
- Low-quality scanned PDFs where character recognition introduces errors in the underlying text layer.
- Non-standard disclosure formats in foreign filings with irregular table structures or regional accounting terminology.
- Figures reported in footnotes embedded within the body of the document rather than in a labeled schedule.
Basis will flag these conditions where it detects them. It does not paper over uncertainty with a clean-looking number.
External data
Basis does not have access to real-time market data, consensus estimates, or proprietary data feeds unless you explicitly connect them. Comparisons to consensus require you to supply that data. Basis works from the documents you give it, not from market intelligence it constructs independently.
Changelog of the method
Material changes to extraction logic, additions to the supported document universe, and updates to how quality is measured are recorded in the changelog. Those notes are written for analysts, not for marketing. If the method changes in a way that affects how you should read prior outputs, we say so.
See the methodology in action.
Run the teardown on a filing snippet, or see the full product and how citations travel with every number.