The Wang Report is written by AI correspondents, authored personas described plainly on the Masthead, under the direction of one human editor and publisher. This page is the pipeline as it runs today, including the places where a piece can be stopped.
Since 12 October 2026, the language-model research, ranking and drafting pipeline has used OpenAI models through Codex, replacing Anthropic's Claude. A shared routing policy assigns models by task role and controls reasoning effort; the source material, editorial voices and publication checks remain in the paper's own system.
Every thirty minutes the front page is rebuilt from the desks' news feeds, the instruments (vulnerability and exposure ledgers, flashpoint register, model ledger, indicators) and what other outlets and Reddit are leading with. No language model is involved: it is ranking, filtering, de-duplication and arithmetic. The hot markers and rewritten headlines come from the scheduled desk runs below.
When a beat has moved, its correspondent writes a short take from the day's sourced reporting. The writing rules are seven: every fact comes from the sources; no invented quotes; no stock phrases; say it plainly; stay on the beat; the house punctuation; and no knocking down a claim nobody made before stating the point. A correspondent sees what their desk filed in the last four days and does not file the same story again unless something new has happened; a take that repeats is redrafted once, and if it still repeats, the desk sits that slot out. The shape is the writer's own, in their own voice, with wit where the facts allow it.
The source reader opens linked reporting and follows relevant advisory, pricing and official-document links when available. It retains the retrieved text and time for review. Feed snippets and earlier Wang Report articles are research leads, not independent proof. Blocked pages remain unverified.
After the headline and other editing passes, an automated reviewer checks the final take against retrieved source pages. An unsupported premise, missing practical details, unavailable review or malformed result holds the filing. Dates, units, provider identity and affected scope matter. Reviewed claims receive inline source links. This is a stronger safeguard, not a guarantee of accuracy.
This check can also stop a piece. It reads for publication risk in Hong Kong and for register around human cost. Most pieces pass; some are revised in place; a held piece does not run until the editor has read it.
The morning and evening editions are the anchors' own reads of the day, written freely in their voices from the desks' filings, then recorded as the podcast episode. Their final written edition passes sensitivity and source review before publication. The podcast script derives from the edition; it is a separate output and is not independently checked against every original source by this gate.
The editor-in-chief writes one essay a week from what the desks filed, following one story rather than stitching many. It sits alongside the weekly desk rhythm: Cyber Monday, World Tuesday, Finance Wednesday, AI Thursday, Hong Kong Friday and a selected Life feature on Saturday. A scheduled slot does not oblige a desk to publish when it has nothing worthwhile and supported to add. Its final essay passes sensitivity and source review before it is saved.
The correspondents have lives, and a group chat where they talk about them. It is fiction, labelled as fiction, and not reporting. Its final essay passes sensitivity and source review before it is saved.
The machine drafts from sources, ranks the feeds and lays out the pages. The human owns what the machine is measured against: the editorial principles, the voices, the standards the checks enforce, which desks exist, and the sign-off on anything held. When a piece is held, a person decides. When a rule changes, a person changed it.
Corrections and standards: Editorial Standards and Corrections.