Narrative vs. Numbers

Corporate strategy is easy to describe and much harder to verify. Annual filings contain carefully framed priorities, while the evidence of execution is spread across capital expenditure, operating metrics, margins, and several years of reported facts.

Narrative vs. Numbers brings those two records together. It asks what management said was strategically important, which financial signals ought to move if the claim is being acted on, and whether the observed history supports, weakens, or leaves that claim unresolved.

The decision is company- and filing-specific. The system does not label management generally credible or incredible, and it does not claim that alignment makes the strategy good. It tests whether stated priorities and visible execution point in the same direction.

The latest 10-K supplies the current strategy. A language model is limited to returning a priority label, one category from a fixed taxonomy, the source quote, and extraction confidence. Categories outside the taxonomy are discarded rather than improvised.

Each accepted category maps to approved financial signals with a required direction and threshold. SEC XBRL facts are normalized into annual metrics, then action evidence, outcome evidence, and persistence are calculated. With all components present, alignment equals 0.50 times action evidence, 0.30 times outcome evidence, and 0.20 times persistence.

A component with no data is not scored as zero. Its weight is redistributed across available components and the gap is carried into the limitations. Analytical confidence is calculated separately from extraction certainty, evidence coverage, and available history.

The system is deployed and can analyze one public company and filing at a time. The live page includes a Microsoft example drawn from its 10-K for the year ended 30 June 2026, placing a capacity-expansion claim beside changes in capital expenditure, property and equipment, and capital-expenditure intensity.

That example demonstrates the product's claim-by-claim structure and the distinction between management language and computed evidence. It is not a general recommendation about Microsoft, and the tool's overall judgment depends on every extracted priority, available metric, limitation, and confidence value in the run.

The current evidence establishes that the ingestion, normalization, mapping, scoring, contradiction, and confidence paths run over SEC data. It does not establish that the taxonomy captures every strategically relevant signal or that the score predicts performance.

The central distinction is deliberate: AI interprets the filing; data grounds the claim; code performs the financial judgment. A model swap can change which priorities are extracted, but it cannot silently change the arithmetic used to score them.

The most useful disagreement is therefore specific. A reader can challenge the extracted claim, its category, the approved metric, the direction, the threshold, the observed value, or the confidence instead of accepting or rejecting one opaque verdict.

Current strategy is read from the most recent 10-K. Historical years primarily provide financial persistence and do not yet show how the strategy narrative itself changed over time.

Evidence depends on the configured strategy-category-to-metric mapping. Relevant qualitative execution or unmapped operational signals may be absent from the score.

Missing components are shown and their weights redistributed, language-model extraction can be imperfect, and analytical confidence must remain distinct from alignment. Alignment is not proof of causality or strategic quality.

Create a human-reviewed set of strategic-priority extractions across several industries and report category-level misses and disagreements before broadening the taxonomy.

Have strategy and finance reviewers inspect the same claim-level reports, record which part of the evidence chain they dispute, and test whether the exposed arithmetic makes those disagreements more precise.