How PräferenzPilot calculates — verifiable.
“Verified engine” should not be a marketing phrase. This page shows the path of every rule from the official source to the verdict — with figures taken straight from the system, and a source register with checksums.
- EU agreements
- 20
- destination countries
- 78
- verified list rules
- 9,433
- rule entries (per commodity code)
- 21,092
- official source documents
- 20
Two figures, because they measure different things: a list rule is one set of rules from the official annex for one heading of one agreement. A rule entry is how it is stored per commodity code — a heading rule applies to every subheading beneath it and is therefore held several times in the database. The entry is what matters for a check; the list rule is what matters for the question of how much official text lies behind it.
From the official source to the verdict — in four steps
Official source
Every list rule comes from the official annex of its agreement (EUR-Lex / Official Journal of the EU). On ingestion, a SHA-256 fingerprint of the source document is recorded — years later it is still provable which document a rule came from. Currently: 20 source documents, all listed in the register below.
Deterministic ingestion — without AI
The annexes are read by software parsers, not by a language model: same input, guaranteed same result, nothing is “interpreted”. Rules whose structure cannot be mapped beyond doubt are marked as drafts — they never feed into a verdict.
Independent verification
A second, separate checking program compares every ingested rule entry against its official wording: is the percentage stated verbatim in the text? Is the change of heading covered? Is an additional condition missing? Only entries that pass this cross-check are given the status “verified” — currently 21,092 of 26,668 entries, which amounts to 9,433 distinct list rules. The rest stays a draft and honestly leads to “review manually”.
Three honest answers
The engine knows three verdicts: passed, not passed — and “review manually”, when a rule cannot be mapped with certainty or details are missing. Missing data is neither a yes nor a no. This third verdict is the heart of the methodology: no answer is better than a wrong one.
No false “passed” — by design
The system is built so that the most dangerous error — origin wrongly confirmed — is structurally ruled out: only verified rules give verdicts. If the engine finds no verified rule, it says “review manually” instead of falling back on a more lenient catch-all rule.
If an agreement offers an alternative rule that cannot be mapped with certainty, failing the main rule also becomes “review manually” — because the alternative might well be met. Honesty in both directions.
So-called ex headings (a rule applies to only part of an HS heading) are something no serious system answers from the HS code alone — here PräferenzPilot always says “review manually”.
AI explains. The engine decides.
The AI assistant reads documents, explains results in plain language and suggests entries. What it can never do: reach an origin verdict, set a rule status or soften a failure. The binding verdict comes exclusively from the deterministic engine.
This limit is not a promise but technically enforced, and secured by a dedicated adversarial test suite that runs on every change.
Four automatic controls — and one official cross-reference
Guard cases (golden master)
For every combination of agreement and rule type, test cases with their expected result are frozen. If a code change alters even one of those verdicts unintentionally, the test fails — before anything goes live.
Coverage guard
The number of verified rules must never quietly fall for any agreement. A snapshot comparison on every test run prevents creeping loss of quality.
AI safety invariants
A dedicated adversarial test suite makes sure no AI output can ever set a verdict, raise a status to “verified” or soften a failure — the limit is technically enforced, not merely promised.
Weekly source monitoring
13 of the 20 official sources are automatically re-fetched and their fingerprint (SHA-256) compared — if the EU amends a document, the check catches it and the affected rules are re-ingested instead of quietly going stale. The remaining 7 sources do not provide a file that can be fetched reliably by script (access protection or a broken text layer in the official document); they are registered with fingerprint and file size and are cross-checked manually.
Official cross-reference: WuP online
Our verification runs against the EU official texts (EUR-Lex/CELLAR). As an independent second reference we use WuP online of the German customs administration: it lets you look up the preferential arrangement, the type of proof and the cumulation matrix for each country officially — every article analysis links straight there, so you can cross-check every result independently of us.
Verified rules per agreement
All 78 countries with an EU preferential agreement in force are covered. How often the engine can give a verdict differs by agreement — where a rule is not (yet) verified, the answer is “review manually”, never an estimate. This table is deliberately unvarnished.
| Agreement | Destination countries | Verified rules | Source as of |
|---|---|---|---|
| EU–Mercosur | 4 | 6,588 | 2026-08-02 |
| EU–Canada | 1 | 3,773 | 2026-06-30 |
| EU–New Zealand | 1 | 3,235 | 2026-08-02 |
| EU–UK (TCA) | 1 | 1,793 | 2026-06-30 |
| EU–Japan | 1 | 1,560 | 2026-06-30 |
| EU–Chile | 1 | 884 | 2026-08-02 |
| EU–South Korea | 1 | 363 | 2026-06-30 |
| EU–SADC | 6 | 271 | 2026-08-02 |
| EU–Central America | 6 | 252 | 2026-08-02 |
| PEM zone | 23 | 243 | 2026-06-29 |
| EU–Ghana | 1 | 242 | 2026-08-02 |
| EU–Andean countries | 3 | 238 | 2026-08-02 |
| EU–Pacific | 4 | 237 | 2026-08-02 |
| EU–ESA | 5 | 219 | 2026-08-02 |
| EU–Singapore | 1 | 217 | 2026-07-01 |
| EU–Mexico | 1 | 206 | 2026-08-02 |
| EU–CARIFORUM | 14 | 206 | 2026-08-02 |
| EU–Cameroon/Kenya | 2 | 192 | 2026-08-02 |
| EU–Côte d'Ivoire | 1 | 191 | 2026-08-02 |
| EU–Vietnam | 1 | 182 | 2026-06-30 |
Every source document, with its checksum
The SHA-256 fingerprint proves which document an agreement's rules came from — byte for byte. For the sources checked automatically every week you can download the official file and recompute the fingerprint yourself. The documents marked “checked manually” do not return a byte-stable file when fetched (access protection or dynamic delivery) — their fingerprint documents exactly the version we ingested.
EU-Mexiko Globalabkommen, Beschluss Nr. 2/2000, Anhang III, Appendix II — Liste der erforderlichen Be- oder Verarbeitungen (EUR-Lex, konsolidierte Fassung 01.01.2021)
checked automatically every weekAs of: 2026-08-02SHA-256: 207b5d5506ebcec77c35e6b5df55dba64061d3bebc0c6d4f65111795ff3090b3Official source
EU-Mercosur Interims-Handelsabkommen, Anhang 3-B — Produktspezifische Ursprungsregeln (OJ L 2026/184)
checked automatically every weekAs of: 2026-08-02SHA-256: 4ff77dee8d17f66504c32ac97155299003c4ccc8799e01a36df37d0e49fbfe06Official source
EU-Côte d'Ivoire Interim-WPA, Beschluss Nr. 2/2019, Protokoll 1, Anhang II (ABl. L 49/2020)
checked manually — fetch not byte-stableAs of: 2026-08-02SHA-256: be5a3561c1e9b94484279ea94c83b64e19fbd5434656b9c09c82a24d8d6abeafOfficial source
EU-Ghana Interim-WPA, Beschluss Nr. 1/2020, Protokoll Nr. 1, Anhang II (ABl. L 350/2020)
checked manually — fetch not byte-stableAs of: 2026-08-02SHA-256: cedad5dd306971396d7765803cd6db100fcf01f4e1a9710bde41c4227f3fefbeOfficial source
Marktzugangsverordnung (EU) 2016/1076, Anhang II, Anlage 2 — maßgeblich für Kamerun und Kenia (ABl. L 185/2016)
checked manually — fetch not byte-stableAs of: 2026-08-02SHA-256: a4844cb15023acd13e5c20294ee48da9ec70d2b86ee5607486bfb98ee270f6d8Official source
EU-Pazifik-WPA, Beschluss Nr. 1/2026 des Handelsausschusses, Protokoll II, Anhang II (ABl. 18.03.2026)
checked manually — fetch not byte-stableAs of: 2026-08-02SHA-256: 17e945c2ad42d904d06013426d1f461a10a7763d0d9adcb580557ed7a4c037cdOfficial source
EU-ESA Interim-WPA, Beschluss Nr. 1/2020 des WPA-Ausschusses, Protokoll 1, Anhang II (ABl. L 93/2020; XHTML-Fassung — der PDF-Textlayer ist defekt)
checked manually — fetch not byte-stableAs of: 2026-08-02SHA-256: 707bdcad47736b1443c70aae0aec0bc71b14274ebde4373defdcd685e684555dOfficial source
EU-CARIFORUM WPA, Protokoll I, Anhang II (ABl. L 289/2008; XHTML-Fassung von CELLAR, CELEX 22008A1030(01))
checked manually — fetch not byte-stableAs of: 2026-08-02SHA-256: fdd759a22e2dd49e64cd304cbe9ce56049b3124838b80dde2567429f115e18d3Official source
EU-Zentralamerika Assoziierungsabkommen, Anhang II zu Anlage II — Liste der erforderlichen Be- oder Verarbeitungen (EUR-Lex, OJ L 346/2012)
checked automatically every weekAs of: 2026-08-02SHA-256: 5f50a30920ed0a3f43b115c1b315d35720cad4e3f6028408142433de640bd939Official source
EU-Kolumbien/Peru/Ecuador Handelsübereinkommen, Anhang II zu Anlage II — Liste der erforderlichen Be- oder Verarbeitungen (EUR-Lex, OJ L 354/2012)
checked automatically every weekAs of: 2026-08-02SHA-256: 47021529e929e34e23b2b5e043e7d4f17f9f525cf97386036d787d268fc523d9Official source
EU-Neuseeland Freihandelsabkommen, Anhang 3-B — Produktspezifische Ursprungsregeln (EUR-Lex, OJ L 2024/866)
checked automatically every weekAs of: 2026-08-02SHA-256: ee074a0ebcd0c3c5c2968826cabe57bf6f32ae60a3ac73d0bebbc559f5f14eb9Official source
EU-Chile Interims-Handelsabkommen (ITA, ABl. 2024), Anhang II — Produktspezifische Ursprungsregeln (Europäische Kommission, GD TAXUD)
checked automatically every weekAs of: 2026-08-02SHA-256: ff6e505a73ff4826d0963f15afecac856076456be0aeacd1a427d13c2775e657Official source
EU-SADC Wirtschaftspartnerschaftsabkommen, Protokoll 1, Anhang II — Liste der erforderlichen Be- oder Verarbeitungen (EUR-Lex, OJ L 250/2016)
checked automatically every weekAs of: 2026-08-02SHA-256: dd5d09beca072593613bbd0236bf532bd3ffc15b590cfd7c7bae47713248c7cbOfficial source
Revidierte PEM-Konvention (Beschluss Nr. 1/2023), Anhang II — HS2022-Transposition (EU-Kommission, GD TAXUD)
checked automatically every weekAs of: 2026-06-29SHA-256: f0b274c8ece39e277ccec6f02800abab16273567bf0b90483615d18289c7eb7cOfficial source
EU-UK Handels- und Kooperationsabkommen (TCA), Anhang 3 / ORIG-2 (EUR-Lex, konsolidiert 2024)
checked automatically every weekAs of: 2026-06-30SHA-256: 26e9602f38234005b99b80a5ee775caaabbfc37b996d99170553cb5dcaa01c84Official source
CETA (EU-Kanada), Protokoll über Ursprungsregeln, Annex 5 — Product-Specific Rules (saubere Annex-5-Fassung, Government of Canada, 2016)
checked manually — fetch not byte-stableAs of: 2026-06-30SHA-256: 13584c38594f8f5185758eb6265d2601a6dd5ce8897f0bdc95377737cb07bea5Official source
EU-Vietnam Freihandelsabkommen, Protokoll 1, Annex II — List of working or processing required (EUR-Lex, OJ L 186/2020)
checked automatically every weekAs of: 2026-06-30SHA-256: c1a85248c90c0a6e0d5d0fa9a0884f2f904a936e27e4fc164c4cbd7c10e6b3deOfficial source
EU-Singapur Freihandelsabkommen (EUSFTA), Protokoll 1, Annex B — List of working or processing required (EUR-Lex, OJ L 294/2019)
checked automatically every weekAs of: 2026-07-01SHA-256: ec883450c92f8b3f01f03b4b9f15705da9d82c44337063fa5c9e84113a9426ddOfficial source
EU-Japan Wirtschaftspartnerschaftsabkommen (EPA), Anhang 3-B — Product Specific Rules (Access2Markets / EU-Kommission, tradoc_157231) (ABl. L 330/2018)
checked automatically every weekAs of: 2026-06-30SHA-256: 496a085b828f0486dd23bb30d0d182d85f579424668bac618c8e388c2d05c729Official source
EU-Korea Freihandelsabkommen, Protokoll über Ursprungserzeugnisse, Anhang II (EUR-Lex, OJ L 127/2011)
checked automatically every weekAs of: 2026-06-30SHA-256: c0ee616e517145b92b25376867b31e144f45c4880b6bca28d6f0d75f791ce705Official source
What PräferenzPilot cannot do
- · Results are only as good as the inputs: the bill of materials, values and origin details come from the user. The system calculates deterministically on that basis — it cannot guess that an input is wrong.
- · Ex headings and unverified rules are always answered by the engine with “review manually” — deliberately.
- · PräferenzPilot produces review-ready drafts and calculations, but not legal or customs advice. Approving a proof remains the user's own responsibility.
We are happy to answer questions about the methodology: kontakt@praeferenzpilot.de