SWFT wants readers to understand both the claim and how much confidence to place in it. Every article shows its sources, the kind of evidence behind each important statement, where the work came from, and the date of its source window.
What do SWFT's evidence labels mean?
Observed artifact (OBS)
SWFT inspected a public artifact or dated event, such as a page, transcript, repository, release, or document. “Observed” proves what the artifact showed on its recorded access date. It does not prove every claim inside that artifact.
Independent study (STUDY)
The claim comes from research with a described method, sample, evaluation, or analysis. SWFT keeps the population and limits visible. One benchmark or study does not become a universal rule.
First-party report (SELF-REPORT)
A person or organization is describing its own product, system, customers, or results. First-party accounts can contain valuable operating detail. SWFT attributes them and does not present them as independent comparisons.
Analysis (INFERENCE)
This is SWFT's own definition, classification, recommendation, or conclusion drawn from named evidence. Readers can inspect the support and disagree with the interpretation.
The short codes stay in Markdown and data exports so machines and repeat readers can identify the category reliably. The visible labels use everyday language.
How does SWFT choose sources?
SWFT prefers primary research and corrections, official documentation or artifacts, and detailed first-party operating accounts. Independent reporting follows. Secondary summaries help us discover a topic, but they should not carry a factual claim when the underlying source is available.
AI Engineer's 2026 program, Latent Space, and Smol News are useful radar for current vocabulary and debates. SWFT follows those signals to the talk, paper, repository, release, specification, or operator account behind a claim whenever possible.
Commercial affiliation does not disqualify a source. It changes what the source can establish. Product capabilities and company metrics remain first-party reports until independent evidence supports a stronger statement.
How does SWFT choose company cases?
A company enters the comparison when public first-party evidence describes a repeatable path from incoming work through agent action, verification, and a handoff or decision. General use of a coding assistant, one impressive demo, or a product announcement is not enough by itself.
SWFT prefers cases with several dated sources, concrete operating detail, visible artifacts, or reported outcomes. A case can still qualify when the implementation is private, but the access limit and missing evidence must be stated. Inclusion is not a ranking, endorsement, or claim that the archive is complete. It means the public record is detailed enough to teach and compare without inventing the missing parts.
How does SWFT measure scale and industry adoption?
The scale reference records observations, each tied to a source and a defined unit. Concurrent agents are workers active at the same time. Cumulative runs count attempts over a period. Deployed agents may mean reusable systems or configurations. Adoption measures people, teams or eligible work. Product capacity describes what a vendor says a system can support; it is not an observed customer run.
Each observation preserves the task, setting, measurement window, source date, check date, result and known limits. Approximate values remain approximate. Unknown counts and costs stay unknown. We do not estimate agent-hours from a peak worker count or turn tokens into agents. A research result is attributed to its authors until a separate review supports stronger wording.
The industry overview includes operating accounts, pilots, research experiments and documented platforms, with those distinctions visible. The detailed company comparison has the narrower workflow criteria above. Neither collection is a representative sample, so its size cannot establish an industry penetration rate. Sector labels describe the organization's business, not its level of AI maturity.
Historical measurements remain attached to their original period. A new editorial date does not imply an old source was rechecked or a system still operates at the same scale. New measurements should preserve earlier observations where they help readers understand change.
What rules apply to claims?
- Numbers name the organization, population, date, and method available in the source.
- Benchmarks explain what tasks and scoring rules they cover.
- Changing maps and product descriptions show their source-window date.
- Quotations stay short unless SWFT has permission to republish the work.
- “Unknown” stays unknown. Missing public evidence is not proof that something does not exist.
- A model-generated summary never replaces the underlying source.
How does SWFT use AI in editorial work?
SWFT Editorial guides, comparisons, references, and company cases are AI-drafted from cited public sources. Their institutional byline and authorship disclosure appear in HTML, Markdown, feeds, and structured metadata. Ben Guo's personal byline appears only on his essay, which SWFT republishes at his request.
Each admitted page has a dated review record naming the second AI editorial-review agent, Codex or Devin, the scope of that check, the evidence type, the closest overlap, the risk if a claim is wrong, and the maintenance trigger. SWFT does not imply that a human or site owner reviewed a page, and an AI editorial check is not qualified clinical, legal, or financial review.
AI drafting does not change the evidence standard. Every factual claim still needs a source that supports it, first-party reports stay labeled, and SWFT Editorial remains responsible for corrections.
What can discovery automation publish?
It can propose sources. Discovery automation can find, clean up, and group possible sources, but it cannot create or change a published page. Before a new page goes live, a separate review record has to name what the page does for readers, what its claims can support, how its evidence fits, how it differs from existing pages, who reviewed it, the risk, and the plan for keeping it current.
One record drives the HTML page, Markdown version, metadata, structured data, feeds, and source list. A question section is added only when it resolves a follow-up decision the article body does not already answer, and the same wording then powers visible and structured representations. SWFT does not maintain hidden answer-engine copy.
The scale data preserves the same observations, units, reporting windows and limitations shown in the scale reference. Its schema is versioned. Approximate, minimum and capacity claims keep their qualifiers; missing measurements stay unstated.
Corrections
How do I request a correction?
Send @hraness a public message on X and include the SWFT page, the disputed sentence or record, and the strongest source you have.
SWFT checks the disputed claim and its evidence, then corrects the page and every copy of it: Markdown, feeds, structured data, and data exports. A meaningful correction changes the modified date. If the earlier version could materially mislead, the page receives a visible note. Spelling and presentation fixes that do not change meaning can remain silent.
Correction history
September 8, 2026: clarified that Shopify's shared Slack conversations are visible inside the company, not necessarily to the public. Made explicit that Vercel's count of deployed agents is not a count of simultaneous workers. Company comparisons now show each case's own source-check date instead of suggesting that every source was refreshed together.
September 23, 2026: Does AI coding actually work? described Stripe's roughly 7,000 Minions pull requests in one week as merged work; Stripe did not say whether all of them were merged, so the page now counts pull requests. What is harness engineering? described HarnessTax as a five-fold cost spread across model–harness pairs with one-attempt scoring; the study reports that the same model can reach similar success at up to five times the token cost, with three attempts per task. The coding-agent harness is a cost lever now notes that the authors cite a Portkey post for the name “harness tax”.
The DORA 2025 errata is one useful model: keep corrections close to the work they qualify and preserve the history.