Methodology: how plots and documents are checked

EUDR Insighter supports the due diligence that Regulation (EU) 2023/1115 (EUDR) asks of operators and traders. For each plot of land, and for the documents of a delivery, it gathers evidence from maps, official registers and the documents themselves, and reports what it found and where each finding comes from.

What you get is context and warnings, not a rating. Every answer is a short summary, the sources it read, and whether there is a warning. Insighter does not decide whether a product may be placed on the market. The report supports your due diligence; it does not replace it, and the responsibility stays with your company.

The check has three steps:

• Step 1, GeoJSON Identifier (per plot, within seconds): reads the outline or point, locates it, and says whether the plot can be worked further.
• Step 2, GeoJSON Analyzer (the land check, per plot): the full check of the land in seven parts, A to G, ending with a report and its conclusion for the plot.
• Step 3, Documents (per delivery): reads the documents, checks each one, links them to each other and to the plots, and adds them to the report and its conclusion.

Every result is one of three words:

• Warning: something was found. The answer says what, where and how much, and names the sources.
• No warning: the check ran and found nothing. The answer says what was looked for and which sources were read (name, publisher, year or version, licence).
• Not verified: the check could not give a full answer. The answer always says why: which data, for which place or which years, and the reason (no data held for the place, partial data, a map that ends in an earlier year, a service that did not answer in time).

The Identifier can also stop a plot (it is not worked further). A missing answer never counts as "no warning".

Fixed rules decide. Every result comes from written rules: the same inputs always give the same result. AI is used only to read documents and, in the news search, to find items and class their tone, always with the exact words it rests on. AI never gives a conclusion: fixed rules decide what each reading counts for.

Sources must agree. Where two sources describe the same thing and disagree, the answer shows both values and says "sources disagree". It does not average them and does not pick one.

On the website and through the dedicated API. On this website you send one GeoJSON file of up to 50 plots, with up to 10 documents of up to 20 pages each, and you receive the report of each plot and one printable page for the whole check. Larger volumes, other file formats and the outline correction are available through the dedicated API: contact us.

The Identifier is step 1 and is built to be fast: it answers from reference files held on its own machine, calls no outside service, so the first answer for each plot comes within seconds.

What you send

• The plot: a Point, a Polygon, a MultiPolygon, a Feature holding one of them, or a FeatureCollection. On the website each Feature of the file is one plot. Coordinates in longitude and latitude (WGS 84), longitude first.
• A point stands for a circle of the area you declare (0.01 ha to 10,000 ha); without a declared area, 4 ha.
• The country you declare (optional): turns on the test for swapped coordinates.
• The commodity (cattle, cocoa, coffee, oil palm, rubber, soya or wood): give it for the Analyzer; without it part A of the Analyzer is not verified.

Limits of an outline

• Up to 20,000 points (all rings and parts counted) and up to 10,000 ha per plot.
• Refused as they are: lines, several points in one plot, mixed collections, coordinates outside -180..180 / -90..90, outlines crossing the 180th meridian, and coordinates in another system than longitude and latitude.
• A broken outline is never repaired silently. If the border crosses itself, parts overlap, a hole lies outside the outline or there is no area, the answer names the fault in words and the plot is stopped. You can fix the outline yourself, or have it repaired by the outline correction of the dedicated API (see "Outline correction and other formats").

What the Identifier checks

Where two sources describe the same thing (built-up land from two maps, water from satellite against mapped lakes and rivers), the answer shows both readings when they disagree.

Result of the Identifier

• Stop: the outline cannot be read, or half or more of the plot is surely in water ("surely" means no land lies within about 500 m of the sea part; a lake or water seen from satellite always counts as sure). A stopped plot is not worked further.
• Warning: water on 1% to under half of the plot; sea close to a shore (it may be land at the precision of the maps); the plot crosses a border; the centre lies in a town; a point given for more than 4 ha (the regulation asks for an outline above 4 ha); coordinates that look swapped; sources that disagree.
• Not verified: a map the Identifier needs could not be read this time; the answer names it.
• No warning: none of the above. The answer lists what was looked for: the outline, water, sea, border, town, a point over 4 ha, sources that disagree and swapped coordinates.

GeoJSON Analyzer (land check)

The Analyzer is step 2. It runs on the plots the Identifier did not stop; give the commodity (without it part A is not verified), the declared country stays optional. It reads data we hold on our own machines (maps, registers and statistics we keep copies of), so its answer does not depend on another service being up; the one exception is the elevation of part F, read from the publisher's open copy where ours is not held yet. Its parts run side by side; each part stands alone, and a part that cannot answer does not stop the others. The Analyzer checks plots up to about 55 km across; a larger outline gets the Identifier's answer only.

Each part lists its sources with the role each one plays: a voting source decides together with the other voting sources of the part; an evidence source is shown with its own value and compared with them (it can make "sources disagree", never a result on its own); a context source is shown to help you read the plot and never changes the result. A voting source that applies to the plot but could not be read, or is not held yet, makes the part "not verified" for what it would have answered: never a "no warning" by silence.

Read as published. No source is altered, and Insighter adds no thresholds of its own to a publisher's definitions. Official registers are read by kind (land-parcel cadastre, protected-area register, forest-management plan, and so on); the Coverage page lists the kinds read in each country.

A. Forest cover 2020

Step A uses the regulation's own definition of forest (Article 2(4)): land of more than 0.5 hectares with trees higher than 5 metres and a canopy cover of more than 10%, or trees able to reach those thresholds in situ, excluding land that is mainly in agricultural or urban use. Orchards and tree crops in agricultural use are not forest; plantations of trees for wood are forest.

A plot can be forest in 2020 even without trees standing on that day: land cut, burnt or blown down shortly before and left to grow back is still forest. Insighter therefore finds forest in 2020 by three ways, on a common grid of about 10 m; a place is forest in 2020 where any of them finds it:

• The reference forest map of 2020 says forest: a map made for this regulation, with its definition.
• An official forest record (forest-management unit, forest stand, forest parcel of the cadastre) says forest land, and the land was not mainly cropland or built land in 2020.
• Forest temporarily without trees in 2020: trees were there before (tree cover of 10% or more in 2000, or forest in an earlier year of the yearly land-cover maps), a loss of trees is recorded between 2010 and 2020, and the land in 2020 was neither cropland nor built land.

Other forest maps of 2020 vote beside the reference map. When their forest shares differ by more than 20 points, or one finds forest on 0.5 ha or more and another finds none, the answer says "sources disagree" and shows both. Maps of tree crops in 2020 are shown beside the forest maps: where they see the crop (for example oil palm or cocoa) on land counted as forest in 2020, the answer says so, to be explained with the documents. They do not take land out of the forest of 2020.

The commodity decides the result. The same land means different things for different commodities:

• Wood: forest in 2020 is expected (no warning). Wood from land that was not forest in 2020 is a warning (trees outside forest, or a plantation), to be explained.
• An agricultural commodity (soya, oil palm, cocoa, coffee, rubber, cattle): forest in 2020 on at least 10% of the plot or 0.5 ha, whichever is larger, is a warning: the goods may come from land deforested after 2020. Less forest than that: no warning.
• Without a commodity, part A is not verified.

The answer gives the forest share and hectares in 2020, the hectares each way found, the shares of crops, grass, built land and water in 2020, every map read with its own value, and whether the sources agree.

B. Land use since 2020

Question: was the forest of 2020 lost after 31 December 2020, and what is the land now?

• Records of authorities: official records of deforestation after 2020, where a country publishes them. A register can say "none" only when it is general, complete and holds every year from 2021 to its newest year; otherwise the registers cannot say "none" for the place, and the answer says so (with the missing years named). An official alert after 2020 on the plot cancels a "none". Records given only as points are not counted.
• Loss on the maps: loss of tree cover by year after 2020 on the land that was forest in 2020, from global forest-loss maps, joined in the humid tropics with a forest-change map of the tropics.
• What the land is now: the newest yearly land-cover map that reaches the plot, and what the land became after each loss (a change of use seen, regrowth seen, open land, or too recent to judge).

Result: a warning when an authority records deforestation on the plot after 2020, or when the maps show loss of the forest of 2020 after 2020 (seen by the maps only, it is said as possible deforestation, not documented); no warning when the maps find no loss after 2020 (and no register records one; where no register can say "none", the answer says that the "no" rests on the maps); not verified when the loss maps could not be read for the plot. The answer also gives the forest share now, as an estimate (the forest of 2020 minus the loss seen since), the land-cover classes of the newest year read, with that year, and the share of the lost land where a change of use or regrowth is seen.

Loss maps also show harvest, storms and fire, so the answer says what the land became afterwards; the fire years of step C are shown beside it. Yearly land cover reaches 2024, so a loss in 2025 or later cannot be judged yet and is said so. Clearings smaller than one map cell (about 30 m) can be missed.

C. Signals since 2020

Question: do fire maps, fire registers, disturbance alerts, yearly vegetation readings or radar show a change on the plot after 31 December 2020? When, and on how much of it? These are signals, not proof of a clearing.

• Fire: monthly burned-area maps and, where held, national and regional registers of burned areas (they list smaller fires than the maps). A fire counts from 1% of the plot or 0.5 ha. A month counts only when at least half of the plot's land was observed; recent months not yet published are "not verified".
• Disturbance alerts: optical and radar alerts of forest loss with their date. A confirmed alert after 2020 counts from 1% of the plot or 0.5 ha; a low-confidence alert is shown and not counted.
• Vegetation by year: a greenness index and a canopy-moisture index from yearly satellite composites, compared with the plot's own earlier years; a drop is reported only when it is clearly above the plot's own noise. Our copy reaches 2024: the answer says "data up to 2024; 2025, 2026: no data".
• Radar by year, where we hold it: a fall of 1 dB or more from 2020 is shown as evidence and can make "sources disagree"; it is never a warning on its own.

A signal on land that was not forest in 2020 is information only, except fire, which is reported on any land. Result: a warning for a fire on the plot after 2020, or a confirmed alert or vegetation drop on forest land, with its date and share; not verified for months or years that could not be read; no warning when nothing was found and every year and month was read.

D. Protected areas and indigenous peoples' land

Two parts, answered side by side.

• Protected areas: the plot is compared with the national and regional registers of protected areas held for its countries, and with an international register of protected areas. The answer gives the share of the plot inside protected areas (all sources joined, never counted twice), how many sources identify the plot and whether they agree, the share in zones of strict protection where the source says so, and areas within 1 km. Proposed areas, buffer and transition zones are reported apart and not counted as protected.
• Indigenous peoples' land: the plot is compared with the registers of indigenous and customary land held for its countries and with world collections of community-mapped land: the share of the plot on such land, the number of sources, and the legal status where known (titled, recognised, in process or claimed). Records given only as points are approximate.

Result: a warning when the plot lies in a protected area or on indigenous land (with the share and the number of sources): a point to look at, not a refusal. A plain "no warning" needs a general, complete official register for every country of the plot; where the data held for the country is partial, "not found" is "not verified". Where customary tenure is the general regime, no map can show that a plot is not on customary land, and the answer says so. In this version the indigenous part is answered "not verified" wherever no general, complete register of indigenous land is held for the place, which in this version is the case in nearly every country (the Coverage page shows where part D is live), so part D then says either warning or not verified.

E. Cadastre and official registers

Question: what do official records say about the plot?

• Parcels and zones: land-parcel cadastres, rural property registers and official land-use and forest zoning; where an edition of 2020 is held, the parcels at 31 December 2020 too. The land-use class is taken from the record, never guessed.
• Official forest: forest-management units, forest maps and forest stands. A stand, a forest map or a forest parcel of the cadastre is forest; ground of a management unit without a stand is "not verified", never "not forest".
• Declared farm parcels: the crop declared by farmers, where these registers are held.

A plain "lies in no parcel" is given only where the registers held cover at least 97% of the country's land, and a plain "not forest" only where the records held cover at least 95% of the forest area the country reports. Otherwise the answer is "not verified" for that part. Where a register's class and the land itself differ, the answer says so; the land decides forest (step A).

F. Surroundings

Question: what lies on the plot or near it that you should be aware of?

Each feature is given with its type, its name where mapped and its distance (0 when inside). Roads and water are shown as context. Features are those mapped: a place not mapped is not seen, so "nothing mapped" never means "nothing there". Works we do not hold yet (dams and dikes drawn as lines, power lines, pipelines, airfields) are said as not verified.

The land of the plot (context): altitude, slope and the share steeper than 30 degrees, the main orientation of the slopes, long-term climate averages, soil at 0-30 cm, and an estimate of soil erosion in classes of tonnes per hectare and year. A large share of very steep ground or of high erosion is a point to be aware of. Where a soil property is not held yet, the erosion estimate is not verified.

Does the commodity fit the place? One plain answer from five readings: whether the country records production of the commodity, the climate, the altitude and slope, the soil acidity, and what the land-cover maps show. It is fits, unusual for this place (one or more readings outside the usual range, with the evidence) or unlikely (the country records no production of the commodity and the climate does not suit it). Wood and cattle are judged on the recorded production and the land cover only.

Result: a warning when one or more points to be aware of were counted, or the commodity is unusual or unlikely for the place; otherwise not verified, with what could not be read named. Because the works listed above are not held yet in this version, step F always names them as not verified and does not answer "no warning". Step F never gives a finding of deforestation or of illegality.

G. Carbon (supplementary)

An estimate of what the plot lost and gained in carbon, per year and per hectare, for 2001-2020 and for 2021 to the newest year of the loss map: losses from forest lost by year, gains from default growth of the forest still standing, and the net (a sink or a source), with a low and a high value. Not counted, and said in every answer: soil, litter and dead wood; gases other than carbon dioxide; degradation without loss of canopy; growth after a loss. Below it, the forest-land balance the country reports, compared with its reference years: it may read "warning" when the balance got worse beyond the country's own uncertainty, and it stays outside the conclusion like the rest of part G.

Carbon is information only. It is outside the conclusion, it is not the regulation's test, and it is not the official verification under the Renewable Energy Directive.

Online checks (a separate call of the dedicated API)

Some sources can only be read online at the moment of the check: an independent assessment of the plot made by another method, recent satellite images over the plot, radar by year for the months our copies do not hold yet, and national services asked online where we keep no copy. They are read in a separate call, side by side; a source that does not answer is asked once more before it is reported as "not verified". The answer lists every source with its publisher, its licence, whether it answered and after how many attempts. These sources are evidence: they are shown beside the Analyzer's parts and can make "sources disagree", never a "no warning" on their own. They are not part of the website check.

The land report and its conclusion

When the Analyzer has run to its end, the report of the plot is made from the Identifier and the Analyzer. It gives a short summary, the result of each part with its evidence, every source read (named, with publisher and licence), and every part that could not be verified with its reason. It ends with one of four conclusions:

The fixed rules, in this order (the same results always give the same conclusion; no AI takes part):

• Warning: the Identifier stopped the plot or did not answer "no warning" (a warning, or a map it could not read), or any part of the Analyzer A to F gave a warning (land that does not match the commodity declared, and a point to be aware of in part F, included).
• Not verified: the land check did not run to its end, no part answered "no warning", or all three parts the deforestation question rests on (A, B and C) are not verified.
• No warning, evidence partial: any part is not verified, or a "no warning" rests on thin evidence: fewer than two sources, or data not graded complete for the country, or a part of it left unverified.
• No warning, sources agree: in every other case.

Carbon (G) stays outside the conclusion. Since parts D and F are mostly "not verified" in this version (see above), a plot with nothing found usually ends "No warning, evidence partial". Where the online checks were run (dedicated API), an online source that finds something where the Analyzer part it speaks for gave no warning makes the report say "sources disagree", with both named; this is a warning.

Documents are sent with the plots: for the whole delivery. Each document is scanned, read and checked on its own, then linked to the other documents and to the plots. The companies named in the documents are checked in official registers and lists (and, through the dedicated API, in the news). The documents part gives context and warnings: what each document says, with the exact words it was read from, how the documents link, where something does not fit, and which sources were asked.

Files

• PDF, PNG or JPEG, recognised by their content, not their name. Up to 200 pages a file on the dedicated API, 20 on the website; images up to 40 million pixels.
• For each document you choose what kind it is: transport document (delivery note, CMR, consignment note); invoice; sale and purchase contract; title deed or land register extract; lease or concession contract; harvesting authorisation or permit; management plan; certificate (FSC, PEFC, other); due diligence statement (EUDR); customs document; weighing or reception slip; something else, with a short description.

Safety and reading

• Antivirus first. Every file is scanned for viruses before anything else. A file that cannot be scanned is refused, as is a file with a virus. The file is also checked for active content and for text written to mislead an automatic reader; such text is never handed on.
• Text: the text layer of a PDF is read; scans are read by a text reader (OCR) that runs inside our service.
• AI reading, with quotes: an AI reading service reads every page. Every value it gives must come with the exact words it was read from, and a value counts only when those words are also found by the second reading (the text layer or the OCR). A value read by one reader only is "to check" and never makes a warning on its own. AI never gives a result or a conclusion.
• Partial reading: pages that could not be read are named; a partly read document is "not verified".

Checks of each document

• Kind: the kind read is compared with the kind you chose; a difference is a warning.
• File history: the dates, program and saves recorded in the file. A file changed after it was signed, or made before the date written on it, is a warning. The author recorded in the file is not kept.
• Plausibility: the check digit of tax and registration codes (for the countries whose codes carry one), dates that do not exist or lie in the future, placeholder names, marks such as SPECIMEN, VOID or DRAFT, and template text. A failed rule is a warning, with the value and its quote.
• Commodity: the product named in the document is mapped to its customs code and to an EUDR commodity (Annex I of the regulation), and compared with the commodity you declared. A document of another EUDR commodity (wood declared, a soya document received) is a warning. Wood with wood products (furniture, paper, pulp) is one chain. Goods outside the regulation (packaging, freight, services) are named as such and not counted.
• Short summary: a few statements about each document, each with the exact words it rests on. A statement whose words are not found in the document is dropped. The summary is context only.

The same file sent again by the same key is not read again: its stored check is reused.

The documents together

Every document of a delivery is linked to the companies, permits, plots, certificates, containers and products it names. Links are made only on hard keys: tax and registration codes (with their check digit tested); permit, waybill, invoice, certificate, container and lot numbers; the plot identifiers you declared. A name alone is never a link: two documents that share only a company name get a "candidate" link for a person to confirm, never used as evidence. The same number with different content (two invoices with one number and different totals) is a warning.

On these links, fixed rules follow the chain of the goods:

• Product: one commodity across the delivery (wood with wood products is one chain).
• Parties: each seller other than the first holder bought the goods on another document of the delivery.
• Dates: the permit comes before delivery and invoice; a certificate is valid on the date of each goods document.
• Plot coverage: each goods document with a quantity reaches a plot through hard-key links.
• Claims: no party sells a higher certification claim than it bought (a lower claim is allowed, a higher one never).
• Consistency: the country of origin, and the species, agree between the documents and with your declaration.
• References: a document number referred to is in the delivery, and the kinds of document every delivery needs are there.
• Quantities across documents (what leaves is not more than what came in) are not checked in this version, and the answer says so.

Permit and parcel numbers read on the documents are compared with the plot and with the identifiers you declared; place names and the country of origin are compared with the plot.

Companies, certificates and permits

Each company found by its tax or registration code is checked against official records:

• Exists and is active on the date of the document, in the company register of its country. A company struck off, inactive or bankrupt on that date is a warning. "Not found" is a warning only in a register that holds every company of its country, and it is said as of the date of our copy; where we hold no register for the country, the check is "not verified".
• Name: the name on the document against the name the register holds for that code.
• Banned parties: each tax code and company name against international sanctions lists and official lists of banned or embargoed parties. An equal tax code is a warning; an exact name in the same country is "possible match, to check".
• Certificates: a certificate code is valid on the date of the document, held by a company of the document, and the company is not disassociated from the scheme.
• Permits and harvest tickets: the number is found in the official permit register of its country, with the document date within the permit's dates.

A company is looked up once; the answer is stored with its date and source and reused by later documents. Where we keep no copy of a register, an online register can be asked as a backup (a separate call of the dedicated API): only identifiers read from the documents (a tax or registration code) leave our service, never a file, a page, a name or the plot. Before anything is sent, the answer lists every item and the service it goes to; sending it again with your confirmation is your consent. A register that blocks automatic reading is not worked around: the check is "not verified".

News

The news is searched for each company named in the documents, and for the commodity in the country and in the region of the plot, back to the end of 2020, in an open global news index and news publishers' feeds, and by an AI search on open sources. The AI finds and quotes: each item it returns carries its link, publisher, date and exact words, and an item whose words cannot be found at its link is dropped. The AI also classes the tone of each item (negative, positive or neutral) with the words of the title it rests on; a class whose words are not in the title is dropped. An item is about the company only when it holds the exact name and the country matches; any other item is context. A negative item on an exact match is a warning ("negative news to read"), by a fixed rule; everything else is context. News never makes a "no warning" on its own. The news search is a separate call of the dedicated API: the queries are shown before anything is sent, and run only on your confirmation. It is not part of the website check.

The full report and its conclusion

With documents, the report of each plot is made from the Identifier, the Analyzer and the documents together (the documents can be read before the Analyzer; the full report waits until the Analyzer of the plot has run to its end). To the land report it adds the documents part (warning, no warning or not verified) and cross-checks between the documents and the plot, each answered warning, no warning, not verified or not applicable (nothing to compare):

• the country of the permit (or the country of origin on the documents) against the country the plot lies in; where they differ, the report says so plainly;
• the commodity declared against the documents: a document of another commodity is a warning;
• the plot identifier written on the documents against the outline;
• the harvest period of the permits against the years of tree-cover loss the Analyzer saw on the plot: a loss year outside every permit period is a warning;
• the largest quantity declared on one document against the plot's expected production: not verified in this version (the expected production is not given);
• the species on the documents against the forest type of the plot: not verified in this version;
• the land findings of the plot carried to the deliveries linked to it: a warning per delivery.

The conclusion follows the rules of the land report, with the documents added:

• Warning also when the documents part is "warning", any document check gives a warning, or a cross-check gives a warning.
• No warning, evidence partial also when a document, a document check or a cross-check that applies is not verified, or when a check is not made in this version (the quantities across documents).

The report names every source it read and every part it could not verify, with the reason. It gives context and warnings, never a rating: you decide what they mean for your due diligence.

Available through the dedicated API: contact us. The website reads GeoJSON only.

An outline is the border of your plot drawn as a list of points. Outlines often come in another file format, in another coordinate system, or with drawing mistakes (a border that crosses itself, a point clicked twice). The outline correction turns such an outline into one Insighter can check, and can also cut out ground that should not be in the plot. In both cases you first see what would change, and nothing is applied until you accept it.

Other formats. An outline in KML or KMZ, a zipped Shapefile, a GeoPackage, WKT, or a CSV of points is read and converted to GeoJSON (and written back as KML, a zipped Shapefile, a GeoPackage, WKT or CSV, if you wish). An outline in another coordinate system (for example Stereo 70 in Romania, or UTM) is converted to longitude and latitude when the file or you name the system. When only bare numbers are given, the usual systems of the declared country are tried and the one that puts the plot inside the country is proposed, with a sentence asking you to check it; when several fit, you choose. Heights are dropped, a closed line is read as the border of an outline, and several outlines of one file can be joined into one plot. Drawings (DWG) and GPX are not read.

Repair of a broken outline. An outline that cannot be read as a valid shape is repaired, keeping every part:

• a border that crosses itself (for example a figure of eight) or touches itself;
• parts that overlap or touch along a line;
• a hole lying outside its outline.

Points repeated one after the other, and points lying on a straight edge (including spikes of no width that go out and come back on the same line), are dropped when any outline is read: they change no land. Longitude and latitude given the wrong way round are swapped back, when a latitude is beyond 90 degrees or the plot lies inside the declared country only once swapped. Every change is named in one sentence, and what could not be checked is said. An outline that cannot be repaired is refused, and nothing is done.

Cutting parts with a warning. Some plots hold ground that should not be in a delivery's plot. On a plot you have already checked, a proposal lists these parts, each with its area, its share of the plot, its outline and how precise it is ("exact: the outline as its register draws it", or "cells of about 10 m"):

• water (lakes, sea, water seen from satellite);
• protected areas (the ground part D counts);
• built-up land on the newest land-cover map;
• forest loss after 2020 on the maps (not for wood);
• land not identified as forest on any map (for wood only).

Water is always looked for; each other kind only where the part of the land check that finds it has been run on the plot. You choose the parts to cut and first see the result; only when you confirm is a new plot made, in its own package. Small slivers are handled by fixed rules, and each is said: patches of map cells under 0.1 ha are not offered (they stay in the plot), and pieces of the result under 0.1 ha left by the cut are dropped; holes left by a cut are not filled. The new plot gets its Identifier answer at once and is looked at again for parts left. When the original plot had its land check, the land check of the new plot is worked again on the new outline. The original plot and its package are never changed. A point cannot be cut.

Both work the same way: you see what would change first, and only what you accept is applied.

The Coverage page shows, for each country, the parts of the Analyzer that are live: the part runs there on data that covers the country. A part that is not live for a country answers "not verified" where its data does not cover the plot. The page lists the kinds of official register read in each country, never the registers themselves.

• Coverage is not the same in every country. Where data is missing or partial, the answer is "not verified", with the reason.
• Every map has its own errors: small clearings, young plantations, shaded tree crops and the edges of fields. Where maps disagree, the answer shows it.
• A forest cut shortly before 2020 and converted later is found only through a recorded loss before 2020 or an official forest record. Cuts narrower than about 30 m, selective felling and loss hidden by cloud can be missed.
• We hold registers as they are now, not their history: a register that took land out of the forest category after 2020 is not seen unless the change shows on the maps.
• Burned-area maps miss small fires; where no register lists small fires, a small fire can be missed.
• Yearly vegetation readings end in 2024 in this version.
• The protected-areas status of a country is that of its main territory.
• Satellite readings are readings, not verdicts. Carbon is not the regulation's test.

• This methodology is model version v1.3 beta. Production serves one model at a time. A change of which checks run, or of how a result or the conclusion is made, is a new model version; a published version is never edited.
• The map and register data of each theme carry their own data version. Adding or fixing a source changes the data version, not the model version.
• Coming in the next version: documents are checked only after the land check of the plot has run, so every document check can use its result.