We don't generate articles. We staff a research desk to write them.
Nine stages. Ten specialists. A source registry tiered from peer-reviewed down to news, with the bottom tier dropped from the body. Every quoted sentence is matched verbatim against the source it cites. Every DOI is cross-checked at Crossref. Every URL is fetched live before sign-off. Bilingual English / 简体中文 from outline through publish. The crew runs in WaterDoctor's editorial backend, shipping research-spotlight pieces on aquaculture nitrogen cycling, SND bacteria and IoT water quality.
The rules a PhD writes under. Every claim. Every quote. Every link.
Eight tiers. Peer-reviewed at the top. News at the bottom. Other dropped from the body. Academic claims are restricted to peer-reviewed, pre-print, standards or government primary — industry sources may support, never carry, a load-bearing claim.
Every prose sentence of substance must trace to a citation. The gate sits at 95%. Below the threshold, the draft goes back to the writer for re-citation, never forward to the editor.
Every quoted sentence is matched verbatim against the cited source. Every DOI is resolved at Crossref and dropped on mismatch. Dead URLs block publish. No fabricated quotes. No invented references.
Nine stages, in order. No skips. No stage is optional.
- 01 · topic
The operator briefs the desk — angle, audience, target length, category.
- 02 · brainstorm
An interactive Q&A — up to three rounds — until the scope, audience and angle are pinned down.
- 03 · outline
A section-by-section structure with explicit evidence requirements per section.
- 04 · research
Per-section source bundles with verbatim excerpts. The bundle is what the drafter reads — never the open web.
- 05 · draft
First-pass prose, citations inline, every load-bearing claim tied back to the bundle the researcher delivered.
- 06 · refine
Voice, structure, redundancy, transitions. The draft is tightened — not rewritten — against the house rubric.
- 07 · humanize
The AI tone is stripped: no "in today's fast-paced world", no over-hedging, no triplet listicles for their own sake.
- 08 · finalize
Claim-coverage gate runs. References are reconciled. Mismatched DOIs are dropped. Dead URLs block the publish.
- 09 · translate
Parallel English / 简体中文. The translation reuses the same citation map — no new claims, no drift.
Ten agents. Each one narrow. One human editor with the final word.
- writers.agent (×7)Stage-specialist crew
One agent per stage from brainstorm through translate. Each is prompted, scoped, evaluated and remembered separately. The drafter never does research; the researcher never drafts. Boundaries are how false confidence is kept out of the prose.
- fact-check.agentClaim coverage, sentence by sentence
Maps every numerical and named claim back to a citation in the references. Returns failures as a numbered list, not a verdict. Below 95% coverage, the draft is sent back to the writer that owns the failing section.
- reference-verifier.agentDOIs, URLs, source tiers
Resolves every DOI at Crossref and drops mismatches. Fetches every URL live and drops the dead ones. Enforces the source-tier allowlist on academic claims. References without primary attribution don't make the article.
- editor (human)Reads every word, then ships
A senior reviewer reads the piece end-to-end. Approves, edits, sends back, or kills. The crew does the legwork; a person owns the byline. We will publish slow and right.
Each agent has its own memory — core principles, long-term knowledge, short-term context — scoped to its role. A correction made on one piece carries forward. A house style decision survives the next run. The crew gets sharper with use; we don't restart from zero each time.
- Articles where a number couldn't be sourced.
- Pieces that quote a "study" without naming the study.
- Top-10 listicles that don't draw on direct project experience.
- "What is [buzzword]?" explainers that don't have an opinion.
- Anything that sounds like a generic AI-agency LinkedIn post.
Every draft is graded before sign-off on four axes: claim coverage, source-tier authority, quote fidelity, and reference liveness. Anything below threshold goes back to the agent that owns the failure — never to the editor. The pipeline runs per piece; the source-tier rules and the eval rubric are reviewed quarterly.
Brief us →