Skip to content

Editorial process

AI models draft OpenPWA entries under editorial direction. Entries cite sources and pass automated checks before publication; those written by the publishing pipeline also pass an independent AI fact-check of each claim. No person signs off each entry, and corrections are recorded publicly.

Reference and guide entries show a publication date and byline and name a reviewer in structured data. Ecosystem and compatibility entries do too, as do news articles. For device tests, a human operator performs each step on a real device. AI drafts the analysis and write-up, which undergo the same checks as other entries.

OpenPWA Editorial, this project’s editorial function, publishes OpenPWA. Entries use Carey Gill as their byline and reviewer name. This pen name represents the whole editorial team, not an individual. It provides an accountable signature in bylines and structured data as different people revisit topics and maintain their facts over time. The editors page describes the review step behind the name.

  1. Topics come from a coverage matrix of the installable web (manifest and service workers; installation and capabilities; storage and notifications; performance and platforms) and observed search demand. Entries link to MDN for the API surface. They cover how platforms launch an installed web app, which store rules apply, and what device tests measured.
  2. Sources should be primary documents, including specifications, browser-vendor documentation and release notes, bug trackers, and store and platform policies. An evergreen entry declares at least two source URLs in frontmatter; a news article declares at least three. Some compatibility rows still cite the aggregator caniuse.com and carry low confidence (see Methodology).
  3. Large language models draft entries from the sources gathered in step 2. A written authoring contract sets the structure, depth standard, and sourcing rules. AI authoring supports the breadth of topics covered.
  4. Each deploy runs automated checks without a language model. They check source reachability, agreement between compatibility rows and the browser-compat-data they cite, English and Chinese pairing, and complete, internally consistent publication metadata. The automated pipeline also checks the body of pages it writes against the template for their kind. A failed check blocks publication, and publishing workflows have no switch to skip it.
  5. Before the pipeline publishes an entry, a different AI model reads the English and Chinese versions separately and checks each claim against its cited sources, including the title and description. One unsupported claim withholds the entry. There is no per-entry human sign-off. After publication, a share of new entries is queued for an audit by a person or a stronger model. The reviewer name represents this process; see Carey Gill.
  6. The lastVerified date determines when an entry is queued for re-checking: after 180 days, or 90 days for entries whose facts depend on store or platform policy. Store terms and platform policies can change without notice, and browser support changes with each release.

A substantive correction appears in the changelog, labelled as a correction, with the date and a one-line summary. News articles also get a dated Correction (YYYY-MM-DD) section appended to the article, stating what was wrong and what it now says. Evergreen entries are corrected in place, with the changelog recording the change.

The entry’s lastVerified date moves. Its publication date stays fixed: an automated check compares it with the first date recorded in repository history, preventing corrections from appearing as fresh reporting. Typographical and formatting fixes that leave the meaning unchanged need no correction note.

To report an error, name the page and claim, and supply a source. See Report a problem.

Coverage, placement, and ranking in site lists are unpaid. Vendor announcements serve as sources; their framing alone does not establish a conclusion for reporting. Claims without sources receive no generated illustrations, charts, or data.

← Back to About. Methodology gives the verification details.