v1.18 · 14 October 2026. The manual for building a cited, jurisdiction-by-jurisdiction reference database of economic countermeasures and trade defence action. It is written to be applied to any acting state or bloc. The Chinese set in data/cn/ is the first instance and the worked example; nothing in this document is specific to it.
The codebook is meant to be revised. Part 10 says how, and every rule carries the failure that produced it, so a later reader can judge whether the rule still earns its place.
#1. What this builds, and what it refuses to build
A record of what official documents say, in their own terms, about measures a state has taken against foreign parties, products or jurisdictions.
It does not assess legality, motive, proportionality or effect. It does not predict. It does not rate a country's behaviour. Those are the analyst's job, and mixing them into the record is what makes a database unusable as evidence, because a reader can no longer tell which parts are the document and which are the compiler.
The discipline is not modesty. A reference that says only what its sources say can be checked, and one that cannot be checked cannot be relied on in advice.
The three questions every record must answer. Who was named. What was done to them, on what stated ground. Which numbered official document says so. A record that cannot answer all three is not published.
#2. The data model
Eight record types. The action is the join: everything else hangs off it, and nothing else carries state.
jurisdiction ─── the acting state or bloc
│
├── instrument ─────── a law, regulation or rule, by normative rank
│ │
│ └── provides for
│ ▼
├── mechanism ──────── an operative tool (list, procedure, regime)
│ │
│ ├── produces ──▶ register ──── the roster a list-based tool maintains
│ │
│ └── produces ──▶ ACTION ◀────── the atomic unit: one dated official act
│ │
│ ┌──────┼───────┐
│ ▼ ▼ ▼
│ target measure ground
│ (entity or (quoted from
│ product-origin) the document)
│ │
└── series ──────────────────┴──▶ source ──── the numbered official document,
(the coverage claim) preserved and hashed
#2.1 The eight types
| Type | What it is | Key fields |
|---|---|---|
| jurisdiction | The acting state or bloc | code, name, native name, authoritative language |
| instrument | A legal instrument | rank, native title, English title, issuer, adopted, effective, status |
| mechanism | An operative tool | kind, pillar, legal basis, administering body, trigger conditions, measure types, action types, exit pathway |
| register | The roster a list-based mechanism maintains | mechanism, native name, counts derived from actions |
| target | Who or what a measure bites on | See 2.3 and 2.3a |
| action | One dated official act | date, type, mechanism, targets, measures, grounds, register effect, source |
| source | A numbered official document | number, native title, issuer, doc type, published, effective, official URL, preserved copy, hash |
| series | A declared source series | coverage, how enumerated, documents held, date range, completeness grade |
#2.2 Status is derived, never stored
An entity's status is computed by replaying every action that names it, in date order, and reading the state at the end. It is never a field a compiler sets.
A stored status is wrong the moment a suspension is published and someone forgets to update the row. A derived one cannot be, because there is no row to forget. This is also what makes a correction cheap: fix the action, and every view of every entity that action touches moves with it.
The same rule governs every count on the site. A register's totals are the result of a replay, not a number anyone typed.
#2.3 Two shapes of target, and why both are needed
This is the generalisation the Chinese model alone does not give you.
Entity-designation regimes name a party. China's Unreliable Entity List, the Anti-Foreign Sanctions Law countermeasure list, export-control recipient lists, the US Entity List and SDN list, the EU's restrictive-measures annexes. The target is an organisation or an individual, and the identity problem is name resolution.
Product-proceeding regimes name a product from an origin. EU and UK anti-dumping, anti-subsidy and safeguard cases, and their equivalents everywhere. The target is a product-origin pair, the identity key is the tariff classification plus the country of origin, and individual exporters attach to it with their own duty rates.
ENTITY-DESIGNATION PRODUCT-PROCEEDING
────────────────── ──────────────────
target = organisation | individual target = product × origin
key = published name(s) key = HS/CN codes + origin
state = listed / suspended / state = initiated / provisional /
removed definitive / expired /
terminated / under review
the the
question: is this party named? question: does my product fall in scope?
A database that models only the first cannot hold trade defence, which is most of what a trade practice actually meets. A database that models only the second cannot hold sanctions. Build the action to carry either, from the start: action.targets[] holds typed targets, and the type decides which fields are required.
#2.3a A third shape: the scoped measure
Some measures name no party and no product-origin pair. China's destination-based export controls name a destination and a class of goods (dual-use items to the United States). The April 2025 rare-earth controls name a class of goods and no destination at all, so they bite erga omnes. EU safeguards behave the same way, with tariff-rate quotas rather than per-exporter duties.
The record is therefore a scope, not a target: a set of destinations, where empty means every destination, a set of item categories, and the customs classifications the document states. There is no register, because there is no roster to keep, and a mechanism of this kind says so rather than rendering an empty list.
Three shapes is the working set. The test for whether a fourth is needed is not "does this measure look different" but "does the question a reader brings to it have a different key". Entity: is this party named. Product-proceeding: is my product from this country dutiable. Scoped: are these goods, to this destination, controlled.
What does not change between any of them: the action is still the atomic unit, status is still derived, every fact still cites one numbered document, and the coverage claim still governs.
Exercised, and what it cost. The EU set put 525 regulations and 272 product-origin proceedings through this shape, back to 2015. Three things the entity shape does not need turned out to be necessary. A proceeding's identity survives changes to its own description, so the key folds the product string rather than matching it exactly. A single regulation routinely acts on several origins at once and sometimes acts differently on each: one 2025 regulation imposes a definitive duty on Egypt, Japan and Vietnam while terminating the proceeding for India, which is 5.1 again but with the targets split rather than the parties. And exporters hang beneath the pair with their own rates and their own start dates, so the duty rate is a property of the exporter-in-a-proceeding, not of either alone.
#2.4 An action's register effect
Naming a party is not the same as listing it. Every action states, explicitly, what it does to the register:
| Register effect | Meaning |
|---|---|
listed | The action places the target on the register, or changes the measures applying to it |
suspended | Measures are suspended, the target stays on the record |
removed | The target comes off, and stays visible in the history |
none | The target is named but not placed on any register |
none is the one people forget, and it is the one that produces defamatory errors. See 5.2.
#3. Sourcing rules
3.1 One document per fact. Every substantive fact traces to one numbered official document. Not to a news report of it, not to a law firm's client alert, not to another tracker.
3.2 Search may locate; only the official source may supply. Finding a document through a search engine, a commentary or a secondary tracker is fine and often necessary. Reading its content from there is not. Fetch it from the issuing authority's own domain. The pipeline enforces this: the fetcher holds a list of official hosts and refuses any other, so the rule cannot be broken by accident.
3.3 Preserve and hash. Government pages move and are deleted, and the deletion is often the interesting part. Store the bytes served, with their SHA-256 beside them. A stored copy without a hash is an assertion, not evidence. When a re-fetch returns different bytes, that is a fact to record, not an update to apply silently.
3.4 Do not take another compiler's compilation. If a comparable database exists, read it to understand the field, and build your own from primary sources. Extracting its rows is both a terms breach in most cases and a methodological failure: you inherit its errors with no way to see them, and your database is then a copy with an independent-looking front page. Corroboration against a published total is legitimate and useful. Extraction is not.
3.5 A fact not stated is not a fact. Where a document does not state a party's country, the country is unspecified and stays unspecified. Do not infer it from the name, the industry or the context. In the Chinese set this left 16 of 73 entities without a country, and that is the honest number: two announcements state no country at all.
3.6 One narrow enrichment exception, declared. A registry lookup to establish which legal person a document means, where the document names an organisation without stating its jurisdiction, is legitimate. It is limited to identity and jurisdiction, it never extends to activities, ownership, products or conduct, it never applies to individuals, and it is shown on the record with a link to the registry used. Everything else about the party comes from the official document.
3.7 Translation is a marked act. Where the document supplies an English name or title, use it exactly, including its own inconsistencies. Where it does not, and a translation is needed to be usable, mark it as unofficial and keep the native text one click away. The native text is authoritative and is never replaced by the translation.
3.8 Translate once, in a vocabulary, and gate it. Substantive text repeats: 24 distinct measure strings cover 66 occurrences across the Chinese corpus, and most of the variation is a party count written into an otherwise identical sentence. Translating in a single table rather than per page means one careful reading instead of many careless ones, and it makes every rendering auditable in one file. A gate then refuses to publish any measure or document title with no English rendering (gate 9), so nothing substantive can reach a reader in the source language alone.
3.10 Individuals are recorded, not profiled. A countermeasure list can name people, and the source may state more about them than the question requires: the Chinese countermeasure list entries give date of birth, gender and nationality alongside the name. Record the name in both scripts as published, the nationality where stated, and the role only where the role is how the document identifies the person. Do not lift date of birth or gender into structured fields, even though the document states them. They do not help a reader answer "is this person designated, under what instrument, with what measures", and republishing them as queryable data turns a legal reference into a personal-data aggregator. Nothing is concealed by this: the full official text sits behind the native-text toggle and the source link, so the facts remain checkable where the issuer put them. The narrow jurisdiction-verification exception at 3.6 never applies to a person.
3.9 Type the measures as well as quoting them. Alongside the quoted text, give each measure a type from a small controlled vocabulary, and say plainly that the type is the compiler's arrangement while the quoted text is the document's. This is what lets a reader see at a glance that a party is barred from import and export, from new investment and from entry, instead of reading four paragraphs to find out. It does not replace the quotation and it never becomes the record.
3.11 A server that answers is not a source that supplied. EUR-Lex answers a request it will not serve with HTTP 202 and an empty body. The fetcher returned that as a successful download, so a harvest of 25 safeguard regulations wrote 25 documents containing nothing, hashed them, and recorded them as preserved. Nothing in the pipeline objected: an empty file hashes true, and a page with no text renders as a page with no text. Two rules follow. A response that is empty, or that carries a status meaning the document is not ready, is a failed fetch and is retried and then raised, never returned. And a preserved copy with no text in it fails a gate, because the failure to notice is the expensive part, not the failure to fetch.
3.11a An index that caps its own page does not always say so. The Chinese content system accepts any page size and returns about a hundred rows whatever is asked for. A single request for five hundred comes back with ninety-nine and looks exactly like a complete enumeration, so a column of a thousand documents would have been collected at a tenth of its size with nothing to show that anything was missing. The response does carry the truth: its pagination block states the total. So the rule is to read the count the server itself reports, walk the pages until the rows match it, and fail loudly when they do not. Applying that check to a column already in use found it was reading fifty-six of fifty-seven rows, because the last row of a payload is followed by the pagination block rather than another row and the row pattern had no terminator for that case. A completeness claim is only as good as the enumeration under it, and an enumeration that cannot count itself is not one.
#4. The coverage model
This is the most transferable idea in the whole design, and the one that separates a reference from a collection of clippings.
4.1 Declare series, not documents. Coverage is claimed over a series: a named, enumerable run of official documents, such as "announcements of the Unreliable Entity List Working Mechanism" or "Commission implementing regulations imposing definitive anti-dumping duties". A series states how it is enumerated, so a reader can check the claim themselves.
4.2 The registry file is the coverage model. Series live in code (collect/registry.py), and the coverage page is generated from it. Nothing can be published from an undeclared source, and a completeness claim cannot drift from what was actually collected, because both come from the same place.
4.3 Four completeness grades.
| Grade | Claim |
|---|---|
complete | Every document in the stated series, through the stated date |
boundary | Complete within a stated window; nothing claimed before it |
partial | A collected subset; not a claim of historical completeness |
none | In scope, nothing collected yet |
Publishing none matters. A visible empty series tells a reader what the database does not yet cover, which is information. An absent series tells them nothing and lets them assume coverage that does not exist.
4.4 Say what is out of scope, and why. The boundary is part of the product. Name the adjacent things a reader might expect and is not getting, and where they sit instead.
4.5 A numbered series with a gap is not complete. Where documents carry sequential numbers, verify the run. A missing number is either a document not yet collected or one that does not exist, and you must know which. The Chinese UEL series runs 2023〔1〕, 2024〔1-3〕, 2025〔1-10〕 with no holes, which is what lets it be graded complete rather than partial. This check is automated (gate 4 in collect/verify.py).
#5. Reading legal text into records
Five findings from the Chinese corpus. Each one produced a wrong record that looked right, and each generalises.
5.1 One document can carry more than one action. UEL announcement 2024 No. 1 designates two companies in its part 一 and issues a circumvention risk advisory naming a third in its part 二. Parsing a document as a single action either loses the third party or, worse, lists a party the document never listed. Segment the document and parse each part separately.
5.2 Naming a party is not listing it. The advisory above names Caplugs, warns industry about diversion risk through it, and threatens measures if it does not act. It does not place Caplugs on the Unreliable Entity List. A pipeline that treats every named party as a designation puts a company on a sanctions list that no government put there. This is the single most damaging error the whole design exists to prevent, which is why register_effect is a required field on every action rather than something inferred at display time.
5.3 A separator inside a name looks exactly like a separator between names. 美国波音防务、空间与 安全集团 is one company, Boeing Defense, Space & Security. 美国PVH集团、因美纳公司 is two. The enumeration mark is identical. What separates them is that a complete Chinese organisation name ends in a legal or institutional form suffix, so a fragment that does not is the first half of a name whose own separator was misread. Every language has some version of this problem; find the form rule before writing the splitter, and do not split on punctuation alone.
5.4 Refuse rather than guess. The extractor writes what it cannot place to a review field, and a record under review is not published. This is why the numbers can be trusted: on the first run every one of the 14 Chinese documents was flagged, and each flag was a real defect. A parser that silently produces its best guess produces a database nobody can audit.
5.5 Verify by hand at least once, against the text. After the flags cleared, a hand check of the extracted names against the source caught four further defects the automated checks had passed: grounds text absorbed into a company name, a truncated name, a silently dropped annex roster, and lost spacing in names broken across lines. Machine checks catch what you thought to check. Read the output against the source before you believe it.
5.6 One reader per series, not one per jurisdiction. Each declared series has its own document grammar, and the differences are not cosmetic. Chinese UEL announcements name parties as 中文名(English Name); MOFCOM countermeasure orders name them in Latin script with no Chinese form at all; export-control announcements carry annexes with addresses, postcodes and 常用名称 common names under a heading naming the list and its date. A single parser tuned to the first silently mangles the other two. Extraction is therefore profile-driven: the series decides the reader.
5.7 An action states which register it affects. Once a jurisdiction has more than one list, the register is not implied by the mechanism at display time; it is a field on the action. The same company can be designated on two lists on different dates under different instruments with different measures, and only the action knows which one it changed.
5.8 An action can name its targets by reference to a prior document. A suspension or lifting order often identifies the parties only by the order that imposed the measures: "suspend the measures of MOFCOM Order 2025 No. 6". Copying the earlier document's parties is the faithful reading, and the only one, but the record must say it was done by reference and name the document it came from. A pipeline that reads such an order literally finds no parties and either drops the action, which loses a suspension, or guesses, which is worse. Resolution runs as a second pass over the series, after every document has been read once.
5.9 A series may need its own segmenter. Splitting a document into its operative parts is not a universal rule. The Chinese MFA decisions read 一、对<parties>,<measures>, so their parts split on the numeral-plus-對 marker; an export-control announcement numbers its measures the same way, and the same rule would shred it. The segmenter belongs to the series profile, like the rest of the grammar.
5.10 The unit lives in the header, not the cell. An Official Journal duty table prints its unit once, in the column header: "Definitive anti-dumping duty (%)", "AD duty rate EUR per tonne net", "Minimum import price (EUR/tonne net product weight)". The cell beneath is a bare number, "7,2", and means nothing on its own. The first reader demanded a percent sign in the cell, so it read only the minority of tables that repeat the sign on every row: 323 of 465 duty-imposing regulations and 136 of 256 live proceedings showed no rate at all. The naive fix, accepting the bare number as a percentage, would have stored fixed amounts per tonne and minimum import prices as percentages, which is worse than showing nothing. The rule: find the header block first, take the unit and the column order from it, classify the table before reading a row, and store every rate with the kind and unit the header gave it. A rate whose unit neither the header nor the caption settles is refused. The residual rate for all other companies is a row with a scope of its own, not an absence; it is the rate most imports pay. A split digit ("7 1,9 %": one digit, a space, then a decimal number) cannot be a thousands grouping and is read as 71,9 with the printed form kept beside it.
5.11 Consignment is not origin. An extension title names the origin and then the countries the goods are consigned from: "originating in the People's Republic of China, as extended to imports consigned from Indonesia, Malaysia, Sri Lanka, Tunisia, Cambodia, Pakistan and the Philippines". 73 of 525 titles carry this language, and reading every country in the title as an origin produced eight proceedings for products no measure covers, bicycles from Sri Lanka among them, and inflated every count built on origins. The origin text is cut at the consignment language, and the countries after the cut go on the target as consigned from, which is what an extension act is about. "Originating in or consigned from X" names X as the origin.
5.12 A column of TARIC codes does not make a duty table. Five kinds of table carry a TARIC additional code column: the duty table proper; a replacement table in an amending act; a deletion table of rows removed from another regulation; an exemption list in an extension act, with no rate column; and rosters of companies added to an annex or released from an undertaking. Reading rows before classifying the table turns an exemption into a rate and a deleted row into a live one. The kind is decided from the header set and the caption before a row is read. Only the first two yield rates; deletions are counted, exemptions are kept as exemptions, and the rest are recorded as seen.
5.13 The act's own verb decides what it is. "Amending Implementing Regulation (EU) No 1106/2013 imposing a definitive anti-dumping duty on imports of biodiesel" is an amendment. Reading a title for the strongest verb anywhere in it typed 127 acts wrongly: 59 amendments and 25 corrections as impositions or continuations, and two repeals following an expiry review as continuations, which kept a repealed measure in force. The leading verb of the title is read first and wins; the phrases that qualify it ("following an interim review", "new exporter review") refine it; the verbs in the titles of the regulations it cites are ignored.
5.16 Not every figure in a duty table is the duty, and not every duty table belongs to the act's own instrument. Two readings failed here, and both published a number a reader would pay on.
A regulation prints, beside the duty it imposes, other figures with their own headers: the rate levied on imports registered during the investigation, the rate that applies once a safeguard is also due, the original duty reprinted for comparison, and the subsidy or dumping margin the investigation found. A margin is what was found and a duty is what was imposed, and they are different numbers. Reading the last table on the page as the duty published a countervailing residual of 8.7 per cent where the act imposed 13.1, the 8.7 being the retroactive rate on registered imports. So each rate carries a rate_basis, read from its own column header, and only duty_in_force reaches the live rate. The others are published as what they are, beneath the rates and never inside them.
And a countervailing regulation routinely amends the parallel anti-dumping table, because imposing a countervailing duty changes the anti-dumping one. Thirteen such tables, carrying 103 rows, were filed under the act's own instrument, which put anti-dumping figures on the countervailing register. That register is the one a reader adds to the other, so the error was a double count waiting to happen. A rate's instrument is therefore read from its own column header where the header names one, and from the act only where it does not. The act's type is unaffected: the act is still what its leading verb says it is (5.13).
5.17 A notice is an act of the case, not an act on the duty. The Official Journal publishes its acts in the L series and its notices in the C series, and a reference that holds only the L series holds the middle of every case and neither end. It cannot say when a proceeding began, that a measure is about to lapse, that it has lapsed, or that a judgment of the Court has reopened the investigation. Forty live proceedings here were more than five years past the last act that restarts their clock with nothing held to explain it, and the explanation was in the C series the whole time.
The notices are enumerable from the same endpoint as the acts: the CELEX pattern changes from sector 3 with descriptor R to sector 5 with descriptor XC, and the title filter does not change at all.
Three rules govern reading them. The leading words decide what the notice is, as the leading verb decides what an act is (5.13), and a title that names no kind this reference models is recorded as held and untyped rather than forced into the nearest one. A notice of initiation opens a proceeding, because the case exists from that notice; no other kind of notice creates one, so a notice about a measure imposed before the collection window is published and attached to nothing. And nothing read from a notice is ever a rate: the notice says a case exists or has ended, and the stage is the fact.
5.18 A date derived from a rule is not a date the document states. Article 11(2) of Regulation (EU) 2016/1036 and Article 18(1) of Regulation (EU) 2016/1037 give a definitive measure five years from its imposition or from the most recent review, unless an expiry review is initiated. Applying that rule to the last act held produces a date a practitioner needs and no act states. It is published, because refusing to publish it would leave the reader to do the same arithmetic without the acts in front of them, and it is marked as derived wherever it appears, cites the act it runs from, and is retired the moment a notice supplies the real date. The three bases are therefore distinct on the record and on the page: stated in a notice of impending expiry, recorded by a notice of expiry, or derived from the statutory default.
5.19 A classification is a key, and ex is part of the fact. A practitioner arrives with the code on the customs entry, not with the Commission's wording for the product. The printed form is the document's own words and is never rewritten; beside it sits a key of the digits alone, the level the digit count implies, and the heading and chapter above it, so 7208, 7208 39 00 and 72083900 reach the same measure. Two things this bought and one it refuses.
It bought a route in: 1,843 classifications across 210 proceedings, by chapter, which is the door a reader with a code and no product name did not have. And it bought cumulation across shapes: a classification stated by a proceeding and also by a scoped measure means one consignment can meet both, which Regulation (EU) 2015/477 governs, and the reference had said cumulation only between an anti-dumping and a countervailing duty.
It refuses to drop ex. An ex code covers part of a heading and not the whole of it, which is the difference between a dutiable consignment and one that is not, so it is a field and it is said on the page. It also refuses a run of digits that is part of a longer run: a recital carries a CAS registry number and a Customs Union and Statistics number beside the classification, and 0011660-9 was read as a chapter 00 code, which does not exist in the nomenclature.
5.20 A series can hold records that are not documents. The Trade Remedy Investigation Bureau keeps one consolidated register of every case China has opened, and a row in it has no document behind it: it is the Bureau's own record that a case exists, carrying the case number it assigns, the product, the respondent, the date the investigation opened, the industry and the state the case is in. That is a series, it is enumerable, it is complete against the total the register reports, and it holds not one document.
Two things follow. The coverage grade is checked against what a series holds rather than against one shape of holding, or a register complete on its own terms reads as empty. And the register is published as what it is: it records that a case exists and what state it is in, and it does not carry the duty rate, which is in the numbered announcement. Saying that on the page is the whole difference between a register and a database of measures.
The case number is the key and it is more than an identifier. A-CNv.JP-20260107-221 states the instrument and the respondent in a shape the Bureau assigns, so neither is read from the product or the title. S-CN-20020520-710 states no respondent, and that is the fact rather than a gap: a safeguard reaches every origin, so the Bureau's own numbering leaves the respondent out of it.
5.21 A table is structure, and structure does not survive the flattener. Every document here is preserved twice: the bytes the official host served, with their hash recorded, and those bytes put through the extractor, which replaces every tag with a newline. A printed table does not survive that second reading. It arrives as a run of orphaned lines, and the owner met one of them on the page: a product, thirteen customs codes, a quota volume, a duty rate, and then a bare number that was the next row's product number and read as part of the row above it.
The obvious repair is to read the table back out of the lines. It cannot be done, and this was established by measurement rather than assumed. The flattener deletes a whitespace-only cell entirely, leaves no trace of a rowspan or a colspan, and marks no row boundary. The information is not hard to recover from the flattened text; it is absent from it. Nine independently constructed text-only readers were tried against this corpus and every one of them publishes a falsehood somewhere: an order number that is empty on six consecutive rows swallows six rows into one and prints one supplier's quota against another's.
So the rule is two-sided, and neither side is sufficient alone:
> the preserved text is the authority for the characters; > the preserved markup is the authority for the structure.
The grid is read from the markup, which is attested by hash. Its cells are then flattened exactly as the extractor flattens, and that sequence must appear, in order and entire, as one contiguous run of lines in the preserved text. Only then are those verbatim lines handed back to the cells they came from, and the rendered table put through the flattener once more to prove the same glyphs return in the same order. A table that fails any step is not shown as a table. It is left as the lines it is today, with no mark and no apology, because a run of lines makes no claim about which column a number sits in and a wrong table does.
Two findings worth carrying to the next jurisdiction. Most <table> elements in a legal corpus are not tables: of 119,102 in the European set, 113,151 are the publisher's layout scaffolding, and rendering one as a grid invents a table the source never printed. And a nested table is usually a list, not a grid: all 2,547 nested here are the publisher's own list scaffold, not one is a real table, so the safe rule is to refuse only a grid inside a grid and to read the rest as the paragraphs of the cell that holds them.
5.22 A figure restated is not a figure changed. A duty rate is set by one act and reprinted by every act that touches the measure afterwards, and the two are different facts. The date a figure began is what a practitioner reads beside the rate, because it answers how long the importer has been paying it. The date it was last stated is how current the reference is on the figure, because it names the most recent act to carry it. Reading every reprint as a change collapsed both into one date and published a figure's start beside a figure that did not yet exist on it: 275 rows carried a date from an act that stated a different rate, and 351 carried a change history whose every entry was the figure printed above it. The comparison is on the figure itself, never on the act it came from, and the history holds changes alone, each entry running from the date it began to the date it was superseded.
The same distinction settles a contradiction the reference was publishing on its own face. A register replays the acts held and can read in force while the lapse date beside it has already passed, because the act that continued the measure, or the notice that ended it, is not held. Both readings are right on what is held and they answer different questions, so the page says which is which: the register states what the acts say, and what became of the measure after that date is not on this record.
5.23 Why a case ended is a sentence the act states, quoted, or it is not read. The practitioner's question about a case that came before is why it was turned down, and the acts that end a proceeding, lapse a measure or withdraw an undertaking state the reason in their recitals, which the reference preserved and never read. It reads them now, into a vocabulary of thirteen grounds every one of which was derived from a held recital before it was named: no review requested, complaint withdrawn, review request withdrawn, transactions not representative, undertaking breached, undertaking withdrawn by the exporter, no circumvention, dumping not likely to recur, injury not likely to recur, imports negligible, no dumping, circumstances unchanged, and a WTO ruling implemented. The ground is carried with the sentence that states it, verbatim and with its recital number, and a gate finds every quoted sentence in the preserved text.
Two refusals. not_read is a value and the page prints it as one: an act whose recitals state a finding without stating the termination, as Regulation (EU) 2025/1919 finds no dumping in its recitals and terminates only in Article 1, is not made to say more than it says. And a complaint rejected before initiation is published nowhere official, so the reference cannot hold it and the outcomes page says that the question has a boundary.
The reading exposed two acts typed by words that were not their own. An amending regulation reproduces the amended regulation's title after its number, and "imposing a definitive anti-dumping duty on X and terminating the proceeding concerning Y" is the cited act's title, so the amending act had been read as ending a proceeding it did not touch. A verb that follows a citation belongs to the citation. And one notice titled as a notice of expiry is written in the impending-expiry form, "will expire unless a review is initiated"; the title decides the kind (5.13) except where the text contradicts it in that one fixed form, and then the text decides and the record says so.
#6. Identity and history
6.1 Resolve on the identifier the document supplies. Where announcements give an English name alongside the native one, that English name is the strongest key, because it is the issuer's own identification. Fold case, punctuation and corporate-form suffixes for matching. Fall back to the native name only where no English name is published.
6.2 Keep every spelling. The same entity appears as 因美纳公司 in one announcement and 美国因美纳公司 in another. Record both, show them as "also published as", and never quietly pick a favourite. The variant spellings are how a reader finds the entity, and they are evidence about how the issuer works.
6.3 Never delete history. A removed designation stays visible, marked removed, with its dates. The question "was this company ever listed, and when" is a real question with real consequences, and a database that answers it only in the present tense cannot be used for diligence on a past period.
6.5 Status is per register, not per entity. An entity can sit on several registers at once, with different measures, different dates and different current states: active on one, suspended on another, removed from a third. Each register's state is replayed from that register's own acts. A single status field cannot express this, and collapsing it destroys the question a reader actually has, which is what applies to this party under which list. Where a single overall status is needed, for a directory row or a search result, derive it as the strongest state across registers and say that is what it is.
6.6 One vocabulary across jurisdictions. A reference that holds two regimes must call the same thing by the same name on both sides, or a reader learns two products. "Status" is the derived state on both sides, never "state" on one. "Register" is the roster on both sides. A target is an entity or a proceeding, a document is a document or a regulation, and the ledger row carries the date, the identifier and the status in the same three places whichever side it is on. The differences that remain are real ones, and they live in one adapter per jurisdiction rather than in the pages.
6.7 A status the held acts cannot give is said as such. A proceeding whose only held acts amend, correct or update an exemption list has a measure imposed before the collection began, or under a product name the identity key does not fold. Replaying those acts gives no status, and showing "under investigation" would be a guess dressed as a derivation. Such a proceeding carries the status "not derived from the regulations held", counted on no tile, until an imposing, extending, terminating or repealing act is collected. "Under investigation" is reserved for a proceeding whose first held act is a registration order. The identity key itself folds what it safely can: product wording differs between regulations on one measure ("butt-welding" and "buttwelding"), so the key drops punctuation, a leading "certain" and a parenthesised alias. It does not fold "continuous filament glass fibre products" into "continuous glass fibre products"; Part 10 carries that.
6.8 The latest regulation's rate wins, and the earlier ones stay. Each company and code pair under a proceeding holds the rate from the latest regulation that states it, with the earlier statements kept as history. A first-wins rule showed a 2016 rate after a 2022 interim review had changed it, which is the kind of wrong that looks right.
6.9 A case renamed is still one case, and the classification is what proves it. The identity of a product proceeding is the product as the act writes it, folded so that punctuation, a parenthesised alias and a leading "certain" do not split it. That fold cannot see a rename, and the Commission renames. "Certain welded tubes and pipes of iron or non-alloy steel" became "certain welded pipes and tubes of iron or non-alloyed steel"; "certain concrete reinforcement bars and rods" became "rebars"; "certain aluminium foils" became "certain aluminium foil"; one pair differed by a single comma. Twelve cases were filed twice.
The damage was not cosmetic. The act continuing the duty landed on the record under the new name while the record under the old name kept a lapse date that had passed, so the reference published a measure as in force beside a date saying it was over, with the expiry review that resolved it sitting on the other page. Five of the eleven such contradictions were this defect. One case showed a provisional duty for a measure that had carried a definitive one for a year, because the initiation and the definitive duty were on one record and the provisional on the other.
What the acts supply is the classification. Two cases on one origin under one instrument that print the same ten digit code are one case, and the wording is what changed. The evidence has to be strong in one of two ways: three or more shared codes, or one shared code and half the shorter product name in common. ex codes are excluded from the test entirely, because an ex code is a slice of a heading (5.19) and two measures can share a heading without sharing a good: corrosion resistant steels and heavy plate overlap on one and ran their own initiations, their own definitive duties and their own expiry reviews, and they stay apart. The surviving wording is the latest document's, which is the rule the expiry clock already runs on, and every retired address redirects to the case that now holds it.
A notice needs a different test, because a notice prints no classification: not one of the 581 notice targets held carries a code, since a notice says a case exists or has ended and never sets a duty. What is available is that these renames are transpositions. The notice opening the 2026 expiry review on welded pipes for Belarus, China and Russia writes a third spelling again, "welded tubes and pipes of iron or non-alloyed steel", and matches neither act. Sorting the words of the product matches all three, and it is used only where the exact key misses and only where exactly one case on that origin answers to it, so an ambiguous fold attaches nothing. Across the 260 cases held it collides with no other, and it placed four notices that had been published attached to nothing, three of them the review that keeps those measures in force.
6.4 Corrections are checked against documents, not commentary. A correction needs the official document that supports it. Until it is verified, the existing cited information stands. Verified corrections appear in the update history rather than being applied invisibly.
#6a. Staying current
A reference that goes quietly stale is worse than none, because people trust it. Collection is not the end of the work; keeping the claim true is.
6a.1 Re-enumerate, do not re-crawl. Each declared series can be re-enumerated from the official index that defined it, and the result compared against what is held. That is cheap, it is the same operation as the original enumeration, and it directly tests the coverage claim rather than approximating it.
6a.2 Two kinds of change, neither applied silently. A new document is a gap to collect. A changed document, where the official page now serves different bytes from the preserved copy, is a fact in its own right: pages are moved, amended and quietly corrected. Report it; never overwrite it. Overwriting destroys the evidence that the text used to say something else, which is precisely what the preserved copy exists to hold.
6a.3 Publish the check as well as the data. The site states when each series was last checked and whether it was in step with its index. A reader can then tell the difference between "nothing has happened" and "nobody has looked", which are the same thing on a page that shows only content.
6a.4 A series that cannot be re-enumerated says so. Where there is no official index to check against, the freshness record says that rather than showing a stale date that implies verification.
6a.5 The scheduled run, and the one thing it may do alone. One process does the whole cycle: re-enumerate, collect what is new, extract, resolve, verify, build and run the repository gates (collect/refresh.py). It runs on a schedule without a person. Decided 7 September 2026 by the owner: a clean run commits to main without a preview, and it is the only process allowed to. The carve-out is narrow. The run never edits code; it may commit sources/, data/ and dist/ and nothing else, and it refuses to commit if anything else changed. It is gated by the same checks the owner runs by hand. Any gate failure, including a new document the reader cannot place, goes to a refresh/<date> branch and a pull request marked for a hand pass, never to main. A series that cannot be reached is recorded as such (6a.4) and the rest proceeds, because a flaky server on one side must not stall publication on the other; the failure is raised as an issue as well. Generated files are never merged by hand: whichever branch merges regenerates. The carve-out is revoked by deleting the schedule block from the workflow.
#7. The gates
Automated, and the build refuses to publish when any fails. Implemented in collect/verify.py.
- Every record traces to a preserved document whose bytes still hash true.
- No record is published while anything on it is under review.
- Every document states a number and a date.
- A numbered series has no gaps, or the gaps are declared.
- Every derived status is replayable from that entity's own events.
- Every stated country carries the basis on which it is stated.
- Every event cites a document the database actually holds.
- A completeness grade matches what is held: a series graded above none holds documents, and one graded none holds nothing.
- Nothing is published in Chinese alone; every measure and every title has an English rendering.
- A regulation held in two series is one record, with identical actions and bytes in both files.
- A rate carries the kind and unit its column header gave it; a fixed amount never carries a percent sign, and a percentage is never above 1000.
- A duty-imposing act whose enacting terms carry a company column beside a TARIC additional code column and a rate column yields rows or says why not, checked against the source text and not against the reader's own report of itself.
- Every rate's country is stated by the table, the row, the title or a hand check, and is an origin the act names.
- Every preserved copy has bytes and has text. An empty preservation hashes true, so nothing else catches it.
- Every status a replay produces lands in a bucket, so a register's parts sum to its own total. A headline of 215 above a breakdown reading 207 stood for a fortnight.
- Every document in a series index is published somewhere, as an act in that series or as an instrument. Two Regulations were fetched, preserved and filed nowhere, and the freshness check could not see it because it compared the official index against the harvest index.
- Only the duty in force is published as a rate, and a rate is filed under the instrument its own column header names. See 5.16.
- A notice carries no rate and states no proceeding state directly. See 5.17.
- Every expiry date says on whose authority it stands: the notice that states it, or the act the statutory five years run from. See 5.18.
- No published classification loses its printed form or its
exqualifier. See 5.19. - A completeness grade is checked against what a series holds, in whatever shape it holds it. A register of cases with no document behind a row is held. See 5.20.
- Nothing about a case is published in Chinese alone: its product, its industry, its state and its instrument each have an English rendering made once in a gated vocabulary.
- Every restored table round trips: put back through the flattener it returns the same glyphs in the same order as the preserved text holds. See 5.21.
24 to 27. Added with v1.17 and described in its version-log entry: the door figures, the grounds quotes found in the preserved text, and the two readers corrected on the way.
- The map restates and derives nothing. Its three headline figures on its tiles and on the home strip equal
door_counts()["map"], its count for China on the Union side equals the origin door's, andmap-data.jsoncarries the same figure as the built page. See 8.9. - The assistant server restates and derives nothing. Every figure
discoverreturns equals the door that publishes it and the register page that carries it; every proceeding and casemeasures_forreturns for an origin is the map's row for that record, with the map's status, rate and expiry; every party's status and registers equal the map's; every document's hash is the hash of the preserved bytes; and every search result resolves to the page it names. The whole run goes through the protocol layer, so the gate proves the endpoint. See 8.10.
Gates are cheap and they are the reason a compiler can leave the work for three months and come back to it. Add one whenever a class of error gets through.
#8. Presentation rules
8.1 Every node opens a cited record. A count, a name, a mechanism or a date is a link to the thing that supports it. A number with no route to its source is decoration.
8.2 The native text is always one click away. A toggle on every quoted passage, showing the original language. This is what lets a reader who reads the language check the compiler, and it is what makes the translation rule (3.7) credible rather than a claim. The corollary is that nothing substantive appears in the source language alone: a reader who cannot read it must still be able to learn what a measure does.
8.3 Show the framework as a map, once. A single page that puts every mechanism, its governing instrument and what it has produced in one view. Practitioners orient on structure before detail, and the map is what a newcomer needs first.
8.4 Separate the register from the mechanism. The mechanism page says when the tool applies, how it proceeds and how a party gets off it. The register page is the roster with filters. Conflating them makes both worse.
8.5 Date the whole thing, visibly. "Information current through" with a date, on the front page. A reference with no as-of date cannot be cited.
8.7 Two jurisdictions are peers, or the reference has a home country. The first build put China at the site root and the EU in a sub-folder with its own vocabulary, no chart and no mechanisms, and the owner's verdict was that the two felt scattered. The neutral landing shows both as equal panels with the same four figures and the same chart, every jurisdiction section uses the same page skeleton, and the shared pages name neither jurisdiction first. Symmetry is a design decision that has to be made once, at the root, or it is unmade on every page.
8.6 The update feed is a first-class page. Every dated act, newest first, each linked to its source. It is the page a returning practitioner opens, and it is how you show the database is alive.
8.8 The home answers the question the reader brings, and the doors are the navigation. The neutral landing of 8.7 presented two jurisdictions and named no reader, and a trade defence lawyer with a client does not arrive holding a jurisdiction. She arrives holding a product and a country of origin and asks three things: has it been challenged, what is the duty, and what happened before. So the home states who it is for in its first sentence, leads with that question and the five doors that answer it, by product, by origin, by sector, by amount of duty and by outcome, then asks the other two questions the fence names, whether a party is named and whether goods to a destination are controlled, and only then shows the two jurisdictions as peers. The doors are the navigation and the jurisdictions are crumbs. Every figure on a door is derived once and a gate holds the home, the door page and the derivation to one number, because three pages once published three totals for one corpus.
8.9 A map fill is a count on goods from an origin, and never a rating of the country. The map page shades a country by the number of measures on goods from that origin, in the band the legend states, with the unit beside it. It states nothing about the country: not a volume, not a value, not a posture. Three sides share one legend and keep three vocabularies, the Combined Nomenclature chapter the Union's acts print, the industry the Bureau files a case under and the register a party sits on, because mapping one onto another would be the compiler's guess beside a cited fact (3.5). The code EU, which the Chinese documents use for the Union as a respondent and as a party's country, fills the twenty-seven as one block, and a member state named on its own says both. A designation whose document states no country is counted and shown as not stated. Every figure on the map restates a door figure and a gate holds them to one number (gate 28). The page is written whole for its default view and the script only re-fills it, so no figure lives behind a script.
8.10 An assistant reads the record the page shows, and no other. A lawyer asks the assistant she already uses, so the reference answers there as well as on its pages: a read-only server (mcp/) over the Model Context Protocol, with the tools assistants already know, search, resolve, document and discover, and two of the reference's own, a product or code from an origin and a named party. It reads one built file (dist/mcp-data.json), the search index and the preserved text, all written by the build from the same profiles as the pages, so no status, rate or count is computed a second time and gate 29 holds the server to the doors. Every answer carries the official document's address and the page here that shows the same record, a field no document states is returned as not stated and the assistant is told not to fill it in, and the server states the editorial rule in its own instructions: it records what documents say and assesses nothing. The page connect.html says where, how, what comes back and what the server refuses.
#9. Applying this to a new jurisdiction
The order matters. Each step depends on the one before it, and skipping to collection is the common way to waste a fortnight.
- Map the pillars. What families of measure exist, and which instrument provides for each. Four or five pillars is normal. Do this before touching data; it decides the shape of everything.
- Identify the official series. Where each family is published, in what numbered form, and how a run of it can be enumerated. This is the step that determines whether a completeness claim is possible at all.
- Find the enumeration route. A column index, a register, an official gazette search, an OJ series. Prefer whatever the issuer treats as canonical. Record its parameters in the registry so the next run reproduces it.
- Collect one series end to end, with preservation and hashing, before starting a second. A pillar finished completely teaches you the document grammar; six pillars started teaches you nothing and produces six partial claims.
- Write the extractor against the grammar you found, with refusals. Expect the first run to flag everything.
- Hand-verify against the source text. Once, properly. See 5.5.
- Derive, gate, publish. Then declare coverage honestly, including what is not in.
Jurisdiction notes to carry forward.
- China. Documents are published in a JPaaS content system whose lists render client-side; the column page carries the query parameters for the endpoint that returns the rows. The authoritative text is Chinese and English names are supplied in parentheses where they exist. Documents number per year per issuing mechanism, which makes gap checking straightforward.
- EU. The Official Journal is the canonical series, and the Publications Office SPARQL endpoint at
publications.europa.eu/webapi/rdf/sparqlenumerates it properly, which is far better than scraping the TRON case front end (an Angular shell that serves the same page for every route). CELEX is the stable key. Titles are highly structured and carry the action type, the product and the origins; Article 1 carries the CN and TARIC codes and the per-exporter rates. Queries take around 30 seconds, so enumerate once per run and cache. - US. Multiple registers with different publishers, and the Federal Register is canonical for the rulemaking but not for the list contents. Expect to reconcile a published list file against the notices that changed it, which is a different problem from the Chinese one.
#10. Keeping this document alive
How to revise. Change the rule, bump the version, and add a line to the log below saying what changed and what made you change it. A rule whose failure is no longer recorded should be reviewed for deletion; if nobody can remember what it prevents, it may not prevent anything.
What belongs here is anything that would otherwise be rediscovered: a document grammar, a class of error, a sourcing boundary, a modelling decision and its reason. What does not belong is anything specific to one jurisdiction's data, which lives with that jurisdiction.
Open questions, carried forward.
- The scoped shape (2.3a) now runs in both directions. China's export controls scope a measure by destination; EU safeguards scope one by origin, because the trade runs the other way. The two records therefore use different country keys for the same idea, which is a wart: a later pass should give the scoped target one country field and a direction beside it.
- The notices answer when a measure lapses and they do not answer everything. A notice of impending expiry names the measures whose five years are running out in about a year, so a measure imposed outside the collection window appears in a notice and has no proceeding here to attach to; those notices are published and attached to nothing rather than to the nearest plausible pair. And a notice of initiation opens a proceeding with no duty, which exercises the
under_investigationstate that the acts alone never produced. - Safeguard quota volumes are not read. Each amendment sets a volume per product category and per supplying country in annex tables that are replaced every few months, so the measure, its scope, its customs classifications and its dates are held and the tables are not.
- Both jurisdictions now hold both halves of their own toolkit, defensive and retaliatory, and the depth is not symmetrical. The European side holds the acts, the notices, the rates and the classifications. The Chinese side holds the case register and not the announcements behind it, so it can say that a case exists, against whom, since when and in what state, and it cannot say what rate the case produced. The 1,036-announcement column at trb.mofcom.gov.cn is the next tranche and it is what turns the register into a measure record.
- The Union's Anti-Coercion Instrument and its Blocking Statute produce almost no stream. The instruments exist and no measure has been imposed under the first; the second has produced two acts since 2015. An instrument that has never been used is a fact worth publishing, and it belongs in the toolkit rather than as a series with no rows.
- The measure vocabulary (3.9) is currently Chinese. An EU duty rate is a measure too, and whether the two should share a vocabulary so they can be compared is unresolved. An EU rate now carries a typed kind (ad valorem, fixed amount, minimum import price), which is the first step toward a vocabulary the two jurisdictions could share.
- The identity key for a proceeding is the product wording plus the origin, folded for punctuation (6.7). Regulations on one measure still differ in wording enough to split a proceeding and leave one half with no derivable status. Keying on the tariff classification is the codebook's own answer at 2.3 and the next thing to try; CN codes are shared across products often enough that it is not free.
- A safeguard that sets its scope in an annex table states no classification in its articles, so the cumulation match at 5.19 cannot see it. The steel safeguard is the case that matters, and its pages say that an absent overlap is not an assurance that none exists.
- Annex rosters are not expanded into rows. An "other cooperating companies listed in Annex I" row carries one rate for the companies the annex names; the annex is preserved but not read, so a search for one of those companies finds nothing.
- Fixed amounts stated per product line and CN code, as for ammonium nitrate, are recorded as seen but not read, because the product description spans lines.
- An extension act that states its rate by reference to an earlier regulation is not resolved to a figure (5.4); where the earlier regulation predates 2015 the figure is not held at all.
- Cross-reference resolution (5.8) currently copies the parties of the referenced document's listing actions. It has not been tested against an order that acts on only some of them, which is the obvious next shape and would need the reference to carry a party filter.
- Individuals as targets now have a position (3.10), taken when the MFA decisions brought 76 named people into the set. Whether a person should ever be removed from the record once a designation is lifted is a harder question, and the current answer, that history is never deleted (6.3), is the same one applied to companies. It may not be the right answer for people.
- Measures are now typed into a controlled vocabulary as well as quoted (3.9). Whether the vocabulary should be shared across jurisdictions, so that a Chinese measure and an EU one can be compared as the same kind of thing, is the next question and a harder one: the types are currently chosen to fit what China does.
- Multi-jurisdiction entity resolution, meaning the same company on a Chinese list and an EU list, is deliberately not attempted yet. It is the obvious next capability and the easiest place to introduce errors that look authoritative.
#Version log
| Version | Date | Change |
|---|---|---|
| 1.19 | 27 Oct 2026 | An assistant reads the reference. A read-only server over the Model Context Protocol, Streamable HTTP and no authentication, hosted as a function beside the pages, answering from one built file the site writes from the same profiles as every door: the measures in force on a product or a customs code from an origin, with every rate and its regulation; a named party's status on each register with the acts that set it; one preserved document with its hash and its text; a search ranked as the site's own; and what is held and how current it is. Added 8.10 and gate 29, which puts every tool through the protocol layer and holds every figure, status, rate and hash it returns to the door, the map row and the preserved bytes. The connect page says how Claude, ChatGPT and Copilot add it. |
| 1.18 | 14 Oct 2026 | The map: one page shading every origin by the count of measures on its goods, on three sides with one legend, with a sector-by-origin matrix per side, initiations by year and the twelve months ahead. Added 8.9 and gate 28. The world is the Natural Earth 1:110m layer, decoded and projected once by build/worldmap.py (Equal Earth), so the build stays offline. |
| 1.17 | 19 Sep 2026 | The reader's fifth question, why a case was turned down. The recitals of every act that ends a proceeding, lapses a measure or withdraws an undertaking are read into a bounded vocabulary of thirteen grounds, each derived from a held recital, each carried with the sentence that states it and its recital number, and not_read printed as a value where the act states its finding without its termination. Added 5.23 and gates 26 and 27. Two readers corrected on the way: an amending act is no longer typed by the terminating clause of the title it cites, and a notice written in the impending-expiry form under an expiry title is typed by its text. The same release adds the five doors a practitioner uses, by product, origin, sector, duty and outcome, each a cross-jurisdictional page, and 8.8: the home answers the reader's question first and the doors are the navigation. |
| 1.16 | 12 Sep 2026 | Twelve cases were filed twice, because the Commission renames a product between acts and the identity fold reads the wording. The act continuing a duty landed on the record under the new name while the record under the old one kept a lapse date already passed, which is five of the eleven registers that read in force beside a date saying they were over; one case showed a provisional duty for a measure that had carried a definitive one for a year. Added 6.9: the classification the acts print is the evidence, ex codes are outside the test because a slice of a heading is shared by measures that are genuinely two, the surviving wording is the latest document's, and every retired address redirects. A notice carries no classification, so a notice matches on the product's words with their order dropped, used only where the exact key misses and only where one case answers to it. Eleven registers contradicted themselves at the start and three do now, and those three hold nothing after the date and say so. Gate 24. |
| 1.15 | 12 Sep 2026 | A duty rate now dates from the act that set the figure rather than from the act that first named the company. Every later act reprinting an unchanged rate had been read as changing it, so 275 rows published a start date taken from an act that stated a different figure, and 351 published a change history whose every entry was the figure above it. Added 5.22 on a figure restated against a figure changed, and gate 21. The same entry settles the contradiction on eleven registers that read in force beside a lapse date already passed: the page now says what the register answers and what the reference does not know. |
| 1.14 | 12 Sep 2026 | Added 5.21 and gate 23, after the owner sent a screenshot of a tariff quota table that had been flattened to a run of orphaned lines, the last of which was the next row's product number. The extractor turns every tag into a newline, so structure does not survive it; the preserved markup does, and it is attested by hash. Tables are now read from the markup, proved against the text line for line, and proved again by flattening the result. 5,819 of the 5,951 printed tables in the European corpus are restored across 781 documents; the rest are left as they were. |
| 1.13 | 12 Sep 2026 | China's own trade defence, first tranche: the Trade Remedy Investigation Bureau's consolidated case register, 345 cases from 1997, complete against the total the register reports. Until now the Chinese side was a countermeasures record and not a trade defence one, which Part 10 called the largest single gap left. Added 5.20 on a series whose records are not documents, and on a case number that states the instrument and the respondent, or states no respondent because a safeguard has none. The announcement stream behind the register is declared and not collected, so the reference holds every case and no duty rate. |
| 1.12 | 12 Sep 2026 | The other end of a case. The Official Journal's C series notices are collected: the notice that starts a proceeding, the notice warning that measures are about to lapse, the notice recording that they have, and the notices implementing a judgment of the Court or a ruling of the Dispute Settlement Body. Same endpoint as the acts, sector 5 with descriptor XC. Added 5.17 on reading a notice as an act of the case rather than an act on the duty, and 5.18 on a date derived from the statutory five years, which is published, marked as derived, cites the act it runs from, and is retired when a notice supplies the real one. Gates 18 and 19. |
| 1.11 | 12 Sep 2026 | Added 5.16, after two readings published a figure a reader would pay on: a retroactive rate on registered imports was live as a countervailing duty four points below the one the act imposed, and thirteen anti-dumping duty tables inside countervailing regulations were filed on the countervailing register, which is the register a reader adds to the other. A rate now carries the basis and the instrument its own column header states. Gates 15 to 17: every replayed status lands in a bucket so a register's parts sum to its total, every document in a series index is published somewhere, and only the duty in force is published as a rate. The Union's safeguard and rebalancing mechanisms added to the framework map, which had been publishing two of its four. |
| 1.10 | 10 Sep 2026 | The Union's own countermeasures collected: 18 rebalancing regulations under the Enforcement Regulation, resolved into 6 measures, most of the stream being the suspensions that follow an imposition. Added 3.11a on enumeration that counts itself (3.11a), after the column reader was found to cap silently at about a hundred rows and to drop the last row of every payload. Routes recorded for China's own trade remedies and for both extraterritoriality regimes, declared and not collected. |
| 1.9 | 10 Sep 2026 | EU safeguards collected, which closes the last declared and uncollected European series: 25 regulations resolved into 4 measures, the scoped shape (2.3a) exercised on the import side for the first time. The Publications Office Cellar was added as the route when EUR-Lex refuses to render a work, so a source that answers 202 no longer stops the collection. |
| 1.8 | 10 Sep 2026 | Added 3.11 and gate 14, after a harvest wrote 25 empty documents and recorded them as preserved: EUR-Lex answered every request that day with HTTP 202 and no body, including for regulations it had served three days earlier. EU safeguards are now enumerated, 25 regulations by CELEX and title, which is the filter problem that blocked the series solved; collection waits on the source. |
| 1.7 | 7 Sep 2026 | Added 6a.5, the scheduled run and its carve-out, after the owner decided that a clean refresh may commit to main without a preview. The EU duty-table reader rebuilt against the grammar found (5.10 to 5.13), after 323 of 465 duty-imposing regulations were found to carry no rate: rows on the regulations rose from 792 to 1,979 and regulations with rows from 146 to 357, and every remaining empty act carries its reason. Consignment countries no longer count as origins and wording differences are folded, so proceedings fell from 271 to 222; 127 acts retyped by their leading verb; a status the held acts cannot give is said as such (6.7); the latest regulation's rate wins (6.8); gates 11 to 13. |
| 1.6 | 7 Sep 2026 | The parity rebuild: 6.6 (one vocabulary across jurisdictions) and 8.7 (jurisdictions as peers, decided at the root), after the owner found the first two-jurisdiction site scattered. EU regulations de-duplicated across series: 525, not 537. |
| 1.5 | 6 Sep 2026 | Added 2.3a, the scoped-measure target shape, exercised on China's destination-based export controls; and part 6a on staying current, after the first freshness run. EU coverage widened from two years to 2015, which is what a five-year measure cycle requires. |
| 1.4 | 6 Sep 2026 | The product-proceeding target shape (2.3) exercised on 127 EU trade defence regulations and 109 proceedings; it is no longer provisional. Safeguards and destination-based controls identified as needing a third, scoped-measure shape. |
| 1.3 | 6 Sep 2026 | Added 3.10, the position on named individuals, and 5.9 on per-series segmenters, after the MFA decisions brought 76 people and a fourth document grammar into the set. |
| 1.2 | 6 Sep 2026 | Added 3.8 and 3.9 on translating in a gated vocabulary and typing measures, after the site was found to be showing substantive Chinese with no English rendering. |
| 1.1 | 6 Sep 2026 | Two further Chinese pillars collected: MOFCOM countermeasure orders and the export-control recipient lists. Added 5.6 (one reader per series), 5.7 (register on the action), 5.8 (targets named by reference to a prior document) and 6.5 (status is per register). Each came from a record that was wrong before the rule existed. |
| 1.0 | 6 Sep 2026 | First statement, written alongside the Chinese Unreliable Entity List instance. Parts 5.1 to 5.5 are findings from that corpus. |
How to cite this page
The codebook. Countermeasures Reference, Lawcrosse. https://lawcrosse-countermeasures.vercel.app/codebook.html.