EUDR Legality Document Assessment Methodology

What the score is, and what it is not

The TRACT EUDR legality analyser produces a single assessed confidence score from 0 to 100 for a document submitted as evidence of legality under Regulation (EU) 2023/1115.

It answers this question: how much weight should a careful reviewer place on this document as Article 9(1)(h) evidence, given what is visible on its face and what is known about the origin?

It does not answer: is this document genuine? Nothing is checked against an issuing register. For most producer countries no public register exists to check. A high score means the instrument is of a strong type, well formed, from a plausible authority, in a context where it can be corroborated. It does not mean the document has been verified.

Legal basis

Every element of the method is anchored in a provision of the Regulation. The table sets out which provision drives which part of the assessment.

Inputs

The assessment runs on three types of input:

1. The document itself. A PDF, JPG or PNG of the instrument submitted as legality evidence. Scans and photographs are accepted, in any language.
2. The origin reference library. The document types, issuing authorities and administrative levels recognised for each covered origin, together with what can be independently checked and how.
3. The country risk profile. Benchmarking tier, corruption indicator, curated concerns and registry verifiability for the origin.

Country risk data

Country risk data is held for six origins: cocoa in Côte d'Ivoire and Ghana, coffee in Brazil and Vietnam, and oil palm in Indonesia and Malaysia. Where a document originates outside these six, no curated profile exists. The model is instructed to cap country context accordingly and to say so in the assessment.

Benchmarking provenance

Tiers were read from the Annex to the instrument itself on 8 September 2026, not from secondary reporting. An earlier edition of this document said EUR-Lex blocks automated retrieval; it served both the consolidated Regulation and Implementing Regulation 2025/1093 on request, so that assumption was retested and dropped.

Reading the Annex directly resolved all three entries that had carried a confidence caveat. Ghana is listed under low risk countries and was corrected from standard. Indonesia and Malaysia appear in neither list, so they are standard by default, as recorded. Every tier-confidence field now states the date the Annex was read and what it said, and the interface no longer shows a conflict caveat for any of the six, because there is no longer a conflict to show. The residual risk is staleness, not sourcing: the first scheduled review falls in 2026, so re-read the Annex before relying on a tier for a compliance decision.

Status of the list

The European Parliament objected to the benchmarking list on 9 July 2025 by 373 votes to 289, criticising the methodology and the classification of major deforestation fronts as standard rather than high risk. The objection did not block the instrument and it remains in force. The first scheduled review falls in 2026, so tiers may change.

Effect of the tier

A low-risk tier permits Article 13 simplified due diligence and lowers the competent authority check rate to 1% of operators, against 3% standard and 9% high. It does not remove the Article 9(1) information obligation and does not lower the Article 9(1)(h) evidential standard. The tier is therefore an input to one component at 0.15 weight, not a shortcut.

The two-pass pipeline

Assessment runs as two sequential model passes followed by deterministic aggregation in code. Pass 2 depends on Pass 1's output, so the two cannot be parallelised. The document itself is re-sent in Pass 2 so that scoring is grounded in the artefact, not only in Pass 1's description of it.

Both passes use Claude Opus 5 with adaptive thinking at high effort and JSON-schema constrained structured output. All generated text follows a fixed style rule set in the prompts: plain English, short sentences, roughly an 8th-grade reading level, and exact legal citations.

The model scores each component from 0 to 100, with a written finding citing specifics from the document. The components are designed to separate cleanly. A genuine, well-formed instrument can still be weak evidence for the purpose it was submitted for, and a single-dimension score would collapse that distinction.

Weights sum to exactly 1.00. They were set from the structure of Article 10(2), not fitted to labelled outcome data. See limitation 5.

3.1 Document integrity · weight 0.20

Is the document complete and properly executed on its face? We look for expected fields, signatures and seals where the instrument requires them, all pages supplied, legibility, no signs of alteration, and no partial extract presented as a whole.

Lowers the score: missing signature blocks, absent seals, a single page of a multi-page instrument, referenced annexes not attached, poor legibility, evidence of tampering.

3.2 Internal consistency · weight 0.20

Do the document's own contents agree with each other? Dates in sequence, areas and percentages that add up, the same party named consistently, identifiers in the correct format, an issuing authority matching the instrument type, and arithmetic that holds.

3.3 Issuing authority plausibility · weight 0.20

Does the named issuing body exist, and could it lawfully issue this instrument for this place? This is checked against the origin reference library: the right authority for the document type, the right administrative level, and territorial competence over the stated location.

3.4 Evidential sufficiency for Art. 9(1)(h) · weight 0.25

How far does this document actually go towards "adequately conclusive and verifiable information"? This carries the highest weight because it speaks most directly to the statutory test.

The instrument is judged for what it is, not for what it is labelled. The governing hierarchy is set out below.

3.5 Country context, Art. 10(2)(h) · weight 0.15

How much does the operating environment reduce the weight this document can carry? This component scores the environment, not the document. A flawless document from a low-verifiability, high-falsification origin still scores lower here. The inputs are:

• Article 29 benchmarking tier: low, standard or high
• Corruption Perceptions Index 2025 score
• Curated country and sector concerns: documented falsification prevalence, enforcement gaps, tenure-system characteristics, sector-specific risks
• Registry verifiability: whether documents of this kind can be independently checked at all

Computed, not judged

Of the five, this is the only component that does not look at the document. Every input is a property of the origin, held with a source and a date, so the model contributed no information to it and contributed variance, to 15 percent of every score. It is now calculated from a published formula, and the component's written finding is the derivation itself rather than prose about it.

The computed values are Brazil 67, Malaysia 59, Indonesia 52, Ghana 53, Vietnam 42, Côte d'Ivoire 35. The term weights are a published judgement, not an empirical fit, exactly as the component weights are. The model is told the computed value so that its rationale can take account of it, and a component entry returned for it anyway is discarded rather than merged, so a model ignoring the instruction cannot reach the score.

The inputs previously included whether documents of this kind can be independently checked, which made this a function of the document class as well as the origin. That clause is gone: it was blurring two components, because how far a particular instrument goes towards Article 9(1)(h) is what evidential sufficiency scores at weight 0.25. What it gives up is resolution. The model could previously distinguish a Ghanaian registered title, which is searchable, from a Ghanaian allocation note, which is not. That distinction belongs in evidential sufficiency, and putting it there properly needs a verifiability field per document type in the reference library.

Aggregation and bands

The overall score is computed in code, not by the model. A reviewer can reproduce the number by hand from the five published component scores, which is the point.

The band is derived from the computed score, never supplied independently. One further rule is then applied.

How to read the score

The band is the output. The number supports it. Repeated runs of the same document move the overall by about 5 points, so a score sitting within 5 of a threshold could report either side of it next time. Every assessment therefore records how far it is from the nearest band boundary, and says so on the face of the result. The thresholds are derived from the band definitions rather than hardcoded, so they cannot drift apart, and the interface leads with the band and shows the score in support of it, because a figure accurate to about 5 points presented as a large number invites a reviewer to read it as a measurement.

The element checklist

Every document type in the reference library already lists what a reviewer should find on the face of an instrument of that kind. There are 172 of these items across the 49 types, and until now they were prose in a prompt that the model could answer or ignore. They are now a required output: for the matched document type, every element is answered with a verdict, the page it appears on, and the text that establishes it quoted as printed.

Component findings must cite the element ids they rest on, because a finding that cites nothing is an assertion a reviewer cannot check. The Ghana specimen's designed defect used to be a sentence inside a component finding; it is now an element marked absent, with the note that ground rent is mentioned as a future obligation while no Lands Commission concurrence appears anywhere on the document. A reviewer can take that to the page.

Structured outputs generate in the order the schema declares, so the schema is ordered elements, then gates, then component scores, then the rationale. That sequences reading before judging inside a single call. Element ids are positional within a document type and derived in code from the existing checklist text, so no reference data had to be rewritten, and the schema's enum is exactly the matched type's checklist, so the model can neither answer an element that does not exist nor quietly skip one that does. Where no library entry matches the document type there is no checklist, and the assessment says so.

Disqualifying gates

A weighted mean cannot express a finding that ends the enquiry. With issuing authority plausibility at zero the arithmetic still permits 80, the top of moderate, so an instrument its named issuer could not lawfully issue scored around 57 and was reported as limited: weak evidence, needs corroboration. The specimen rule had the mirror-image problem: it craters two components carrying 0.45 between them, so its maximum effect is 45 points, and a clean specimen from a plausible authority landed near 41, also inside limited. In both cases the arithmetic contradicted this document's own prose.

The score is unaffected. Both bands are published: the score band is what the weighted sum alone gives, so the arithmetic remains reproducible by hand, and the band is what the assessment concludes once the gates are applied. Where they differ, the gate that did it is named. The cap is one stated rule applied to published numbers, not a second hidden calculation.

Only findings that are unambiguous on the face of the document are gates. An expired term deliberately is not: Article 9(1)(h) asks whether production accorded with the relevant legislation, so a term that ended after the production it covers is not a defect. Expiry remains a high-severity red flag. And because the gate carries the disqualification, the components are no longer instructed to crater on a specimen. They score the instrument on what is visible, which tells a reviewer more: a specimen can be a well executed example of a strong instrument type, and the gate says plainly that it evidences nothing about a real supply chain.

The determinism boundary

Deciding what belongs to the model and what belongs to code is the central design choice of this methodology.

Why the arithmetic moved into code

In development the model was asked to compute the weighted overall itself. On a test document it returned 26 where its own component scores implied 34. Both fell in the same band, so the conclusion was unaffected. But a score a reviewer cannot reproduce is not defensible in a compliance file. The model now supplies judgement only.

Robustness handling

The model has been observed to emit the same component twice. Post-processing keeps the first occurrence of each, restores canonical order, and fills any omitted component with a zero score and an explicit "not assessed" finding. A malformed response degrades visibly rather than silently skewing the total.

Schema constraint

Structured outputs reject minimum and maximum on integer fields, so the 0 to 100 range is stated in the field description and enforced by clamping after parsing.

Calibration anchors

Anchors are specified in the prompt to prevent clustering:

• A well-executed, internally consistent registered title from a moderate-risk origin sits at 75 to 85, not 95. The residual reflects that nothing has been verified.
• A self-declared registry printout offered as tenure evidence belongs at 25 to 40, however cleanly it is printed.
• Above 85 is reserved for instruments that are both intrinsically strong and independently checkable.
• There is an explicit instruction not to cluster in the middle, and to distinguish clearly between strong and weak instruments.

Outputs beyond the score

The score alone is not actionable. Each assessment also returns four further outputs.

Red flags

Specific indicators that the document may not be what it purports to be, or may not support what it is being used to support. Each is rated low, medium or high severity, with an explanation of why it matters. The model is instructed not to manufacture flags.

Verification steps

Concrete next actions, each tagged by feasibility.

Each step and each corroboration request also carries an impact rating of high, medium or low. High means it addresses the decisive factor, or could confirm or refute the document itself. The model is told that most assessments have only one or two high-impact steps. The interface sorts and groups by these ratings, so a reviewer sees the items that move the assessment first, packaged by who can act on them.

Steps must be specific. "Verify the title" is rejected because it names no source, no identifier and no comparison. The standard names all three: the register or dataset, the identifier taken from the document, and the result that would confirm or contradict what the document says. For a Brazilian CAR that is a SICAR consultation on the document's own CAR number, with the returned polygon area compared against the area the document declares.

The prompt teaches that form without naming a source or a value. It used to carry one hardcoded example, a CAR number and a hectare figure both taken from the Brazil specimen, injected into every assessment of every document from every origin. A specific figure in the cached prefix of every request is a value that can leak into an unrelated assessment, and Brazil is the only one of the six with a public land register, so the example modelled a step most documents cannot support and invited the model to invent portals for the others. The origin's real checkable sources are listed in the country profile, which is where the content belongs. Earlier editions quoted a specific CAR number here; it was drawn from a synthetic specimen, and SICAR is a genuinely public consultation, so anyone who tried it got nothing back.

Corroboration needed

The documents that should sit alongside this one to close the Article 9(1)(h) gap for the categories it touches.

Verification status

A plain sentence recording that no issuing register was contacted, and what that means for reading the score.

Validation to date

Three synthetic specimens were assessed, each constructed with a known defect. Calibration mode was used so that the scoring exercised instrument strength rather than bottoming out on the specimen watermark. These are the current pre-computed example results, which are what the tool displays.

The Brazil row demonstrates the components separating cleanly. It is a genuine, well-formed federal instrument, with integrity at 74 and authority at 85, that is nonetheless weak evidence for the purpose it was submitted for, with evidential sufficiency at 32.

Calibration mode

A synthetic specimen declares on its face that it has no legal effect, which correctly craters two components and made the fixtures useless for testing whether the scoring distinguishes a strong instrument from a weak one. Calibration mode is now much narrower than it was: it used to instruct the model to score the instrument on its merits as if genuine, and the gates removed that need, because components score the instrument on its merits either way and the no_legal_effect gate carries the disqualification. So calibration suppresses that one gate and the matching red flag, so a fixture produces a band describing the document type. Everything else is unchanged, including the issuer_without_competence gate.

Known limitations

1. No authentication. No document is checked against an issuing register. The score reflects intrinsic quality and context only.
2. Absolute calibration unproven. Component discrimination is demonstrated. The pitch of the absolute numbers is not yet validated against real documents.
3. Benchmarking tiers are point-in-time. No longer secondary-sourced: all six were read from the official Annex on 8 September 2026, which corrected Ghana from standard to low. What remains is that nothing re-checks them. The first scheduled review falls in 2026, and a tier change would not be noticed by the tool.
4. Six origins only. Documents from elsewhere are assessed without a curated country profile and without an authority reference library, which materially weakens two components.
5. Weights are a judgement, not an empirical fit. They were set from the structure of Article 10(2), not derived from labelled outcome data. They are a defensible starting point, to be revised once real-world review outcomes exist.
6. Single-document scope. Each document is scored in isolation. The tool does not yet assess whether a set of documents collectively satisfies Article 9(1)(h) for a given plot or consignment, which is the actual compliance question.
7. Non-determinism. The model may return slightly different component scores on repeated runs of the same document. Aggregation is deterministic. The inputs to it are not.

Change control

Any change to weights, bands or component definitions changes the meaning of previously issued scores. Bump the version at the head of this document, record the change, and do not compare scores across versions without noting it. Re-generate the pre-computed example results after any such change.

Corrections to reference data

A correction to the origin reference library or the country risk profile changes scores without changing the method, so it is recorded here rather than as a version bump. Do not compare a score issued before a correction with one issued after it for the affected origin.