Check the vendor's story
against primary sources.

A vendor's narrative is not, by itself, material for a decision. You check it against primary sources first, then use it. 3-01: Companies Don't Write Their Own Code showed office and core standing in parallel, with the tax levied twice over. What decides whether to keep paying that tax or to leave is not the story a vendor tells; it is the primary sources. This chapter lays out the procedure for that check. The conclusions of 3-03 and 3-04 were all obtained with it.

AI Leans Toward the Story Too

Checking a narrative was long the work of reporters, researchers, and lawyers, because it takes time and effort. AI lowers that effort considerably. Arrange five years of statements in chronological order; see whether a 2020 claim and a 2024 claim sit on the same line — that kind of sweep is genuinely fast.

But AI leans toward the story too. That is the starting point.

• The training data is unevenly thick — centered on English, on Western sources, on online discussion. Microsoft Learn, GitHub READMEs, and large vendors' blogs are in it in volume. Critical forum posts, regional papers, print, and non-English discussion are thin.
• Weight settles on authority — official documentation, CEO statements, and corporate blogs get placed first as trustworthy sources. GitHub issues and Reddit posts are handled lightly.
• Weight settles on the majority — when many articles make the same claim, AI treats it as fact. The volume of discussion stands in for how certain something is.
• The story of whoever owns the information infrastructure is amplified — the sources AI learns from are GitHub, Stack Overflow, technical blogs, and corporate documentation. Microsoft owns GitHub, owns LinkedIn, owns npm, and has invested in OpenAI. When the teller owns the pathway, the telling comes back through AI.

There is one more. AI also matches its questioner. A model trained on human feedback leans toward answers the user will be satisfied with. Ask "is this claim correct?" and affirmation comes back; ask "isn't this exaggerated?" and the opposite comes back. The shape of the question sets the direction of the answer.

So You Design the Checking

When an AI drawn from one body of training data checks that same body, the same lean stays on both sides. AI checking AI only reinforces the shared assumption (2-17). So you design the checking first.

• Combine AIs built by different methods — add a Deep Research / Deep Search type, which requires citations to primary sources, to the ordinary chat type (Gemini Pro, Claude, ChatGPT). The former emits source URLs, so the lean toward authority drops.
• Ask more than one vendor — put the same question to Anthropic, Google, and OpenAI, and lay the answers side by side. Where they agree, confidence rises; where they split is where a human looks.
• State a critical hypothesis — ask explicitly, "is this claim exaggerated?" "is there evidence on the other side?" A question that expects agreement draws agreement.
• End at primary sources — AI output is a hypothesis. Official release notes, GitHub issues, court filings, charters and bylaws — in the end a human reads the original.

The four cases below were checked with this procedure. In all four, the shape of the narrative is different.

Case: WordPress — When Authority Concentrates in One Person

Matt Mullenweg — co-founder of WordPress and CEO of Automattic — has been publicly at odds since 2024 with WP Engine, a major hosting company. Public statements, court filings, blog posts, and conference talks exist in volume, so the material for checking is there.

His narrative's main claims come to roughly four. WP Engine free-rides on the WordPress ecosystem. WP Engine distorts the form of WordPress. WordPress.org is his personal property. Blocking WP Engine is a legitimate act to protect the community.

Run the procedure. First, separate the claims into claims of fact, opinions and evaluations, and figures of speech. "Free-riding" is an evaluation; "improper use of the trademark" is a claim of fact; "protecting the community" is an evaluation. Only claims of fact can be checked.

Next, apply primary sources. The WordPress Foundation's trademark policy, ten years of community discussion, statements at WordCamps. Lay out the timeline too — 2014, when WP Engine brought in Heather Brunner as CEO; 2018, when Lee Wittlinger of Silver Lake joined the WP Engine board; and the 2024 block. There were periods when WP Engine was welcomed as a WordCamp sponsor, and periods when it was listed on the WordPress.org recommended-hosting page.

Finally, apply third-party records. In WP Engine v. Automattic (from October 2024), both sides file their claims under oath. The preliminary order of December 2024 made Automattic's blocking of WP Engine from WordPress.org subject to injunction. That draws a different line from the claim that "WordPress.org is my personal property, so I decide freely."

What comes into view is that the boundary between the individual and the foundation — the relationship among WordPress.org, Automattic, and the WordPress Foundation — is re-placed with each telling. This is not a conclusion that makes anyone a villain. Everyone tells a story shaped to their situation. The more influence a person has, the further that story reaches. Which is why you check.

Case: Node.js — When No One Holds the Whole

WordPress was the shape where authority concentrates in one person. There is an opposite shape: no one is placed to hold the whole. Take whether to adopt Node.js for production work, and it becomes visible.

Checking gives this.

• The Node.js runtime belongs to the OpenJS Foundation; the npm registry is owned by Microsoft. The "vendor neutral" story applies to the first half.
• Founder Ryan Dahl gave a 2018 talk, "10 Things I Regret About Node.js," and moved on to build Deno.
• event-stream, ua-parser-js, colors, SAP — supply-chain incidents occur almost yearly.
• Inside the story of "backed by major companies" there are sections where a billion-dollar business rests on one or two unpaid volunteers.

WordPress is the shape where one person holds too much; Node.js is the shape where no one is placed to hold the whole. Both are questions about the shape of governance, pointing in opposite directions.

Case: Linux Distributions — When the Promise Changes Mid-Way

The third shape is a corporate steward changing the promise partway through. It bears directly on choosing a Linux distribution for server use.

• CentOS 8 — on December 8, 2020, the support period, which had been stated as running to 2029, was changed to end at the close of 2021. Eight years earlier.
• Red Hat's statements — 2014, "we will protect CentOS's independence"; 2019, "independence is preserved after the IBM acquisition"; 2020, "CentOS 8 is ending." Placed on a timeline, the three do not sit on one line.
• Ubuntu / Canonical — Snap as the default (from 2020), the Ubuntu Pro registration requirement (from 2022), the Mir and Unity independent paths and their withdrawal. Direction has changed every five to ten years, on a private company's business judgment.

WordPress is the shape where one person holds too much; Node.js is the shape where no one holds the whole; Linux distributions are the shape where the promise changes mid-way.

Case: Microsoft's "Return to Native Apps" — Measuring the Slogan's Reach

The fourth is the shape where the owner of the information infrastructure tells the story. The subject is the "return to native apps" that Satya Nadella announced on April 29, 2026.

Checking makes the scope visible.

• At the OS shell layer (Start menu, taskbar, File Explorer), the move to "100% native" with WinUI 3 and .NET 10 Native AOT is under way. The May 2026 Windows 11 update KB5083631 is the result of it.
• *Inside the same Microsoft, the Microsoft 365 division has chosen another path.* The new Outlook and Teams are both WebView2 wrappers. Migration from classic Outlook was postponed a year (April 2026 → March 2027), and support continues through 2029.
• At the third-party layer, Microsoft itself is advancing React Native for Windows (0.81 / 0.82 in 2026). As long as the developer-side economics hold, Electron and React Native continue.

So the "return to native" is genuinely happening, within the limited scope of the OS shell layer. The reach of the slogan and the reach of the implementation are not the same size. When you hear "100% X" or "return to Y," first ask what the 100% applies to.

Gemini Pro Leaned Toward Microsoft's Story

This case contains a moment where AI's lean appeared directly.

The sources are the Gemini Pro document on that subject, and Microsoft's official documentation. The former was part of a study following the 2026 "return to native"; the latter is text still present in the EF Core documentation.

What happened here is a combination of the leans listed at the start of the chapter. Weight settled on an authoritative source, the industry's repeated telling that "Native AOT has matured" passed through unchecked, and the story of the party that owns the information pathway came back via AI. The gap appeared only once an AI built by a different method — a Deep Search type that requires citations — was applied. This is where designing the checking pays.

Checking Debian — The Ground on Which 2-02 Stands

The Linux distribution case has another result from the same procedure: Debian.

• The 1997 social contract — what the distribution promises its users and free software is published as a document.
• The 1998 constitution — who holds which decision, how the project leader is chosen, and how a decision can be overturned are all fixed in a document.
• No owning company — there is no corporate entity to be acquired in the first place.
• Assets are held by SPI (Software in the Public Interest) — the trademark and the funds sit with a non-profit, and no pathway is in place for direction to change on a business judgment.

What happened with CentOS was that a corporate steward could change the term of the promise. What repeated with Ubuntu was that direction changes on a private company's business judgment. In Debian that pathway is absent. What makes the promise is not the technology but the structure of governance.

2-02: Give the AI a PC of Its Own decides to turn one company PC into Debian and hand it over, and this check is why. A foundation is better when it keeps the same shape for a long time. So you pick the one whose structure keeps the same shape for a long time. It sits there as the result of a check, not because it is pleasant to use.

The Five Practices, Generalized

The procedure common to the four cases comes to five practices.

1. Extract and Separate the Claims

Before taking a narrative in, write out what it claims. Have AI do it. Separate each claim into (a) a claim about objective fact, (b) opinion and evaluation, (c) figures of speech and expressions of feeling. Only (a) can be checked.

2. Apply Primary Sources to Claims of Fact

Take the extracted claims of fact and put them against the originals.

News articles are secondary sources. Where possible, read the statements and numbers inside an article in the original. Asking AI where an article's quote came from gets you to the original faster.

3. Lay Out the Timeline and Check the Fit

Do past statements and present statements sit on the same line? If not, is the reason explained? Laying out Red Hat's 2014, 2019, and 2020 statements is this practice. The most checkable face of any narrative is the time axis. People speak to fit the present, and past statements stay on record.

4. Apply Third-Party Records

Beyond the parties' own statements, apply third-party records. Court filings, audit reports, legislative and committee testimony, exit interviews, citations in academic work. Applying the preliminary order in the WordPress case is this. Third-party records surface the faces a party's own telling does not touch.

5. Separate What Is Known from What Is Not Known

At the end, write separately: facts confirmed, facts that could not be confirmed, points where the opposite claim has weight, and points where information is missing. If information is missing, write that it is missing. Not rushing the conclusion is the practice here.

These five are not a tool for denouncing anyone. *Use them to keep your own judgment from going wrong.* What you write is the contrast — "the official document says this" — and nothing more or less.

The Conclusions of the Next Two Chapters Came from This Practice

The stance the Shift part takes toward vendor narratives rests on this chapter.

link:/en/ai-native-software/sovereignty/[3-03: Digital Sovereignty — The Microsoft Problem and the Trump Problem] is the result of checking the narrative that Microsoft 365 was the default on both economics and safety. The trajectory of per-seat pricing, the text of the CLOUD Act, published telemetry information, the official description of where Copilot sends content — it says the premise has inverted after applying primary sources.

link:/en/ai-native-software/sier-uneconomic/[3-04: The Structural Uneconomy of the SIer Model] does the same. It opens the narrative that "outsourcing is cheaper" into checkable form: where the upstream judgment stays, and what one turn of the loop requires.

Neither is a conclusion taken from a vendor's telling as given. Both passed through the five practices of this chapter. Which means a reader can run the same check.

Summary

This chapter laid out the procedure for checking a narrative against primary sources.

• AI leans toward narratives too — the uneven thickness of training data, authority, the majority, and the structure that amplifies the story of whoever owns the information infrastructure
• So you design the checking — combine AIs built by different methods, ask more than one vendor, state a critical hypothesis, and end at primary sources
• Four cases — WordPress (authority concentrating in one person), Node.js (no one placed to hold the whole), Linux distributions (the promise changing mid-way), Microsoft's "return to native" (slogan and implementation not the same size)
• Gemini Pro summarized EF Core's AOT support as "nearly perfect"; Microsoft's official documentation says "recommend against" and "highly experimental"
• Debian's 1997 social contract, 1998 constitution, absence of an owning company, and SPI holding the assets — this is the ground on which 2-02 places Debian as the foundation
• Five practices — extract the claims → primary sources → timeline → third-party records → separate what is known from what is not known

AI is good at making narratives. AI is good at checking them too. But when the AI on the making side and the AI on the checking side come from the same training data, the lean stays. Placing an AI built by a different method on the checking side is a human choice.

The next chapter applies these practices to Microsoft. Per-seat pricing, the CLOUD Act, telemetry, and dependence on the US government — it examines where the premise on the office side inverted.

Related articles

• 3-01: Companies Don't Write Their Own Code — Office and Core, Two Parallel Worlds
• 2-02: Give the AI a PC of Its Own — The Machine the Independence Part Runs On
• 3-03: Digital Sovereignty — The Microsoft Problem and the Trump Problem