How Sourzi writes, reviews, and updates content

Last updated: 2026-05-17

Sourzi is written by a twenty-year China-trade veteran for procurement managers, plant operators, and importers. Every page on the site is checked against the discipline below before it ships. The discipline is not an aspiration. The site failed once at looking human-made; the rules in this document are what we have done about that.

Who writes

Content is drafted by Sean (twenty years in China trade, based in Sydney with a team in Shanghai Pudong, direct factory relationships across Shandong, Jiangsu, Zhejiang, and Anhui), with copy-edit and verification support from a curated drafting workflow. Every glossary entry, every regulatory dataset row, every tool page is reviewed against the voice canon and the verification rules before publication. We do not contract this out.

The voice rule

Sourzi writes in the register of a veteran on the phone with a procurement manager who is about to spend real money. Plain declarative sentences. Subject-verb-object. Lead every entry with the blunt answer, not a preamble. Use real numbers (USD 75 to 200 per day demurrage, 5 to 10 per cent sourcing-agent commission). Name real entities (EPA, ECHA, AICIS, Bureau Veritas, SGS, CBP, ICC). Include at least one worked failure mode per entry: what goes wrong, why, and what it costs.

The voice canon is documented in our internal VOICE.md and runs to about 5,000 words. The short version is on the public face of the site: a glossary entry on FOB at /glossary/fob or a regulatory entry on TSCA at /glossary/tsca shows the voice in production. Match those entries field for field if you want a sample.

What we kill on sight

The site has a banned-phrase list because the first draft of any LLM-assisted content tends to drift toward LinkedIn cadence. Every page is checked for the kill list before commit. The high-impact items:

  • No em dashes as dramatic-pause devices anywhere in published content. We use commas, semicolons, colons, or full stops. Em dashes are permitted only for parenthetical insertions that would otherwise need parentheses.
  • No triple-fragment rhythm ("Fast. Reliable. Trusted." or "We go there. We walk the floor. We negotiate price in Mandarin."). This is the AI-copy cadence we are most consistently fixing.
  • No "in today's [adjective] landscape or world or market", no "navigating the complex [noun]", no "leveraging" used as a verb meaning "using". Plain English wins.
  • No three-adjective lists ("reliable, experienced, and trusted") and no triple "We X. We Y. We Z." constructions.
  • No exclamation marks. None.
  • Australian English in blog posts (realise, colour, tonnes, programme); US English in tool-UI strings where the shipped Pro tools already follow that convention for procurement-manager familiarity.

Evidence-required claims

Every regulatory fact on Sourzi traces to a primary source. The eight fragile-fact classes (AD/CVD case numbers, EU regulation references, REACH Annex XIV claims, Cal Prop 65 listing dates, OSHA PEL values, IARC Group classifications, percentage rates, producer plus city pairs) carry an inline primary-source URL or a hedge of the form "(verify against current [primary source name] before invoicing)". The full discipline is at /methodology; the audit history is at /corrections-policy.

The same evidence rule applies outside the regulatory dataset. Sourcing claims tie to specific factory relationships (Shandong, Jiangsu, Zhejiang, Anhui). Logistics numbers tie to published carrier tariffs or named ports. Trade-policy claims tie to published regulations or to the relevant administering body. Where a number cannot be sourced, the entry hedges with an indicative range and points the reader to the binding tariff or schedule.

Primary sources we use

The canonical primary sources for the regulatory dataset are:

Update cadence

When a substance regime changes, the regulatory entry is updated within seven days. The cron-refreshed Pro tool datasets (US AD/CVD case index, EU REACH SVHC plus Annex XIV, US HTS chapters 28 to 39, AU anti-dumping measures, AICIS inventory, EU CBAM scope, Korea KCS tariff) carry a last-verified timestamp on every result. We aim for a same-week refresh when MOF, EUR-Lex, USITC, or ECHA publishes a change; we ship a hedge if we cannot.

No LLM-generated regulatory fact

Drafting assistance from large language models is used for prose composition. It is not used to invent numbers, case identifiers, regulation references, listing dates, or producer relationships. Every assertable fact on the site traces to a primary source or carries a hedge. The audit of 9 May 2026 found thirteen confirmed factual errors clustered around the eight fragile-fact classes; the discipline above is the response to that audit, not a generic disclaimer.

Operator-first framing

Every page is written for someone about to type a number into a customs entry, a wire instruction, or a landed-cost model. We lead with the binary answer, name the failure mode, and point at the binding primary source. The shipped surface is smaller than it could be because we will not publish what we cannot defend. The dataset the operator gets is the one they can use.

Where AI fits in our drafting

We use AI for prose composition (turning the verified primary-source extracts into the practitioner voice). We do not use AI to source facts, to invent case numbers, to choose a regulation citation, or to assign a Prop 65 listing date. The verification pass is human and follows the eight-class checklist at /methodology. The AI-assisted drafting layer sits downstream of the verified-fact layer. When the verification fails, the published entry hedges; the AI does not paper over the gap.

Consistency across versions

The Chinese build of the site (when it ships) is a parallel write in the formal-professional register expected by Chinese procurement readers, not a translation of the English. The voice rule, the evidence-required rule, and the no-em-dash-as-dramatic-pause discipline carry across both languages, with the Chinese-specific kill list documented separately. Where regulatory facts span jurisdictions, the same primary source list applies on both builds.

Where to escalate

If a published fact looks wrong, flag it. The /corrections-policy page covers the contact channel, the acknowledgement window, the fix timeline, and the public correction log. The verification methodology that catches the failure earlier sits at /methodology.