Narrative vs. Numbers Strategy–capital alignment from SEC filings

Management wrote

We continue to identify and evaluate opportunities to expand our datacenter locations and increase our server capacity to meet the evolving needs of our customers, particularly given the growing demand for AI services.

Microsoft, 10-K for the year ended 30 June 2026 0001193125-26-323660

The filings recorded

Capital expenditure+79.6%
Property & equipment+52.7%
CapEx share of revenue+12.03 pp

Three signals, all moving the way the claim implies, all past the threshold at which a move counts as decisive. That claim scores 100. Two claims in the same filing score 48.6.

Reads the latest 10-K for the strategy and the XBRL facts for the years behind it. A first, uncached analysis typically takes about a minute; a recent matching report is returned immediately.

What the model is allowed to do

Read Item 1 and Item 7 and return, for each priority it finds, four things: a label, one category from a fixed taxonomy, the sentence it came from, and how sure it is. A category outside the taxonomy is discarded rather than guessed at.

What the model is not allowed to do

Produce a number. Alignment, confidence, every metric and every threshold come from arithmetic over XBRL facts. If the model were swapped out tomorrow, the scoring would be unchanged.

What it will admit

Missing data is listed, never scored as zero — a component with no evidence has its weight redistributed and the gap is reported. Alignment is kept apart from confidence, so a company can be shown to look aligned on very thin evidence.

Ticker
CIK
Periods scored
Filings read
Result

Alignment, confidence-weighted across claims

Analytical confidence

Reading this report

“quoted prose” is management’s own language, located and labelled by the language model. It produces a label, a category, a quote and a confidence — nothing else.

0.0–100.0 is arithmetic over XBRL facts. Every score on this page was computed by code you can read.

Claims on the docket

Claim Action, outcome, persistence Score

Financial record

Reported figures come straight from XBRL company facts. Derived rows are the normalized quantities the scoring engine actually reads, so no comparison depends on company size.

Contradictions

Raised only where an action metric moved against the claim by more than the threshold that defines a decisive move — not merely the wrong sign.

What this analysis could not see

Gaps are listed rather than scored as zero. A missing component has its weight redistributed over the components that survived.

    Method

    1. saidLatest 10-K, Item 1 and Item 7 located by title, not by item number alone
    2. saidPriorities extracted: label, category, verbatim quote, confidence
    3. saidCategory forced onto a fixed taxonomy; anything outside it is dropped
    4. didXBRL tags mapped to canonical metrics, then to growth rates and ratios
    5. didCategory determines which metrics may be scored, and which way each must move
    6. didEach metric scored 0–100 against its threshold; components weighted
    Alignment
    0.50 action evidence + 0.30 outcome evidence + 0.20 persistence. A component with no data has its weight spread over the rest.
    Action evidence
    Did the money move. CapEx, R&D, SG&A, debt, distributions.
    Outcome evidence
    Did the result move. Revenue, margin, cash generation.
    Persistence
    Share of scored years in which the action metric moved the implied way. A move under 0.5 pp counts as half agreement, since flat neither supports nor contradicts.
    Metric score
    50 + 50 × clamp(direction × value / threshold, −1, +1). 50 means no movement. The threshold is 15% for a year-over-year rate and 1 pp for a change in share of revenue.
    Confidence
    Held apart from alignment: extraction certainty, share of mapped metrics found, and years of history. A company can look aligned on very thin evidence.
    Bands
    90–100strongly supported
    75–89supported
    55–74mixed evidence
    35–54weak support
    0–34contradicted