diff --git a/DESCRIPTION b/DESCRIPTION
index 3c00a79..c895f59 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -14,13 +14,24 @@ Encoding: UTF-8
 LazyData: true
 Imports: 
         jsonlite,
-        httr
+        httr,
+        stats,
+        tools,
+        utils
 Suggests:
         knitr,
+        data.table,
+        nanoparquet,
         rmarkdown,
+        rWCVP,
+        rWCVPdata,
+        sf,
+        terra,
         testthat,
         devtools,
         BIEN,
+        TNRS,
         vcr (>= 0.6.0)
+Additional_repositories: https://matildabrown.github.io/drat
 VignetteBuilder: knitr
-RoxygenNote: 7.3.2
+RoxygenNote: 7.3.3
diff --git a/NAMESPACE b/NAMESPACE
index 12632d5..0d3cc0f 100644
--- a/NAMESPACE
+++ b/NAMESPACE
@@ -4,6 +4,11 @@ export(NSR)
 export(NSR_citations)
 export(NSR_data_dictionary)
 export(NSR_from_coordinates)
+export(NSR_local)
+export(NSR_local_build)
+export(NSR_local_by_region)
+export(NSR_local_remove)
+export(NSR_local_status)
 export(NSR_metadata)
 export(NSR_political_divisions)
 export(NSR_simple)
diff --git a/NEWS b/NEWS
index a985c9d..8015925 100644
--- a/NEWS
+++ b/NEWS
@@ -4,6 +4,24 @@
 
 ## NEW FEATURES
 * Added new function to check native status using species name and coordinates (NSR_from_coordinates)
+* Added an offline implementation: NSR_local() answers without an internet connection from a
+  local cache built by NSR_local_build(), with NSR_local_status(), NSR_local_remove() and
+  NSR_local_by_region() alongside it.  The checklist sources are acquired on the user's
+  machine; nothing derived ships with the package.  The offline path needs data.table,
+  nanoparquet, terra, sf and TNRS, all of which are Suggests.
+* NSR_local() gains exclude_extinct (default TRUE).  WCVP flags a distribution extinct
+  without a date, so the cache keeps those records and the query decides: the default
+  answers about the present day, FALSE models a past distribution.  Either way a region
+  a taxon has been lost from stays part of its native range, so such a record is
+  reported absent rather than introduced.
+* NSR_local() answers from political division names by default.  Coordinates are
+  echoed but not used unless use_coordinates = TRUE: placing every record by point
+  costs a raster lookup per call, and turning coordinates into political divisions is
+  what GVS already does.
+* Where a record gives both coordinates and division names and the two resolve to
+  different places, NSR_local() no longer pools them.  Neither is treated as
+  authoritative: place_conflict flags the record, native_status is UNK, and the two
+  bases are answered separately in native_status_coordinates and native_status_names.
 
 ## Bug fixes
 
diff --git a/R/NSR_local.R b/R/NSR_local.R
new file mode 100644
index 0000000..f830787
--- /dev/null
+++ b/R/NSR_local.R
@@ -0,0 +1,740 @@
+#' Determine native status without an internet connection
+#'
+#' Offline Native Status Resolver.  For each taxon and place, the status every built
+#' checklist gives it, reduced to one answer.
+#'
+#' Output carries \code{\link{NSR}}'s columns, and adds: \code{isEndemic};
+#' \code{native_status_conflict} and \code{native_status_conflict_type}, saying whether
+#' and how the sources disagreed; \code{native_status_scope}, which of the polygons
+#' bearing on the place produced the answer; \code{n_subpolygons_native} and
+#' \code{n_subpolygons_introduced}; \code{native_status_opinions}, every opinion consulted;
+#' \code{native_status_coordinates}, \code{native_status_names} and \code{place_conflict}
+#' (see Place); \code{regions_matched}, what the place was matched on; and
+#' \code{taxon_evaluable}, whether any source holds native status for the taxon at all -
+#' which is what separates \code{A} (absent) from \code{UNK} (no data).  The coordinates
+#' given are echoed as \code{latitude} and \code{longitude}.
+#'
+#' \strong{Place.}  By default the answer comes from political division names
+#' (\code{country}, \code{state_province}, \code{county_parish}), which are resolved to
+#' GADM units through the GNRS backbone and carried to each source's own geography by the
+#' spatial link table.  Any \code{latitude} and \code{longitude} present are echoed but
+#' not used: placing every record by point costs a raster lookup per call, and turning
+#' coordinates into political divisions is what GVS already does.  Resolve them there and
+#' pass the divisions here, or set \code{use_coordinates = TRUE} to have this function do
+#' it.
+#'
+#' With \code{use_coordinates = TRUE}, a record is placed by its point in each source's
+#' own geography (WCVP against WGSRPD level-3 areas, VASCAN and Flora do Brasil against
+#' GADM states), with no crosswalk between them - the finer question, where the data
+#' supports it.  Given coordinates and names together, the two are independent claims
+#' about where the record is and either can be wrong: a transposed longitude and a
+#' mistyped province are equally easy mistakes, so neither is authoritative.  Where they
+#' agree, both are used.  Where they disagree, \code{place_conflict} is \code{TRUE},
+#' \code{native_status} is \code{UNK}, and the two bases are answered separately in
+#' \code{native_status_coordinates} and \code{native_status_names} for you to judge.  The
+#' service has no combined mode to follow here: \code{\link{NSR}} takes names and
+#' \code{\link{NSR_from_coordinates}} takes coordinates.
+#'
+#' \strong{Precedence.}  If any source says native, the answer is native: asserting
+#' nativity is a positive claim, whereas "introduced" and "absent" are often artefacts of
+#' a list's scope, age or purpose.  Disagreement is recorded, not hidden, in
+#' \code{native_status_conflict} and \code{native_status_opinions}.
+#'
+#' \strong{Rank.}  A distribution recorded against an accepted infraspecific taxon counts
+#' for its species: a subspecies of \emph{X} native somewhere means \emph{X} is native
+#' there.  The reverse is not done.  A species with no opinion of its own is not given its
+#' genus's status, because a genus being native to a region says nothing about which of its
+#' species are - it is the introduced one as readily as the native one.  Such a species is
+#' \code{UNK}, or \code{A} where a comprehensive source covers the place;
+#' \code{taxon_evaluable} distinguishes them.  A genus \emph{query} is still answered
+#' directly, from the distributions WCVP records against genera.  \code{\link{NSR}} does
+#' consult higher ranks, so the two differ here by design.
+#'
+#' \strong{Which polygons answer.}  A place is judged by the polygons it lies IN.  An
+#' opinion about a polygon containing the place applies to it (POWO's finest statement
+#' about Guadeloupe is "native in the Leeward Islands"), and among those any native
+#' opinion wins.  Polygons INSIDE the place describe only parts of it, so they are not its
+#' status: they answer only when they all agree, and otherwise the answer is \code{P} with
+#' the reason saying the status varies and how many sub-polygons say what
+#' (\code{n_subpolygons_native}, \code{n_subpolygons_introduced}).  Give coordinates and
+#' the question does not arise: the record is judged on the ground it sits on.
+#' \code{native_status_scope} records which of these produced the answer.
+#'
+#' \strong{Endemism is the exception}, deliberately: \code{Ne} and \code{Ie} are claims
+#' about the taxon's whole range rather than about one polygon, so they draw on evidence
+#' from elsewhere.  A record of a taxon confined to California, found in Michigan, is
+#' introduced there however little Michigan's own checklists say.
+#' @param occurrence_dataframe A data.frame with \code{species}, and either coordinates
+#'   or political division names (see Place).
+#' @param dir Cache directory, shared with GNRS and GVS.
+#' @param resolve_names Resolve submitted names against WCVP with
+#'   \code{TNRS::TNRS_local()}?  Names already matching WCVP accepted names need none.
+#' @param use_coordinates Place each record by its \code{latitude} and \code{longitude}
+#'   as well as by its division names?  Default \code{FALSE}: coordinates cost a raster
+#'   lookup per call, and resolving them to political divisions is GVS's job.  See Place.
+#' @param min_overlap Ignore region links covering less than this share of a region.
+#' @param exclude_extinct Ignore distribution records the source marks extinct?  Default
+#'   \code{TRUE}, which answers about the present day.  Set \code{FALSE} to model a past
+#'   distribution, where a region the taxon has since been lost from still counts.  See
+#'   Extinct records.
+#' @param quiet Suppress progress messages?
+#' @return A data.frame, one row per input row, carrying \code{user_id} from the input
+#'   (or sequential ids where the input has none), as \code{\link{NSR}} does.
+#' @section Extinct records:
+#' WCVP marks a distribution extinct with a bare flag and no date, so the record says
+#' only "considered no longer present as of this release".  Whether it should count is
+#' therefore a property of the question rather than of the data: a present-day status
+#' wants it out, a historical distribution wants it in, and the cache keeps both so
+#' \code{exclude_extinct} can decide per call.
+#'
+#' Excluding it never makes the taxon introduced there.  The endemism rules read the
+#' taxon's whole native range, extinct records included, whatever this argument says -
+#' a region a taxon has been lost from is still a region it was native to, so \code{Ie}
+#' ("introduced, inferred from endemism elsewhere") cannot fire against it.  The answer
+#' for such a record is \code{A}: gone, not foreign.
+#' @section County and parish:
+#' \code{county_parish} is accepted and echoed, but not resolved: the political-division
+#' backbone stops at state and province, so \code{native_status_county_parish} is always
+#' \code{NA} and the answer is given at the finest level that could be matched.  A
+#' warning says so.  Coordinates are the way to ask a finer question.
+#' @export
+NSR_local <- function(occurrence_dataframe, dir = nsr_cache_dir(), resolve_names = TRUE,
+                      use_coordinates = FALSE, min_overlap = 0.01, exclude_extinct = TRUE,
+                      quiet = FALSE) {
+  if (!inherits(occurrence_dataframe, "data.frame")) {
+    stop("occurrence_dataframe should be a data.frame", call. = FALSE)
+  }
+  x <- occurrence_dataframe
+  col <- function(nm) if (nm %in% names(x)) as.character(x[[nm]]) else rep(NA_character_, nrow(x))
+  num <- function(nm) if (nm %in% names(x)) suppressWarnings(as.numeric(x[[nm]])) else rep(NA_real_, nrow(x))
+  species <- col("species")
+  country <- col("country")
+  state <- col("state_province")
+  county <- col("county_parish")
+  lon <- num("longitude")
+  lat <- num("latitude")
+  n <- nrow(x)
+
+  # user_id on NSR()'s terms, so a result can be joined back to its input either way
+  user_id <- if ("user_id" %in% names(x)) x[["user_id"]] else rep(NA, n)
+  if (all(is.na(user_id))) user_id <- seq_len(n)
+  if (any(duplicated(user_id))) {
+    stop("user_id should be either null or populated by unique values", call. = FALSE)
+  }
+
+  # No county polygons exist offline, so county_parish cannot be answered.  The service
+  # answers it, so say plainly that this one does not rather than returning a silent NA.
+  if (any(!is.na(county) & nzchar(county))) {
+    warning("county_parish is not resolved offline: the backbone carries no county ",
+            "polygons, so native_status_county_parish is NA and the answer is given at ",
+            "the finest level that could be matched. Give coordinates for a finer answer.",
+            call. = FALSE)
+  }
+
+  db <- nsr_local_db(dir)
+
+  # ---- taxa -------------------------------------------------------------------------
+  # WCVP records distributions for genera too, so a genus query is answerable directly.
+  # What must not happen is a bare genus being RESOLVED to some species of that genus,
+  # which would answer confidently about the wrong taxon.
+  above_species <- !is.na(species) & nzchar(species) &
+    lengths(strsplit(trimws(species), "[[:space:]]+")) < 2
+  taxon_id <- db$name_index$taxon_id[match(species, db$name_index$species_name)]
+  need <- unique(species[is.na(taxon_id) & !is.na(species) & nzchar(species) & !above_species])
+  if (resolve_names && length(need)) {
+    if (!quiet) message("Resolving ", format(length(need), big.mark = ","), " names ...")
+    r <- nsr_resolve_names(need, quiet = quiet, species_index = db$name_index)
+    taxon_id[is.na(taxon_id)] <- r$taxon_id[match(species[is.na(taxon_id)], r$name)]
+  }
+
+  # ---- place ------------------------------------------------------------------------
+  if (!use_coordinates && any(is.finite(lon) & is.finite(lat)) && !quiet) {
+    message("Coordinates present but not used: answering from division names. ",
+            "Set use_coordinates = TRUE to place each record by its point instead.")
+  }
+  places <- nsr_query_regions(lon, lat, country, state, county, dir, db, use_coordinates)
+
+  # ---- opinions ---------------------------------------------------------------------
+  use_set <- !is.null(db$chk_dt) && nrow(occurrence_dataframe) >= getOption("NSR.set_min", 200)
+  resolve <- function(tid, pl) if (use_set)
+    nsr_resolve_status_set(tid, pl, db, min_overlap, exclude_extinct) else
+    nsr_resolve_status(tid, pl, db, min_overlap, exclude_extinct)
+  res <- resolve(taxon_id, places)
+
+  # ---- records whose two statements of place disagree ---------------------------------
+  # Neither basis is taken as authoritative, so the pooled answer - which drew on
+  # evidence for both places and describes neither - is withdrawn, and each basis is
+  # answered on its own.  The two columns are populated only for these rows; elsewhere
+  # they are NA because there is nothing to compare, and native_status is the answer.
+  status_xy <- rep(NA_character_, n)
+  status_nm <- rep(NA_character_, n)
+  cf <- places$conflict
+  if (any(cf)) {
+    status_xy[cf] <- resolve(taxon_id[cf], nsr_split_keys(places$xy[cf]))$code
+    status_nm[cf] <- resolve(taxon_id[cf], nsr_split_keys(places$nm[cf]))$code
+    res$code[cf] <- "UNK"
+    res$reason[cf] <- paste0(
+      "Coordinates and division names resolve to different places, so no single answer ",
+      "is given; see native_status_coordinates and native_status_names")
+    res$scope[cf] <- "none"
+    res$sources[cf] <- NA_character_
+    res$opinions[cf] <- NA_character_
+    res$country_code[cf] <- NA_character_
+    res$state_code[cf] <- NA_character_
+    res$cultivated[cf] <- NA_integer_
+    res$conflict[cf] <- FALSE
+    res$conflict_type[cf] <- "none"
+    res$n_sub_native[cf] <- 0L
+    res$n_sub_introduced[cf] <- 0L
+    if (!quiet) message(sum(cf), " record(s) whose coordinates and division names ",
+                        "disagree; see place_conflict.")
+  }
+
+  out <- data.frame(
+    family = db$taxa$family[match(taxon_id, db$taxa$taxon_id)],
+    genus = db$taxa$genus[match(taxon_id, db$taxa$taxon_id)],
+    species = species,
+    country = country, state_province = state, county_parish = county,
+    latitude = lat, longitude = lon,
+    poldiv_full = places$label,
+    poldiv_type = places$level,
+    native_status_country = res$country_code,
+    native_status_state_province = res$state_code,
+    native_status_county_parish = rep(NA_character_, n),
+    native_status = res$code,
+    native_status_reason = res$reason,
+    native_status_sources = res$sources,
+    isIntroduced = as.integer(res$code %in% c("I", "Ie")),
+    isEndemic = as.integer(res$code == "Ne"),
+    isCultivatedNSR = res$cultivated,
+    # NA, not production's 0: the flag is dropped (a taxon-level use flag with no
+    # polygon attached says nothing about the record in hand) and a never-populated
+    # column should not read as a real negative. See dev_notes/01-offline-nsr-design.md,
+    # open question 3. isCultivatedNSR above is the version that has a geography.
+    is_cultivated_taxon = NA_integer_,
+    native_status_conflict = res$conflict,
+    native_status_conflict_type = res$conflict_type,
+    native_status_scope = res$scope,
+    n_subpolygons_native = res$n_sub_native,
+    n_subpolygons_introduced = res$n_sub_introduced,
+    native_status_opinions = res$opinions,
+    native_status_coordinates = status_xy,
+    native_status_names = status_nm,
+    place_conflict = places$conflict,
+    regions_matched = places$matched,
+    taxon_evaluable = !is.na(taxon_id) & taxon_id %in% db$evaluable,
+    user_id = user_id,
+    stringsAsFactors = FALSE
+  )
+  rownames(out) <- NULL
+  out
+}
+
+nsr_session <- new.env(parent = emptyenv())
+
+#' Load the local tables once per call
+#' @keywords internal
+#' @noRd
+nsr_local_db <- function(dir) {
+  nsr_need("nanoparquet")
+  stamp <- paste(normalizePath(dir), file.mtime(nsr_table_path("checklist", dir)),
+                 file.mtime(nsr_table_path("region-links", dir)))
+  hit <- nsr_session$db
+  if (!is.null(hit) && identical(hit$stamp, stamp)) return(hit$value)
+  need <- c("sources", "taxa", "regions", "checklist")
+  miss <- vapply(need, function(t) !file.exists(nsr_table_path(t, dir)), logical(1))
+  if (any(miss)) {
+    stop("The local NSR is not built in ", dir, ".\nRun NSR_local_build().", call. = FALSE)
+  }
+  db <- lapply(need, function(t) as.data.frame(nanoparquet::read_parquet(nsr_table_path(t, dir))))
+  names(db) <- need
+  lf <- nsr_table_path("region-links", dir)
+  db$links <- if (file.exists(lf)) as.data.frame(nanoparquet::read_parquet(lf)) else
+    data.frame(from_region = character(0), to_region = character(0), relation = character(0),
+               fraction = numeric(0), stringsAsFactors = FALSE)
+  db <- nsr_index_db(db)
+  nsr_session$db <- list(stamp = stamp, value = db)
+  db
+}
+
+#' Derive every lookup index the resolvers use from the four cache tables
+#'
+#' Internal.  Split out from \code{nsr_local_db()} so the resolvers can be exercised on
+#' tables held in memory, without a built cache: the row path and the set path must give
+#' the same answer for the same query, and that is only testable if a db can be made
+#' without a multi-gigabyte build.
+#' @keywords internal
+#' @noRd
+nsr_index_db <- function(db) {
+  # Caches built before extinct records were retained have no such column; they simply
+  # hold no extinct rows, so 0 is the truth for every row they do hold.
+  if (is.null(db$checklist$is_extinct)) db$checklist$is_extinct <- 0L
+  db$link_idx <- split(seq_len(nrow(db$links)), db$links$from_region)
+  db$chk_idx <- split(seq_len(nrow(db$checklist)),
+                      paste(db$checklist$taxon_id, db$checklist$region_key))
+  db$chk_env <- list2env(db$chk_idx, envir = new.env(hash = TRUE, parent = emptyenv()))
+  db$link_env <- list2env(db$link_idx, envir = new.env(hash = TRUE, parent = emptyenv()))
+  # GADM's key carries its own hierarchy: gid_1 "BRA.25_1" sits in gid_0 "BRA".  That
+  # containment is exact by construction, which is why it is kept OUT of the link table:
+  # link rows carry overlap fractions and are filtered by min_overlap, and a state is a
+  # legitimately tiny share of its country (Distrito Federal is 0.07% of Brazil), so as
+  # a fraction it would be discarded as a sliver.  Only regions that actually carry
+  # checklist rows are listed, since those are the only ones an answer can draw on.
+  g1 <- unique(db$regions$region_key[db$regions$system == "gadm1"])
+  g1 <- g1[g1 %in% db$checklist$region_key]
+  parent <- if (length(g1)) paste0("gadm0:", sub("\\..*$", "", sub("^gadm1:", "", g1))) else character(0)
+  db$gadm_parent <- stats::setNames(parent, g1)
+  db$gadm_children <- if (length(g1)) split(g1, parent) else list()
+  if (requireNamespace("data.table", quietly = TRUE)) {
+    db$gadm_edges_dt <- if (length(g1)) {
+      data.table::data.table(from_region = c(g1, parent), to_region = c(parent, g1),
+                             relation = rep(c("within", "contains"), each = length(g1)))
+    } else {
+      data.table::data.table(from_region = character(0), to_region = character(0),
+                             relation = character(0))
+    }
+    data.table::setkeyv(db$gadm_edges_dt, "from_region")
+    db$chk_dt <- data.table::as.data.table(db$checklist)
+    data.table::setkeyv(db$chk_dt, c("taxon_id", "region_key"))
+    db$links_dt <- data.table::as.data.table(db$links)
+    data.table::setkeyv(db$links_dt, "from_region")
+    nat <- unique(db$chk_dt[status == "native", list(taxon_id, region_key)])
+    db$native_dt <- nat
+    data.table::setkeyv(db$native_dt, "taxon_id")
+    db$confined_dt <- nsr_confined_ranges(nat, db)
+  }
+  # WCVP sometimes carries the same name at two ranks (a variety row also called
+  # "Pinus ponderosa"), and only one of them holds the distributions: index names to the
+  # species-rank id, and among those to the one with opinions
+  tx <- db$taxa
+  tx$is_species <- tolower(tx$rank) %in% c("species", "")
+  tx$has_rows <- tx$taxon_id %in% db$checklist$taxon_id
+  tx <- tx[order(!tx$has_rows, !tx$is_species), , drop = FALSE]
+  db$name_index <- tx[!duplicated(tx$species_name), c("species_name", "taxon_id")]
+  db$chk_status <- db$checklist$status
+  db$chk_source <- db$checklist$source_name
+  db$chk_cult <- db$checklist$is_cultivated
+  db$chk_extinct <- as.integer(db$checklist$is_extinct) %in% 1L
+  db$chk_region <- db$checklist$region_key
+  db$link_to <- db$links$to_region
+  db$link_rel <- db$links$relation
+  db$link_frac <- db$links$fraction
+  db$chk_idx_taxon <- split(seq_len(nrow(db$checklist)), db$checklist$taxon_id)
+  db$evaluable <- unique(db$checklist$taxon_id)
+  db$comprehensive <- db$sources$source_name[db$sources$is_comprehensive %in% TRUE]
+  db$covered <- unique(db$checklist$region_key[db$checklist$source_name %in% db$comprehensive])
+  db
+}
+
+#' The regions a query refers to, in every system
+#'
+#' Internal.  With coordinates, each system is looked up directly.  With names, the GADM
+#' unit comes from the GNRS backbone and other systems are reached through the link
+#' table.  Returns, per row, the regions to consult and how each relates to the query.
+#' @keywords internal
+#' @noRd
+nsr_query_regions <- function(lon, lat, country, state, county, dir, db,
+                              use_coordinates = FALSE) {
+  n <- length(lon)
+  direct <- vector("list", n)
+  fine <- vector("list", n)
+  ctry <- vector("list", n)
+  xy <- vector("list", n)
+  nm <- vector("list", n)
+  conflict <- rep(FALSE, n)
+  label <- rep(NA_character_, n)
+  level <- rep("country", n)
+  matched <- rep("none", n)
+
+  loc <- if (use_coordinates && any(is.finite(lon) & is.finite(lat)))
+    nsr_locate_regions(lon, lat, dir) else NULL
+  bb <- try(nsr_gnrs_backbone(dir), silent = TRUE)
+  pd <- if (!inherits(bb, "try-error")) nsr_match_poldiv(country, state, bb) else NULL
+  pd0 <- if (!inherits(bb, "try-error")) nsr_match_poldiv(country, NULL, bb) else NULL
+  # a declared state that belongs to another division system (a Norwegian county from
+  # after the 2018 or 2020 reform, a lan under the name records write) contributes the
+  # GADM units it covers, so the question is answered below country level instead of
+  # falling back to the country
+  iso0 <- if (!is.null(pd0)) bb$country$iso[match(pd0$country_id, bb$country$country_id)] else rep(NA_character_, n)
+  alt_keys <- nsr_altdiv_keys(iso0, state, dir)
+
+  at <- function(k, p) grep(p, k, value = TRUE)
+  for (i in seq_len(n)) {
+    kx <- if (!is.null(loc))
+      as.character(stats::na.omit(c(loc$wgsrpd3[i], loc$gadm[i], loc$gadm0[i]))) else character(0)
+    kn <- character(0)
+    if (!is.null(pd)) {
+      gid1 <- if (!is.na(pd$state_province_id[i]))
+        bb$state$gid_1[match(pd$state_province_id[i], bb$state$state_province_id)] else NA_character_
+      gid0 <- if (!is.na(pd0$country_id[i]))
+        bb$country$gid_0[match(pd0$country_id[i], bb$country$country_id)] else NA_character_
+      if (!is.na(gid1)) kn <- c(kn, paste0("gadm1:", gid1))
+      if (is.na(gid1) && length(alt_keys[[i]])) kn <- c(kn, alt_keys[[i]])
+      if (!is.na(gid0)) kn <- c(kn, paste0("gadm0:", gid0))
+    }
+    xy[[i]] <- kx
+    nm[[i]] <- kn
+    conflict[i] <- !nsr_places_agree(kx, kn)
+    keys <- unique(c(kx, kn))
+    keys <- keys[!is.na(keys)]
+    # the finest place the query actually names, kept apart from its country: a question
+    # about Amazonas must not inherit Brazil's answer
+    ctry[[i]] <- at(keys, "^gadm0:")
+    fine[[i]] <- grep("^gadm0:", keys, value = TRUE, invert = TRUE)
+    direct[[i]] <- keys
+    has_state <- any(grepl("^gadm1:", keys)) || (!is.null(pd) && !is.na(pd$state_province_id[i]))
+    level[i] <- if (has_state) "state_province" else "country"
+    matched[i] <- if (!length(keys)) "none" else if (conflict[i]) "conflict" else
+      if (!is.null(loc) && !is.na(loc$gadm[i])) "coordinates" else "names"
+    label[i] <- paste(stats::na.omit(c(country[i], if (!is.na(state[i]) && nzchar(state[i])) state[i])),
+                      collapse = ":")
+    if (!nzchar(label[i]) && length(keys)) label[i] <- paste(keys, collapse = " + ")
+  }
+  list(direct = direct, fine = fine, country = ctry, xy = xy, nm = nm,
+       conflict = conflict, label = label, level = level, matched = matched)
+}
+
+#' Do coordinates and division names describe the same place?
+#'
+#' Internal.  Coordinates and names are two independent claims about where a record is,
+#' and either can be wrong: a transposed longitude and a mistyped province are equally
+#' easy mistakes, so neither is authoritative.  They are compared in the geography they
+#' share, GADM, and at each level separately, so a right country with a wrong state is
+#' caught as readily as a wrong country.  A level only one of them speaks to is not a
+#' disagreement.  Where they agree, pooling the keys costs nothing and the wider set
+#' answers; where they do not, \code{\link{NSR_local}} answers each on its own.
+#' @keywords internal
+#' @noRd
+nsr_places_agree <- function(kx, kn) {
+  lvl <- function(p) {
+    a <- grep(p, kx, value = TRUE); b <- grep(p, kn, value = TRUE)
+    !length(a) || !length(b) || identical(sort(a), sort(b))
+  }
+  lvl("^gadm0:") && lvl("^gadm1:")
+}
+
+#' Split a set of region keys into the finest level named and its country
+#' @keywords internal
+#' @noRd
+nsr_split_keys <- function(keys) {
+  list(fine = lapply(keys, function(z) grep("^gadm0:", z, value = TRUE, invert = TRUE)),
+       country = lapply(keys, function(z) grep("^gadm0:", z, value = TRUE)))
+}
+
+#' Gather and reduce every opinion bearing on each row
+#' @keywords internal
+#' @noRd
+nsr_resolve_status <- function(taxon_id, places, db, min_overlap = 0.01,
+                               exclude_extinct = TRUE) {
+  n <- length(taxon_id)
+  blank <- rep(NA_character_, n)
+  out <- list(code = blank, reason = blank, sources = blank, opinions = blank,
+              country_code = blank, state_code = blank, scope = blank,
+              conflict = rep(FALSE, n), conflict_type = rep(NA_character_, n),
+              cultivated = rep(NA_integer_, n), n_sub_native = rep(NA_integer_, n),
+              n_sub_introduced = rep(NA_integer_, n))
+  for (i in seq_len(n)) {
+    tid <- taxon_id[i]
+    fine <- places$fine[[i]]
+    ctry <- places$country[[i]]
+    keys <- if (length(fine)) fine else ctry          # answer at the finest level named
+    if (is.na(tid) || !length(c(fine, ctry))) {
+      out$code[i] <- "UNK"
+      out$reason[i] <- if (is.na(tid)) "Taxon not matched to a species in the backbone" else
+        "Place not matched to any region"
+      # nothing was consulted, which is a definite "no disagreement, no sub-polygons"
+      # rather than an unknown; the set path says the same for these rows
+      out$scope[i] <- "none"
+      out$conflict_type[i] <- "none"
+      out$n_sub_native[i] <- 0L
+      out$n_sub_introduced[i] <- 0L
+      next
+    }
+    op <- nsr_opinions_for(tid, keys, db, min_overlap, exclude_extinct)
+    # a coarser polygon the place sits in (its country) also contains it, so opinions
+    # recorded ON that polygon apply; its OTHER sub-polygons do not, so only direct rows
+    # are taken, never its links
+    anc <- setdiff(ctry, keys)
+    if (length(anc)) {
+      rows <- unlist(mget(paste(tid, anc), db$chk_env, ifnotfound = list(NULL)), use.names = FALSE)
+      if (exclude_extinct && length(rows)) rows <- rows[!db$chk_extinct[rows]]
+      if (length(rows)) {
+        add <- list(status = db$chk_status[rows], source_name = db$chk_source[rows],
+                    is_cultivated = db$chk_cult[rows], relation = rep("within", length(rows)))
+        op <- if (is.null(op)) add else Map(c, op, add)
+      }
+    }
+    # the ancestor country supplies opinions above, so absence has to be read against it
+    # too, or a state with no coverage of its own inside a comprehensively listed country
+    # answers UNK here and A in the set path, which reads Q's country row
+    consulted <- nsr_consulted_regions(c(keys, anc), db, min_overlap)
+    ev <- tid %in% db$evaluable
+    r <- nsr_reduce(op, consulted, ev, db)
+    if (identical(r$code, "N") && nsr_is_endemic(tid, keys, db, min_overlap)) {
+      r$code <- "Ne"
+      r$reason <- paste0(sub("^Native", "Native and endemic", r$reason))
+    }
+    if (identical(r$code, "A")) {
+      ee <- nsr_endemic_elsewhere(tid, keys, db, min_overlap)
+      if (!is.na(ee)) {
+        r$code <- "Ie"
+        r$reason <- paste0("Absent from this region and endemic to ", ee,
+                           ", so introduced here (inferred)")
+      }
+    }
+    for (f in names(r)) out[[f]][i] <- r[[f]]
+    # per-level codes, for the service's columns
+    out$country_code[i] <- if (length(ctry)) {
+      nsr_reduce(nsr_opinions_for(tid, ctry, db, min_overlap, exclude_extinct),
+                 nsr_consulted_regions(ctry, db, min_overlap), ev, db)$code
+    } else NA_character_
+    out$state_code[i] <- if (length(fine)) {
+      nsr_reduce(nsr_opinions_for(tid, fine, db, min_overlap, exclude_extinct),
+                 nsr_consulted_regions(fine, db, min_overlap), ev, db)$code
+    } else NA_character_
+  }
+  out
+}
+
+#' The GADM units directly below, or directly above, the regions given
+#'
+#' Internal.  Exact containment read off the key, so no overlap fraction applies and
+#' \code{min_overlap} does not filter it.  See \code{nsr_index_db()}.
+#' @keywords internal
+#' @noRd
+nsr_gadm_children <- function(keys, db) {
+  if (!length(keys) || !length(db$gadm_children)) return(character(0))
+  k <- keys[startsWith(keys, "gadm0:")]
+  if (!length(k)) return(character(0))
+  setdiff(as.character(unlist(db$gadm_children[k], use.names = FALSE)), keys)
+}
+
+#' @keywords internal
+#' @noRd
+nsr_gadm_parents <- function(keys, db) {
+  if (!length(keys)) return(character(0))
+  k <- keys[startsWith(keys, "gadm1:")]
+  p <- if (length(k) && length(db$gadm_parent))
+    as.character(unname(db$gadm_parent[k])) else character(0)
+  setdiff(p[!is.na(p)], keys)
+}
+
+#' The countries a set of regions lies in, a country counting as its own
+#'
+#' Internal.  For the confined-range test only.  A taxon native both to a country row
+#' and to one of that country's states is confined to it, so gadm0 has to answer for
+#' itself here - it has no parent to look up, and its links only reach other geographies.
+#' @keywords internal
+#' @noRd
+nsr_gadm_countries <- function(keys, db) {
+  unique(c(nsr_gadm_parents(keys, db), keys[startsWith(keys, "gadm0:")]))
+}
+
+#' Opinions about one taxon bearing on a set of regions
+#'
+#' Internal.  Direct opinions, plus opinions about regions the query's regions are inside
+#' of or contain, with the relation recorded so the reducer can apply the one-way
+#' inheritance rules.
+#' @keywords internal
+#' @noRd
+nsr_opinions_for <- function(taxon_id, keys, db, min_overlap = 0.01, exclude_extinct = TRUE) {
+  if (!length(keys)) return(NULL)
+  live <- function(r) if (exclude_extinct && length(r)) r[!db$chk_extinct[r]] else r
+  rows <- live(unlist(mget(paste(taxon_id, keys), db$chk_env, ifnotfound = list(NULL)),
+                      use.names = FALSE))
+  rel <- if (length(rows)) rep("same", length(rows)) else character(0)
+  li <- unlist(mget(keys, db$link_env, ifnotfound = list(NULL)), use.names = FALSE)
+  if (length(li)) {
+    to <- db$link_to[li]; rl <- db$link_rel[li]; fr <- db$link_frac[li]
+    known <- unique(sub(":.*", "", keys))        # geographies this place is located in
+    good <- !(to %in% keys) & fr >= min_overlap & !(sub(":.*", "", to) %in% known)
+    if (any(good)) {
+      to <- to[good]; rl <- rl[good]
+      idx <- lapply(mget(paste(taxon_id, to), db$chk_env, ifnotfound = list(NULL)), live)
+      n <- lengths(idx)
+      if (sum(n)) {
+        rows <- c(rows, unlist(idx, use.names = FALSE))
+        rel <- c(rel, rep(rl, n))
+      }
+    }
+  }
+  # The states inside a queried country: a source keyed on GADM states (VASCAN, Flora do
+  # Brasil) is otherwise invisible to a country query, which breaks native-up propagation.
+  # A state query does not come back here for its country - that is the ancestor path in
+  # nsr_resolve_status(), which takes direct rows only, so a country's OTHER states never
+  # reach it.
+  kids <- nsr_gadm_children(keys, db)
+  if (length(kids)) {
+    idx <- lapply(mget(paste(taxon_id, kids), db$chk_env, ifnotfound = list(NULL)), live)
+    if (sum(lengths(idx))) {
+      rows <- c(rows, unlist(idx, use.names = FALSE))
+      rel <- c(rel, rep("contains", sum(lengths(idx))))
+    }
+  }
+  if (!length(rows)) return(NULL)
+  list(status = db$chk_status[rows], source_name = db$chk_source[rows],
+       is_cultivated = db$chk_cult[rows], relation = rel)
+}
+
+#' Is the taxon native only inside the queried place?
+#'
+#' Internal.  Endemism (\code{Ne}) is read off the checklist: every region any source
+#' calls it native must be the queried region, inside it, or overlapping it - never a
+#' region lying outside.  Only claimed when at least one source has a native opinion.
+#' @keywords internal
+#' @noRd
+nsr_is_endemic <- function(taxon_id, keys, db, min_overlap = 0.01) {
+  rows <- db$chk_idx_taxon[[as.character(taxon_id)]]
+  if (is.null(rows)) return(FALSE)
+  nat <- db$chk_region[rows][db$chk_status[rows] == "native"]
+  if (!length(nat)) return(FALSE)
+  li <- unlist(mget(keys, db$link_env, ifnotfound = list(NULL)), use.names = FALSE)
+  inside <- if (!length(li)) keys else
+    c(keys, db$link_to[li][db$link_rel[li] %in% c("same", "contains") & db$link_frac[li] >= min_overlap])
+  inside <- c(inside, nsr_gadm_children(keys, db))
+  all(nat %in% inside)
+}
+
+#' Is the taxon endemic to somewhere else?
+#'
+#' Internal.  The service's `Ie`: a taxon absent from the queried region but whose whole
+#' native range lies elsewhere and is confined - endemic to one region, or to one country -
+#' cannot be native here, so an occurrence must be introduced.  A widely native species
+#' merely unrecorded here stays `A`: absence is not evidence of introduction unless the
+#' species could not have been native in the first place.
+#' @return The name of the region it is endemic to, or NA.
+#' @keywords internal
+#' @noRd
+nsr_endemic_elsewhere <- function(taxon_id, keys, db, min_overlap = 0.01) {
+  rows <- db$chk_idx_taxon[[as.character(taxon_id)]]
+  if (is.null(rows)) return(NA_character_)
+  nat <- unique(db$chk_region[rows][db$chk_status[rows] == "native"])
+  if (!length(nat)) return(NA_character_)
+  if (any(nat %in% nsr_consulted_regions(keys, db, min_overlap))) return(NA_character_)
+  nm <- function(k) {
+    i <- match(k, db$regions$region_key)
+    if (is.na(i)) k else db$regions$region_name[i]
+  }
+  if (length(nat) == 1) return(nm(nat))
+  containers <- lapply(nat, function(r) {
+    li <- db$link_env[[r]]
+    if (is.null(li)) return(nsr_gadm_countries(r, db))
+    keep <- db$link_rel[li] %in% c("same", "within") & db$link_frac[li] >= min_overlap &
+      startsWith(db$link_to[li], "gadm0:")
+    c(db$link_to[li][keep], nsr_gadm_countries(r, db))
+  })
+  common <- Reduce(intersect, containers)
+  if (length(common)) nm(common[1]) else NA_character_
+}
+
+#' Every region an answer may draw on: the query's own, and those linked to them
+#'
+#' Internal.  Used to decide whether absence is interpretable: a query about California
+#' finds no source publishing against GADM California, but WGSRPD's CAL is the same place
+#' and is comprehensively listed, so absence there does mean something.
+#'
+#' \code{min_overlap} applies here for the same reason it applies to opinions: a link
+#' covering a sliver of a region is not that region being listed, and letting one through
+#' would turn "no opinion" into \code{A}, or suppress an \code{Ie}, on the strength of a
+#' shared coastline.  The set path filters its links once, when it builds them.
+#' @keywords internal
+#' @noRd
+nsr_consulted_regions <- function(keys, db, min_overlap = 0.01) {
+  if (!length(keys)) return(character(0))
+  li <- unlist(mget(keys, db$link_env, ifnotfound = list(NULL)), use.names = FALSE)
+  if (!length(li)) return(unique(c(keys, nsr_gadm_children(keys, db))))
+  keep <- db$link_rel[li] %in% c("same", "within", "contains") & db$link_frac[li] >= min_overlap
+  unique(c(keys, db$link_to[li][keep], nsr_gadm_children(keys, db)))
+}
+
+#' What kind of disagreement is this?
+#'
+#' Internal.  Two quite different things look like conflict.  Sources can genuinely
+#' disagree about a polygon (VASCAN calls a plant introduced in Ontario where POWO calls it
+#' native).  Or one source can hold both statuses for the same polygon because WCVP records
+#' one subspecies as native and another as introduced, and we roll infraspecific taxa up to
+#' the species - no one disagrees with anyone there.  The flag says which.
+#' @return "none", "sources", "within source" or "both".
+#' @keywords internal
+#' @noRd
+nsr_conflict_type <- function(status, source) {
+  if (!length(status)) return("none")
+  by_src <- split(status, source)
+  within <- any(vapply(by_src, function(z) length(unique(z)) > 1, logical(1)))
+  sets <- vapply(by_src, function(z) paste(sort(unique(z)), collapse = "+"), character(1))
+  between <- length(unique(sets)) > 1
+  if (within && between) "both" else if (between) "sources" else if (within) "within source" else "none"
+}
+
+#' Reduce opinions to one status code
+#'
+#' Internal.  A place is judged by the polygons it lies IN: an opinion about a polygon
+#' containing it applies to it, and among those, any native opinion wins (BM's
+#' precedence).  Polygons INSIDE the place describe only parts of it, so they are not
+#' its status: they answer only when they all agree, and otherwise the answer is that it
+#' depends where the record falls - report `P` and say so.  Polygons merely overlapping
+#' describe neither.  This keeps a record's verdict tied to the evidence for the ground
+#' it actually sits on, which is what coordinates give directly.
+#'
+#' Internal.  Inheritance: an opinion about a region the query sits INSIDE ("within"
+#' from the query's side) carries introduced downward but not native; an opinion about a
+#' region inside the query ("contains") carries native upward but not introduced;
+#' partial overlap carries neither.  Then: any native wins, else introduced, else
+#' present; else absent where a comprehensive source covers the region and the taxon is
+#' evaluable, else unknown.
+#' @keywords internal
+#' @noRd
+nsr_reduce <- function(op, keys, evaluable, db) {
+  if (is.null(op)) {
+    st <- src <- rel <- character(0); cult <- integer(0)
+  } else {
+    st <- op$status; src <- op$source_name; rel <- op$relation; cult <- op$is_cultivated
+  }
+  opinions <- if (!length(st)) NA_character_ else
+    paste(paste0(src, ":", st, ifelse(rel == "same", "", paste0("(", rel, ")"))), collapse = "; ")
+  cultivated <- if (!length(cult)) NA_integer_ else as.integer(any(cult %in% c(1, "1", TRUE)))
+  inc <- rel %in% c("same", "within")        # polygons the place lies in
+  sub <- rel == "contains"                   # polygons inside the place
+  ovl <- rel == "overlaps"
+  n_sub_n <- sum(sub & st == "native"); n_sub_i <- sum(sub & st == "introduced")
+  out <- function(code, reason, scope, srcs = character(0), conflict = FALSE,
+                  conflict_type = "none") {
+    list(code = code, reason = reason, scope = scope,
+         sources = if (!length(srcs)) NA_character_ else paste(sort(unique(srcs)), collapse = ", "),
+         opinions = opinions, conflict = conflict, conflict_type = conflict_type,
+         cultivated = cultivated, n_sub_native = n_sub_n, n_sub_introduced = n_sub_i)
+  }
+  if (any(inc)) {
+    si <- st[inc]
+    ct <- nsr_conflict_type(si, src[inc])
+    code <- if ("native" %in% si) "N" else if ("introduced" %in% si) "I" else "P"
+    want <- c(N = "native", I = "introduced", P = "present")[[code]]
+    stated <- any(rel == "same" & st == want)
+    word <- c(N = "Native", I = "Introduced", P = "Present")[[code]]
+    reason <- if (ct != "none" && code == "N") {
+      paste0("Native to this polygon as per checklist (", ct, " disagree; any native opinion is taken)")
+    } else if (stated) paste0(word, " in this polygon, as per checklist")
+    else paste0(word, " in a polygon containing this place, as per checklist")
+    return(out(code, reason, if (stated) "polygon" else "containing polygon", src[inc],
+               ct != "none", ct))
+  }
+  if (any(sub)) {
+    ss <- unique(st[sub])
+    if (length(ss) == 1) {
+      code <- c(native = "N", introduced = "I", present = "P")[[ss]]
+      word <- c(native = "Native", introduced = "Introduced", present = "Present")[[ss]]
+      return(out(code, paste0(word, " in every listed polygon within this place, as per checklist"),
+                 "sub-polygons agree", src[sub]))
+    }
+    return(out("P", paste0("Status varies among the polygons within this place (", n_sub_n,
+                           " native, ", n_sub_i, " introduced); give coordinates or a finer division"),
+               "sub-polygons differ", src[sub]))
+  }
+  if (any(ovl)) {
+    return(out("UNK", "Only polygons partly overlapping this place have a status; none covers it",
+               "overlapping only", src[ovl]))
+  }
+  if (!evaluable) return(out("UNK", "No source holds native status information for this taxon", "none"))
+  if (any(keys %in% db$covered)) return(out("A", "Absent from the comprehensive checklists for this polygon", "polygon"))
+  out("UNK", "No comprehensive checklist covers this polygon", "none")
+}
diff --git a/R/NSR_local_set.R b/R/NSR_local_set.R
new file mode 100644
index 0000000..582d3fc
--- /dev/null
+++ b/R/NSR_local_set.R
@@ -0,0 +1,387 @@
+# data.table is used inside the package without being attached; this tells it the
+# package knows what it is doing, so [.data.table keeps its own semantics here.
+.datatable.aware <- TRUE
+
+# Set-based resolution: every query answered in a handful of joins rather than one pass
+# per row. Same semantics as the row path in NSR_local.R - a place is judged by the
+# polygons containing it, sub-polygons answer only when unanimous, endemism is the
+# exception that may look outside - but it scales to the millions of taxon x polygon
+# pairs an occurrence pipeline asks about.
+
+#' Which taxa have a native range confined to one region or one country
+#'
+#' Internal.  Computed once per build, for the `Ie` rule: a taxon absent from the place
+#' queried but confined elsewhere cannot be native there.  A taxon native in a single
+#' region is confined to it; otherwise it is confined to a country if every region it is
+#' native in lies inside that one country.
+#' @keywords internal
+#' @noRd
+nsr_confined_ranges <- function(nat, db) {
+  dt <- data.table::data.table
+  cnt <- nat[, list(n_nat = .N), by = "taxon_id"]
+  single <- merge(nat, cnt[n_nat == 1L], by = "taxon_id")
+  nm <- function(k) {
+    i <- match(k, db$regions$region_key)
+    data.table::fifelse(is.na(i), k, db$regions$region_name[i])
+  }
+  out <- dt(taxon_id = single$taxon_id, confined_to = nm(single$region_key))
+  multi <- cnt[n_nat > 1L]
+  if (nrow(multi)) {
+    lk <- db$links_dt[relation %in% c("same", "within") &
+                        startsWith(to_region, "gadm0:"),
+                      list(region_key = from_region, country = to_region)]
+    if (!is.null(db$gadm_edges_dt) && nrow(db$gadm_edges_dt)) {
+      lk <- unique(data.table::rbindlist(list(
+        lk,
+        db$gadm_edges_dt[relation == "within",
+                         list(region_key = from_region, country = to_region)]),
+        use.names = TRUE))
+    }
+    # a country row is its own country: a taxon native both to gadm0:CAN and to one of
+    # Canada's states is confined to Canada, and without this it has no common container
+    own <- unique(nat$region_key[startsWith(nat$region_key, "gadm0:")])
+    if (length(own)) {
+      lk <- unique(data.table::rbindlist(
+        list(lk, dt(region_key = own, country = own)), use.names = TRUE))
+    }
+    m <- merge(nat[taxon_id %in% multi$taxon_id], lk, by = "region_key", allow.cartesian = TRUE)
+    per <- m[, list(n_in = data.table::uniqueN(region_key)), by = c("taxon_id", "country")]
+    per <- merge(per, multi, by = "taxon_id")
+    conf <- per[n_in == n_nat]
+    conf <- conf[!duplicated(conf$taxon_id)]
+    if (nrow(conf)) out <- data.table::rbindlist(list(out,
+      dt(taxon_id = conf$taxon_id, confined_to = nm(conf$country))))
+  }
+  data.table::setkeyv(out, "taxon_id")
+  out
+}
+
+#' Native status for taxa and polygons you have already resolved
+#'
+#' A pipeline that has located its records itself - in the WGSRPD raster and the GADM
+#' index, as an occurrence workflow does once for all of its coordinates - already knows
+#' the polygons each record sits in.  This takes those directly, skipping name resolution
+#' and point-in-polygon, and answers with the same rules as \code{\link{NSR_local}}.
+#'
+#' @param taxon_id WCVP accepted ids (character), one per query.
+#' @param region_keys The polygons each record sits in, as a list of character vectors
+#'   (\code{"wgsrpd3:BZL"}, \code{"gadm1:BRA.25_1"}), or a single character vector for
+#'   one polygon each.
+#' @param country_keys Optional country polygons (\code{"gadm0:BRA"}), same shape.
+#' @param dir Cache directory.
+#' @param min_overlap Ignore region links covering less than this share of a region.
+#' @param exclude_extinct Ignore distribution records the source marks extinct?  See
+#'   \code{\link{NSR_local}}.
+#' @return A data.frame with \code{native_status} and the status columns
+#'   \code{NSR_local()} returns, keyed on \code{taxon_id}.  It does not echo the input
+#'   columns \code{NSR_local()} carries through (\code{species}, the division names,
+#'   \code{user_id}), since this entry point is given ids and polygon keys rather than
+#'   a record.
+#' @export
+NSR_local_by_region <- function(taxon_id, region_keys, country_keys = NULL,
+                                dir = nsr_cache_dir(), min_overlap = 0.01,
+                                exclude_extinct = TRUE) {
+  if (!is.list(region_keys)) region_keys <- as.list(region_keys)
+  if (is.null(country_keys)) country_keys <- vector("list", length(taxon_id))
+  if (!is.list(country_keys)) country_keys <- as.list(country_keys)
+  stopifnot(length(taxon_id) == length(region_keys),
+            length(taxon_id) == length(country_keys))
+  db <- nsr_local_db(dir)
+  clean <- function(z) { z <- z[!is.na(z) & nzchar(z)]; if (!length(z)) character(0) else z }
+  places <- list(fine = lapply(region_keys, clean), country = lapply(country_keys, clean))
+  res <- if (!is.null(db$chk_dt))
+           nsr_resolve_status_set(as.character(taxon_id), places, db, min_overlap, exclude_extinct)
+         else nsr_resolve_status(as.character(taxon_id), places, db, min_overlap, exclude_extinct)
+  data.frame(taxon_id = as.character(taxon_id),
+             native_status_country = res$country_code,
+             native_status_state_province = res$state_code,
+             native_status = res$code, native_status_reason = res$reason,
+             native_status_sources = res$sources, native_status_opinions = res$opinions,
+             native_status_scope = res$scope,
+             native_status_conflict = res$conflict,
+             native_status_conflict_type = res$conflict_type,
+             isIntroduced = as.integer(res$code %in% c("I", "Ie")),
+             isEndemic = as.integer(res$code == "Ne"),
+             isCultivatedNSR = res$cultivated,
+             n_subpolygons_native = res$n_sub_native,
+             n_subpolygons_introduced = res$n_sub_introduced,
+             stringsAsFactors = FALSE)
+}
+
+#' Resolve every query at once, with joins
+#'
+#' Internal.  Needs \code{data.table}; \code{\link{NSR_local}} falls back to the row path
+#' without it.  Returns the same list of per-row vectors as \code{nsr_resolve_status()},
+#' \emph{including} the per-level codes: the reduction is run once for the answer and once
+#' more against each political level on its own, which is what the row path does per row.
+#' Endemism is deliberately not applied to the per-level codes - \code{Ne} and \code{Ie}
+#' are claims about the taxon's whole range rather than about one level - so that the two
+#' implementations agree column for column.
+#' @keywords internal
+#' @noRd
+nsr_resolve_status_set <- function(taxon_id, places, db, min_overlap = 0.01,
+                                   exclude_extinct = TRUE) {
+  n <- length(taxon_id)
+  out <- nsr_resolve_status_set1(taxon_id, places, db, min_overlap, endemism = TRUE,
+                                 exclude_extinct = exclude_extinct)
+  empty <- rep(list(character(0)), n)
+  # a level reports a code only when it was asked about AND there is a taxon to ask
+  # about: with no match in the backbone no lookup happened at any level, which is what
+  # the row path records
+  asked <- !is.na(taxon_id)
+  nc <- lengths(places$country)
+  nf <- lengths(places$fine)
+  if (any(nc > 0)) {
+    lv <- nsr_resolve_status_set1(taxon_id, list(fine = empty, country = places$country),
+                                  db, min_overlap, endemism = FALSE,
+                                  exclude_extinct = exclude_extinct)
+    out$country_code <- ifelse(asked & nc > 0, lv$code, NA_character_)
+  }
+  if (any(nf > 0)) {
+    lv <- nsr_resolve_status_set1(taxon_id, list(fine = places$fine, country = empty),
+                                  db, min_overlap, endemism = FALSE,
+                                  exclude_extinct = exclude_extinct)
+    out$state_code <- ifelse(asked & nf > 0, lv$code, NA_character_)
+  }
+  out
+}
+
+#' One pass of the set resolver, against the polygons it is given
+#' @keywords internal
+#' @noRd
+nsr_resolve_status_set1 <- function(taxon_id, places, db, min_overlap = 0.01,
+                                    endemism = TRUE, exclude_extinct = TRUE) {
+  dt <- function(...) data.table::data.table(...)
+  n <- length(taxon_id)
+  blank <- rep(NA_character_, n)
+  out <- list(code = blank, reason = blank, sources = blank, opinions = blank,
+              country_code = blank, state_code = blank, scope = blank,
+              conflict = rep(FALSE, n), conflict_type = rep(NA_character_, n),
+              cultivated = rep(NA_integer_, n), n_sub_native = rep(NA_integer_, n),
+              n_sub_introduced = rep(NA_integer_, n))
+
+  fine <- places$fine
+  ctry <- places$country
+  nf <- lengths(fine)
+  nc <- lengths(ctry)
+  has_fine <- nf > 0
+  # the polygons the query names: its own ("same"), and its country, which contains it
+  Q <- data.table::rbindlist(list(
+    dt(qid = rep(seq_len(n), nf), region_key = unlist(fine), relation = "same"),
+    dt(qid = rep(seq_len(n), nc), region_key = unlist(ctry),
+       relation = "same")))                       # fixed below for rows that have a finer key
+  has_place <- nf > 0 | nc > 0
+  if (!nrow(Q)) return(nsr_set_finish(out, taxon_id, db, n, endemism = endemism,
+                                      has_place = has_place))
+  Q[, relation := data.table::fifelse(has_fine[qid] & region_key %in% unlist(ctry), "within", relation)]
+  Q[, taxon_id := taxon_id[qid]]
+  Q <- Q[!is.na(taxon_id) & !is.na(region_key)]
+
+  # polygons related to the ones named: only the query's OWN polygons follow links, so a
+  # country's other sub-polygons never reach a state-level query
+  own <- Q[relation == "same"]
+  L <- db$links_dt[own, on = c(from_region = "region_key"), nomatch = 0L,
+                   allow.cartesian = TRUE]
+  if (nrow(L)) {
+    known <- own[, list(sys = unique(sub(":.*", "", region_key))), by = "qid"]
+    L[, to_sys := sub(":.*", "", to_region)]
+    L <- L[!paste(qid, to_sys) %in% paste(known$qid, known$sys)]
+    L <- L[fraction >= min_overlap & to_region != from_region,
+           .(qid, region_key = to_region, relation = relation, taxon_id)]
+    Q <- data.table::rbindlist(list(Q, L), use.names = TRUE)
+  }
+  # GADM's exact parent/child containment, joined on the query's OWN polygons only so a
+  # country's other states never reach a state-level query.  Deliberately after the
+  # min_overlap filter: these edges carry no fraction because the containment is exact.
+  if (!is.null(db$gadm_edges_dt) && nrow(db$gadm_edges_dt)) {
+    H <- db$gadm_edges_dt[relation == "contains"][own, on = c(from_region = "region_key"),
+                                                  nomatch = 0L, allow.cartesian = TRUE]
+    if (nrow(H)) {
+      Q <- data.table::rbindlist(
+        list(Q, H[to_region != from_region, .(qid, region_key = to_region, relation, taxon_id)]),
+        use.names = TRUE)
+    }
+  }
+  # keep the strongest relation per (query, polygon)
+  ord <- c(same = 1L, within = 2L, contains = 3L, overlaps = 4L)
+  Q[, r_ord := ord[relation]]
+  data.table::setorder(Q, qid, region_key, r_ord)
+  Q <- unique(Q, by = c("qid", "region_key"))
+
+  # the opinions themselves
+  O <- db$chk_dt[Q, on = c("taxon_id", "region_key"), nomatch = 0L, allow.cartesian = TRUE]
+  # after the join, so the checklist is never copied just to drop 0.14% of its rows
+  if (exclude_extinct && nrow(O) && "is_extinct" %in% names(O)) O <- O[is_extinct == 0L]
+  if (!nrow(O)) return(nsr_set_finish(out, taxon_id, db, n, Q, endemism = endemism,
+                                      has_place = has_place))
+
+  O[, `:=`(inc = relation %in% c("same", "within"), sub = relation == "contains",
+           ovl = relation == "overlaps")]
+
+  # --- verdict from the polygons the place lies in -------------------------------------
+  inc <- O[inc == TRUE]
+  agg <- inc[, .(
+    has_nat = any(status == "native"), has_int = any(status == "introduced"),
+    stated_nat = any(relation == "same" & status == "native"),
+    stated_int = any(relation == "same" & status == "introduced"),
+    stated_pre = any(relation == "same" & status == "present"),
+    n_src_sets = data.table::uniqueN(.SD[, paste(sort(unique(status)), collapse = "+"), by = source_name]$V1),
+    within_src = any(.SD[, data.table::uniqueN(status), by = source_name]$V1 > 1),
+    srcs = paste(sort(unique(source_name)), collapse = ", "),
+    cult = as.integer(any(is_cultivated %in% c(1, "1", TRUE)))
+  ), by = qid]
+
+  sub <- O[sub == TRUE]
+  subagg <- sub[, .(n_sub_nat = sum(status == "native"), n_sub_int = sum(status == "introduced"),
+                    sub_sets = data.table::uniqueN(status), sub_status = status[1],
+                    sub_srcs = paste(sort(unique(source_name)), collapse = ", ")), by = qid]
+  ovlagg <- O[ovl == TRUE, .(ovl_srcs = paste(sort(unique(source_name)), collapse = ", ")), by = qid]
+  allop <- O[, .(opinions = paste(paste0(source_name, ":", status,
+                                         data.table::fifelse(relation == "same", "",
+                                                             paste0("(", relation, ")"))),
+                                  collapse = "; "),
+                 cult_any = as.integer(any(is_cultivated %in% c(1, "1", TRUE)))), by = qid]
+
+  res <- dt(qid = seq_len(n))
+  res <- merge(res, agg, by = "qid", all.x = TRUE)
+  res <- merge(res, subagg, by = "qid", all.x = TRUE)
+  res <- merge(res, ovlagg, by = "qid", all.x = TRUE)
+  res <- merge(res, allop, by = "qid", all.x = TRUE)
+
+  res[, conflict_type := data.table::fcase(
+    is.na(has_nat), NA_character_,
+    within_src & n_src_sets > 1, "both",
+    n_src_sets > 1, "sources",
+    within_src == TRUE, "within source",
+    default = "none")]
+  res[, code := data.table::fcase(
+    !is.na(has_nat) & has_nat, "N",
+    !is.na(has_int) & has_int, "I",
+    !is.na(has_nat), "P",
+    !is.na(sub_sets) & sub_sets == 1L,
+      c(native = "N", introduced = "I", present = "P")[sub_status],
+    !is.na(sub_sets), "P",
+    !is.na(ovl_srcs), "UNK",
+    default = NA_character_)]
+  res[, scope := data.table::fcase(
+    !is.na(has_nat) & ((code == "N" & stated_nat) | (code == "I" & stated_int) |
+                         (code == "P" & stated_pre)), "polygon",
+    !is.na(has_nat), "containing polygon",
+    !is.na(sub_sets) & sub_sets == 1L, "sub-polygons agree",
+    !is.na(sub_sets), "sub-polygons differ",
+    !is.na(ovl_srcs), "overlapping only",
+    default = NA_character_)]
+  res[, reason := data.table::fcase(
+    code == "N" & conflict_type != "none",
+      paste0("Native to this polygon as per checklist (", conflict_type,
+             " disagree; any native opinion is taken)"),
+    code == "N" & scope == "polygon", "Native in this polygon, as per checklist",
+    code == "N" & scope == "containing polygon",
+      "Native in a polygon containing this place, as per checklist",
+    code == "N", "Native in every listed polygon within this place, as per checklist",
+    code == "I" & scope == "polygon", "Introduced in this polygon, as per checklist",
+    code == "I" & scope == "containing polygon",
+      "Introduced in a polygon containing this place, as per checklist",
+    code == "I", "Introduced in every listed polygon within this place, as per checklist",
+    code == "P" & scope == "sub-polygons differ",
+      paste0("Status varies among the polygons within this place (", n_sub_nat, " native, ",
+             n_sub_int, " introduced); give coordinates or a finer division"),
+    code == "P" & scope == "polygon", "Present in this polygon, as per checklist",
+    code == "P", "Present in a polygon containing this place, as per checklist",
+    code == "UNK", "Only polygons partly overlapping this place have a status; none covers it",
+    default = NA_character_)]
+  res[, srcs_any := data.table::fcase(!is.na(srcs), srcs, !is.na(sub_srcs), sub_srcs,
+                                      !is.na(ovl_srcs), ovl_srcs, default = NA_character_)]
+
+  for (f in c("code", "reason", "scope", "conflict_type")) out[[f]] <- res[[f]]
+  out$sources <- res$srcs_any
+  out$opinions <- res$opinions
+  out$cultivated <- res$cult_any
+  out$n_sub_native <- data.table::fifelse(is.na(res$n_sub_nat), 0L, as.integer(res$n_sub_nat))
+  out$n_sub_introduced <- data.table::fifelse(is.na(res$n_sub_int), 0L, as.integer(res$n_sub_int))
+  out$conflict <- !is.na(res$conflict_type) & res$conflict_type != "none"
+  # "none" is the contract value for a result nothing disagreed about; NA here would be
+  # read as unknown, and the row path says "none" for the same query
+  out$conflict_type[!is.na(out$code) & is.na(out$conflict_type)] <- "none"
+  nsr_set_finish(out, taxon_id, db, n, Q, endemism = endemism, has_place = has_place)
+}
+
+#' Fill in the answers that need no opinions: absence, unknowns, and endemism
+#' @keywords internal
+#' @noRd
+nsr_set_finish <- function(out, taxon_id, db, n, Q = NULL, endemism = TRUE,
+                           has_place = TRUE) {
+  dt <- function(...) data.table::data.table(...)
+  evaluable <- !is.na(taxon_id) & taxon_id %in% db$evaluable
+  # the regions each query consulted, for the coverage test
+  covered <- rep(FALSE, n)
+  if (!is.null(Q) && nrow(Q)) {
+    cv <- Q[relation %in% c("same", "within", "contains") & region_key %in% db$covered,
+            .(any = TRUE), by = qid]
+    covered[cv$qid] <- TRUE
+  }
+  no_answer <- is.na(out$code)
+  out$code[no_answer] <- data.table::fcase(
+    !evaluable[no_answer], "UNK",
+    covered[no_answer], "A",
+    default = "UNK")
+  out$reason[no_answer] <- data.table::fcase(
+    !evaluable[no_answer], "No source holds native status information for this taxon",
+    covered[no_answer], "Absent from the comprehensive checklists for this polygon",
+    default = "No comprehensive checklist covers this polygon")
+  out$scope[no_answer & out$code == "A"] <- "polygon"
+  out$scope[no_answer & out$code == "UNK"] <- "none"
+  out$conflict_type[no_answer] <- "none"
+  out$n_sub_native[is.na(out$n_sub_native)] <- 0L
+  out$n_sub_introduced[is.na(out$n_sub_introduced)] <- 0L
+  # absence carries no opinion, so it carries no cultivation flag either: NA, matching
+  # nsr_reduce().  A 0 here would read as a checked negative and would differ from the
+  # row path for the same query, on nothing but batch size.
+
+  # A taxon that was matched but has nowhere to look is not the same as a place no
+  # source covers, and the row path distinguishes them; the generic absence logic above
+  # would otherwise report this as uncovered.
+  noplace <- !is.na(taxon_id) & !rep_len(has_place, n)
+  if (any(noplace)) {
+    out$code[noplace] <- "UNK"
+    out$reason[noplace] <- "Place not matched to any region"
+    out$scope[noplace] <- "none"
+  }
+
+  # rows with no place at all, or no taxon
+  none <- is.na(taxon_id)
+  if (any(none)) {
+    out$code[none] <- "UNK"
+    out$reason[none] <- "Taxon not matched to a species in the backbone"
+    out$scope[none] <- "none"
+  }
+
+  # --- endemism, the one rule allowed to look outside the queried polygon --------------
+  if (endemism && !is.null(Q) && nrow(Q)) {
+    inside <- Q[relation %in% c("same", "contains"), .(qid, region_key)]
+    nat <- db$native_dt
+    tq <- dt(qid = seq_len(n), taxon_id = taxon_id)[!is.na(taxon_id)]
+    # how many of the taxon's native regions fall inside the queried place
+    m <- nat[tq, on = "taxon_id", nomatch = 0L, allow.cartesian = TRUE]
+    if (nrow(m)) {
+      m[, in_place := paste(qid, region_key) %in% paste(inside$qid, inside$region_key)]
+      per <- m[, .(n_nat = .N, n_in = sum(in_place)), by = qid]
+      endemic <- per[n_nat > 0 & n_in == n_nat, qid]
+      i <- endemic[out$code[endemic] == "N"]
+      out$code[i] <- "Ne"
+      out$reason[i] <- sub("^Native", "Native and endemic", out$reason[i])
+      # Ie: absent here, and the whole native range lies elsewhere and is confined
+      consulted <- Q[relation %in% c("same", "within", "contains"), .(qid, region_key)]
+      m[, consulted_here := paste(qid, region_key) %in% paste(consulted$qid, consulted$region_key)]
+      elsewhere <- m[, .(n_nat = .N, n_here = sum(consulted_here)), by = qid][n_nat > 0 & n_here == 0]
+      conf <- db$confined_dt
+      elsewhere <- merge(elsewhere, tq, by = "qid")
+      e <- merge(elsewhere, conf, by = "taxon_id", all.x = FALSE, sort = FALSE)
+      j <- e$qid[out$code[e$qid] == "A" & !is.na(e$confined_to)]
+      out$code[j] <- "Ie"
+      out$reason[j] <- paste0("Absent from this region and endemic to ",
+                              e$confined_to[match(j, e$qid)], ", so introduced here (inferred)")
+    }
+  }
+  out
+}
diff --git a/R/local_altdiv.R b/R/local_altdiv.R
new file mode 100644
index 0000000..530f474
--- /dev/null
+++ b/R/local_altdiv.R
@@ -0,0 +1,69 @@
+# ===========================================================================
+# Divisions that belong to another division system
+#
+# A large part of the record stream declares a state that is not a GADM division:
+# Swedish landskap and lappmarker, Watsonian vice-counties, Norwegian counties from
+# after the 2018 and 2020 reforms. GNRS records these as units in their own right
+# and, where it knows which GADM units they cover, writes that extent to the shared
+# cache (altdiv-*). NSR reads those tables the same way it reads the GNRS reference:
+# straight from the cache, without depending on the package.
+#
+# Only a unit whose extent is known is of any use here: it turns a declared name
+# that matched nothing into the set of divisions the question is really about, so
+# native status can be answered below country level. A unit with no extent yet adds
+# no regions, which leaves the query where it was - answered at country level.
+# ===========================================================================
+
+#' The alternative divisions whose extent is known, keyed by country and name
+#'
+#' Internal.  NULL when the component has not been built.  Cached per directory
+#' for the session, like the other reference tables.
+#' @keywords internal
+#' @noRd
+nsr_altdiv <- function(dir) {
+  key <- paste0("altdiv:", dir)
+  if (!is.null(nsr_session[[key]])) return(nsr_session[[key]])
+  f <- function(x) file.path(dir, paste0("altdiv-", x, ".gz.parquet"))
+  if (!all(file.exists(f(c("units", "names", "extent"))))) return(NULL)
+  units <- as.data.frame(nanoparquet::read_parquet(f("units")))
+  names_ <- as.data.frame(nanoparquet::read_parquet(f("names")))
+  extent <- as.data.frame(nanoparquet::read_parquet(f("extent")))
+  # only exact names, and only units that have an extent: a regex system (the
+  # vice-counties) has none yet, and a unit without one adds no regions
+  names_ <- names_[names_$match == "exact" & names_$entity_key %in% extent$entity_key, , drop = FALSE]
+  iso <- units$country_iso[match(names_$entity_key, units$entity_key)]
+  out <- list(
+    key = paste(toupper(iso), tolower(trimws(names_$name)), sep = "\u001f"),
+    entity = names_$entity_key,
+    extent = split(extent$gid, extent$entity_key),
+    level = vapply(split(extent$level, extent$entity_key), function(x) as.integer(x[1]), integer(1))
+  )
+  nsr_session[[key]] <- out
+  out
+}
+
+#' Region keys for a declared division that belongs to another system
+#'
+#' Internal.  Vectorised over rows; returns a list of character vectors, empty
+#' where the declared name is not such a division or its extent is unknown.
+#' @keywords internal
+#' @noRd
+nsr_altdiv_keys <- function(country_iso, name, dir) {
+  n <- length(name)
+  out <- vector("list", n)
+  for (i in seq_len(n)) out[[i]] <- character(0)
+  a <- nsr_altdiv(dir)
+  if (is.null(a)) return(out)
+  nm <- tolower(trimws(ifelse(is.na(name), "", name)))
+  cc <- toupper(ifelse(is.na(country_iso), "", country_iso))
+  hit <- match(paste(cc, nm, sep = "\u001f"), a$key)
+  for (i in which(!is.na(hit))) {
+    k <- a$entity[hit[i]]
+    gid <- a$extent[[k]]
+    # NSR's regions are countries and states; a county-level extent has nothing to map
+    # onto yet, so it is skipped rather than answered at the wrong level
+    if (!length(gid) || a$level[[k]] != 1L) next
+    out[[i]] <- paste0("gadm1:", gid)
+  }
+  out
+}
diff --git a/R/local_build.R b/R/local_build.R
new file mode 100644
index 0000000..0de0ed0
--- /dev/null
+++ b/R/local_build.R
@@ -0,0 +1,292 @@
+#' Build the local native-status reference
+#'
+#' Reads each checklist source, resolves its names against WCVP with
+#' \code{TNRS::TNRS_local()} and its political divisions against the GNRS backbone, and
+#' writes the shared cache tables.  Sources are acquired on the user's machine; nothing
+#' derived ships with the package.
+#'
+#' \strong{The archives are not downloaded for you.}  \code{vascan} and \code{flbr} must
+#' be supplied through \code{files}, and each stops with the URL to fetch if they are
+#' not; \code{wcvp} is the exception, since it reads the WCVP archive already in the
+#' TNRS cache when there is one.  This is why \code{sources} defaults to \code{"wcvp"}
+#' alone: it is the only source that can build unattended.
+#'
+#' \code{"powo"} is accepted as a synonym for \code{"wcvp"} - it is the name the live
+#' service reports for the same Kew dataset - but \code{wcvp} is what gets written to
+#' \code{native_status_sources}.
+#'
+#' Tables written: \code{nsr-sources} (one per source, with licence and whether it is
+#' comprehensive), \code{nsr-taxa} (one per taxon, keyed on the WCVP accepted id),
+#' \code{nsr-regions} (one per region, in the geography its source publishes against) and
+#' \code{nsr-checklist} (one per taxon x region x source).  \code{nsr-region-links}, the
+#' spatial relations between geographies, is built too when the GADM index from GVS is in
+#' the cache.
+#'
+#' @param sources Which sources to build.  See \code{NSR_local_status()}.
+#' @param dir Cache directory, shared with GNRS and GVS.
+#' @param files Named list of local archives to read, e.g.
+#'   \code{list(flbr = "flbr_dwca.zip")}.  Required for \code{vascan} and \code{flbr}.
+#'   For \code{wcvp}, the WCVP zip; if absent, the copy in the TNRS cache is used when
+#'   there is one.
+#' @param overwrite Rebuild sources that are already built?
+#' @param quiet Suppress progress messages?
+#' @return Invisibly, \code{NSR_local_status()}.
+#' @export
+NSR_local_build <- function(sources = "wcvp",
+                            dir = nsr_cache_dir(create = TRUE),
+                            files = list(), overwrite = FALSE, quiet = FALSE) {
+  for (pkg in c("nanoparquet", "TNRS")) {
+    if (!requireNamespace(pkg, quietly = TRUE)) {
+      stop("Building the local NSR needs the '", pkg, "' package.", call. = FALSE)
+    }
+  }
+  reg <- nsr_builtin_registry()
+  sources <- match.arg(nsr_canonical_source(sources), names(reg), several.ok = TRUE)
+  if (length(files)) names(files) <- nsr_canonical_source(names(files))
+  msg <- function(...) if (!quiet) message(...)
+
+  existing <- lapply(nsr_tables(), function(t) {
+    p <- nsr_table_path(t, dir)
+    if (file.exists(p)) as.data.frame(nanoparquet::read_parquet(p)) else NULL
+  })
+  names(existing) <- nsr_tables()
+  done <- if (is.null(existing$sources)) character(0) else existing$sources$source_name
+  todo <- setdiff(sources, if (overwrite) character(0) else done)
+  if (!length(todo)) {
+    msg("Nothing to build (", paste(sources, collapse = ", "), " already built).")
+    return(invisible(NSR_local_status(dir)))
+  }
+
+  bb <- nsr_gnrs_backbone(dir)
+  raw <- list()
+  for (s in todo) {
+    msg("Importing ", s, " ...")
+    t0 <- Sys.time()
+    raw[[s]] <- switch(s,
+      wcvp = nsr_import_wcvp(files$wcvp, bb, quiet),
+      vascan = nsr_import_vascan(files$vascan, bb, quiet),
+      flbr = nsr_import_flbr(files$flbr, bb, quiet)
+    )
+    msg("  ", format(nrow(raw[[s]]), big.mark = ","), " records (",
+        round(as.numeric(difftime(Sys.time(), t0, units = "mins")), 1), " min)")
+  }
+
+  # ---- names --------------------------------------------------------------------
+  # sources keyed on WCVP already carry the taxon id (POWO is the backbone); the rest
+  # are resolved once, together, against the same backbone
+  need <- unique(unlist(lapply(raw, function(x) x$taxon_name[is.na(x$taxon_id)])))
+  need <- need[!is.na(need) & nzchar(need)]
+  res <- if (length(need)) {
+    msg("Resolving ", format(length(need), big.mark = ","), " distinct names against WCVP ...")
+    # POWO's species table, when it is being built in this run, keys the others
+    sp_index <- do.call(rbind, lapply(raw, function(d) {
+      if (!all(c("taxon_id", "species_name") %in% names(d))) return(NULL)
+      x <- unique(d[!is.na(d$taxon_id), c("taxon_id", "species_name")])
+      x[!duplicated(x$species_name), , drop = FALSE]
+    }))
+    if (is.null(sp_index) && !is.null(existing$taxa)) {
+      sp_index <- existing$taxa[, c("taxon_id", "species_name")]
+    }
+    r <- nsr_resolve_names(need, quiet = quiet, species_index = sp_index)
+    msg("  matched ", format(sum(!is.na(r$taxon_id)), big.mark = ","), " (",
+        round(100 * mean(!is.na(r$taxon_id)), 1), "%)")
+    r
+  } else {
+    data.frame(name = character(0), taxon_id = character(0), species_name = character(0),
+               family = character(0), genus = character(0), rank = character(0),
+               stringsAsFactors = FALSE)
+  }
+
+  # ---- assemble ---------------------------------------------------------------------
+  keep_old <- function(x, col) if (is.null(x)) NULL else x[!x[[col]] %in% todo, , drop = FALSE]
+  # A cache built before extinct records were retained has no is_extinct column, and the
+  # rows kept from it are rbind()ed with new rows that do; give it the column first, or
+  # adding a source to an older cache fails on the bind.
+  if (!is.null(existing$checklist) && is.null(existing$checklist$is_extinct)) {
+    existing$checklist$is_extinct <- 0L
+  }
+  chk <- lapply(names(raw), function(s) {
+    d <- raw[[s]]
+    fill <- is.na(d$taxon_id)
+    d$taxon_id[fill] <- res$taxon_id[match(d$taxon_name[fill], res$name)]
+    d <- d[!is.na(d$taxon_id) & !is.na(d$region_key), , drop = FALSE]
+    data.frame(taxon_id = d$taxon_id, region_key = d$region_key, status = d$status,
+               is_cultivated = d$is_cultivated,
+               is_extinct = if (is.null(d$is_extinct)) 0L else as.integer(d$is_extinct),
+               source_name = s, stringsAsFactors = FALSE)
+  })
+  checklist <- unique(do.call(rbind, c(list(keep_old(existing$checklist, "source_name")), chk)))
+
+  taxa_res <- res[!is.na(res$taxon_id), c("taxon_id", "family", "genus", "species_name", "rank")]
+  taxa_src <- do.call(rbind, lapply(raw, function(d) {
+    if (!all(c("family", "genus", "species_name", "rank") %in% names(d))) return(NULL)
+    x <- d[!is.na(d$taxon_id), c("taxon_id", "family", "genus", "species_name", "rank")]
+    x[!duplicated(x$taxon_id), , drop = FALSE]
+  }))
+  taxa_new <- rbind(taxa_res, taxa_src)
+  taxa <- unique(rbind(existing$taxa, taxa_new))
+  taxa <- taxa[!duplicated(taxa$taxon_id), , drop = FALSE]
+
+  reg_new <- unique(do.call(rbind, lapply(raw, function(d) {
+    x <- d[!is.na(d$region_key), c("region_key", "region_name")]
+    x[!duplicated(x$region_key), , drop = FALSE]
+  })))
+  reg_new$system <- sub(":.*", "", reg_new$region_key)
+  reg_new$code <- sub("^[^:]*:", "", reg_new$region_key)
+  regions <- unique(rbind(existing$regions, reg_new))
+  regions <- regions[!duplicated(regions$region_key), , drop = FALSE]
+
+  src_new <- do.call(rbind, lapply(names(raw), function(s) {
+    r <- reg[[s]]
+    n <- sum(checklist$source_name == s)
+    data.frame(source_name = s, source_name_full = r$source_name_full, version = r$version,
+               url = r$url, licence = r$licence, citation = r$citation,
+               is_comprehensive = r$is_comprehensive, region_system = r$region_system,
+               coverage = r$coverage, date_accessed = as.character(Sys.Date()),
+               n_records = n, n_taxa = length(unique(checklist$taxon_id[checklist$source_name == s])),
+               stringsAsFactors = FALSE)
+  }))
+  srcs <- rbind(keep_old(existing$sources, "source_name"), src_new)
+
+  for (nm in nsr_tables()) {
+    x <- switch(nm, sources = srcs, taxa = taxa, regions = regions, checklist = checklist)
+    nanoparquet::write_parquet(x, nsr_table_path(nm, dir), compression = "gzip")
+  }
+  saveRDS(list(built = as.character(Sys.Date()), sources = srcs$source_name,
+               n_records = nrow(checklist), n_taxa = nrow(taxa), n_regions = nrow(regions)),
+          nsr_provenance_path("nsr", dir))
+
+  # ---- geography ---------------------------------------------------------------------
+  # The WGSRPD raster is what places coordinates in a WCVP region, so it is needed
+  # whenever a source publishes against WGSRPD - not only when there is a second
+  # geography to link it to.  A wcvp-only build needs it just as much.
+  systems <- unique(regions$system)
+  raster_ok <- TRUE
+  if ("wgsrpd3" %in% systems && !file.exists(nsr_raster_path("wgsrpd3", dir))) {
+    msg("Building the WGSRPD level-3 raster ...")
+    r <- try(nsr_build_wgsrpd_raster(dir = dir, quiet = quiet), silent = TRUE)
+    raster_ok <- !inherits(r, "try-error")
+    if (!raster_ok) {
+      warning("The WGSRPD level-3 raster could not be built: ",
+              conditionMessage(attr(r, "condition")), "\n",
+              "The checklist tables are written, but until this raster exists no query ",
+              "can reach a WCVP opinion: coordinates cannot be placed in a WCVP region, ",
+              "and division names resolve to GADM, which needs the link table below. ",
+              "Install 'rWCVPdata' (or 'rWCVP'), then re-run NSR_local_build(overwrite = TRUE).",
+              call. = FALSE)
+    }
+  }
+
+  # Names always resolve to GADM through the GNRS backbone, whatever geography the
+  # sources use, so a WGSRPD source is unreachable by name until the two are linked.
+  # That holds for a wcvp-only build as much as a mixed one - hence no test on the
+  # number of systems, which is what used to leave the default build unanswerable.
+  needs_links <- "wgsrpd3" %in% systems || length(systems) > 1
+  gadm_index <- file.path(dir, "gadmindex-units-30s.tif")
+  if (needs_links && raster_ok && file.exists(gadm_index)) {
+    msg("Linking region systems (", paste(c(systems, "gadm"), collapse = ", "), ") ...")
+    r <- try(nsr_build_region_links(dir = dir, quiet = quiet), silent = TRUE)
+    if (inherits(r, "try-error")) {
+      warning("The region link table could not be built: ",
+              conditionMessage(attr(r, "condition")), "\n",
+              "The cache is written, but queries by division name will not reach sources ",
+              "published against another geography. Re-run NSR_local_build(overwrite = TRUE) ",
+              "once the cause is fixed.", call. = FALSE)
+    }
+  } else if (needs_links && raster_ok) {
+    warning("GVS's GADM index is not in ", dir, ", so the geographies cannot be linked. ",
+            "Queries by division name resolve to GADM and will not reach sources published ",
+            "against WGSRPD; coordinates still work. Build it with GVS, then re-run ",
+            "NSR_local_build(overwrite = TRUE).", call. = FALSE)
+  }
+  msg("Built: ", format(nrow(checklist), big.mark = ","), " checklist records, ",
+      format(nrow(taxa), big.mark = ","), " taxa, ", format(nrow(regions), big.mark = ","),
+      " regions.")
+  invisible(NSR_local_status(dir))
+}
+
+#' The GNRS political-division backbone, as NSR keys on it
+#' @keywords internal
+#' @noRd
+nsr_gnrs_backbone <- function(dir) {
+  f <- function(x) file.path(dir, paste0("gnrs-", x, ".gz.parquet"))
+  if (!file.exists(f("country")) || !file.exists(f("state_province"))) {
+    stop("The GNRS backbone is not in ", dir, ".\n",
+         "Build it first with GNRS::GNRS_local_build(); NSR keys its divisions on GNRS ids.",
+         call. = FALSE)
+  }
+  list(country = as.data.frame(nanoparquet::read_parquet(f("country"))),
+       state = as.data.frame(nanoparquet::read_parquet(f("state_province"))))
+}
+
+#' Political division ids for a set of country (and optionally state) names
+#'
+#' Internal.  Exact, case- and accent-insensitive matching against the GNRS backbone;
+#' anything unmatched comes back NA and is reported by the caller, rather than being
+#' guessed at.
+#' @keywords internal
+#' @noRd
+nsr_match_poldiv <- function(country, state = NULL, bb) {
+  norm <- function(x) {
+    x <- iconv(as.character(x), to = "ASCII//TRANSLIT")
+    tolower(trimws(gsub("[^A-Za-z0-9 ]", "", ifelse(is.na(x), "", x))))
+  }
+  ci <- match(norm(country), norm(bb$country$country))
+  out <- data.frame(
+    country_id = bb$country$country_id[ci], country = bb$country$country[ci],
+    state_province_id = NA_integer_, state_province = NA_character_,
+    stringsAsFactors = FALSE
+  )
+  if (!is.null(state)) {
+    key <- paste(norm(country), norm(state))
+    # the backbone spells some divisions bilingually ("New Brunswick/Nouveau-Brunswick"),
+    # so every variant is a valid target
+    variants <- function(col) {
+      parts <- strsplit(ifelse(is.na(col), "", col), "/", fixed = TRUE)
+      data.frame(row = rep(seq_along(parts), lengths(parts)),
+                 name = norm(unlist(parts)), stringsAsFactors = FALSE)
+    }
+    v <- unique(rbind(
+      data.frame(row = seq_len(nrow(bb$state)), name = norm(bb$state$state_province)),
+      data.frame(row = seq_len(nrow(bb$state)), name = norm(bb$state$state_province_std)),
+      variants(bb$state$state_province), variants(bb$state$state_province_std)))
+    v$key <- paste(norm(bb$state$country)[v$row], v$name)
+    v <- v[!duplicated(v$key), , drop = FALSE]
+    si <- v$row[match(key, v$key)]
+    out$state_province_id <- bb$state$state_province_id[si]
+    out$state_province <- bb$state$state_province[si]
+  }
+  out$poldiv_level <- ifelse(!is.na(out$state_province_id), "state_province",
+                             ifelse(!is.na(out$country_id), "country", NA_character_))
+  out$poldiv_id <- ifelse(!is.na(out$state_province_id), paste0("S", out$state_province_id),
+                          ifelse(!is.na(out$country_id), paste0("C", out$country_id), NA_character_))
+  out
+}
+
+#' Resolve checklist names to WCVP accepted species
+#'
+#' Internal.  The id returned is the SPECIES' id, not an infraspecific one, so every
+#' source keys on the same unit as POWO after its infraspecifics are rolled up.
+#' @keywords internal
+#' @noRd
+nsr_resolve_names <- function(names_in, quiet = FALSE, species_index = NULL) {
+  r <- TNRS::TNRS_local(taxonomic_names = names_in, sources = "wcvp", matches = "best",
+                        build_missing = FALSE, quiet = quiet)
+  acc <- r$Accepted_species
+  acc[is.na(acc) | !nzchar(acc)] <- NA_character_
+  # prefer the species' own id where we know it, so subspecies land on their species
+  id <- ifelse(is.na(acc), NA_character_, r$Accepted_name_id)
+  if (!is.null(species_index)) {
+    sid <- species_index$taxon_id[match(acc, species_index$species_name)]
+    id <- ifelse(!is.na(sid), sid, id)
+  }
+  data.frame(
+    name = r$Name_submitted,
+    taxon_id = id,
+    species_name = acc,
+    family = r$Accepted_family,
+    genus = sub(" .*", "", acc),
+    rank = r$Accepted_name_rank,
+    stringsAsFactors = FALSE
+  )
+}
diff --git a/R/local_cache.R b/R/local_cache.R
new file mode 100644
index 0000000..b0762a4
--- /dev/null
+++ b/R/local_cache.R
@@ -0,0 +1,176 @@
+#' Cache directory for the offline NSR
+#'
+#' Internal.  Shared with GNRS and GVS, so the three services build one cache: the
+#' political divisions NSR keys on are GNRS's, and a user who has built either of the
+#' others already has them.  \code{options(NSR.cache_dir=)} overrides, then
+#' \code{options(GNRS.cache_dir=)}, then \code{tools::R_user_dir("GNRS", "cache")}.
+#'
+#' \code{tools::R_user_dir()} arrived in R 4.0.0, and the package supports R 3.5 for the
+#' API functions.  Rather than invent a different default on old R - which would put the
+#' cache somewhere GNRS and GVS do not look, and silently give each package its own copy -
+#' this says what is missing and how to set it.  The API functions never reach here.
+#' @keywords internal
+#' @noRd
+nsr_cache_dir <- function(create = FALSE) {
+  fallback <- function() {
+    if (!is.null(getNamespace("tools")[["R_user_dir"]])) {
+      return(tools::R_user_dir("GNRS", which = "cache"))
+    }
+    stop("The local NSR needs R >= 4.0.0 for the shared cache location, or an explicit ",
+         "one: set options(GNRS.cache_dir = ) to the directory GNRS and GVS use, or pass ",
+         "dir = to this function.", call. = FALSE)
+  }
+  dir <- getOption("NSR.cache_dir", getOption("GNRS.cache_dir", fallback()))
+  if (create && !dir.exists(dir)) dir.create(dir, recursive = TRUE, showWarnings = FALSE)
+  dir
+}
+
+#' Paths of the local NSR tables
+#' @keywords internal
+#' @noRd
+nsr_table_path <- function(table, dir = nsr_cache_dir()) {
+  file.path(dir, paste0("nsr-", table, ".gz.parquet"))
+}
+
+#' @keywords internal
+#' @noRd
+nsr_provenance_path <- function(component, dir = nsr_cache_dir()) {
+  file.path(dir, paste0("nsr-", component, "-provenance.rds"))
+}
+
+#' The tables the local NSR is made of
+#' @keywords internal
+#' @noRd
+nsr_tables <- function() c("sources", "taxa", "regions", "checklist")
+
+#' Resolve a source's accepted synonyms to its canonical name
+#'
+#' Internal.  The Kew dataset is read as WCVP and reported as \code{wcvp}, which is what
+#' it is and what \code{TNRS_local()} calls the same data.  \code{powo}, the portal the
+#' live service names it after, is accepted so existing calls and notes keep working.
+#' @keywords internal
+#' @noRd
+nsr_canonical_source <- function(x) {
+  if (!length(x)) return(x)
+  alias <- c(powo = "wcvp")
+  i <- match(x, names(alias))
+  unname(ifelse(is.na(i), x, alias[i]))
+}
+
+#' Require the optional packages a local-NSR path needs
+#'
+#' Internal.  The offline implementation is all Suggests: the API functions must keep
+#' working on an install that has none of it, so every entry point says what is missing
+#' rather than failing inside a \code{::} call.
+#' @keywords internal
+#' @noRd
+nsr_need <- function(..., what = "The local NSR") {
+  for (pkg in c(...)) {
+    if (!requireNamespace(pkg, quietly = TRUE)) {
+      stop(what, " needs the '", pkg, "' package.", call. = FALSE)
+    }
+  }
+  invisible(TRUE)
+}
+
+#' Checklist sources the local NSR can build
+#'
+#' Internal.  One entry per source: what it covers, whether it is comprehensive for its
+#' area (so that absence there is informative), where it comes from and under what
+#' licence.  Sources are fetched on the user's machine at build time; nothing derived is
+#' shipped with the package.
+#'
+#' \code{region_system} is the geography the source publishes against, kept as it is.
+#' \code{is_comprehensive} says the source aims to list everything in its area; it does
+#' not say which taxa it covers, which is measured from the data instead.
+#' @keywords internal
+#' @noRd
+nsr_builtin_registry <- function() {
+  list(
+    wcvp = list(
+      source_name = "wcvp",
+      source_name_full = "Plants of the World Online / World Checklist of Vascular Plants (WCVP)",
+      version = "v15",
+      url = "https://sftp.kew.org/pub/data-repositories/WCVP/",
+      licence = "CC BY 4.0",
+      citation = "Govaerts, R. et al. The World Checklist of Vascular Plants (WCVP). Royal Botanic Gardens, Kew.",
+      is_comprehensive = TRUE,
+      region_system = "wgsrpd3",
+      coverage = "global, WGSRPD level 3 (the geography WCVP publishes against)",
+      local_file = "wcvp-v15.zip (from the TNRS cache, or supplied)"
+    ),
+    vascan = list(
+      source_name = "vascan",
+      source_name_full = "Database of Vascular Plants of Canada (VASCAN)",
+      version = NA_character_,
+      url = "https://data.canadensys.net/ipt/archive.do?r=vascan",
+      licence = "CC0 1.0",
+      citation = "Brouillet, L. et al. VASCAN, the Database of Vascular Plants of Canada.",
+      is_comprehensive = TRUE,
+      region_system = "gadm1",
+      coverage = "Canada, by province and territory",
+      local_file = "vascan_dwca.zip"
+    ),
+    # usda is deliberately absent: the USDA no longer serves the per-state native-status
+    # export the service was built from. The GBIF copy of PLANTS is names only, and the
+    # current API gives status by region (CAN / L48 / AK / HI), not by state, which adds
+    # nothing over WCVP. See dev_notes/01-offline-nsr-design.md.
+    flbr = list(
+      source_name = "flbr",
+      source_name_full = "Flora e Funga do Brasil (Lista Oficial)",
+      version = NA_character_,
+      url = "https://api.checklistbank.org/dataset/2031/archive",
+      licence = "CC BY 4.0",
+      citation = "Flora e Funga do Brasil. Jardim Botanico do Rio de Janeiro.",
+      is_comprehensive = TRUE,
+      region_system = "gadm1",
+      coverage = "Brazil, by state; the level WGSRPD cannot reach (five level-3 regions)",
+      local_file = "flbr_dwca.zip"
+    )
+  )
+}
+
+#' Which parts of the local NSR are built
+#'
+#' @param dir Cache directory.  Defaults to the one shared with GNRS and GVS.
+#' @return A data.frame with one row per source: whether it is built, how many checklist
+#'   records and taxa it contributed, and when it was built.
+#' @export
+NSR_local_status <- function(dir = nsr_cache_dir()) {
+  nsr_need("nanoparquet")
+  reg <- nsr_builtin_registry()
+  built_tables <- vapply(nsr_tables(), function(t) file.exists(nsr_table_path(t, dir)), logical(1))
+  src <- if (built_tables[["sources"]]) {
+    as.data.frame(nanoparquet::read_parquet(nsr_table_path("sources", dir)))
+  } else NULL
+  out <- do.call(rbind, lapply(reg, function(s) {
+    row <- if (!is.null(src) && s$source_name %in% src$source_name) src[src$source_name == s$source_name, ] else NULL
+    data.frame(
+      source = s$source_name, coverage = s$coverage, licence = s$licence,
+      built = !is.null(row),
+      n_records = if (is.null(row)) NA_integer_ else as.integer(row$n_records),
+      n_taxa = if (is.null(row)) NA_integer_ else as.integer(row$n_taxa),
+      date_built = if (is.null(row)) NA_character_ else as.character(row$date_accessed),
+      stringsAsFactors = FALSE
+    )
+  }))
+  rownames(out) <- NULL
+  attr(out, "cache_dir") <- dir
+  out
+}
+
+#' Remove the local NSR tables
+#'
+#' @param dir Cache directory.
+#' @param quiet Suppress messages?
+#' @return Invisibly, the files removed.
+#' @export
+NSR_local_remove <- function(dir = nsr_cache_dir(), quiet = FALSE) {
+  f <- c(vapply(nsr_tables(), nsr_table_path, character(1), dir = dir),
+         nsr_table_path("region-links", dir), nsr_raster_path("wgsrpd3", dir),
+         nsr_raster_path("wgsrpd3_index", dir), nsr_provenance_path("nsr", dir))
+  f <- f[file.exists(f)]
+  unlink(f)
+  if (!quiet) message("Removed ", length(f), " file(s) from ", dir)
+  invisible(f)
+}
diff --git a/R/local_globals.R b/R/local_globals.R
new file mode 100644
index 0000000..25288e3
--- /dev/null
+++ b/R/local_globals.R
@@ -0,0 +1,14 @@
+# The offline resolver is written against data.table, which refers to columns by bare
+# name.  R CMD check cannot tell those from undefined globals, so they are declared here
+# rather than silenced one call at a time.  Keep this list in step with the set path.
+utils::globalVariables(c(
+  ".", ".N", ".SD", ":=",
+  "code", "conflict_type", "consulted_here", "fraction", "from_region",
+  "has_int", "has_nat", "in_place", "is_cultivated", "is_extinct",
+  "n_here", "n_in", "n_nat", "n_src_sets", "n_sub_int", "n_sub_nat",
+  "ovl", "ovl_srcs", "qid", "r_ord", "reason", "region_key", "relation",
+  "scope", "source_name", "srcs", "srcs_any",
+  "stated_int", "stated_nat", "stated_pre", "status",
+  "sub_sets", "sub_srcs", "sub_status",
+  "taxon_id", "to_region", "to_sys", "within_src"
+))
diff --git a/R/local_import.R b/R/local_import.R
new file mode 100644
index 0000000..fff73aa
--- /dev/null
+++ b/R/local_import.R
@@ -0,0 +1,233 @@
+# Source importers for the local NSR.
+#
+# Each returns a data.frame with one row per taxon x division x opinion:
+#   taxon_name        the source's own name string (resolved later against WCVP)
+#   taxon_id          the WCVP accepted id, where the source supplies one directly
+#   status            "native", "introduced" or "present"
+#   is_cultivated     1 where the source says the taxon is cultivated in that division
+#   region_key        "<system>:<code>", the source's OWN geography
+#   region_name       the source's name for it
+#
+# Each source keeps the regions it publishes against: WCVP records WGSRPD level-3 areas
+# (biogeographic, e.g. "Borneo"), VASCAN and Flora do Brasil record provinces and states
+# (GADM level 1). Queries are resolved against those polygons, not against names; see
+# local_regions.R.
+
+#' Read a delimited file quickly when data.table is installed
+#' @keywords internal
+#' @noRd
+nsr_read_delim <- function(path_or_conn, sep = "\t", quote = "\"", select = NULL) {
+  if (requireNamespace("data.table", quietly = TRUE) && is.character(path_or_conn)) {
+    return(as.data.frame(data.table::fread(path_or_conn, sep = sep, quote = quote,
+                                           showProgress = FALSE, data.table = FALSE,
+                                           select = select, colClasses = "character")))
+  }
+  x <- utils::read.delim(path_or_conn, sep = sep, quote = quote, colClasses = "character",
+                         stringsAsFactors = FALSE, na.strings = "")
+  if (!is.null(select)) x <- x[, intersect(select, names(x)), drop = FALSE]
+  x
+}
+
+#' Extract one member of a zip to a temporary file
+#' @keywords internal
+#' @noRd
+nsr_unzip_member <- function(zip, member) {
+  td <- file.path(tempdir(), paste0("nsr-", basename(zip)))
+  dir.create(td, showWarnings = FALSE, recursive = TRUE)
+  utils::unzip(zip, files = member, exdir = td, overwrite = TRUE)
+  file.path(td, member)
+}
+
+#' POWO / WCVP: global native and introduced ranges at WGSRPD level 3
+#'
+#' Internal.  Distributions are kept against the WGSRPD level-3 areas WCVP publishes
+#' them against; the polygons do the rest.  Matching those areas to political divisions
+#' by NAME loses a quarter of them (Borneo, Sulawesi, Maluku, New Guinea, Lesser Sunda
+#' Is. are not country or state names) and cannot express "Brazil Southeast" containing
+#' four states, so it is not done.
+#'
+#' Names need no resolver: WCVP is the backbone, so \code{accepted_plant_name_id} is the
+#' taxon id directly.
+#' @keywords internal
+#' @noRd
+nsr_import_wcvp <- function(zip = NULL, bb, quiet = FALSE) {
+  if (is.null(zip)) zip <- nsr_find_wcvp_zip()
+  if (is.null(zip) || !file.exists(zip)) {
+    stop("Could not find the WCVP archive. Supply it as files = list(wcvp = \"wcvp-v15.zip\"), ",
+         "or build the TNRS 'wcvp' source, whose cache holds one.", call. = FALSE)
+  }
+  dist_f <- nsr_unzip_member(zip, "wcvp_distribution.csv")
+  names_f <- nsr_unzip_member(zip, "wcvp_names.csv")
+  if (!quiet) message("  reading WCVP ...")
+  dist <- nsr_read_delim(dist_f, sep = "|",
+                         select = c("plant_name_id", "area_code_l3", "area", "introduced",
+                                    "extinct", "location_doubtful"))
+  nm <- nsr_read_delim(names_f, sep = "|",
+                       select = c("plant_name_id", "accepted_plant_name_id", "parent_plant_name_id",
+                                  "taxon_name", "taxon_rank", "taxon_status", "family", "genus"))
+
+  # distributions are recorded against a name; carry them to its accepted taxon
+  i <- match(dist$plant_name_id, nm$plant_name_id)
+  acc <- nm$accepted_plant_name_id[i]
+  acc[is.na(acc) | !nzchar(acc)] <- nm$plant_name_id[i][is.na(acc) | !nzchar(acc)]
+  # WCVP records distributions against accepted infraspecific taxa too (60k of them), so a
+  # subspecies' range hangs off the subspecies. Roll those up to the species: it is NSR's
+  # own rule (native propagates upward) and it is what a species-level query needs.
+  sp <- nsr_species_of(acc, nm)
+  j <- match(sp, nm$plant_name_id)
+  dist$taxon_id <- sp
+  dist$taxon_name <- nm$taxon_name[j]
+  dist$rank_published <- nm$taxon_rank[match(acc, nm$plant_name_id)]
+
+  flag <- function(x) !is.na(x) & x %in% c("1", 1, TRUE, "TRUE")
+  # Extinct records are KEPT and flagged rather than dropped here.  WCVP's `extinct` is a
+  # bare 0/1 with no date - it means "considered no longer present as of this release" -
+  # so whether it should count is a property of the question, not of the build: present-day
+  # status wants it out, a historical distribution wants it in.  NSR_local(exclude_extinct)
+  # decides at query time, which is only possible if the build keeps the row.  Dropping it
+  # here would also lose the fact that the taxon was ONCE native there, and the Ie rule
+  # needs that fact or it will call a natively extirpated record introduced.
+  # A DOUBTFUL record is kept as "present": too weak to support native or introduced, and
+  # dropping it would let the region read as a confident absence when the source in fact
+  # records a maybe.
+  keep <- !is.na(dist$taxon_id) & !is.na(dist$taxon_name)
+  dist <- dist[keep, , drop = FALSE]
+  j <- j[keep]                      # j indexes nm per dist row, so it is subset with it
+
+  out <- data.frame(
+    taxon_name = dist$taxon_name, taxon_id = dist$taxon_id,
+    species_name = dist$taxon_name, family = nm$family[j], genus = nm$genus[j],
+    rank = nm$taxon_rank[j],
+    status = ifelse(flag(dist$location_doubtful), "present",
+                    ifelse(flag(dist$introduced), "introduced", "native")),
+    is_cultivated = 0L, is_extinct = as.integer(flag(dist$extinct)),
+    region_key = paste0("wgsrpd3:", dist$area_code_l3), region_name = dist$area,
+    stringsAsFactors = FALSE
+  )
+  out[!is.na(out$region_key), , drop = FALSE]
+}
+
+#' Walk accepted infraspecific taxa up to their species
+#'
+#' Internal.  Up to three steps (a variety under a subspecies), stopping at the first
+#' ancestor of species rank; anything already at species rank or above is left alone.
+#' @keywords internal
+#' @noRd
+nsr_species_of <- function(ids, nm, max_steps = 3L) {
+  rank <- tolower(nm$taxon_rank)
+  is_infra <- !rank %in% c("species", "genus", "family", "")
+  out <- ids
+  for (step in seq_len(max_steps)) {
+    i <- match(out, nm$plant_name_id)
+    up <- !is.na(i) & is_infra[i]
+    if (!any(up)) break
+    par <- nm$parent_plant_name_id[i[up]]
+    par[is.na(par) | !nzchar(par)] <- out[up][is.na(par) | !nzchar(par)]
+    out[up] <- par
+  }
+  out
+}
+
+#' Where TNRS keeps its copy of WCVP, if it has one
+#' @keywords internal
+#' @noRd
+nsr_find_wcvp_zip <- function() {
+  if (!requireNamespace("TNRS", quietly = TRUE)) return(NULL)
+  d <- try(TNRS:::tnrs_cache_dir(), silent = TRUE)
+  if (inherits(d, "try-error") || !dir.exists(d)) return(NULL)
+  f <- list.files(d, pattern = "^wcvp.*\\.zip$", full.names = TRUE)
+  if (!length(f)) NULL else f[order(file.mtime(f), decreasing = TRUE)][1]
+}
+
+#' VASCAN: Canada, by province and territory
+#' @keywords internal
+#' @noRd
+nsr_import_vascan <- function(zip, bb, quiet = FALSE) {
+  if (is.null(zip) || !file.exists(zip)) {
+    stop("Supply the VASCAN archive as files = list(vascan = \"vascan_dwca.zip\"); ",
+         "download it from https://data.canadensys.net/ipt/archive.do?r=vascan", call. = FALSE)
+  }
+  tx <- nsr_read_delim(nsr_unzip_member(zip, "taxon.txt"))
+  di <- nsr_read_delim(nsr_unzip_member(zip, "distribution.txt"))
+  # carry each distribution to the accepted name of its taxon
+  acc <- tx$acceptedNameUsageID
+  acc[is.na(acc) | !nzchar(acc)] <- tx$id[is.na(acc) | !nzchar(acc)]
+  i <- match(di$id, tx$id)
+  j <- match(acc[i], tx$id)
+  di$taxon_name <- tx$scientificName[j]
+  # VASCAN splits the two questions: occurrenceStatus is present / excluded / doubtful /
+  # irregular / absent, establishmentMeans is native / introduced. Nativeness comes from
+  # the latter; the former only says whether the record counts at all.
+  occ <- tolower(ifelse(is.na(di$occurrenceStatus), "", di$occurrenceStatus))
+  est <- tolower(ifelse(is.na(di$establishmentMeans), "", di$establishmentMeans))
+  status <- ifelse(est %in% c("native", "endemic"), "native",
+                   ifelse(est %in% c("introduced", "naturalized", "naturalised"), "introduced",
+                          "present"))
+  status[occ %in% c("absent", "excluded")] <- NA_character_
+  # locationID is "ISO3166-2:CA-NB"; the code beats the name, since VASCAN's spellings and
+  # GADM's differ and the list also covers places outside Canada
+  code <- sub("^ISO3166-2:", "", ifelse(is.na(di$locationID), "", di$locationID))
+  gid1 <- bb$state$gid_1[match(ifelse(nzchar(code), sub("-", ".", toupper(code)), NA_character_),
+                               bb$state$hasc_full)]
+  # the three VASCAN entries carrying no code: two are one GADM province, one another country
+  loc <- tolower(ifelse(is.na(di$locality), "", di$locality))
+  gid1[is.na(gid1) & loc %in% c("newfoundland", "labrador")] <-
+    bb$state$gid_1[match("CA.NF", bb$state$hasc_full)]
+  gid0 <- rep(NA_character_, nrow(di))
+  gid0[is.na(gid1) & loc == "greenland"] <- "GRL"
+  out <- data.frame(taxon_name = di$taxon_name, taxon_id = NA_character_, status = status,
+                    is_cultivated = 0L, is_extinct = 0L,
+                    region_key = ifelse(!is.na(gid1), paste0("gadm1:", gid1),
+                                        ifelse(!is.na(gid0), paste0("gadm0:", gid0), NA_character_)),
+                    region_name = di$locality, stringsAsFactors = FALSE)
+  if (!quiet) {
+    miss <- unique(di$locality[is.na(out$region_key)])
+    if (length(miss)) message("  ", length(miss), " VASCAN divisions unmatched: ",
+                              paste(utils::head(miss, 6), collapse = ", "))
+  }
+  out[!is.na(out$status) & !is.na(out$region_key) & !is.na(out$taxon_name), , drop = FALSE]
+}
+
+#' Flora e Funga do Brasil: Brazil, by state
+#'
+#' Internal.  \code{establishmentMeans} is Portuguese: NATIVA, NATURALIZADA, CULTIVADA.
+#' Cultivated rows are kept as "present" and flagged, which is what the service's
+#' \code{isCultivatedNSR} reports.  Divisions come from the ISO 3166-2 code in
+#' \code{locationID} (BR-SP ...), matched to the GNRS backbone's HASC code.
+#' @keywords internal
+#' @noRd
+nsr_import_flbr <- function(zip, bb, quiet = FALSE) {
+  if (is.null(zip) || !file.exists(zip)) {
+    stop("Supply the Flora do Brasil archive as files = list(flbr = \"flbr_dwca.zip\"); ",
+         "download it from https://api.checklistbank.org/dataset/2031/archive", call. = FALSE)
+  }
+  tx <- nsr_read_delim(nsr_unzip_member(zip, "taxon.txt"))
+  di <- nsr_read_delim(nsr_unzip_member(zip, "distribution.txt"))
+  acc <- tx$acceptedNameUsageID
+  acc[is.na(acc) | !nzchar(acc)] <- tx$id[is.na(acc) | !nzchar(acc)]
+  i <- match(di$id, tx$id)
+  j <- match(acc[i], tx$id)
+  di$taxon_name <- tx$scientificName[j]
+
+  em <- toupper(ifelse(is.na(di$establishmentMeans), "", di$establishmentMeans))
+  status <- ifelse(em == "NATIVA", "native",
+                   ifelse(em == "NATURALIZADA", "introduced", "present"))
+  cult <- as.integer(em == "CULTIVADA")
+
+  # BR-SP -> GNRS hasc_full "BR.SP" -> GADM gid_1
+  hasc <- sub("-", ".", toupper(ifelse(is.na(di$locationID), "", di$locationID)))
+  k <- match(hasc, bb$state$hasc_full)
+  out <- data.frame(
+    taxon_name = di$taxon_name, taxon_id = NA_character_, status = status, is_cultivated = cult,
+    is_extinct = 0L,
+    region_key = ifelse(is.na(bb$state$gid_1[k]), NA_character_, paste0("gadm1:", bb$state$gid_1[k])),
+    region_name = bb$state$state_province[k],
+    stringsAsFactors = FALSE
+  )
+  if (!quiet) {
+    miss <- unique(di$locationID[is.na(k)])
+    if (length(miss)) message("  ", length(miss), " Brazilian location codes unmatched: ",
+                              paste(utils::head(miss, 8), collapse = ", "))
+  }
+  out[!is.na(out$region_key) & !is.na(out$taxon_name), , drop = FALSE]
+}
diff --git a/R/local_regions.R b/R/local_regions.R
new file mode 100644
index 0000000..962f4cf
--- /dev/null
+++ b/R/local_regions.R
@@ -0,0 +1,200 @@
+# Regions, and the spatial relations between them.
+#
+# Sources publish against different geographies: WCVP against WGSRPD level-3 areas
+# (biogeographic), VASCAN and Flora do Brasil against GADM level-1 units (political).
+# Both are kept as they are. A coordinate is located in each system directly; a query
+# given only division names is carried between systems by a link table derived from the
+# polygons themselves, so nothing is matched by name.
+
+#' Paths of the region rasters
+#' @keywords internal
+#' @noRd
+nsr_raster_path <- function(what, dir = nsr_cache_dir()) {
+  switch(what,
+    wgsrpd3 = file.path(dir, "nsr-wgsrpd3-30s.tif"),
+    wgsrpd3_index = file.path(dir, "nsr-wgsrpd3-index.gz.parquet"),
+    stop("unknown raster: ", what)
+  )
+}
+
+#' Rasterize the WGSRPD level-3 polygons
+#'
+#' Internal.  30 arc-seconds, to match the GADM index GVS builds, so the two can be
+#' cross-tabulated cell for cell.  Polygons come from \code{rWCVPdata}, the published
+#' TDWG shapes; nothing is downloaded here.
+#' @keywords internal
+#' @noRd
+nsr_build_wgsrpd_raster <- function(dir = nsr_cache_dir(create = TRUE), resolution = 1 / 120,
+                                    quiet = FALSE) {
+  for (pkg in c("sf", "terra")) {
+    if (!requireNamespace(pkg, quietly = TRUE)) stop("Needs the '", pkg, "' package.", call. = FALSE)
+  }
+  pol <- nsr_wgsrpd_polygons()
+  idx <- data.frame(unit = seq_len(nrow(pol)), area_code_l3 = pol$LEVEL3_COD,
+                    area_name = pol$LEVEL3_NAM, stringsAsFactors = FALSE)
+  nanoparquet::write_parquet(idx, nsr_raster_path("wgsrpd3_index", dir), compression = "gzip")
+  if (!quiet) message("Rasterizing ", nrow(pol), " WGSRPD level-3 areas at ",
+                      round(resolution * 3600), " arc-seconds ...")
+  pol$unit <- idx$unit
+  tmpl <- terra::rast(terra::ext(-180, 180, -90, 90), resolution = resolution, crs = "EPSG:4326")
+  terra::rasterize(terra::vect(pol), tmpl, field = "unit",
+                   filename = nsr_raster_path("wgsrpd3", dir), overwrite = TRUE,
+                   wopt = list(datatype = "INT2U", gdal = c("COMPRESS=DEFLATE", "TILED=YES")))
+  if (!quiet) message("  done")
+  invisible(nsr_raster_path("wgsrpd3", dir))
+}
+
+#' The TDWG level-3 polygons
+#' @keywords internal
+#' @noRd
+nsr_wgsrpd_polygons <- function() {
+  nsr_need("sf", what = "The WGSRPD polygons")
+  if (requireNamespace("rWCVPdata", quietly = TRUE)) {
+    p <- try(rWCVPdata::wgsrpd3, silent = TRUE)
+    if (!inherits(p, "try-error")) return(sf::st_as_sf(p))
+  }
+  if (requireNamespace("rWCVP", quietly = TRUE)) {
+    e <- new.env()
+    utils::data("wgsrpd3", package = "rWCVP", envir = e)
+    return(sf::st_as_sf(get("wgsrpd3", envir = e)))
+  }
+  stop("The WGSRPD level-3 polygons need the 'rWCVPdata' (or 'rWCVP') package.", call. = FALSE)
+}
+
+#' Spatial links between region systems
+#'
+#' Internal.  Cross-tabulates the WGSRPD level-3 raster against the GADM unit index GVS
+#' builds, in blocks of rows, and records for each overlapping pair how much of each
+#' region lies in the other.  \code{fraction_from} is the share of the FROM region inside
+#' the TO region, so \code{within} means the from-region is (almost) entirely inside.
+#'
+#' @return Invisibly, the link table.
+#' @keywords internal
+#' @noRd
+nsr_build_region_links <- function(dir = nsr_cache_dir(), gadm_raster = NULL, block_rows = 500,
+                                   within_threshold = 0.95, quiet = FALSE) {
+  if (!requireNamespace("terra", quietly = TRUE)) stop("Needs the 'terra' package.", call. = FALSE)
+  if (is.null(gadm_raster)) gadm_raster <- file.path(dir, "gadmindex-units-30s.tif")
+  if (!file.exists(gadm_raster)) {
+    stop("The GADM unit index is not in ", dir, ".\n",
+         "Build it with GVS (GVS_local_build(\"index\")); NSR reuses it rather than making its own.",
+         call. = FALSE)
+  }
+  if (!file.exists(nsr_raster_path("wgsrpd3", dir))) nsr_build_wgsrpd_raster(dir, quiet = quiet)
+
+  w <- terra::rast(nsr_raster_path("wgsrpd3", dir))
+  g <- terra::rast(gadm_raster)
+  if (!isTRUE(all.equal(as.vector(terra::ext(w)), as.vector(terra::ext(g)))) ||
+      terra::ncol(w) != terra::ncol(g) || terra::nrow(w) != terra::nrow(g)) {
+    stop("The WGSRPD and GADM rasters do not align; rebuild one at the other's resolution.",
+         call. = FALSE)
+  }
+  units <- as.data.frame(nanoparquet::read_parquet(file.path(dir, "gadmindex-units.gz.parquet")))
+  widx <- as.data.frame(nanoparquet::read_parquet(nsr_raster_path("wgsrpd3_index", dir)))
+
+  nr <- terra::nrow(w)
+  nc <- terra::ncol(w)
+  terra::readStart(w); terra::readStart(g)
+  on.exit({terra::readStop(w); terra::readStop(g)}, add = TRUE)
+  acc <- new.env(parent = emptyenv())
+  if (!quiet) message("Cross-tabulating WGSRPD x GADM (", nr, " rows) ...")
+  for (s in seq(1L, nr, by = block_rows)) {
+    e <- min(nr, s + block_rows - 1L)
+    wv <- terra::readValues(w, row = s, nrows = e - s + 1L, col = 1L, ncols = nc)
+    gv <- terra::readValues(g, row = s, nrows = e - s + 1L, col = 1L, ncols = nc)
+    ok <- !is.na(wv) & !is.na(gv) & wv > 0 & gv > 0
+    if (!any(ok)) next
+    key <- paste(wv[ok], gv[ok])
+    tb <- table(key)
+    for (k in names(tb)) {
+      acc[[k]] <- (if (is.null(acc[[k]])) 0L else acc[[k]]) + as.integer(tb[[k]])
+    }
+    rm(wv, gv, ok, key, tb)
+  }
+  keys <- ls(acc)
+  if (!length(keys)) stop("No overlap found between the two rasters.", call. = FALSE)
+  parts <- do.call(rbind, strsplit(keys, " ", fixed = TRUE))
+  cells <- vapply(keys, function(k) acc[[k]], integer(1))
+  d <- data.frame(w_unit = as.integer(parts[, 1]), g_unit = as.integer(parts[, 2]),
+                  cells = as.integer(cells), stringsAsFactors = FALSE)
+  d$w_key <- paste0("wgsrpd3:", widx$area_code_l3[match(d$w_unit, widx$unit)])
+  gi <- match(d$g_unit, units$unit)
+  d$g_key <- ifelse(!is.na(units$gid_1[gi]), paste0("gadm1:", units$gid_1[gi]),
+                    paste0("gadm0:", units$gid_0[gi]))
+  # A query about a country must find links too, so the level-0 pairs are built as well -
+  # but each level needs its own denominators. Pooling them counts every cell twice (once
+  # for its state, once for its country), which halves every fraction and turns even
+  # coextensive regions into "overlaps".
+  d$g_key0 <- paste0("gadm0:", units$gid_0[gi])
+  fracs <- function(gk) {
+    dd <- stats::aggregate(cells ~ w_key + g_key,
+                           data = data.frame(w_key = d$w_key, g_key = gk, cells = d$cells,
+                                             stringsAsFactors = FALSE), FUN = sum)
+    wt <- stats::aggregate(cells ~ w_key, data = dd, FUN = sum)
+    gt <- stats::aggregate(cells ~ g_key, data = dd, FUN = sum)
+    dd$frac_w <- dd$cells / wt$cells[match(dd$w_key, wt$w_key)]
+    dd$frac_g <- dd$cells / gt$cells[match(dd$g_key, gt$g_key)]
+    dd
+  }
+  d <- rbind(fracs(d$g_key), fracs(d$g_key0))
+  d <- d[!duplicated(paste(d$w_key, d$g_key)), , drop = FALSE]
+
+  # Coextensive regions (California the GADM state and CAL the WGSRPD area) are the same
+  # place, and must not demote a native opinion to mere presence. The threshold is 0.95,
+  # not 0.99: these fractions come from 30 arc-second rasters of two independent
+  # coastlines, so Austria and WGSRPD's AUT score 0.990/0.988 rather than 1.0.
+  rel <- function(frac_from, frac_to) {
+    ifelse(frac_from >= within_threshold & frac_to >= within_threshold, "same",
+           ifelse(frac_from >= within_threshold, "within",
+                  ifelse(frac_to >= within_threshold, "contains", "overlaps")))
+  }
+  links <- rbind(
+    data.frame(from_region = d$w_key, to_region = d$g_key,
+               relation = rel(d$frac_w, d$frac_g), fraction = d$frac_w, stringsAsFactors = FALSE),
+    data.frame(from_region = d$g_key, to_region = d$w_key,
+               relation = rel(d$frac_g, d$frac_w), fraction = d$frac_g, stringsAsFactors = FALSE)
+  )
+  nanoparquet::write_parquet(links, nsr_table_path("region-links", dir), compression = "gzip")
+  if (!quiet) message("  ", format(nrow(links), big.mark = ","), " links (",
+                      sum(links$relation == "within"), " within, ",
+                      sum(links$relation == "contains"), " contains, ",
+                      sum(links$relation == "overlaps"), " overlaps)")
+  invisible(links)
+}
+
+#' Locate coordinates in both region systems
+#'
+#' Internal.  Raster lookup in the WGSRPD level-3 raster and in the GADM unit index, so a
+#' record with coordinates is placed in each source's own geography without any crosswalk.
+#' @return data.frame with \code{wgsrpd3} and \code{gadm} region keys (and the GADM
+#'   country key), one row per point.
+#' @keywords internal
+#' @noRd
+nsr_locate_regions <- function(longitude, latitude, dir = nsr_cache_dir()) {
+  if (!requireNamespace("terra", quietly = TRUE)) stop("Needs the 'terra' package.", call. = FALSE)
+  n <- length(longitude)
+  out <- data.frame(wgsrpd3 = rep(NA_character_, n), gadm = NA_character_, gadm0 = NA_character_,
+                    stringsAsFactors = FALSE)
+  ok <- is.finite(longitude) & is.finite(latitude)
+  if (!any(ok)) return(out)
+  xy <- cbind(longitude[ok], latitude[ok])
+
+  wf <- nsr_raster_path("wgsrpd3", dir)
+  if (file.exists(wf)) {
+    widx <- as.data.frame(nanoparquet::read_parquet(nsr_raster_path("wgsrpd3_index", dir)))
+    u <- as.integer(terra::extract(terra::rast(wf), xy)[, 1])
+    out$wgsrpd3[ok] <- ifelse(is.na(u) | u == 0, NA_character_,
+                              paste0("wgsrpd3:", widx$area_code_l3[match(u, widx$unit)]))
+  }
+  gf <- file.path(dir, "gadmindex-units-30s.tif")
+  if (file.exists(gf)) {
+    units <- as.data.frame(nanoparquet::read_parquet(file.path(dir, "gadmindex-units.gz.parquet")))
+    u <- as.integer(terra::extract(terra::rast(gf), xy)[, 1])
+    i <- match(u, units$unit)
+    out$gadm[ok] <- ifelse(is.na(i), NA_character_,
+                           ifelse(!is.na(units$gid_1[i]), paste0("gadm1:", units$gid_1[i]),
+                                  paste0("gadm0:", units$gid_0[i])))
+    out$gadm0[ok] <- ifelse(is.na(i), NA_character_, paste0("gadm0:", units$gid_0[i]))
+  }
+  out
+}
diff --git a/dev_notes/01-offline-nsr-design.md b/dev_notes/01-offline-nsr-design.md
new file mode 100644
index 0000000..fab93f7
--- /dev/null
+++ b/dev_notes/01-offline-nsr-design.md
@@ -0,0 +1,573 @@
+# 05. Design: an offline NSR (native status) for the BIEN service stack
+
+Draft 2026-09-18, for BM's review before implementation. Sibling of `04-versioned-divisions-design.md`;
+the political-division reference built there is the backbone here. Lives in `RGNRS/dev_notes`
+only until an `RNSR` clone exists, then moves with the code.
+
+## Why this, and what already exists
+
+NSR is the last BIEN service with no offline implementation:
+
+| Service | Offline | Where |
+|---|---|---|
+| TNRS | yes | `RTNRS`: `TNRS_local()`, `local_build.R`, `local_backbone.R`, `tnrs_cache_dir()` |
+| GNRS | yes | `RGNRS` branch `versioned-divisions`: `GNRS_local()`, history/cshapes components |
+| GVS | yes | `RGVS` branch `local-implementation`: `GVS_local()`, raster index, centroids, CShapes |
+| **NSR** | **no** | `NSR` 0.1.0 is API-only: `NSR()`, `NSR_simple()`, `NSR_sources()`, metadata helpers |
+
+The service's database is built by PHP under `ojalaquellueva/nsr/db` (the separate `nsr_db` repo is
+retired), importing 12 checklists. There is no published build of the compiled database, and the
+repository documents no licence for the source data, so, as with GADM and CShapes, **sources are
+downloaded on the user's machine at build time and nothing derived is shipped**.
+
+## What the service does, and must be reproduced
+
+Output (probed live, `Pinus ponderosa` / United States / California), 18 columns:
+
+    family, genus, species, country, state_province, county_parish, poldiv_full, poldiv_type,
+    native_status_country, native_status_state_province, native_status_county_parish,
+    native_status, native_status_reason, native_status_sources, isIntroduced,
+    isCultivatedNSR, is_cultivated_taxon, user_id
+
+Status codes: **N** native, **Ne** native and endemic, **I** introduced, **Ie** introduced inferred
+from endemism elsewhere, **P** present without status, **A** absent from the checklists, **UNK** no
+data. A status is returned per political level, plus an overall `native_status` taken at the lowest
+division supplied.
+
+Propagation rules, from the service README:
+- **Taxonomic:** native propagates *up* (a native variety makes the species, genus and family native
+  there); introduced propagates *down* (an introduced species makes its varieties introduced). With
+  no opinion for the taxon, higher ranks are consulted. **We reproduce the first and decline the
+  last - see below.**
+- **Political:** native propagates *up* (native in a county implies native in its state and country);
+  introduced propagates *down* (introduced in a country implies introduced in its states).
+- **Absence:** absent from a checklist gives `A`; absent while endemic elsewhere gives `Ie`.
+- `is_comprehensive` per source decides whether absence is informative at all: only a checklist that
+  claims to list everything in its area can turn "not listed" into evidence.
+
+## Architecture
+
+Mirrors the other two packages, so the three share one cache and one idiom.
+
+- Component `nsr` in the shared cache (`GNRS.cache_dir` / `R_user_dir("GNRS")`), tables prefixed
+  `nsr-`, provenance RDS per component, `nanoparquet` `.gz.parquet` files.
+- `NSR_local_build(sources = c("powo", "vascan", "flbr"), dir, ...)`, one importer per
+  source behind a common interface (fetch -> standardise -> resolve names -> resolve divisions ->
+  append), so adding the remaining eight later is additive.
+- `NSR_local(occurrence_dataframe, dir, ...)` with the service's input columns
+  (`species`/`genus`/`family`, `country`, `state_province`, `county_parish`) and its 18 output
+  columns, so existing code swaps one call for the other.
+- **Names** resolve through `TNRS_local(sources = "wcvp")`, the resolver every other layer of this
+  project already uses; checklist names and query names go through the same call, so matching is
+  backbone-consistent by construction rather than by string equality.
+- **Political divisions** resolve through `GNRS_local()`, which gives the country/state/county ids
+  and, new in the versioned-divisions work, **historical divisions**: a record labelled "USSR" or
+  "Zaire" resolves to the entity and its current successors, so native status can be looked up in
+  the successor whose polygon contains it. The live NSR cannot do this.
+
+### Cache tables
+
+| File | One row per | Columns |
+|---|---|---|
+| `nsr-sources` | source | source_name, full name, url, date_accessed, is_comprehensive, licence, citation, poldiv_level, divisions covered |
+| `nsr-checklist` | taxon x division x source | taxon_id, rank, poldiv_id, poldiv_level, status (`native`/`introduced`/`present`), source_id |
+| `nsr-taxa` | taxon | taxon_id (WCVP accepted id), family, genus, species, rank, is_cultivated_taxon |
+| `nsr-regions` | region | region_key (`system:code`), system, code, name, level, GNRS ids and GADM gid where applicable |
+| `nsr-region-links` | pair of regions | from_region, to_region, relation (`within`/`contains`/`overlaps`), fraction |
+| `nsr-endemism` | taxon | taxon_id, endemic_poldiv_id, source (for the `Ie` rule) |
+
+Checklist rows key on resolved ids, never on strings, so a source's own spellings and division names
+are a build-time problem only.
+
+## Sources, first release
+
+All four are public and redistributable-in-principle, and cover the places where WGSRPD level 3 (the
+resolution WCVP gives us) is too coarse to say anything about sub-country status.
+
+| Source | What it gives | Format, size | Licence |
+|---|---|---|---|
+| **wcvp** | global native/introduced by WGSRPD level 3 | WCVP v15 **already in hand** (`data/wcvp_v15/wcvp_distribution.csv`), no download | CC BY 4.0 |
+| ~~usda~~ | USA, by state | **not obtainable**: the per-state export is no longer served; the GBIF copy is names only and the API gives region-level status | CC0 |
+| **vascan** | Canada, by province; native/introduced/ephemeral | DwC-A, `data.canadensys.net/ipt/archive.do?r=vascan`, ~10 MB, updated 2026-08-04 | CC0 |
+| **flbr** | Brazil, by state; native/naturalised/cultivated | DwC-A, `ipt.jbrj.gov.br/jbrj/archive.do?r=lista_especies_flora_brasil`, ~50-100 MB | CC BY 4.0 |
+
+**Division names are the default input; coordinates are opt-in (BM, 2026-09-22):**
+`NSR_local()` answers from `country`/`state_province` unless `use_coordinates = TRUE`.
+Two reasons. Placing every record by point costs a raster lookup per call, which is the
+single most expensive thing the query path can do. And resolving coordinates to political
+divisions is GVS's job: a pipeline that needs it already runs GVS, and duplicating that
+step here would mean two implementations of the same lookup drifting apart. The
+coordinate path stays available because it asks a finer question - it consults each
+source in its own geography with no crosswalk - but it is a deliberate choice rather than
+a default.
+
+**Coordinates and names are compared, not pooled (BM, 2026-09-22):** the first
+implementation unioned the region keys from both, so a record whose point and whose
+division names disagreed was answered from evidence for two places and described neither.
+Coordinates are not taken as authoritative instead: a transposed longitude and a mistyped
+province are equally easy mistakes, and the existing `NSR_from_coordinates()` example
+already renames `country` to `country_declared` precisely so declared and inferred
+divisions can be compared. The two are therefore compared in the geography they share,
+GADM, at each level separately - a right country with a wrong state is caught as readily
+as a wrong country - and a level only one of them speaks to is not a disagreement. Where
+they agree the wider key set answers, as before. Where they do not, `place_conflict` is
+`TRUE`, `native_status` is `UNK`, and `native_status_coordinates` and
+`native_status_names` carry the two answers. There is no service behaviour to follow
+here: `NSR()` takes names only and `NSR_from_coordinates()` takes coordinates only, so
+the API never meets the case.
+
+**Extinct records are kept and filtered at query time (BM, 2026-09-22):** WCVP's
+`extinct` is a bare 0/1 with no date or year - checked against the v15 distribution table,
+whose only columns are `plant_locality_id, plant_name_id, continent_code_l1, continent,
+region_code_l2, region, area_code_l3, area, introduced, extinct, location_doubtful`. It
+therefore means "considered no longer present as of this release" and cannot be compared
+against an occurrence's own date. Whether it should count is a property of the question,
+so the build keeps the row with an `is_extinct` flag and `NSR_local(exclude_extinct = TRUE)`
+decides per call: the default answers about the present day, `FALSE` models a past
+distribution. 2,701 of 1,970,252 rows are affected, over 2,393 taxa; 906 of those taxa have
+no surviving native record at all.
+
+Independently of that switch, the endemism rules always read the taxon's whole native
+range, extinct records included. A region a taxon has been lost from is still a region it
+was native to, so `Ie` must not fire against it - the answer for a natively extirpated
+record is `A` (gone), never `I`/`Ie` (arrived). Filtering at build time, as the garden
+pipeline's script 02 does, would have destroyed the evidence needed for that distinction.
+
+**The source is named `wcvp`, not `powo` (BM, 2026-09-22):** it is the WCVP archive that is
+read, and it is what `TNRS_local(sources = "wcvp")` calls the same data, so the stack uses one
+name for one dataset. `"powo"` is accepted as a synonym in `sources =` and in `files =`. The
+cost is that `native_status_sources` no longer matches the live service's label for this source,
+which is a known diff when validating `NSR_local()` against `NSR()`.
+
+POWO is the service's global backbone and is the same Kew dataset as WCVP, at the same resolution;
+we already apply its native filter in the garden pipeline (script 02 keeps `introduced == 0 &
+extinct == 0 & location_doubtful == 0`). The other three add state-level status in the USA, Canada
+and Brazil. Brazil matters most: WGSRPD divides it into five regions, so level 3 cannot distinguish
+a species native to Amazonia from one introduced to São Paulo.
+
+**usda is dropped from the release (BM, 2026-09-19): the USDA no longer serves the file.**
+NSR's own importer expects `usda_plants_native_status.csv`, an export of PLANTS' native status
+by state. Checked 2026-09-18: the GBIF-hosted copy of the PLANTS database
+(`hosted-datasets.gbif.org/datasets/usda.zip`, CC0, 2.1 MB) is a COLDP archive of names only -
+`NameUsage.tsv` and `VernacularName.tsv`, no distribution file; the `csvdownload` endpoint the
+old site used now returns the site shell; and the current API
+(`plantsservices.sc.egov.usda.gov/api/PlantProfile?symbol=...`) returns native status only by
+region - CAN, L48, AK, HI, PR - not by state, which is the resolution that would have added
+anything over POWO. No public bulk route to per-state status was found.
+
+**Consequence, to state in the write-up:** in the USA the build has POWO alone, whose introduced
+ranges are thin there - WCVP v15 holds no California record for *Quercus robur* at all, though it
+is naturalised across the state. Five of the fifteen disagreements with the live service in the
+API validation cited `usda`. US records will therefore skew towards `A` (absent) rather than `I`
+where a species is introduced but unlisted, and `Ie` only rescues those whose native range is
+confined elsewhere. Weakley's southeastern flora would cover part of the same ground if a route to
+it appears.
+
+Deferred to a later release: conosur, fwi, ipane, mab, mexico, newguinea, tropicos, weakley. Each
+needs its own acquisition route (several are web pages or an API with a key), and two (`mab`
+endemic genera, `ipane` introduced-only) are small special-purpose lists rather than checklists.
+
+## Regions are polygons, not names (BM, 2026-09-18)
+
+**Each source keeps its own geography, and queries are resolved against it spatially.**
+WCVP records distributions against WGSRPD level-3 areas, which are biogeographic, not political;
+VASCAN and Flora do Brasil record provinces and states, which are GADM units. Crosswalking either
+onto the other by *name* loses data and invents detail:
+
+- A first implementation matched level-3 unit names to GNRS divisions. Result: 79 of 370 units
+  matched a state, 194 fell back to country, and **97 (26%) matched nothing at all** - Borneo,
+  Sulawesi, Maluku, New Guinea, Lesser Sunda Is., Santa Cruz Is. Every POWO opinion for those areas
+  would have been dropped.
+- The reverse fails too: "Brazil Southeast" is one level-3 unit covering four states, so a name
+  match can say nothing about Sao Paulo, although the containment is exact and knowable.
+
+Instead:
+
+| Source | Region system | Key |
+|---|---|---|
+| powo | WGSRPD level 3 (TDWG polygons, `rWCVPdata::wgsrpd3`) | `wgsrpd3:BZL` |
+| vascan, flbr | GADM level 1, via the GNRS backbone's `gid_1` | `gadm1:BRA.25_1` |
+| (later sources) | whichever they publish against | `<system>:<code>` |
+
+**A query with coordinates is resolved natively in both systems**, by raster lookup: the WGSRPD
+level-3 raster (built for the garden pipeline) and the GADM unit index (built for GVS), both at 30
+arc-seconds with an exact fallback in boundary cells. No crosswalk is involved, which is the case
+that matters for occurrence records.
+
+**A query with only division names** resolves through GNRS to GADM units, then to WGSRPD areas
+through a **spatial link table** computed once at build time (`nsr-region-links`: from, to, relation,
+fraction of the smaller unit's area inside the larger). The link table is derived by cross-tabulating
+the two rasters, so it needs no new downloads and costs one pass over the cells.
+
+**Inheritance (revised 2026-09-18, BM): a place is judged by the polygons it lies IN.**
+Evidence attaches to the polygon it was recorded for, and is not transferred to other polygons:
+
+- an opinion about a polygon **containing** the place applies to it (POWO's finest statement about
+  Guadeloupe is "native in the Leeward Islands"), and among those any native opinion wins;
+- polygons **inside** the place describe only parts of it, so they are not its status. They answer
+  only when they all agree; otherwise the answer is `P`, the reason says the status varies, and
+  `n_subpolygons_native` / `n_subpolygons_introduced` give the split;
+- polygons that merely **overlap** describe neither the place nor its parts, and answer `UNK`;
+- `native_status_scope` records which of these produced the answer.
+
+Give coordinates and the question does not arise: the record is judged on the ground it sits on,
+which is how the garden pipeline will use it.
+
+**Endemism is the deliberate exception** (BM): `Ne` and `Ie` are claims about the taxon's whole
+range, not about one polygon, so they draw on evidence from elsewhere. A taxon confined to
+California, found in Michigan, is introduced there whatever Michigan's checklists say.
+
+The earlier design let a sub-polygon set the whole place's status, which made *Polycarpon
+tetraphyllum* native in the USA on one `powo:native` area against eleven `powo:introduced` ones.
+That is the failure this rule removes.
+while "introduced in Brazil Southeast" answers it as `I`. Nothing is invented and nothing is lost.
+
+## Resolution semantics
+
+For each row, after name and division resolution:
+
+1. Gather every checklist opinion for the taxon in the division (or a containing/contained one,
+   per the political rule), from all built sources. Opinions recorded against a taxon's own
+   infraspecifics are already part of it, having been rolled up at build time; opinions recorded
+   against a *higher* rank are not consulted (see "Higher ranks are not consulted", below).
+2. Reduce to one status per political level. **Precedence (BM, 2026-09-18): if any source says
+   native, the taxon is native there.** Native is the hardest claim for a checklist to make by
+   accident, whereas "introduced" and "absent" are often an artefact of a list's scope or age. Where
+   sources disagree, the answer is native and the disagreement is recorded rather than hidden, in
+   `native_status_conflict` (whether opinions differed) and `native_status_opinions` (the per-source
+   verdicts, e.g. `powo:native; usda:introduced`). This mirrors `TNRS_local()`'s `Source_conflict`.
+   Failing a native opinion, prefer the finer political level, then explicit `introduced` over
+   `present`, then a comprehensive source over a non-comprehensive one.
+3. Absence only yields `A`/`Ie` where a comprehensive source covers the division **and** the taxon
+   is evaluable, i.e. some source holds native-status information about it (see below).
+4. `native_status` is the status at the lowest division supplied, and `isIntroduced` is 1 for `I`
+   and `Ie`.
+
+New columns beyond the service, marked as extensions so output stays a superset:
+`poldiv_is_historical`, `resolved_in_successor` (which current division answered a historical one),
+and `native_status_basis` (the source and division that decided it).
+
+## Absence: when "not listed" is evidence
+
+**Decisions (BM, 2026-09-18): POWO counts as comprehensive for the species it covers; and a taxon
+about which no source holds any native-status information cannot be called Absent or Present at
+all.**
+
+The live service does not make the second distinction, and it matters. Probed 2026-09-18:
+
+| Query | Live NSR | Comment |
+|---|---|---|
+| *Sphagnum palustre*, California | **A**, "Absent from all checklists for region", sources powo, usda | a moss; POWO covers vascular plants only |
+| *Marchantia polymorpha*, California | **A**, same | a liverwort |
+| *Quercus robur*, California | I, sources powo, usda, weakley | correct |
+| *Pinus ponderosa*, Brazil | I, source flbr, `isCultivatedNSR` 1 | correct |
+| *Zea mays*, Mexico | N, sources mexico, powo | correct |
+
+A bryophyte in California is reported Absent, when the truth is that no consulted checklist has an
+opinion about bryophytes. Absence of evidence is returned as evidence of absence, and downstream
+that becomes `isIntroduced` through the `Ie` rule.
+
+**The rule (BM's formulation): go by which taxa the sources hold native-status information about,
+not by what the sources claim to cover.** A taxon is *evaluable* if it has at least one
+native-status record in any built source, anywhere in the world. Absence is interpretable only for
+evaluable taxa; for the rest the answer is `UNK`, with `native_status_reason` "no source holds
+native-status information for this taxon".
+
+Why this rather than a declared taxonomic scope per source:
+- **It does not need to know what a moss is.** *Sphagnum palustre* has no row in WCVP's
+  distributions, so it is not evaluable, and no `A` is emitted. The server never reasons about
+  taxonomic groups.
+- **It survives new and user-supplied sources.** A curated scope field would have to be written, and
+  kept right, for every source anyone adds. Coverage here is a property of the data, computed at
+  build time as the distinct taxa in `nsr-checklist`.
+- **It protects newly described species.** A species accepted in WCVP but with no distribution
+  records yet is not evaluable, so it returns `UNK` instead of being flagged Absent (and then
+  introduced) everywhere it is found. A declared-scope rule would get this wrong, since the species
+  is squarely inside POWO's stated scope.
+
+Implementation is one lookup: `evaluable(taxon_id)` is true when the taxon appears anywhere in
+`nsr-checklist`. No scope column, no taxonomic-group inference, and the semantics stay the same as
+sources are added: each new checklist can only widen the evaluable set.
+
+This is a deliberate divergence from the service, and one to report upstream: the same query that
+returns `A` there returns `UNK` here.
+
+## Validation
+
+1. **Against the live service**, on a stratified sample of BIEN and GBIF records (target ~50,000
+   rows spanning all four source regions and a tail of elsewhere): agreement on `native_status`,
+   a confusion matrix of codes, and disagreements broken down by source, division level and region.
+   The four-source build cannot match the service everywhere, so the honest target is: agreement
+   where only these four sources apply, and a measured, explained gap elsewhere.
+2. **Against BIEN's stored `is_introduced`** on the 284M-record BIEN pull, the same check we ran for
+   geovalidity and centroids.
+3. **Historical divisions**: records whose country no longer exists resolve and receive a status
+   through the successor, which the live service cannot do; report how many records that recovers.
+4. **Garden pipeline**: the occurrence layer currently keeps a record only if it falls in the
+   species' native WGSRPD level-3 range (POWO). Re-run with NSR and report how many cells change,
+   which is the concrete answer to "does the finer native status matter for the paper?"
+
+## Built, 2026-09-18 (milestones 1-3)
+
+Branch `local-implementation`: `R/local_cache.R`, `R/local_build.R`, `R/local_import.R`,
+`R/local_regions.R`, `R/NSR_local.R`, `tests/testthat/test-local-nsr.R` (24 expectations, all
+passing). Built against the shared GNRS/GVS cache in 24 minutes:
+
+| Source | Records | Taxa | Geography |
+|---|---|---|---|
+| powo | 1,986,877 | 443,168 | 375 WGSRPD level-3 areas |
+| flbr | 154,169 | 38,326 | 27 Brazilian states |
+| vascan | 25,459 | 6,086 | 13 Canadian provinces + Greenland |
+
+Plus 415 regions and **13,742 spatial links** (3,330 within, 3,330 contains, 6,684 overlaps).
+POWO needs no download (the WCVP archive is already in the TNRS cache) and no name resolution
+(WCVP is the backbone, so `accepted_plant_name_id` is the taxon id). The other two resolve through
+`TNRS_local()`: VASCAN matches 99.5% of names; FLBR 75%, the residue being the fungi and algae that
+a vascular-plant backbone cannot match - which is exactly the evaluability rule's case.
+
+### Reference cases
+
+| Query | Local NSR | Live service |
+|---|---|---|
+| *Pinus ponderosa*, US/California | N (powo) | N |
+| *Araucaria angustifolia*, Brazil/Sao Paulo | N (flbr + powo) | - |
+| *Araucaria angustifolia*, Brazil/Amazonas | A | - |
+| *Araucaria angustifolia*, by coordinates in Sao Paulo | N, resolved in both geographies | n/a |
+| *Acer saccharum*, Canada/New Brunswick | N (vascan + powo) | - |
+| *Zea mays*, Mexico | N (powo) | N |
+| *Welwitschia mirabilis*, Namibia | N (powo) | N |
+| *Sphagnum palustre* / *Marchantia polymorpha*, US/California | **UNK** | **A** |
+| *Quercus robur*, US/California | **A** | **I** (powo, usda, weakley) |
+
+The bryophyte divergence is the intended one. The *Quercus robur* divergence is not: **WCVP holds no
+California record for it at all** (22 introduced records elsewhere, none in CAL), so the service's
+answer came from USDA and Weakley. POWO's introduced ranges are thinnest exactly where USDA would
+cover, which is an argument for keeping `usda` in the first release rather than dropping it (see
+Open questions).
+
+### Bugs worth remembering
+
+- **Pooling administrative levels in the link table.** Aggregating cells to both level 1 and level 0
+  before computing fractions counted every cell twice, so California-the-state and CAL-the-WGSRPD-area
+  each looked half-inside the other and their relation degraded from `same` to `overlaps` - which
+  correctly demoted every native opinion to mere presence. Each level now has its own denominators.
+- **A state query inheriting its country's answer.** Keys for both levels sat in one pool, so
+  Amazonas inherited Brazil's "native". The finest named place now answers alone; the country key
+  only fills the country column.
+- **VASCAN's two status columns are the reverse of the obvious guess**: `occurrenceStatus` is
+  present/excluded/doubtful/irregular/absent, `establishmentMeans` is native/introduced.
+- **The GNRS backbone spells some divisions bilingually** ("New Brunswick/Nouveau-Brunswick") and
+  Newfoundland's HASC is `CA.NF`, not `CA.NL`.
+
+### Validation against the live service (milestone 4)
+
+`scratchpad/nsr_validate.R`: a stratified sample of species x division (POWO natives, POWO
+introduced, Brazilian states, Canadian provinces, and a tail of random species in random
+countries), sent to both. The service drops roughly a third of API batches whatever their size,
+so 150 of 300 rows were compared.
+
+**84% same decision** (78% identical code), against a service with four times as many sources:
+
+| local \ service | absent | introduced | native | present | unknown |
+|---|---|---|---|---|---|
+| absent | 6 | 4 | 0 | 0 | 0 |
+| introduced | 5 | 49 | 1 | 1 | 1 |
+| native | 9 | 1 | 71 | 0 | 1 |
+| present | 1 | 0 | 0 | 0 | 0 |
+
+By stratum: powo 86.7% (n=120), vascan 86.7% (n=15), flbr 60% (n=5), no-source tail 60% (n=10).
+
+Disagreement classes, all explained:
+- **local native, service absent (9).** Four cite no source at all (*Monanthotaxis schweinfurthii*
+  in DR Congo, *Galium spurium* in Syria, *Phleum* and *Trifolium spadiceum* in Russia); we cite
+  POWO v15. The service's POWO import is older, so these look like ours being more current.
+- **local introduced, service absent (5).** US and Australian cases where POWO records the species
+  as introduced in an area inside the country and we inherit that upward; the service answers
+  absent. Ours is the more informative reading of the same data.
+- **local absent, service introduced (4)** and **one each way on native/introduced**: sources we
+  lack (`usda`, `weakley`, `mab`, `fwi`, `conosur`) or POWO version differences.
+
+Two rule changes came out of reading the disagreements, both now in the code:
+- **Inheritance is symmetric.** Any native opinion about a related region makes it native (BM's
+  precedence); failing that, any introduced opinion makes it introduced. Inferred answers say so in
+  `native_status_reason`, and `native_status_opinions` records the relation.
+- **Links under 1% of a region are ignored**, so raster disagreement along a coastline carries
+  nothing. The `within` threshold is 0.95, not 0.99: these fractions come from 30 arc-second
+  rasters of two independently drawn coastlines, so coextensive regions score 0.987-0.990.
+
+Also added: **endemism** (`Ne`, `isEndemic`), read off the checklist - every region any source calls
+the taxon native lies inside the queried place. Verified on single-area natives, and correctly
+declining for *Araucaria angustifolia* (also Argentina, Paraguay) and *Welwitschia mirabilis* (also
+Angola). Note it is a statement about the checklists, not a conservation claim: *Oryza sativa* in
+China returns `Ne`, because POWO gives rice one native area. **Genus queries** are answered
+directly (WCVP records distributions for 14,126 genera); what is refused is resolving a bare genus
+onto some species of that genus.
+
+Bugs fixed this round, worth remembering: WCVP carries the same name at two ranks (a *variety* row
+also called "Pinus ponderosa"), and only one holds the distributions, so names index to the
+species-rank id that has opinions; and accepted infraspecific taxa (60,278 of them) keep their own
+distributions, so they are rolled up to their species through `parent_plant_name_id`.
+
+### Status model completed: `Ie` (2026-09-18)
+
+Absence becomes introduction only where the taxon could not have been native: its whole native
+range lies elsewhere **and is confined**, to one region or to one country. A widely native species
+merely unrecorded here stays `A`, because absence alone is not evidence of introduction.
+
+| Query | Answer | Why |
+|---|---|---|
+| *Lepechinia calycina* (endemic to California) in Michigan | `Ie`, isIntroduced 1 | it could not be native there |
+| *Lepechinia calycina* in California | `Ne`, isEndemic 1 | native and confined to it |
+| *Homonoia* (widely native) in Michigan | `A` | absent, but could have been native |
+| *Washingtonia* (California, Arizona, NW Mexico) in Michigan | `A` | native range not confined |
+
+Final validation: **85.3% same decision** (79.3% identical code) on 150 compared rows; by stratum
+powo 86.7%, vascan 86.7%, no-source tail 80%, flbr 60% (n=5). `Ie` moved four records from absent
+to introduced.
+
+The nine remaining "we native, service absent" rows were checked individually and are not artefacts
+of loose inheritance: *Monanthotaxis schweinfurthii* in DR Congo is a **stated** `powo:native` while
+the service returns absent citing nothing; *Galium spurium* in Syria is native within WGSRPD's
+Lebanon-Syria unit, POWO's finest statement about Syria; *Trifolium spadiceum* is native in a
+Russian sub-area. Over 1,500 country-level queries, inferred answers come through `within` and
+`contains`, not weak overlaps.
+
+**One consequence of the any-native precedence worth stating in the write-up:** at country level a
+single native area carries the whole country. *Polycarpon tetraphyllum* in the USA answers `N` from
+one `powo:native(contains)` against eleven `powo:introduced(contains)`. The counts are visible in
+`native_status_opinions`, and the alternative (majority, or requiring a stated opinion) would lose
+the cases the rule exists for, but users filtering on `native_status` at country level should know
+it.
+
+### Validation under the containment rule
+
+**82.0% same decision** (76.7% identical code), against 85.3% under the old inheritance. The whole
+drop is the deliberate change: 9 rows where we now answer `P` (the status varies among the polygons
+inside the queried place) while the service commits to one, and 2 where only partly overlapping
+polygons have an opinion and we answer `UNK`. Excluding those eleven, agreement is ~89%. Nothing
+accidental changed: the same `usda`/`weakley` gaps and POWO-version differences remain.
+
+By stratum: powo 82.5%, vascan 86.7%, no-source tail 80%, flbr 60% (n=5).
+
+This is a case where agreement with the service is the wrong target. The service propagates a
+sub-region's status upward; we decline to, and say so. For records with coordinates - every record
+in the garden pipeline - the two approaches coincide, because the point lies inside exactly one
+polygon per system.
+
+### Higher ranks are not consulted (BM, 2026-09-24)
+
+The service's taxonomic rule has two halves and they are not the same claim.
+
+Upward is sound and **is** implemented, at build time in `nsr_species_of()`: a distribution recorded
+against an accepted infraspecific taxon is rolled up to its species, because a subspecies of *X*
+occurring natively in a region means *X* occurs natively in that region. The subspecies is an *X*.
+Roughly 60k WCVP distribution rows hang off infraspecifics and reach the species this way.
+
+Downward from a higher rank is **not** implemented, and deliberately. "With no opinion for the
+taxon, higher ranks are consulted" means answering about a species from its genus's range, and
+knowing that a genus is native to a region tells us nothing about any particular species in it:
+that is inferring a property of a part from a property of the whole - the ecological fallacy, in
+its division form. A genus native to Quebec means *some* species of it is native to Quebec, and the
+species in hand is as likely to be the introduced one. Native is the strongest claim the resolver
+can make, and making it at genus level would manufacture it at scale for exactly the taxa about
+which the sources are silent.
+
+So a species with no opinion of its own is `UNK` (or `A` where a comprehensive source covers the
+place), never native-by-genus. `taxon_evaluable` says which it is. A genus *query* is still answered
+directly, because WCVP records distributions against genera and those rows are kept keyed on the
+genus - that is answering the question asked, not inferring downward from it.
+
+**This is a known and permanent divergence from the service**, and belongs with the `wcvp`/`powo`
+source-name difference in any validation against `NSR()`: where the live service returns a
+genus-derived status for an unlisted species, we return `UNK` or `A`.
+
+## Milestones
+
+1. Cache scaffolding, `nsr-sources`, and the POWO importer from WCVP (no download); `NSR_local()`
+   returning service-shaped output for country-level queries. Validates against the live API on
+   country-level rows.
+2. USDA, VASCAN and Flora do Brasil importers, with the state-level resolution and the political
+   propagation rules.
+3. Taxonomic roll-up of infraspecifics, endemism (`Ie`) and conflict handling. (`is_cultivated_taxon`
+   was dropped - see open question 3; consulting higher ranks was declined - see "Higher ranks are
+   not consulted".)
+4. Validation 1-3; write-up.
+5. Garden re-run (validation 4) and, separately, the remaining eight sources.
+
+## Open questions for BM
+
+1. **Conflict precedence.** Is the proposal above (finer division, explicit over present,
+   comprehensive over not, all opinions recorded) how the service behaves? If BM has the service's
+   SQL to hand it is faster than inferring it from probes.
+2. **POWO as comprehensive.** Treating WCVP's absence as evidence of non-nativeness globally is what
+   makes `A`/`Ie` possible at all outside the three national checklists. Reasonable, or too strong?
+3. Settled (BM, 2026-09-19): **`is_cultivated_taxon` is dropped.** It was already dead in
+   production - filled from table `cultspp`, loaded from `cultspp_staging` only
+   `if (exists_table(...))`, with no importer in the public repo populating it; probes return
+   `is_cultivated_taxon = 0` for both *Zea mays* and *Triticum aestivum*. The candidate
+   replacements were examined 2026-09-19 and none is worth carrying.
+
+   **Why dropped rather than substituted (BM):** a use flag is not much use without a polygon
+   attached to it. Everything else NSR reports is an assertion about a taxon *in a place*, and
+   the containment semantics exist precisely so that polygon-level evidence is not spread to
+   other polygons. "Grown somewhere by someone" is not evidence that the record in hand was
+   planted, so a taxon-level flag would either sit unused or be read as if it were
+   place-specific. `isCultivatedNSR` already answers the version of the question that has a
+   geography, and is unaffected by this: it comes from checklists carrying a cultivated status
+   (FLBR flagged *Pinus ponderosa* in Brazil), reproduced from the same source field.
+
+   What was examined, so it need not be examined again:
+
+   - **Kew's World Checklist of Useful Plant Species** (Diazgranados et al. 2020, 40,292 species,
+     CC BY 4.0) is published only as an 11.3 MB PDF plus its EML (KNB, doi:10.5063/F1CV4G34) -
+     no data table. (The DOI in the earlier draft of this note, 10.34885/172, was wrong; that is
+     the State of the World's Plants and Fungi report.) **The PDF was parsed successfully on
+     2026-09-19** - it encodes rank in the font, so the hierarchy is recoverable exactly - and the
+     result is `garden_variety_traits/data/nsr_sources/wcups_2020.csv`, produced by
+     `R_scripts/48_parse_wcups_pdf.py`. It reproduces every published control total (40,292
+     records; 3 kingdoms / 6 phyla / 14 classes / 101 orders / 433 families / 6,737 genera;
+     40,239 LSIDs; all ten use-category counts; the top five families and genera; 70 species with
+     all ten uses; 91 families and 2,790 genera with a single species). So availability is no
+     longer the obstacle - but the data are still ten use categories per taxon with **no
+     geography**, so the reason for dropping the flag is unchanged.
+   - **GRIN Taxonomy** (CC0, COLDP, `hosted-datasets.gbif.org/datasets/grin.zip`, 10.7 MB) does
+     carry economic uses in `TaxonProperty.tsv`: 17,111 taxa with a use, 9,836 in
+     cultivation-like classes (ornamental 9,155, plus food, forage, materials, fuel). Machine-
+     readable and usable - but taxon-level with **no geography at all**, which is the objection
+     above. Note too that economic use is not cultivation: the largest class after ornamental is
+     folklore medicine (4,812), largely wild-harvested.
+   - **Wiersema & Leon's World Economic Plants** has the same taxon-level-only shape.
+
+   GRIN's region-level `cultivated` distribution status was checked as a way to give the flag a
+   geography, and is not fit for it: 1,437 taxa worldwide, 5,260 rows, 229 ISO3 countries and
+   nothing finer. Coverage is not the heavily cultivated combinations - *Triticum aestivum*,
+   *Oryza sativa*, *Glycine max* and *Malus domestica* have no distribution rows at all, and
+   *Zea mays*, *Manihot esculenta*, *Sorghum bicolor*, *Vitis vinifera* and *Theobroma cacao*
+   carry native ranges with zero cultivated entries, while *Hordeum vulgare* has 91. The reason
+   is that GRIN's distribution field documents germplasm provenance, not cultivation extent, so a
+   cultigen with no natural range often gets nothing; country totals (China 528, US 348, India
+   299, Europe thin) track collecting effort rather than area under cultivation. The GBIF COLDP
+   export also flattens GRIN's sub-country geography onto ISO3 - visible as 71 identical
+   `iso:RUS` rows for a single taxon, one per Russian region - so even where it has content the
+   resolution is country-level.
+
+   In the code: `NSR_local()` returns `is_cultivated_taxon = NA_integer_` rather than production's
+   `0`, so the column keeps its place in the service-shaped output without a never-populated
+   field being read as a real negative. `NSR_local_by_region()` omits it.
+
+4. Settled (BM, 2026-09-18): for the garden paper NSR is a **robustness check** on the existing
+   WCVP native-range filter, not a gate on the occurrence layer, to be revisited if validation 4
+   shows it makes a material difference.
+5. Settled: sources downloaded 2026-09-18 to `garden_variety_traits/data/nsr_sources/` (gitignored):
+   usda.zip (2.1 MB), VASCAN DwC-A, Flora do Brasil DwC-A. POWO comes from the WCVP v15 files
+   already in the repo's data directory.
+6. **GRIN as a fourth source?** Open. Examining GRIN for the cultivated flag turned up
+   something more useful than the flag: `Distribution.tsv` holds 579,184 native/introduced rows
+   for 65,196 taxa across 229 countries, CC0, independent of POWO. It is country-level only (the
+   COLDP export flattens GRIN's sub-country geography onto ISO3), so it does not recover the
+   per-state US resolution that `usda` would have given, and the documented POWO-alone US skew
+   towards `A` stands either way. `iso:` maps directly onto the `gadm0:` keys the link table
+   already carries. Awaiting BM's call; not built.
+
+7. Settled: BM cloned `EnquistLab/RNSR` on 2026-09-19; work is on branch
+   `local-implementation` and this note lives in the clone.
diff --git a/man/NSR_local.Rd b/man/NSR_local.Rd
new file mode 100644
index 0000000..20d5889
--- /dev/null
+++ b/man/NSR_local.Rd
@@ -0,0 +1,132 @@
+% Generated by roxygen2: do not edit by hand
+% Please edit documentation in R/NSR_local.R
+\name{NSR_local}
+\alias{NSR_local}
+\title{Determine native status without an internet connection}
+\usage{
+NSR_local(
+  occurrence_dataframe,
+  dir = nsr_cache_dir(),
+  resolve_names = TRUE,
+  use_coordinates = FALSE,
+  min_overlap = 0.01,
+  exclude_extinct = TRUE,
+  quiet = FALSE
+)
+}
+\arguments{
+\item{occurrence_dataframe}{A data.frame with \code{species}, and either coordinates
+or political division names (see Place).}
+
+\item{dir}{Cache directory, shared with GNRS and GVS.}
+
+\item{resolve_names}{Resolve submitted names against WCVP with
+\code{TNRS::TNRS_local()}?  Names already matching WCVP accepted names need none.}
+
+\item{use_coordinates}{Place each record by its \code{latitude} and \code{longitude}
+as well as by its division names?  Default \code{FALSE}: coordinates cost a raster
+lookup per call, and resolving them to political divisions is GVS's job.  See Place.}
+
+\item{min_overlap}{Ignore region links covering less than this share of a region.}
+
+\item{exclude_extinct}{Ignore distribution records the source marks extinct?  Default
+\code{TRUE}, which answers about the present day.  Set \code{FALSE} to model a past
+distribution, where a region the taxon has since been lost from still counts.  See
+Extinct records.}
+
+\item{quiet}{Suppress progress messages?}
+}
+\value{
+A data.frame, one row per input row, carrying \code{user_id} from the input
+  (or sequential ids where the input has none), as \code{\link{NSR}} does.
+}
+\description{
+Offline Native Status Resolver.  For each taxon and place, the status every built
+checklist gives it, reduced to one answer.
+}
+\details{
+Output carries \code{\link{NSR}}'s columns, and adds: \code{isEndemic};
+\code{native_status_conflict} and \code{native_status_conflict_type}, saying whether
+and how the sources disagreed; \code{native_status_scope}, which of the polygons
+bearing on the place produced the answer; \code{n_subpolygons_native} and
+\code{n_subpolygons_introduced}; \code{native_status_opinions}, every opinion consulted;
+\code{native_status_coordinates}, \code{native_status_names} and \code{place_conflict}
+(see Place); \code{regions_matched}, what the place was matched on; and
+\code{taxon_evaluable}, whether any source holds native status for the taxon at all -
+which is what separates \code{A} (absent) from \code{UNK} (no data).  The coordinates
+given are echoed as \code{latitude} and \code{longitude}.
+
+\strong{Place.}  By default the answer comes from political division names
+(\code{country}, \code{state_province}, \code{county_parish}), which are resolved to
+GADM units through the GNRS backbone and carried to each source's own geography by the
+spatial link table.  Any \code{latitude} and \code{longitude} present are echoed but
+not used: placing every record by point costs a raster lookup per call, and turning
+coordinates into political divisions is what GVS already does.  Resolve them there and
+pass the divisions here, or set \code{use_coordinates = TRUE} to have this function do
+it.
+
+With \code{use_coordinates = TRUE}, a record is placed by its point in each source's
+own geography (WCVP against WGSRPD level-3 areas, VASCAN and Flora do Brasil against
+GADM states), with no crosswalk between them - the finer question, where the data
+supports it.  Given coordinates and names together, the two are independent claims
+about where the record is and either can be wrong: a transposed longitude and a
+mistyped province are equally easy mistakes, so neither is authoritative.  Where they
+agree, both are used.  Where they disagree, \code{place_conflict} is \code{TRUE},
+\code{native_status} is \code{UNK}, and the two bases are answered separately in
+\code{native_status_coordinates} and \code{native_status_names} for you to judge.  The
+service has no combined mode to follow here: \code{\link{NSR}} takes names and
+\code{\link{NSR_from_coordinates}} takes coordinates.
+
+\strong{Precedence.}  If any source says native, the answer is native: asserting
+nativity is a positive claim, whereas "introduced" and "absent" are often artefacts of
+a list's scope, age or purpose.  Disagreement is recorded, not hidden, in
+\code{native_status_conflict} and \code{native_status_opinions}.
+
+\strong{Rank.}  A distribution recorded against an accepted infraspecific taxon counts
+for its species: a subspecies of \emph{X} native somewhere means \emph{X} is native
+there.  The reverse is not done.  A species with no opinion of its own is not given its
+genus's status, because a genus being native to a region says nothing about which of its
+species are - it is the introduced one as readily as the native one.  Such a species is
+\code{UNK}, or \code{A} where a comprehensive source covers the place;
+\code{taxon_evaluable} distinguishes them.  A genus \emph{query} is still answered
+directly, from the distributions WCVP records against genera.  \code{\link{NSR}} does
+consult higher ranks, so the two differ here by design.
+
+\strong{Which polygons answer.}  A place is judged by the polygons it lies IN.  An
+opinion about a polygon containing the place applies to it (POWO's finest statement
+about Guadeloupe is "native in the Leeward Islands"), and among those any native
+opinion wins.  Polygons INSIDE the place describe only parts of it, so they are not its
+status: they answer only when they all agree, and otherwise the answer is \code{P} with
+the reason saying the status varies and how many sub-polygons say what
+(\code{n_subpolygons_native}, \code{n_subpolygons_introduced}).  Give coordinates and
+the question does not arise: the record is judged on the ground it sits on.
+\code{native_status_scope} records which of these produced the answer.
+
+\strong{Endemism is the exception}, deliberately: \code{Ne} and \code{Ie} are claims
+about the taxon's whole range rather than about one polygon, so they draw on evidence
+from elsewhere.  A record of a taxon confined to California, found in Michigan, is
+introduced there however little Michigan's own checklists say.
+}
+\section{Extinct records}{
+
+WCVP marks a distribution extinct with a bare flag and no date, so the record says
+only "considered no longer present as of this release".  Whether it should count is
+therefore a property of the question rather than of the data: a present-day status
+wants it out, a historical distribution wants it in, and the cache keeps both so
+\code{exclude_extinct} can decide per call.
+
+Excluding it never makes the taxon introduced there.  The endemism rules read the
+taxon's whole native range, extinct records included, whatever this argument says -
+a region a taxon has been lost from is still a region it was native to, so \code{Ie}
+("introduced, inferred from endemism elsewhere") cannot fire against it.  The answer
+for such a record is \code{A}: gone, not foreign.
+}
+
+\section{County and parish}{
+
+\code{county_parish} is accepted and echoed, but not resolved: the political-division
+backbone stops at state and province, so \code{native_status_county_parish} is always
+\code{NA} and the answer is given at the finest level that could be matched.  A
+warning says so.  Coordinates are the way to ask a finer question.
+}
+
diff --git a/man/NSR_local_build.Rd b/man/NSR_local_build.Rd
new file mode 100644
index 0000000..18e9cf7
--- /dev/null
+++ b/man/NSR_local_build.Rd
@@ -0,0 +1,55 @@
+% Generated by roxygen2: do not edit by hand
+% Please edit documentation in R/local_build.R
+\name{NSR_local_build}
+\alias{NSR_local_build}
+\title{Build the local native-status reference}
+\usage{
+NSR_local_build(
+  sources = "wcvp",
+  dir = nsr_cache_dir(create = TRUE),
+  files = list(),
+  overwrite = FALSE,
+  quiet = FALSE
+)
+}
+\arguments{
+\item{sources}{Which sources to build.  See \code{NSR_local_status()}.}
+
+\item{dir}{Cache directory, shared with GNRS and GVS.}
+
+\item{files}{Named list of local archives to read, e.g.
+\code{list(flbr = "flbr_dwca.zip")}.  Required for \code{vascan} and \code{flbr}.
+For \code{wcvp}, the WCVP zip; if absent, the copy in the TNRS cache is used when
+there is one.}
+
+\item{overwrite}{Rebuild sources that are already built?}
+
+\item{quiet}{Suppress progress messages?}
+}
+\value{
+Invisibly, \code{NSR_local_status()}.
+}
+\description{
+Reads each checklist source, resolves its names against WCVP with
+\code{TNRS::TNRS_local()} and its political divisions against the GNRS backbone, and
+writes the shared cache tables.  Sources are acquired on the user's machine; nothing
+derived ships with the package.
+}
+\details{
+\strong{The archives are not downloaded for you.}  \code{vascan} and \code{flbr} must
+be supplied through \code{files}, and each stops with the URL to fetch if they are
+not; \code{wcvp} is the exception, since it reads the WCVP archive already in the
+TNRS cache when there is one.  This is why \code{sources} defaults to \code{"wcvp"}
+alone: it is the only source that can build unattended.
+
+\code{"powo"} is accepted as a synonym for \code{"wcvp"} - it is the name the live
+service reports for the same Kew dataset - but \code{wcvp} is what gets written to
+\code{native_status_sources}.
+
+Tables written: \code{nsr-sources} (one per source, with licence and whether it is
+comprehensive), \code{nsr-taxa} (one per taxon, keyed on the WCVP accepted id),
+\code{nsr-regions} (one per region, in the geography its source publishes against) and
+\code{nsr-checklist} (one per taxon x region x source).  \code{nsr-region-links}, the
+spatial relations between geographies, is built too when the GADM index from GVS is in
+the cache.
+}
diff --git a/man/NSR_local_by_region.Rd b/man/NSR_local_by_region.Rd
new file mode 100644
index 0000000..58befab
--- /dev/null
+++ b/man/NSR_local_by_region.Rd
@@ -0,0 +1,44 @@
+% Generated by roxygen2: do not edit by hand
+% Please edit documentation in R/NSR_local_set.R
+\name{NSR_local_by_region}
+\alias{NSR_local_by_region}
+\title{Native status for taxa and polygons you have already resolved}
+\usage{
+NSR_local_by_region(
+  taxon_id,
+  region_keys,
+  country_keys = NULL,
+  dir = nsr_cache_dir(),
+  min_overlap = 0.01,
+  exclude_extinct = TRUE
+)
+}
+\arguments{
+\item{taxon_id}{WCVP accepted ids (character), one per query.}
+
+\item{region_keys}{The polygons each record sits in, as a list of character vectors
+(\code{"wgsrpd3:BZL"}, \code{"gadm1:BRA.25_1"}), or a single character vector for
+one polygon each.}
+
+\item{country_keys}{Optional country polygons (\code{"gadm0:BRA"}), same shape.}
+
+\item{dir}{Cache directory.}
+
+\item{min_overlap}{Ignore region links covering less than this share of a region.}
+
+\item{exclude_extinct}{Ignore distribution records the source marks extinct?  See
+\code{\link{NSR_local}}.}
+}
+\value{
+A data.frame with \code{native_status} and the status columns
+  \code{NSR_local()} returns, keyed on \code{taxon_id}.  It does not echo the input
+  columns \code{NSR_local()} carries through (\code{species}, the division names,
+  \code{user_id}), since this entry point is given ids and polygon keys rather than
+  a record.
+}
+\description{
+A pipeline that has located its records itself - in the WGSRPD raster and the GADM
+index, as an occurrence workflow does once for all of its coordinates - already knows
+the polygons each record sits in.  This takes those directly, skipping name resolution
+and point-in-polygon, and answers with the same rules as \code{\link{NSR_local}}.
+}
diff --git a/man/NSR_local_remove.Rd b/man/NSR_local_remove.Rd
new file mode 100644
index 0000000..c3adef8
--- /dev/null
+++ b/man/NSR_local_remove.Rd
@@ -0,0 +1,19 @@
+% Generated by roxygen2: do not edit by hand
+% Please edit documentation in R/local_cache.R
+\name{NSR_local_remove}
+\alias{NSR_local_remove}
+\title{Remove the local NSR tables}
+\usage{
+NSR_local_remove(dir = nsr_cache_dir(), quiet = FALSE)
+}
+\arguments{
+\item{dir}{Cache directory.}
+
+\item{quiet}{Suppress messages?}
+}
+\value{
+Invisibly, the files removed.
+}
+\description{
+Remove the local NSR tables
+}
diff --git a/man/NSR_local_status.Rd b/man/NSR_local_status.Rd
new file mode 100644
index 0000000..9b63763
--- /dev/null
+++ b/man/NSR_local_status.Rd
@@ -0,0 +1,18 @@
+% Generated by roxygen2: do not edit by hand
+% Please edit documentation in R/local_cache.R
+\name{NSR_local_status}
+\alias{NSR_local_status}
+\title{Which parts of the local NSR are built}
+\usage{
+NSR_local_status(dir = nsr_cache_dir())
+}
+\arguments{
+\item{dir}{Cache directory.  Defaults to the one shared with GNRS and GVS.}
+}
+\value{
+A data.frame with one row per source: whether it is built, how many checklist
+  records and taxa it contributed, and when it was built.
+}
+\description{
+Which parts of the local NSR are built
+}
diff --git a/tests/testthat/test-local-altdiv.R b/tests/testthat/test-local-altdiv.R
new file mode 100644
index 0000000..3e21366
--- /dev/null
+++ b/tests/testthat/test-local-altdiv.R
@@ -0,0 +1,47 @@
+context("divisions that belong to another division system")
+
+# Self-contained: the tables GNRS writes to the shared cache, in a temp directory.
+
+dir <- file.path(tempdir(), "nsr-altdiv")
+unlink(dir, recursive = TRUE)
+dir.create(dir, recursive = TRUE, showWarnings = FALSE)
+w <- function(x, table) nanoparquet::write_parquet(x, file.path(dir, paste0("altdiv-", table, ".gz.parquet")),
+                                                   compression = "gzip")
+w(data.frame(system = c("no-fylke", "se-landskap", "gb-vice-county"),
+             entity_key = c("NO-VIKEN", "SE-LS-Uppland", "GB-VC"),
+             country_iso = c("NO", "SE", "GB"),
+             name = c("Viken", "Uppland", "Watsonian vice-county"),
+             kind = c("superseded", "parallel", "parallel"),
+             extent_known = c(TRUE, FALSE, FALSE), stringsAsFactors = FALSE), "units")
+w(data.frame(entity_key = c("NO-VIKEN", "SE-LS-Uppland", "GB-VC"),
+             name = c("Viken", "Uppland", "^VC ?[0-9]+"),
+             match = c("exact", "exact", "regex"), stringsAsFactors = FALSE), "names")
+w(data.frame(entity_key = "NO-VIKEN", level = 1L,
+             gid = c("NOR.1_1", "NOR.4_1", "NOR.2_1"), stringsAsFactors = FALSE), "extent")
+
+test_that("a division with a known extent contributes the units it covers", {
+  k <- nsr_altdiv_keys("NO", "Viken", dir)
+  expect_equal(k[[1]], c("gadm1:NOR.1_1", "gadm1:NOR.4_1", "gadm1:NOR.2_1"))
+})
+
+test_that("a division whose extent is unknown contributes nothing", {
+  # recognised by GNRS, but there is no set of GADM units to ask about yet, so the
+  # query stays where it was rather than being answered at the wrong level
+  expect_equal(nsr_altdiv_keys("SE", "Uppland", dir)[[1]], character(0))
+  expect_equal(nsr_altdiv_keys("GB", "VC57 Derbyshire", dir)[[1]], character(0))
+})
+
+test_that("an ordinary GADM division is left alone, and matching is per country", {
+  expect_equal(nsr_altdiv_keys("NO", "Akershus", dir)[[1]], character(0))
+  # "Viken" is a Norwegian entity: the same string under another country is not it
+  expect_equal(nsr_altdiv_keys("SE", "Viken", dir)[[1]], character(0))
+  expect_equal(nsr_altdiv_keys(NA_character_, NA_character_, dir)[[1]], character(0))
+})
+
+test_that("without the component built, nothing is claimed", {
+  empty <- file.path(tempdir(), "nsr-altdiv-empty")
+  unlink(empty, recursive = TRUE)
+  dir.create(empty, recursive = TRUE, showWarnings = FALSE)
+  expect_null(nsr_altdiv(empty))
+  expect_equal(nsr_altdiv_keys("NO", "Viken", empty)[[1]], character(0))
+})
diff --git a/tests/testthat/test-local-nsr-paths.R b/tests/testthat/test-local-nsr-paths.R
new file mode 100644
index 0000000..44d12a6
--- /dev/null
+++ b/tests/testthat/test-local-nsr-paths.R
@@ -0,0 +1,302 @@
+# NSR_local() switches from the row resolver to the data.table set resolver at 200 rows
+# (option NSR.set_min).  That switch must be invisible: the same query has to come back
+# the same either way.  It has not been, twice - the set path never filled the per-level
+# columns, and it reported a cultivated flag of 0 where the row path reports NA - so this
+# compares the two directly, on tables held in memory rather than a built cache.
+
+# a small world: Quebec in two geographies, inside Canada, plus a sliver of a region the
+# query only just touches
+fixture_db <- function() {
+  # one row per taxon x region x source; is_extinct only ever set by WCVP
+  spec <- c(
+    "t1 wgsrpd3:QUE    native     wcvp   0 0",
+    "t2 gadm0:CAN      introduced vascan 0 0",
+    "t3 wgsrpd3:ONT    native     wcvp   0 0",
+    "t3 wgsrpd3:BZL    native     wcvp   0 0",
+    "t4 wgsrpd3:QUE    present    wcvp   1 0",
+    "t4 gadm1:CAN.11_1 native     vascan 0 0",
+    "t5 gadm1:CAN.11_1 native     vascan 0 0",
+    "t5 gadm1:CAN.4_1  native     vascan 0 0",
+    # t6 was native in Quebec and has been lost from it; it survives only in Brazil
+    "t6 wgsrpd3:QUE    native     wcvp   0 1",
+    "t6 wgsrpd3:BZL    native     wcvp   0 0",
+    # t7 is the contrast: native ONLY in Brazil, never recorded in Quebec at all
+    "t7 wgsrpd3:BZL    native     wcvp   0 0",
+    # t8 is native to a country row AND to one of that country's states
+    "t8 gadm0:CAN      native     vascan 0 0",
+    "t8 gadm1:CAN.11_1 native     vascan 0 0")
+  f <- do.call(rbind, strsplit(trimws(spec), "[[:space:]]+"))
+  checklist <- data.frame(
+    taxon_id = f[, 1], region_key = f[, 2], status = f[, 3], source_name = f[, 4],
+    is_cultivated = as.integer(f[, 5]), is_extinct = as.integer(f[, 6]),
+    stringsAsFactors = FALSE)
+  links <- data.frame(
+    from_region = c("gadm1:CAN.11_1", "wgsrpd3:QUE",    "gadm1:CAN.11_1"),
+    to_region   = c("wgsrpd3:QUE",    "gadm1:CAN.11_1", "wgsrpd3:SLIVER"),
+    relation    = c("same",           "same",           "within"),
+    fraction    = c(0.99,             0.99,             0.001),
+    stringsAsFactors = FALSE)
+  taxa <- data.frame(
+    taxon_id = paste0("t", 1:8),
+    species_name = paste("Sp", 1:8),
+    family = "Fam", genus = "Gen", rank = "species", stringsAsFactors = FALSE)
+  sources <- data.frame(source_name = c("wcvp", "vascan"), is_comprehensive = TRUE,
+                        stringsAsFactors = FALSE)
+  regions <- data.frame(
+    region_key = c("wgsrpd3:QUE", "wgsrpd3:ONT", "wgsrpd3:BZL", "wgsrpd3:SLIVER",
+                   "gadm1:CAN.11_1", "gadm1:CAN.4_1", "gadm0:CAN"),
+    region_name = c("Quebec", "Ontario", "Brazil S", "Sliver", "Quebec",
+                    "New Brunswick", "Canada"),
+    level = c("state_province", "state_province", "state_province", "state_province",
+              "state_province", "state_province", "country"),
+    stringsAsFactors = FALSE)
+  regions$system <- sub(":.*", "", regions$region_key)
+  NSR:::nsr_index_db(list(sources = sources, taxa = taxa, checklist = checklist,
+                          regions = regions, links = links))
+}
+
+# every query names Quebec, inside Canada
+fixture_places <- function(n_taxa) {
+  list(fine = rep(list("gadm1:CAN.11_1"), n_taxa),
+       country = rep(list("gadm0:CAN"), n_taxa))
+}
+
+test_that("the row path and the set path agree, column for column", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  tid <- c("t1", "t2", "t3", "t4", NA_character_)
+  places <- fixture_places(length(tid))
+
+  row <- NSR:::nsr_resolve_status(tid, places, db)
+  set <- NSR:::nsr_resolve_status_set(tid, places, db)
+
+  for (f in c("code", "country_code", "state_code", "scope", "conflict", "conflict_type",
+              "cultivated", "n_sub_native", "n_sub_introduced", "sources", "reason")) {
+    expect_equal(set[[f]], row[[f]], info = f)
+  }
+})
+
+test_that("the set path fills the per-level codes", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  tid <- c("t1", "t2")
+  set <- NSR:::nsr_resolve_status_set(tid, fixture_places(2), db)
+  # native in Quebec via the WGSRPD twin; nothing said about Canada as a whole
+  expect_equal(set$state_code[1], "N")
+  expect_false(is.na(set$country_code[1]))
+  # introduced at country level propagates down to the state answer
+  expect_equal(set$country_code[2], "I")
+})
+
+test_that("a level that was not asked about has no code", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  places <- list(fine = list(character(0)), country = list("gadm0:CAN"))
+  for (res in list(NSR:::nsr_resolve_status("t2", places, db),
+                   NSR:::nsr_resolve_status_set("t2", places, db))) {
+    expect_true(is.na(res$state_code[1]))
+    expect_equal(res$country_code[1], "I")
+  }
+})
+
+test_that("absence leaves the cultivated flag unknown in both paths", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  # t3 is evaluable (it is native in Ontario and in Brazil, so its range is not confined
+  # and the Ie rule does not fire) and Quebec is comprehensively listed, so a missing
+  # opinion here is a real absence rather than a gap
+  places <- fixture_places(1)
+  row <- NSR:::nsr_resolve_status("t3", places, db)
+  set <- NSR:::nsr_resolve_status_set("t3", places, db)
+  expect_equal(row$code, "A")
+  expect_equal(set$code, "A")
+  expect_true(is.na(row$cultivated))
+  expect_true(is.na(set$cultivated))
+})
+
+# GADM keys carry their own hierarchy (BRA.25_1 sits in BRA).  Without it, a country
+# query cannot see checklist rows published against that country's states, which is how
+# VASCAN and Flora do Brasil publish - so native-up propagation and confined-country Ie
+# both fail at country level.
+test_that("a country query sees the states inside it", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  places <- list(fine = list(character(0)), country = list("gadm0:CAN"))
+  row <- NSR:::nsr_resolve_status("t4", places, db)
+  set <- NSR:::nsr_resolve_status_set("t4", places, db)
+  # t4's only Canadian opinion is VASCAN's, recorded against Quebec the GADM state.
+  # Without the hierarchy the country query cannot see it at all and answers A; with it
+  # the state answers, and since Quebec is t4's only native region it is endemic there.
+  expect_equal(row$code, "Ne")
+  expect_equal(set$code, "Ne")
+  expect_equal(row$scope, "sub-polygons agree")
+  expect_equal(set$scope, "sub-polygons agree")
+  expect_equal(set$sources, row$sources)
+})
+
+test_that("the hierarchy is exact, so min_overlap cannot filter it out", {
+  db <- fixture_db()
+  kids <- NSR:::nsr_gadm_children("gadm0:CAN", db)
+  expect_setequal(kids, c("gadm1:CAN.11_1", "gadm1:CAN.4_1"))
+  expect_equal(NSR:::nsr_gadm_parents("gadm1:CAN.4_1", db), "gadm0:CAN")
+  # a state is a tiny share of its country, so as an overlap fraction it would be
+  # dropped; the parent/child edges carry no fraction and are not filtered
+  expect_setequal(NSR:::nsr_gadm_children("gadm0:CAN", db), kids)
+  # regions with no checklist rows are not children of anything
+  expect_equal(NSR:::nsr_gadm_children("gadm0:BRA", db), character(0))
+})
+
+test_that("a range confined to one country's states is confined to that country", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  # t5 is native in Quebec and New Brunswick and nowhere else; asked about a Brazilian
+  # region that is comprehensively listed, it is absent there and endemic to Canada
+  places <- list(fine = list("wgsrpd3:BZL"), country = list(character(0)))
+  row <- NSR:::nsr_resolve_status("t5", places, db)
+  set <- NSR:::nsr_resolve_status_set("t5", places, db)
+  expect_equal(row$code, "Ie")
+  expect_equal(set$code, "Ie")
+  expect_match(row$reason, "Canada")
+  expect_match(set$reason, "Canada")
+})
+
+# WCVP's extinct flag carries no date, so whether it counts belongs to the question
+# rather than to the build: the cache keeps the row and the query decides.
+test_that("exclude_extinct decides whether a lost population still answers", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  places <- list(fine = list("gadm1:CAN.11_1"), country = list(character(0)))
+  for (f in list(NSR:::nsr_resolve_status, NSR:::nsr_resolve_status_set)) {
+    gone <- f("t6", places, db, 0.01, exclude_extinct = TRUE)
+    past <- f("t6", places, db, 0.01, exclude_extinct = FALSE)
+    # present day: it is not there any more
+    expect_equal(gone$code, "A")
+    # as a historical distribution: it was native there
+    expect_equal(past$code, "N")
+  }
+})
+
+test_that("a lost population is absent, never introduced", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  places <- list(fine = list("gadm1:CAN.11_1"), country = list(character(0)))
+  # t7 is native only in Brazil and was never in Quebec, so Quebec is an inferred
+  # introduction - this is the machinery that must NOT fire for t6
+  expect_equal(NSR:::nsr_resolve_status("t7", places, db)$code, "Ie")
+  expect_equal(NSR:::nsr_resolve_status_set("t7", places, db)$code, "Ie")
+  # t6's surviving range is equally confined to Brazil, but Quebec is part of its native
+  # range whether or not the extinct record answers, so absence there is loss, not arrival
+  expect_equal(NSR:::nsr_resolve_status("t6", places, db)$code, "A")
+  expect_equal(NSR:::nsr_resolve_status_set("t6", places, db)$code, "A")
+})
+
+test_that("min_overlap keeps slivers out of the consulted set", {
+  db <- fixture_db()
+  keys <- "gadm1:CAN.11_1"
+  # the default threshold drops the 0.1% link but keeps the twin
+  expect_true("wgsrpd3:QUE" %in% NSR:::nsr_consulted_regions(keys, db))
+  expect_false("wgsrpd3:SLIVER" %in% NSR:::nsr_consulted_regions(keys, db))
+  # lower the threshold and it comes back, so it is the threshold doing the work
+  expect_true("wgsrpd3:SLIVER" %in% NSR:::nsr_consulted_regions(keys, db, min_overlap = 0))
+})
+
+# Coordinates and names are compared, not pooled, when they disagree: neither is
+# authoritative, and a pooled answer would describe neither place.
+test_that("place agreement is judged per level, in the geography the two share", {
+  xy <- c("wgsrpd3:QUE", "gadm1:CAN.11_1", "gadm0:CAN")
+  # the same place, said twice
+  expect_true(NSR:::nsr_places_agree(xy, c("gadm1:CAN.11_1", "gadm0:CAN")))
+  # right country, wrong province - the case a country-only comparison would miss
+  expect_false(NSR:::nsr_places_agree(xy, c("gadm1:CAN.4_1", "gadm0:CAN")))
+  # wrong country
+  expect_false(NSR:::nsr_places_agree(xy, c("gadm0:BRA")))
+  # a level only one basis speaks to is not a disagreement
+  expect_true(NSR:::nsr_places_agree(xy, c("gadm0:CAN")))
+  expect_true(NSR:::nsr_places_agree("wgsrpd3:QUE", c("gadm1:CAN.11_1", "gadm0:CAN")))
+  # nothing to compare
+  expect_true(NSR:::nsr_places_agree(character(0), character(0)))
+})
+
+test_that("keys split into the finest level named and its country", {
+  s <- NSR:::nsr_split_keys(list(c("wgsrpd3:QUE", "gadm1:CAN.11_1", "gadm0:CAN"),
+                                 "gadm0:BRA", character(0)))
+  expect_equal(s$fine[[1]], c("wgsrpd3:QUE", "gadm1:CAN.11_1"))
+  expect_equal(s$country[[1]], "gadm0:CAN")
+  expect_equal(s$fine[[2]], character(0))
+  expect_equal(s$country[[2]], "gadm0:BRA")
+  expect_equal(s$fine[[3]], character(0))
+})
+
+# The two paths must agree on the answers that need no opinions at all, not just on the
+# ones that do.  Each of these diverged.
+test_that("a matched taxon with nowhere to look is not an uncovered place", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  nowhere <- list(fine = list(character(0)), country = list(character(0)))
+  row <- NSR:::nsr_resolve_status("t1", nowhere, db)
+  set <- NSR:::nsr_resolve_status_set("t1", nowhere, db)
+  expect_equal(row$reason, "Place not matched to any region")
+  expect_equal(set$reason, row$reason)
+  expect_equal(set$code, row$code)
+  expect_equal(set$scope, row$scope)
+})
+
+test_that("nothing disagreeing is \"none\", not unknown", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  # a country-only query is answered by the states inside it, so no opinion is recorded
+  # against the polygon itself and the conflict summary has nothing to aggregate
+  p <- list(fine = list(character(0)), country = list("gadm0:CAN"))
+  expect_equal(NSR:::nsr_resolve_status("t4", p, db)$conflict_type, "none")
+  expect_equal(NSR:::nsr_resolve_status_set("t4", p, db)$conflict_type, "none")
+})
+
+test_that("absence is read against the country a state was given with", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  # CAN.9_1 carries no checklist rows and no links, but it was asked about as part of
+  # Canada, whose rows the answer already draws on - so absence there is interpretable
+  p <- list(fine = list("gadm1:CAN.9_1"), country = list("gadm0:CAN"))
+  expect_equal(NSR:::nsr_resolve_status("t3", p, db)$code, "A")
+  expect_equal(NSR:::nsr_resolve_status_set("t3", p, db)$code, "A")
+})
+
+test_that("a country counts as its own container for a confined range", {
+  skip_if_not_installed("data.table")
+  db <- fixture_db()
+  # t8 is native to Canada the country row and to Quebec the state; asked about Brazil
+  # it is absent, and its whole native range is confined to Canada
+  p <- list(fine = list("wgsrpd3:BZL"), country = list(character(0)))
+  expect_equal(NSR:::nsr_resolve_status("t8", p, db)$code, "Ie")
+  expect_equal(NSR:::nsr_resolve_status_set("t8", p, db)$code, "Ie")
+  expect_match(NSR:::nsr_resolve_status("t8", p, db)$reason, "Canada")
+})
+
+# Knowing a genus is native to a region says nothing about which of its species are, so
+# a species with no opinion of its own is not given its genus's status.  The live service
+# does consult higher ranks; this is a deliberate divergence, pinned here so it cannot
+# drift back in. See dev_notes/01-offline-nsr-design.md, "Higher ranks are not consulted".
+test_that("a species does not inherit its genus's native status", {
+  chk <- data.frame(taxon_id = "g1", region_key = "wgsrpd3:QUE", status = "native",
+                    source_name = "wcvp", is_cultivated = 0L, is_extinct = 0L,
+                    stringsAsFactors = FALSE)
+  db <- NSR:::nsr_index_db(list(
+    sources = data.frame(source_name = "wcvp", is_comprehensive = TRUE,
+                         stringsAsFactors = FALSE),
+    taxa = data.frame(taxon_id = c("g1", "s1"), species_name = c("Gen", "Gen spec"),
+                      family = "Fam", genus = "Gen", rank = c("genus", "species"),
+                      stringsAsFactors = FALSE),
+    checklist = chk,
+    regions = data.frame(region_key = "wgsrpd3:QUE", region_name = "Quebec",
+                         level = "state_province", system = "wgsrpd3",
+                         stringsAsFactors = FALSE),
+    links = data.frame(from_region = character(0), to_region = character(0),
+                       relation = character(0), fraction = numeric(0),
+                       stringsAsFactors = FALSE)))
+  p <- list(fine = list("wgsrpd3:QUE"), country = list(character(0)))
+  # the genus answers when the genus is what was asked about
+  expect_equal(NSR:::nsr_resolve_status("g1", p, db)$code, "Ne")
+  # the species does not borrow it, and is not evaluable on the strength of it
+  expect_equal(NSR:::nsr_resolve_status("s1", p, db)$code, "UNK")
+  expect_false("s1" %in% db$evaluable)
+})
diff --git a/tests/testthat/test-local-nsr.R b/tests/testthat/test-local-nsr.R
new file mode 100644
index 0000000..cb854a2
--- /dev/null
+++ b/tests/testthat/test-local-nsr.R
@@ -0,0 +1,145 @@
+# Offline NSR. The pure helpers run anywhere; the rest need a built cache and are skipped
+# unless the option NSR.test_cache names one.
+
+test_that("division names match through bilingual and standard spellings", {
+  bb <- list(
+    country = data.frame(country_id = 1L, country = "Canada", gid_0 = "CAN",
+                         stringsAsFactors = FALSE),
+    state = data.frame(state_province_id = c(10L, 11L), country_id = 1L, country = "Canada",
+                       state_province = c("New Brunswick/Nouveau-Brunswick", "Québec"),
+                       state_province_std = c("New Brunswick/Nouveau-Brunswick", "Quebec"),
+                       hasc_full = c("CA.NB", "CA.QC"), gid_1 = c("CAN.4_1", "CAN.11_1"),
+                       stringsAsFactors = FALSE))
+  r <- NSR:::nsr_match_poldiv(rep("Canada", 4),
+                              c("New Brunswick", "Nouveau-Brunswick", "Quebec", "Atlantis"), bb)
+  expect_equal(r$state_province_id, c(10L, 10L, 11L, NA_integer_))
+  expect_equal(r$poldiv_level, c("state_province", "state_province", "state_province", "country"))
+})
+
+test_that("a place is judged by the polygons containing it", {
+  db <- list(evaluable = "t1", covered = c("a", "b"))
+  op <- function(status, relation, source = "powo") {
+    data.frame(status = status, relation = relation, source_name = source,
+               is_cultivated = 0L, stringsAsFactors = FALSE)
+  }
+  # the polygon itself, and any polygon containing it, answer directly
+  expect_equal(NSR:::nsr_reduce(op("native", "same"), "a", TRUE, db)$code, "N")
+  expect_equal(NSR:::nsr_reduce(op("introduced", "within"), "a", TRUE, db)$code, "I")
+  r <- NSR:::nsr_reduce(op("native", "within"), "a", TRUE, db)
+  expect_equal(r$code, "N")
+  expect_equal(r$scope, "containing polygon")
+  # among containing polygons, any native opinion wins, and disagreement is recorded
+  r <- NSR:::nsr_reduce(rbind(op("native", "same"), op("introduced", "same", "usda")),
+                        "a", TRUE, db)
+  expect_equal(r$code, "N")
+  expect_true(r$conflict)
+  # polygons INSIDE the place describe parts of it: they answer only when unanimous
+  r <- NSR:::nsr_reduce(rbind(op("native", "contains"), op("native", "contains")),
+                        "a", TRUE, db)
+  expect_equal(r$code, "N")
+  expect_equal(r$scope, "sub-polygons agree")
+  r <- NSR:::nsr_reduce(rbind(op("native", "contains"), op("introduced", "contains")),
+                        "a", TRUE, db)
+  expect_equal(r$code, "P")
+  expect_match(r$reason, "varies among the polygons")
+  expect_equal(r$n_sub_native, 1L)
+  expect_equal(r$n_sub_introduced, 1L)
+  # a containing polygon outranks the sub-polygons
+  r <- NSR:::nsr_reduce(rbind(op("introduced", "within"), op("native", "contains")),
+                        "a", TRUE, db)
+  expect_equal(r$code, "I")
+  # merely overlapping polygons describe neither the place nor its parts
+  r <- NSR:::nsr_reduce(op("native", "overlaps"), "a", TRUE, db)
+  expect_equal(r$code, "UNK")
+  expect_match(r$reason, "partly overlapping")
+})
+
+test_that("absence is only read for taxa the sources know about", {
+  db <- list(evaluable = "t1", covered = "a")
+  # evaluable taxon, comprehensively listed region, no opinion -> absent
+  r <- NSR:::nsr_reduce(NULL, "a", TRUE, db)
+  expect_equal(r$code, "A")
+  # a taxon no source holds information about -> unknown, never absent
+  r <- NSR:::nsr_reduce(NULL, "a", FALSE, db)
+  expect_equal(r$code, "UNK")
+  expect_match(r$reason, "No source holds native status")
+  # evaluable taxon, but no comprehensive source covers the region -> unknown
+  expect_equal(NSR:::nsr_reduce(NULL, "z", TRUE, db)$code, "UNK")
+})
+
+cache <- getOption("NSR.test_cache", "")
+
+test_that("NSR_local answers the reference cases (needs a built cache)", {
+  skip_if(!nzchar(cache) || !file.exists(file.path(cache, "nsr-checklist.gz.parquet")),
+          "no built NSR cache (option NSR.test_cache)")
+  x <- data.frame(
+    species = c("Pinus ponderosa", "Araucaria angustifolia", "Araucaria angustifolia",
+                "Acer saccharum", "Sphagnum palustre", "Araucaria angustifolia"),
+    country = c("United States", "Brazil", "Brazil", "Canada", "United States", NA),
+    state_province = c("California", "Sao Paulo", "Amazonas", "New Brunswick", "California", NA),
+    county_parish = "",
+    latitude = c(NA, NA, NA, NA, NA, -23.5),
+    longitude = c(NA, NA, NA, NA, NA, -47.5),
+    stringsAsFactors = FALSE)
+  r <- NSR_local(x, dir = cache, quiet = TRUE, use_coordinates = TRUE)
+  expect_equal(r$native_status, c("N", "N", "A", "N", "UNK", "N"))
+  # Sao Paulo is answered by the Brazilian flora, which WGSRPD cannot reach
+  expect_match(r$native_status_sources[2], "flbr")
+  # a moss is unknown, not absent: WCVP holds no information about it
+  expect_false(r$taxon_evaluable[5])
+  # coordinates are resolved in each source's own geography
+  expect_equal(r$regions_matched[6], "coordinates")
+
+  # by default coordinates are not consulted at all, so the row that has nothing else
+  # to go on has no place; the rows with division names are unaffected
+  d <- NSR_local(x, dir = cache, quiet = TRUE)
+  expect_equal(d$native_status[1:5], r$native_status[1:5])
+  expect_equal(d$regions_matched[6], "none")
+  expect_equal(d$native_status[6], "UNK")
+})
+
+test_that("region links relate the geographies sensibly (needs a built cache)", {
+  skip_if(!nzchar(cache) || !file.exists(file.path(cache, "nsr-region-links.gz.parquet")),
+          "no built NSR cache")
+  lk <- as.data.frame(nanoparquet::read_parquet(file.path(cache, "nsr-region-links.gz.parquet")))
+  expect_true(all(lk$relation %in% c("same", "within", "contains", "overlaps")))
+  expect_true(all(lk$fraction >= 0 & lk$fraction <= 1 + 1e-9))
+  # Sao Paulo lies inside WGSRPD's Brazil Southeast
+  sp <- lk[lk$from_region == "gadm1:BRA.25_1" & lk$to_region == "wgsrpd3:BZL", ]
+  expect_equal(nrow(sp), 1)
+  expect_true(sp$relation %in% c("within", "same"))
+  # and the relation is recorded from both sides
+  rev <- lk[lk$from_region == "wgsrpd3:BZL" & lk$to_region == "gadm1:BRA.25_1", ]
+  expect_equal(nrow(rev), 1)
+  expect_true(rev$relation %in% c("contains", "same"))
+})
+
+test_that("absence becomes introduction only for taxa confined elsewhere", {
+  skip_if(!nzchar(cache) || !file.exists(file.path(cache, "nsr-checklist.gz.parquet")),
+          "no built NSR cache (option NSR.test_cache)")
+  db <- NSR:::nsr_local_db(cache)
+  nat <- db$checklist[db$checklist$status == "native", ]
+  per <- table(nat$taxon_id)
+  # a species WCVP gives a single native area, queried far away
+  cal <- intersect(names(per)[per == 1], nat$taxon_id[nat$region_key == "wgsrpd3:CAL"])
+  skip_if(!length(cal), "no single-area Californian native in this build")
+  sp <- db$taxa$species_name[match(cal[1], db$taxa$taxon_id)]
+  r <- NSR_local(data.frame(species = c(sp, sp), country = "United States",
+                            state_province = c("Michigan", "California"),
+                            county_parish = "", stringsAsFactors = FALSE),
+                 dir = cache, quiet = TRUE)
+  expect_equal(r$native_status, c("Ie", "Ne"))
+  expect_equal(r$isIntroduced, c(1L, 0L))
+  expect_equal(r$isEndemic, c(0L, 1L))
+  expect_match(r$native_status_reason[1], "endemic to")
+  # a widely native species merely unrecorded there stays absent
+  wide <- names(per)[per > 20]
+  wide <- setdiff(wide, db$checklist$taxon_id[db$checklist$region_key == "wgsrpd3:MIC"])
+  skip_if(!length(wide), "no widespread species absent from Michigan")
+  w <- db$taxa$species_name[match(wide[1], db$taxa$taxon_id)]
+  r2 <- NSR_local(data.frame(species = w, country = "United States", state_province = "Michigan",
+                             county_parish = "", stringsAsFactors = FALSE),
+                  dir = cache, quiet = TRUE)
+  expect_equal(r2$native_status, "A")
+  expect_equal(r2$isIntroduced, 0L)
+})