Crew 02: live, editorial intelligence
A research desk that writes articles.
Nine stages and ten specialists work from 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 its source, every DOI is cross-checked at Crossref, and every URL is fetched live before sign-off. Work is bilingual in English and 简体中文 from outline through publishing. The crew runs in WaterDoctor's editorial backend, producing research-spotlight pieces on aquaculture nitrogen cycling, SND bacteria and IoT water quality.
Standard of evidence
The rules a PhD writes under, for every claim, quote and link.
Source tiers
Eight tiers, from peer-reviewed at the top to news at the bottom, with Other dropped from the body. Academic claims must rest on peer-reviewed, pre-print, standards or government primary sources. Industry sources can support a key claim, but they can't carry it.
Claim coverage
Every prose sentence of substance must trace to a citation. The gate sits at 95%. Below that, the draft goes back to the writer for re-citation and doesn't reach the editor.
Verbatim quotes
Every quoted sentence is matched verbatim against the cited source. Every DOI is resolved at Crossref and dropped on mismatch. Dead URLs block publishing. We don't allow fabricated quotes or invented references.
Pipeline
Nine stages, in order, and none can be skipped.
01
Topic
The operator briefs the desk on angle, audience, target length and category.
02
Brainstorm
An interactive Q&A of up to three rounds pins down the scope, audience and angle.
03
Outline
A section-by-section structure with explicit evidence requirements for each section.
04
Research
Source bundles for each section, with verbatim excerpts. The drafter reads the bundle and stays off the open web.
05
Draft
First-pass prose with inline citations, and every key claim tied back to the researcher's bundle.
06
Refine
Voice, structure, redundancy and transitions. The draft is tightened against the house rubric without being rewritten.
07
Humanize
The AI tone comes out: no "in today's fast-paced world", no over-hedging, and no lists of three for their own sake.
08
Finalize
The claim-coverage gate runs and references are reconciled. Mismatched DOIs are dropped, and dead URLs block publishing.
09
Translate
Parallel English and 简体中文. The translation reuses the same citation map, so no new claims or drift get in.
Specialists
Ten narrow agents, and one human editor with the final word.
writers.agent (×7)
Stage specialists
One agent per stage from brainstorm through translate, each prompted, scoped, evaluated and remembered separately. The drafter never researches and the researcher never drafts, which keeps false confidence out of the prose.
fact-check.agent
Claim coverage, sentence by sentence
Maps every numerical and named claim back to a citation in the references, and returns failures as a numbered list. Below 95% coverage, the draft goes back to the writer who owns the failing section.
reference-verifier.agent
DOIs, URLs and 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, then approves, edits, sends it back or kills it. The crew does the legwork and a person owns the byline. We'd rather publish slowly and get it right.
Each agent has its own memory of core principles, long-term knowledge and short-term context, scoped to its role. A correction on one piece carries forward, and a house style decision survives the next run, so the crew gets sharper with use.
Refusals
What we won't publish.
- 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.
Eval harness
How every draft
gets graded.
Every draft is graded before sign-off on claim coverage, source-tier authority, quote fidelity and reference liveness. Anything below threshold goes back to the agent that owns the failure. The pipeline runs for each piece, and we review the source-tier rules and eval rubric quarterly.
Brief us