Measured evidence for supplier decisions, free to cite with its dates.

For editors and researchersPublication-ready assets

Research media kit

Three findings are packaged below as 1200 × 675 graphics, with the underlying dataset, DOI, copy-ready caption and the limitation that must travel with the number. The PNG is ready for publishing; the SVG remains sharp when resized.

1. Can an English supplier name be confirmed?

Download PNG Download SVG

Frame: 106 Chinese manufacturers registered with NHTSA vPIC, searched using the registered English name held by that regulator. Taken: 12 August 2026; two runs seven hours apart returned the same counts.

Limit: This sector-specific frame is plausibly an upper bound. It is not an estimate for all Chinese suppliers. A matching English name identifies a candidate; it does not prove the company is active, capable, connected to the seller, or safe to pay.

Copy-ready caption:

Currawong found that 47 of 106 registered English names (44.3%) in a China-manufacturer frame produced a candidate whose own registration carried the same English name. The NHTSA vPIC sector frame is plausibly an upper bound, not a population estimate. Measured 12 August 2026.

Read the study and method · Download the CSV · DOI 10.5281/zenodo.21907920

2. How often does registered scope mention import-export?

Download PNG Download SVG

Frame: 264 companies on three official provincial excerpts of MIIT’s sixth ‘Little Giant’ batch: Shandong 136, Anhui 113 and Yunnan 15. This is a census of that frame. It does not sample all Chinese suppliers. Queried: 8 August 2026.

Limit: Scope wording does not prove current export capacity, required licences, product capability, or shipment history. Absence of the wording is a follow-up signal. It is not a verdict.

Copy-ready caption:

Currawong found import-export wording in 197 of 264 registrations (74.6%) across three official provincial excerpts of MIIT’s sixth ‘Little Giant’ batch. Registered scope wording is not proof of current export capacity. Queried 8 August 2026.

Read the study and method · Download the CSV · DOI 10.5281/zenodo.21858946

3. How quickly do Chinese company records move?

Download PNG Download SVG

Frame: the same 264-company ‘Little Giant’ frame, using registered change dates and category labels. Queried: 8 August 2026. Within the frame, 207 of 264 changed within 24 months and the median lifetime change count was 39.

Limit: A registered change is not itself a risk finding. The field that changed, its before-and-after facts and its relevance to the transaction determine whether it matters.

Copy-ready caption:

Currawong found that 145 of 264 registrations (54.9%) in the same ‘Little Giant’ company frame changed within the prior 12 months. A registered change is not itself a risk finding; the changed field and transaction context matter. Queried 8 August 2026.

Read the study and method · Download the CSV · DOI 10.5281/zenodo.21858946

Nineteen original studies

• Before payment: establish the current registered entity and compare it with the invoice issuer and payee. An English name, platform profile or empty risk screen cannot do that alone.
• Before signing: match the registered Chinese name and 18-character code to the contract party. Use the name-change and platform-field studies to see why the trading name is not enough.
• Before inspection: treat business-scope wording and regulator profiles as signals, not proof of current production capability. Decide what the inspection still has to establish.
• Registry versus database. 17 companies compared (registry read 5 September 2026). The same 17 companies read out of the official registry and out of a commercial database eight days apart. 115 of 118 compared field values agreed. No registered-capital amount disagreed. But the two sources worded the same live status differently for 15 of the 17, so a literal text comparison raises 15 false alarms.

Eighteen studies measure something a buyer may need to decide. The last measures our build process and is included for people who ship multilingual interfaces.

Every study indexed here draws on one of these measurements. The frames are deliberately small and completely enumerated: a census of 264 companies answers a narrower question than a survey of thousands, but it answers it without a sampling assumption.

• Can you find a Chinese supplier from its English name? (measured 12 August 2026). 106 Chinese manufacturers taken from the US NHTSA vPIC register, each searched by the registered English name that register holds. 57.5% returned a candidate, and for 77.0% of those the top candidate’s own registration record carried the same English name — an end-to-end confirmable rate of 44.3%. On the 22 companies carrying both, brand or short names returned far more often (90.9% against 59.1%) and were mostly the wrong company.
• 1 in 4 had no import-export wording in registered scope (queried 8 August 2026). a census, not a sample: every company on three official provincial excerpts of MIIT’s “Little Giant” list, read for scope wording, former names, capital fields and status. Findings: 74.6% carry import-export wording (so 25.4% do not), 53.4% have changed legal name, 37% show paid-in below subscribed capital, all 264 resolved as live registrations.
• How often Chinese company records change (queried 8 August 2026). the change logs of the same 264 companies: 54.9% amended their registration within 12 months, 78.4% within 24; median lifetime change count 39; median days since last change 320.
• Five of eight official Chinese sources would not open (measured 8 August 2026). eight official hosts requested three rounds each from inside mainland China with same-session controls. Five refused their front page. None of the three that loaded carried an English-version marker. One court host served a browser and refused a command-line client.
• How much does a check digit protect you? (measured 9 August 2026). 2.9 million deliberate transcription errors run through the USCI (GB 32100-2015) and VIN (49 CFR 565) check algorithms with a fixed seed. USCI caught 100% of single errors and 100% of data-position swaps. VIN caught 92.5% and 86.8%. Every single-substitution miss traced to same-value transliteration.
• What each segment of a Chinese USCI tells before you search (mapped 5 September 2026). The 18-character Unified Social Credit Code encodes registrar authority, entity type, GB/T 2260 jurisdiction and a mod-31 check digit. Each segment mapped to what a foreign buyer can read cold from a code alone, per GB 32100-2015.
• Chinese company entity types: what each one means for buyers (classified 5 September 2026) — 11 distinct entity type strings appeared in 46 manufacturer registry records. Each type classified by liability structure, export eligibility and contract counterparty risk for foreign buyers, with the bracket normalisation problem documented.
• Is one search result confirmation? (measured 18 August 2026). the same 106-company vPIC frame as the English-name study, this time looking at what comes back rather than whether anything does. 42.5% returned no candidate, 50.9% returned exactly one row, 6.6% returned more than one. Of the 84 rows, 13.1% were registered in Hong Kong (ten of eleven carrying no 18-character code) and 7.1% were already deregistered. The hypothesis it set out to test — crowding by investment and holding entities — did not reproduce, and that is reported on the page.
• How many Chinese trailer makers are NHTSA-registered? (captured 9 August 2026) — an exhaustive count of NHTSA's public vPIC manufacturer database: 22,881 registrants across 92 countries, 2,589 from China, and 106 Chinese trailer manufacturers — the full 106-name list published as CSV.
• Certificate registries: who can actually open them? (measured 9 August 2026). eight certificate-verification registries probed three rounds each over two network paths with same-session controls. 5 of 8 answered a mainland-China connection. UL's Product iQ refused a scripted client on every path. Both US government hosts produced no response at all through the proxy egress. On the direct path one served (NHTSA vPIC, 200) and the other refused (FCC, 403). Those are different facts.
• China company registry availability (measured 5 and 8 August 2026). the GSXT question specifically, with control hosts, a user-agent pair experiment, an independent same-method replication three days later, and one published conclusion we withdrew, documented on the page. The same instrument has been turned on its operator: our own company, run through our own pipeline, output unedited.
• What a China FDA registration tells you: 4,973 measured (source export dated 10 August 2026) — all 41,745 China listing records in openFDA deduplicated to 4,973 establishments. Declared roles, name continuity and agent concentration are reported as aggregate data.
• One platform tells you who the company is, the other tells you what it can make (read 28 August 2026) — subject-identity fields on 94 supplier profiles across Alibaba.com and 1688. The registered Chinese name and legal representative were present on all 59 sampled 1688 profiles and absent from all 33 Alibaba.com profiles — absent from the page source, not left blank by suppliers.
• Told to check the regulator’s record? It is 2.1 years old at the median (queried 28 August 2026). all 25 fields the US vehicle regulator holds on 106 China-registered trailer manufacturers. Every record names a person; none states what the company makes; 19.8% have not been touched in five years.
• 264 US recalls, 46 Chinese firms, and 83% are still open (queried 28 August 2026). every device enforcement record openFDA holds for a recalling firm in China. Half the records come from three firms. The median open case has been open 875 days. A recall is evidence that a corrective action happened, and never evidence of a bad supplier.
• Registered capital, status, scope: what the three fields a buyer reads actually guarantee (queried 28 August 2026). the registry record for 46 Chinese manufacturers. Only 34.8% show paid-in capital matching the subscribed figure. Three different status words all mean the company is live. Business scope isolated a manufacturer from a trader in exactly none of the 46.
• Three of the six free risk checks never fire on an ordinary cohort (queried 28 August 2026). six company-level public risk signals run across 46 Chinese manufacturers. Court hearing notices appear on 37.0%, enforcement on 8.7%, and the abnormal-operations directory, serious-violation list and tax-arrears notice on none of the 46.
• One model translated, another checked: 11 of 98 wrong (reviewed 12 August 2026) — 98 interface strings across 15 languages on this site, back-translated by a second, independent model before shipping. 11.2% came back wrong, and the four automated string checks that ran first caught none of them. The most expensive one told Arabic-reading buyers to ask a Chinese factory for a work permit instead of a business licence.

What this desk has actually measured, counted on 22 August 2026

The studies below are the written-up ones. Underneath them sits a larger body of measurement that reaches the guides without ever becoming a study of its own. A fill-rate table sits on one page, a probe result on another. This is a count of the whole thing, taken mechanically from the pages themselves.

Thirteen distinct measurement sets, carried by 99 English pages, with query dates spanning 8–22 August 2026. Every table on this site states which set it came from and the date it was queried. A measurement without a date stops being evidence at some point, and nobody can tell when.

Two honest notes on this count. It is a mechanical inventory. We counted the evidence markers embedded in the pages, not the quality of what they show. And several sets are read along different dimensions on different pages: the 45-manufacturer cohort alone appears on 40 pages, which is one sample answering forty questions, not forty studies.

That distinction matters enough that every page carrying it says so in its own footnote. If you are citing any figure from this site, cite the underlying set and its query date rather than the page you found it on. Method for the largest set: the NHTSA manufacturer census.

Original study94 supplier profiles · two platforms

One platform tells you who the company is. The other tells you what it can make.

If the profile does not show the registered Chinese name and 18-character code, ask for the business licence before payment. We read 94 profiles across Alibaba.com and 1688 to see whether those identity fields were present. The result split cleanly by platform.

Measured 28 August 2026 · Frame: six product categories, 33 profiles on Alibaba.com and 61 on 1688 · Prepared by Bao L. Zhou, Currawong’s China-side verification desk.

Platform subject fields: the numbers

The last three rows are the reason this study distinguishes three states rather than two. On 1688 the code field exists on every profile and eight suppliers simply left it empty. That is a supplier behaviour. On Alibaba.com there is nothing to leave empty: regCode, legalRep and regCapital do not appear anywhere in the page source. That is a platform decision, and only the three-state reading tells them apart.

The badge verifies the factory, not the entity

A reasonable objection is that paid or independently assessed suppliers might carry more identity fields. They did not in this frame. Our Alibaba.com sample spans nine combinations of assessor and tier: SGS, TÜV Rheinland and no assessor at all; a “Verified” supplier badge, a “Verified” factory badge, and no badge. All 33 are identical on these fields.

What the assessed profiles do carry is substantial and worth reading: floor area, annual export revenue, R&D and QC headcount, certification numbers, on-time delivery, order counts. One profile in our sample published US$6,510,025 in annual exports and a 1,364 m² building. That is a real assessment of a factory. It is simply no answer to which registered company will sign your contract and receive your money. A buyer who reads the badge as covering both has been given no reason on the page to think otherwise.

What this means at a desk

If you are sourcing on 1688, the entity data is already in front of you. It is on the factory card, and on the profiles we sampled it was served to a logged-out request from outside China. Most buyers never look, because nothing on the page suggests the 18 characters matter.

If you are sourcing on Alibaba.com, you hold an English trading name and nothing else. That lines up with what we measured separately: an English name alone produced a candidate for 57.5% of 106 companies and an end-to-end confirmable match for 44.3%. Roughly three in five suppliers cannot be resolved from what the international profile gives you.

Ask the supplier for the business licence. The registered name and 18-character code are both printed on it, and the check digit can be validated offline before you pay anyone.

Method, and what this study cannot say

Six product categories chosen before collection and not changed afterwards: phone cases, copper alloy, LED lighting, knitted fabric, stainless-steel kitchenware and plastic packaging. Suppliers were taken in the order the site search returned them, without selection. On Alibaba.com we read the first five per category; on 1688, the first ten. Two 1688 shops whose profile identifier could not be read are recorded as exclusions rather than dropped silently.

Field presence was read from the page source and scored in three states: present, empty, absent. The whole finding is the difference between “this supplier left it blank” and “the platform has no such field”. Every 18-character candidate had to satisfy both the GB 32100-2015 check digit and a six-digit administrative-division segment. That second test earns its place. One Alibaba.com page carried a hexadecimal string that passed the check digit by chance. A check-digit-only method would have counted it as a real code.

What it cannot say

• It is not a count of every supplier. 94 profiles across six categories on one day. Platform layouts change; the date on this study is load-bearing.
• The 1688 figure is an optimistic bound. Only suppliers with a factory card can be sampled this way, and suppliers without one are plausibly the ones carrying less.
• It says nothing about whether the data is correct. The 1688 block is a third-party verified snapshot with its own pass date. We measured whether the page gives you the fields, not whether the values are true or current.
• It does not evaluate the badges. We counted fields. Whether an SGS or TÜV assessment is thorough is a different question and not one this method touches.
• It is not reproducible unattended. Both platforms interrupted collection with slider challenges, which a person cleared by hand. We do not solve challenges automatically, and anyone repeating this will need to clear them too.

Citing the platform subject-field study

You may quote or reproduce these figures, including commercially, provided the measurement date, the frame and the stated limits travel with them. One row per profile, with the three-state reading and the check result for every code candidate, is in the result file (JSON).

Archived copies, each with its own DOI, resolving independently of this site: Zenodo · Harvard Dataverse.

A citation with everything it needs:

Currawong, “One platform tells you who the company is, the other tells you what it can make”. Subject-identity fields read from 94 supplier profiles across Alibaba.com and 1688 in six product categories, 28 August 2026. https://currawongweb.com/verify/alibaba-1688-transparency-study/ Dataset: https://doi.org/10.5281/zenodo.22136065

@dataset{currawong_platform_subject_fields_2026,
  author    = {Bao L. Zhou},
  title     = {{Subject-identity fields on Chinese supplier profiles: Alibaba.com and 1688 compared (94 profiles, August 2026)}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22136065},
  url       = {https://doi.org/10.5281/zenodo.22136065}
}

This study reports field presence, not a service claim. Currawong verifies public records on request and states sources, query dates and limits with every result; it does not certify suppliers.

Original study106 regulator records · 25 fields each

Told to check the regulator’s record? Here is what is actually in it.

Use this regulator record for identity and contact fields, not as proof of production capability. We read every field held on 106 China-registered trailer manufacturers. Every record names a person; none states what the company makes.

What the record holds, and how old it is

Median record age is 2.1 years. The oldest original submission on the frame dates from 11 December 2004, the newest from 8 July 2026; three records have never been amended since the day they were filed.

Two records, opposite failure modes

Read this next to the Chinese side and the pair becomes useful. In a full census of 264 registered Chinese manufacturers, 54.9% amended their registration within twelve months: the Chinese record moves faster than a quarterly check can follow. The regulator record measured here has the opposite problem: 73.6% of it did not move at all in the same kind of window. A sixth of it has been still for eight years.

The two sources fail in opposite directions, and a buyer needs to know which one they are holding. A Chinese registry answer can change, so date it and recheck it before payment. A regulator record can be stale by a decade, but it provides a named person and contact route that a registry search does not. Neither replaces the other, and neither tells you what the factory can actually build.

One more thing the frame shows about record quality: the province field is not normalised. Shandong appears as both Shandong and SHANDONG, Henan as both Henan and HENAN. Anyone counting by province from this source without folding case will split one province into two.

Method, and what this study cannot say

The frame is the 106 China-registered trailer manufacturers already published in our vPIC manufacturer census of 9 August 2026 — the same cohort three earlier studies use, so the results stack. One request per manufacturer to the regulator’s public, unauthenticated manufacturer-detail endpoint, no retry, a quarter-second apart. All 106 returned a record.

Field presence is scored in three states: present, empty, absent. The reason is the same as in our platform study. “The regulator does not collect this” and “the manufacturer did not supply it” are different facts. Dates come from the regulator’s own LastUpdated and SubmittedOn fields.

What it cannot say

• A stale record is not evidence of anything wrong. A manufacturer with no new filings has no reason to update it. Age measures the record, not the company.
• We did not verify a single value. Presence of an address or a principal is not proof that either is current or correct.
• One regulator, one industry, one day. Trailer manufacturers in a US database. Other regulators keep different fields to different standards.
• “Last updated” is the regulator’s own field. We report what it says; we cannot see what changed, or whether an update was substantive.
• It says nothing about capability. That the record omits primary product is a fact about the record. It is not evidence that the company makes nothing, or anything in particular.

Citing the regulator-record study

You may quote or reproduce these figures, including commercially, provided the measurement date, the frame and the stated limits travel with them. One row per manufacturer, with the three-state reading for all 19 scored fields and both dates, is in the result file (JSON).

Archived copies, each with its own DOI, resolving independently of this site: Zenodo · Harvard Dataverse.

A citation with everything it needs:

Currawong, “Told to check the regulator’s record? Here is what is actually in it”, all 25 fields held by NHTSA vPIC on 106 China-registered trailer manufacturers, 28 August 2026. https://currawongweb.com/research/#vpic-record-age Dataset: https://doi.org/10.5281/zenodo.22144964

@dataset{currawong_vpic_manufacturer_records_2026,
  author    = {Bao L. Zhou},
  title     = {{What a US vehicle regulator holds on 106 China-registered manufacturers, and when it was last updated}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22144964},
  url       = {https://doi.org/10.5281/zenodo.22144964}
}

This study reports what a public regulator record contains and when it was last updated. It is not an assessment of any manufacturer, and Currawong does not certify suppliers.

Original study264 enforcement records · 46 firms

When it goes wrong in the US market, what does the record show?

A recall record should prompt a case-specific review, not an automatic supplier verdict. This study covers every device enforcement record the US regulator returned for a recalling firm located in China: 264 records. We counted what the public corrective-action trail shows.

Who the records belong to, and whether they ever close

The median Ongoing record has been open 875 days; the longest has been open 3,600. Initiations run from 8 June 2012 to 21 April 2026, with 43.2% of the whole frame initiated during 2024 alone.

Re-run on 7 September 2026 with the same query and the same de-duplication. All sixteen published measures came back identical. The two days-open figures moved by exactly the ten days that had passed. The reproduction file is fda-china-device-enforcement-2026-09-07.json.

The reading that is easy to get backwards

A recall is not evidence of a bad supplier. It is evidence that a corrective action happened. 90.5% of these records were firm-initiated rather than ordered, which means the firm found the problem, or accepted it when told, and acted. A factory with no recall history has not demonstrated that its products are safer. It has demonstrated nothing on this record either way, and 46 firms across an entire country’s device export trade is far too few for absence to carry weight.

Repeated records support a different inference. 23 of the 46 firms appear more than once and account for 91.3% of all records. That is a pattern rather than an incident, and it is visible before you order from the regulator’s public record.

The status field is the part buyers most often misread. 83.0% Ongoing does not mean 83% of these products are still on shelves; it means the regulator has not recorded the case as closed. A record that stays open for a median of 875 days is neither a live safety signal nor a clearance. Treat it as what it is: an open file.

Method, and what this study cannot say

Every record openFDA returns for country:"China" on the device enforcement endpoint, retrieved in full by paging to the reported total; no sampling and no retry. That field is the recalling firm’s location: not the country of manufacture and not the country of distribution. Dates arrive as YYYYMMDD strings and were converted before any comparison. Firms were matched on the recalling-firm name, case-folded and trimmed.

What it cannot say

• It is not a supplier ranking, and we publish none. The record file carries the firm names the regulator publishes; this page reports distributions only.
• Absence proves nothing. Most Chinese exporters will never appear here. Not appearing is not a clean bill of health.
• Devices only. One regulator, one product class. Other sectors have different records and different reporting cultures.
• Ongoing is a filing state, not a safety state. We report what the status field says; we cannot see whether the underlying issue was fixed.
• Name matching is shallow. One firm operating under two spellings would count as two. The 46-firm figure is an upper bound on distinct firms.

Citing the enforcement-record study

You may quote or reproduce these figures, including commercially, provided the measurement date, the frame and the stated limits travel with them. One row per enforcement record, with classification, status, both dates and the regulator’s stated reason, is in the result file (JSON).

Archived copies, each with its own DOI, resolving independently of this site: Zenodo · Harvard Dataverse.

A citation with everything it needs:

Currawong, “When it goes wrong in the US market, what does the record show?”, all 264 openFDA device enforcement records for recalling firms located in China, 28 August 2026. https://currawongweb.com/research/#fda-enforcement Dataset: https://doi.org/10.5281/zenodo.22144974

@dataset{currawong_fda_china_device_enforcement_2026,
  author    = {Bao L. Zhou},
  title     = {{US device enforcement records for recalling firms located in China: concentration, status and age}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22144974},
  url       = {https://doi.org/10.5281/zenodo.22144974}
}

This study reports the shape of a public enforcement record. It is not an assessment of any manufacturer, and Currawong does not certify suppliers.

Original study46 registry records · three fields

Registered capital, status, scope: what those three fields actually guarantee.

Do not clear a supplier because the record shows large registered capital, active status and manufacturing wording. We read those three fields across 46 Chinese manufacturers to measure what each one establishes. Each answers a narrower question than buyers often assume.

Registered capital is a promise with a deadline, not money in an account

Under China’s subscribed-capital system a founder states an amount and a date by which it will be contributed. The registry publishes the promise. Whether it was kept is a separate field, and often an empty one.

Subscribed capital on this frame runs from ¥100,000 to ¥1,023,855,833, median ¥10,000,000. That thousand-fold spread is the reason the number reads as a size signal, and 65.2% of the frame gives a buyer no confirmed paid-in figure to place inside it.

Who does not disclose, and why our two frames disagree

The companies withholding a paid-in figure are the younger and smaller ones. The 17 that disclose nothing were founded on average in 2017 and carry a median subscribed capital of ¥3,000,000. The 29 that do disclose were founded on average in 2011 with a median of ¥20,000,000. Under the 2024 Company Law a founder has five years to contribute. A company incorporated recently with an unmet subscription has nothing to report yet. There an empty field is the expected state, and never an evasive one.

That also reconciles this study with our own earlier one, which should be read alongside it. In the census of 264 MIIT “Little Giant” manufacturers, only two records carried no paid-in figure at all, and 37.0% of the rest showed paid-in below subscribed. Those are government-designated SMEs: older, larger and already through a selection process. On this frame — ordinary exporters on a US regulator’s roster — disclosure drops to 63.0%. The disclosure rate is a property of which companies you are looking at. It is not a constant. A buyer applying either figure to a supplier of unknown vintage is applying the wrong one.

Status is a state, not a grade, and it is written three ways

Every one of the 46 records carries a status meaning the company is live. It is written as 存续 on 26, 开业 on 19 and 仍注册 on one. A reader who does not read Chinese, translating each separately, gets three different English words for one fact.

Company type is worse. The 46 records carry 11 distinct type strings, and two of them are the same type written with full-width and half-width brackets. Anyone grouping suppliers by company type from this field, without normalising the punctuation first, will split one category into two.

The practical consequence: status tells you the registration has not been revoked or cancelled. It carries no information about trading history, solvency or delivery, and a buyer who reads “active” as “checked” has read a word that was never doing that work.

Business scope cannot tell a factory from a trader

Read the fourth row again. Every company on this frame that claims manufacturing also claims selling, so a scope containing “manufacture” never rules out a trading operation. The reverse test does work a little: 13.0% mention trade and no production at all, and that is a real signal — the only direction in which this field discriminates.

Scope is also long and increasingly formulaic: median 194 characters, longest 591, with 65.2% now using the post-2021 licensed-project and general-project prefix format. Length is not detail; it is a template.

Method, and what this study cannot say

The frame is the 46 companies from our candidate-ordering study that resolved to exactly one registry record carrying an 18-character code. Several earlier studies use the same cohort, so results stack. Each was queried by that code, not by name, because a code is unambiguous and a name is not. Two licensed interfaces were used: company base information and published shareholder contribution. No retry.

A platform response of “no record found” is counted as a legitimate zero rather than a read failure — the same rule the live free-signal layer follows. An empty paid-in capital string is recorded as undisclosed, never as zero.

What it cannot say

• Undisclosed is not unpaid. A company may have contributed in full and simply not published it. The measurement is about what a buyer can see, not about what happened.
• Nothing here evaluates a supplier. These are field-level properties of a record, not judgements about any company.
• One platform, one day. The data comes from a licensed Chinese business-information platform, not from the government portal directly, and platform coverage is not identical to official coverage.
• 46 manufacturers in one industry. The frame is trailer manufacturers on a US regulator’s roster. Scope wording in other industries will differ.
• Keyword matching on scope is coarse. Presence of a word is not proof of the activity, and a company can perform work its scope never names.

Citing the registration-field study

You may quote or reproduce these figures, including commercially, provided the measurement date, the frame and the stated limits travel with them. One row per company, with the parsed capital figures, status, company type and the scope-wording flags, is in the result file (JSON). Natural-person names are not stored: the legal representative and shareholder fields appear as presence and counts only.

Archived copies, each with its own DOI, resolving independently of this site: Zenodo · Harvard Dataverse.

A citation with everything it needs:

Currawong, “Registered capital, status, scope: what those three fields actually guarantee”, live registry records for 46 Chinese manufacturers, 28 August 2026. https://currawongweb.com/verify/china-registration-fields-study/ Dataset: https://doi.org/10.5281/zenodo.22145056

@dataset{currawong_registration_field_meaning_2026,
  author    = {Bao L. Zhou},
  title     = {{What Chinese registry fields guarantee: paid-in capital, registration status and business scope across 46 manufacturers}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22145056},
  url       = {https://doi.org/10.5281/zenodo.22145056}
}

This study reports properties of registry fields. It is not an assessment of any company, and Currawong does not certify suppliers.

Original studysix risk signals · 46 manufacturers

Three of the six free risk checks never fire on an ordinary cohort.

An empty risk screen does not clear a supplier. To show how much weight absence can carry, we ran six company-level risk signals across 46 Chinese manufacturers and measured how often each returned anything.

How often each signal actually fires

Twenty-eight companies, or 60.9%, came back empty on all six. Of the 18 that did not, 14 fired exactly one signal and 4 fired two. Where court hearings exist at all, the median company has 3 notices and the busiest has 52.

What a base rate changes about reading the result

Empty is not clean. When three of six checks return nothing on every single company in a normal cohort, running them and finding nothing tells you almost exactly what you knew beforehand. Their value is asymmetric: they cannot clear a supplier, and they are decisive when they do fire.

That asymmetry has a practical consequence for check order. Court hearing notices, at 37.0%, were the only signal in this frame likely to return something actionable. The abnormal-operations directory is the check most guides put first, and on this frame it never fired once.

Buyers ask one question after a payment goes wrong: is there still an entity to pursue? The honest reading is that 8.7% of an ordinary cohort is already subject to enforcement and 2.2% is already on the dishonest-debtor list. Those are the companies where a court has already tried and the record shows how that went.

Method, and what this study cannot say

Six company-level interfaces on a licensed Chinese business-information platform, queried by 18-character code for each of the 46 companies from our candidate-ordering frame. Every one of the 276 queries returned a readable answer; none was recorded as unreadable. A platform response of no-record-found is counted as a legitimate zero, and a transport failure would have been recorded as unreadable rather than as zero.

Only counts and presence were stored. No case numbers, parties, amounts or individual names were recorded, and no person-level interface was queried.

Before trusting the zeros we ran a positive control: the same interfaces, queried against two large well-known Chinese companies, returned 2,262 and 9,665 hearing notices and 342 enforcement records. The interfaces and the counting were working; the zeros on this frame are real zeros.

What it cannot say

• Zero of 46 does not mean zero in the population. It means the rate is low enough that 46 companies cannot measure it. Do not read it as "this never happens".
• The platform is not the official register. A platform holding no record is not proof the government register holds none.
• Absence never clears a supplier. That is the entire point of publishing the base rate.
• One industry, one day. Trailer manufacturers on a US regulator’s roster; a cohort in another sector may behave differently.
• A hearing notice is not a verdict. It records that a case was scheduled, not who was right, and not whether the company was plaintiff or defendant.

Citing the base-rate study

You may quote or reproduce these figures, including commercially, provided the measurement date, the frame and the stated limits travel with them. One row per company, with the six per-signal states and counts, is in the result file (JSON). No case details or individual names appear in either file.

Archived copies, each with its own DOI, resolving independently of this site: Zenodo · Harvard Dataverse.

A citation with everything it needs:

Currawong, “Three of the six free risk checks never fire on an ordinary cohort”, six company-level public risk signals across 46 Chinese manufacturers, 28 August 2026. https://currawongweb.com/research/#risk-signal-base-rate Dataset: https://doi.org/10.5281/zenodo.22145063

@dataset{currawong_public_risk_signal_base_rate_2026,
  author    = {Bao L. Zhou},
  title     = {{Base rate of six public risk signals across 46 Chinese manufacturers}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22145063},
  url       = {https://doi.org/10.5281/zenodo.22145063}
}

This study reports how often public signals appear on a cohort. It is not an assessment of any company, and Currawong does not certify suppliers.

Original measurement106 companies

Can you find a Chinese supplier from its English name?

Try the English name once, but do not accept an unconfirmed first result. If the candidate’s own record does not carry the same name, ask for the Chinese registered name or business licence. We measured how often that path worked on 12 August 2026.

Measured 12 August 2026 · Frame: 106 Chinese manufacturers in the US NHTSA vPIC register · Prepared by Bao L. Zhou, Currawong’s China-side verification desk.

English-name lookup: the numbers

The frame is a census, not a sample. It is every Chinese manufacturer we had already enumerated in the US NHTSA vPIC register, searched by the English name that register holds for it.

A search returning something is not the same as a search returning the right thing. When the entity you land on independently records the English name you started from, you have a checkable link between the two. Otherwise you have a guess with a company name attached.

A candidate without an 18-character code cannot be taken to a dated registration check, so it is a dead end regardless of how right it looks.

The brand-name trap

Twenty-two companies in the frame carry both a full registered English name and a brand or short name. Searching the same 22 companies both ways:

The higher number is the worse one. Inspecting the brand-name results one by one, most were a different company. A three-letter brand returned three Hong Kong shell companies sharing those letters. The actual manufacturer, a specialist vehicle maker in Hubei, was not among them. Another brand returned a machinery firm in Guangdong when the company sought was a machinery firm in Henan with a homophone name.

The cause is constraint count. A full registered English name usually encodes a place, a company style and an industry word. A brand name encodes only the style, and Chinese company styles repeat heavily across provinces and industries.

This is a general search-interface hazard: a query that returns more is not a query that works better. The measurement that matters is the share of returns you can independently check.

What this means at a desk

• An English name is a starting point for about two suppliers in five. That is worth trying, but a plan that assumes it will work fails three times in five.
• Never take the first result. Ranking is relevance, not identity. In our brand-name runs the top result was frequently unrelated and presented exactly as convincingly as a correct one.
• Ask for the business licence (营业执照). It carries the registered Chinese name and 18-character code, turning a search problem into a lookup.
• A refusal is evidence too. A supplier unwilling to provide a business-licence photo has not proved fraud. It has blocked the shortest route to confirming the entity, and should never be silently treated as a clean result.
• A name match is not a supplier check. It establishes which entity you are discussing. Active status and scope live in the registration record; whether the person emailing you is connected to it lives in no registry.

Method, and what this study cannot say

Each company's registered English name was submitted to a licensed Chinese business-information platform from a mainland network egress. The platform refused requests from outside China during this measurement, so a foreign desk cannot reproduce the run merely by copying the query.

For every search that returned candidates, we retrieved the top candidate's registration record and compared its recorded English name with the input. Normalisation converted the text to uppercase and removed non-alphanumeric characters, so Co., Ltd. and CO.,LTD compare equal.

Two independent runs were executed seven hours apart and returned identical counts on every metric. That is evidence of a stable index, not of correctness.

What it cannot say

• It is one sector. The frame is vehicle and trailer manufacturers registered with a US federal authority. Every company had a reason to record an English name somewhere, so this is plausibly an upper bound for Chinese exporters generally, not an average.
• The paired comparison is 22 companies. The direction — brand names return more and are mostly wrong — is clear and mechanically explicable. The specific percentages are not stable at that size.
• vPIC English names are not Alibaba storefront names. An English name recorded with a US federal regulator is more likely to be the company's registered English name; a storefront name is chosen for marketing and can change. Buyers holding only a storefront name should expect to do worse than 44.3%, not better.
• A matching English name confirms nothing else. Status, scope, export capability and counterparty identity are separate checks.
• One platform, one date. Coverage differs between providers and changes over time. The measurement date travels with the numbers.

English-name lookup questions

Citing the English-name study

You may quote or reproduce these figures, including commercially, provided the measurement date, frame and limits travel with them. Machine-readable metrics with numerators and denominators are in the result file. The frame comes from our census of Chinese manufacturers in NHTSA vPIC, which readers can re-derive from the vPIC API.

Archived copies, each with its own DOI, resolving independently of this site: Zenodo · Harvard Dataverse.

A citation with everything it needs:

Currawong, “Can you find a Chinese supplier from its English name?”. 106 Chinese manufacturers from the US NHTSA vPIC register each searched by the registered English name from a mainland egress, two independent runs, 12 August 2026. https://currawongweb.com/verify/english-name-lookup-study/ Dataset: https://doi.org/10.5281/zenodo.21907920

@dataset{currawong_english_name_lookup_2026,
  author    = {Bao L. Zhou},
  title     = {{English-name lookup rates for Chinese manufacturers (NHTSA vPIC frame, August 2026)}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21907920},
  url       = {https://doi.org/10.5281/zenodo.21907920}
}

This study reports rates, not a service claim. Currawong verifies public records on request and states sources, query dates and limits with every result; it does not certify suppliers.

Engineering practice, not a buyer decision98 strings · 15 languages

One model translated. Another one checked.

Automated string checks did not establish translation accuracy. A second model back-translated every string before release and found errors, including one that would have sent buyers to ask for the wrong document.

Reviewed 12 August 2026 · Frame: 98 interface strings across 15 languages, two rounds · Prepared by Bao L. Zhou, Currawong’s China-side verification desk.

Machine-translation review: the numbers

Flags fell in six of the fourteen non-English languages — Arabic, Vietnamese, Hungarian, Russian, Polish and Portuguese, with one to three each. The other eight came back clean. We are not drawing a conclusion from that split. Eleven flags spread across six languages is far too thin to say which languages a model handles worse. At this size any pattern could be noise, and saying otherwise would be the same overreach this study is about.

The worst one

One string tells a buyer what to do when the tool finds nothing. Ask the supplier for a photo of the business licence (营业执照) — the Chinese registration document that carries the registered company name and the 18-character Unified Social Credit Code.

The Arabic version used a phrase meaning work permit.

It was grammatical. It used the right register. It passed the character-set check, the forbidden-word list and the length budget. A buyer reading it would have asked a Chinese factory for an employment document, . When that produced nothing useful, the buyer would have concluded the supplier was being evasive. The wrong sentence was ours. The failure would have been silent and misattributed.

Two more of the same kind. A Hungarian gloss meant an operating permit for premises. A Polish phrase was not the term for that document at all. Each one was a plausible-looking word in the right semantic neighbourhood. That is exactly what makes them expensive — a translation that is obviously broken gets fixed; a translation that is confidently wrong ships.

Why the automated gates missed all eleven

Before the review, every string had to pass four checks. Each corresponds to a real failure we had already seen or expected:

• Completeness: every string present in every language, so no reader silently falls back to English.
• Script and diacritics: each language must actually show its own characters. This catches stripped accents, which are not cosmetic: in a separate keyword measurement, a Turkish phrase written without its diacritics returned a hundredth of the search volume. The correctly accented form got the rest.
• Forbidden wordings: a per-language list of claims we must never make. Examples: calling an unconfirmed candidate a verified supplier, or calling an outage a company that does not exist.
• Length: a budget per string so nothing overflows a card on a narrow screen.

We verified those gates were not decorative: stripping the Turkish diacritics made the script check fail, and injecting a forbidden phrase made the wording check fail. They work.

And they caught none of the eleven. Every gate here operates on characters, on word lists, or on counts. An error like work permit where business licence belongs is well-formed at every one of those levels. Meaning is simply not visible from that layer, and no amount of stricter character rules would have found it.

This is the finding with the widest application. If your localisation quality assurance consists of automated string checks, you are testing whether the text is intact, not whether it is right.

Method, and what this study cannot say

One language model produced all translations. A second, independent model received each string with the intended English meaning and back-translated it literally. It flagged only three things: a wrong domain term, a wrong or dangerously softened legal meaning, or an error a native speaker would notice immediately. It was not shown the first model's reasoning. We then classified each flag by hand and either corrected the string or recorded the accepted drift.

Using the same model to check its own work would not provide an independent review. That is why the reviewer was a different model.

Of the eleven, eight were corrected and the live interface carries the corrected wording. Three were accepted with the drift recorded. One is a low-risk status word where the reviewer's own preferred phrasing was itself unverified. Further edits in languages we do not read would have added new unchecked risk without removing any.

What it cannot say

• It is not a benchmark. 98 short, domain-specific interface strings, one translating model, one reviewing model, one review date. It measures a workflow, not machine translation as a field.
• Neither model is a native speaker. The reviewer flagged what it could see; it also stated plainly that it could not find authoritative sources for some phrases it was judging. Its flags are findings to investigate, not verdicts.
• The true error rate is a floor, not a ceiling. 11.2% is what a second model caught. Errors both models share — a term they are both confidently wrong about — are invisible to this method by construction. A native-speaker review would plausibly find more.
• The per-language split proves nothing. Six languages with flags and eight without, at one to three flags each, is not evidence about those languages.
• We publish this about our own text. These were our strings and our errors, found before shipping. The corrected wording is live; per our editorial policy the localised interface is marked as machine-assisted and has not been reviewed by native speakers.

Citing the machine-translation review

You may quote or reproduce these figures, including commercially, provided the review date, the frame and the stated limits travel with them. Machine-readable metrics with numerators and denominators are in the result file.

Archived copies, each with its own DOI, resolving independently of this site: Zenodo · Harvard Dataverse.

A citation with everything it needs:

Currawong, “One model translated, another checked: 11 of 98 wrong”, independent back-translation review of 98 interface strings across 15 languages in two rounds, 12 August 2026. https://currawongweb.com/research/#machine-translation-review Dataset: https://doi.org/10.5281/zenodo.21908377

@dataset{currawong_machine_translation_review_2026,
  author    = {Bao L. Zhou},
  title     = {{Independent back-translation review of machine-translated interface text (98 strings, 15 languages, August 2026)}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21908377},
  url       = {https://doi.org/10.5281/zenodo.21908377}
}

This study reports what one review found in our own interface text. It is not an assessment of any named translation product, and it is not advice on whether to use machine translation for regulated content.

What these studies are, and are not

The separate source-availability method page documents paired vantage points, user agents, same-session controls and correction rules; it is a method, not a ninth Dataset. The verification-cost comparison is a dated buyer guide, not a Dataset.

• Frames are stated, not implied. The 264-company frame is three official provincial excerpts of one national batch — elite, state-vetted manufacturers, not a random sample of Chinese suppliers. Rates measured on this frame are conservative bounds for claims like “even top manufacturers change names”; they are not population estimates.
• Registration data was read through licensed commercial data platforms that republish filings originating in the National Enterprise Credit Information Publicity System; person-name fields were discarded at collection. An absence on a platform is not proof of absence in the official record.
• Availability observations are dated, single-connection facts. A status code from August 2026 says nothing about a portal today, and our measurements support no conclusion about reachability from outside China. We have no trustworthy overseas observation point.

Which file version did you use?

Download the file-version manifest (JSON) and keep it with your data. It lists each English research dataset, its study page, observation period, method and reuse terms. Each site download has a byte count and SHA-256 checksum. Compare that checksum with your saved file to check whether the bytes match.

The page version and the observation period are separate. An edit to a report does not mean the data was collected again. The manifest also records a content fingerprint for each report page. This follows our site's content-version rules; it is not a checksum of the raw HTML.

Archive links come from the current study pages. Their status in this manifest is NOT_CHECKED: it does not check whether an archive is reachable or which archived version matches the site file. A DOI and a matching checksum do not prove that a measurement is correct. These are public research assets; they do not show that a customer has been served.

Citing this research

You are welcome to quote or reproduce these figures, including commercially, provided the observation date and the stated limits travel with them. The dates are load-bearing: citing a rate or status code without its date misrepresents it. Link the study page rather than this index so readers reach the method and limits.

The English-name lookup study has its own page. The machine-translation review is an anchored section of this index; link directly to #machine-translation-review. Each study page carries machine-readable Dataset metadata; the contact routes are here for method clarifications.

These pages report measurements, not legal conclusions, audits or safety findings, and none of them evaluates any specific supplier a reader is dealing with. For a specific supplier: the free in-browser screen or a dated China-side record check.