THIS ISN'T ONE COMPANY

Three things the chart is telling you

The bars further down carry numbers. Here are the same numbers in words — the lens Editions I and II were written in, applied to the cohort data instead of a single named deal.

One. Salesforce is the outlier — and they're telling you why

At ~30 hits per 100,000 characters of filing text, Salesforce scores roughly double the next company in the cohort. The mechanism is no secret: they say "strategic investments" about fifty times in every quarterly filing. That's the disclosure footprint of Salesforce Ventures, one of the largest corporate venture funds in tech. The question Edition III raises isn't "is Salesforce doing something wrong?" It's "what fraction of their reported returns is coming through that fund versus the operating business?" The answer is in the same filings. Read them.

Two. Microsoft and Oracle are reading from the Edition I script

Both companies lean heavily on the phrase remaining performance obligations — the GAAP line where future cloud backlog gets disclosed. Edition I traced this directly: Microsoft's $627B and Oracle's $553B in committed pipeline, large fractions tied to OpenAI alone. The EDGAR scan is picking up the same disclosure surface from the opposite direction. One finding. Two methods. That is what corroboration looks like.

Three. Amazon's round-1 score was a methodology bug, not a data point

The first scan ran a regex that looked for the literal phrase "cloud credits." Amazon never writes that. They write equity method investments and unrealized gains on investments — different words for the same OpenAI / Anthropic mark-to-market mechanism. Round 2 broadened the regex and Amazon's raw score jumped 25×, from 4 hits to 100. The lesson is general: when a regex returns near-zero on a company you have prior reason to suspect, the regex is the answer, not the absence. Edition III's chart only goes as far as the patterns. The structure goes further.

Pattern density per company

Each bar is the total number of times any of our sixteen patterns appears across that company's four most recent 10-K and 10-Q filings. The patterns are blunt regexes — "cloud credits," "strategic investments," "mark-to-market," "non-GAAP exclude," "remaining performance obligations." They don't catch every variant. They establish a floor, not a ceiling. The bar chart below is live, pulled from /data/edgar-prevalence.json.

The single biggest read in this chart is the spread itself. Some companies are saturated in the language of the structure. Others barely register. That doesn't mean the structure isn't there — it means the regex isn't finding it in the corpus, which can happen for two reasons: the company describes the same arrangement differently, or the company genuinely doesn't run on it. Edition III treats both possibilities as open. The method is published in full, down to the regex sources. Rebuild it yourself.

Top of the leaderboard

A handful of headline numbers, computed live from the same data file. They will shift the next time the scraper runs — these aren't permanent claims, they're a snapshot of what the SEC corpus looked like on the date this page was generated.

Which pattern lights up which company

Below each company is its dominant pattern — the single regex that accounts for most of its hits. This is where the texture lives. Microsoft, Oracle, and Salesforce all lean hard on remaining performance obligations — the accounting line where future cloud backlog is disclosed. Salesforce still tops the cohort on both raw and normalized scoring, driven by an unusually heavy strategic investments count from its venture arm. Amazon, which was nearly invisible in the first scan, now scores in line with the cohort once the regex was broadened to cover its preferred vocabulary — equity investments and unrealized gain rather than the literal "cloud credits" phrase. That's the lesson of round one: low scores in regex work usually mean the regex, not the absence.

What this is, and what it isn't

This is a floor measurement, not a verdict. A regex hit on "cloud credits" doesn't prove anything by itself. It tells you the company uses the language. The volume tells you how central that language is to how they describe themselves to investors. The interesting comparisons are across the cohort, not within any single filing.

Three honest caveats. First, the patterns are still tuned to a specific structural story; a company describing an identical arrangement in entirely different words may still be invisible to them. Second, the corpus is only the four most recent 10-K / 10-Q filings per company; older filings might tell a different story. Third, even normalized counts conflate "high signal" with "high verbosity about the same disclosed item" — Oracle disclosing one performance obligation eight times in one filing scores the same as eight separate obligations.

How to rebuild this

There is nothing to install and nothing to ask for. The SEC's submissions API needs no account and no key — it needs a User-Agent header that identifies you with an email, and it asks you to stay under ten requests a second. Everything else about this scan is below, and the sixteen regexes are published with their sources in the data file this page reads.

That is the entire instrument. Pick your own cohort — the CIKs are above and every other one is in the same API. Write your own patterns; ours are blunt on purpose and three of them return zero. If your version returns different numbers, the cohort and the regexes you chose are part of the disagreement — that's a feature, not a bug. Open for replication. Open for correction. Corrections to [email protected].