diff --git a/README.md b/README.md
new file mode 100644
index 0000000..78bd424
--- /dev/null
+++ b/README.md
@@ -0,0 +1,24 @@
+# PARAmap API
+
+This repo will contain an API serving two kinds of data to be rendered by PARAmap:
+
+1) data points of surveys of genetic markers, which come to us stored in the [STAVE](https://mrc-ide.github.io/STAVE/index.html) schema;
+2) and a surface of model outputs  imputed from the survey data, which are essentially interpolated prevalences of the difference genetic markers per region.
+
+## How to update the data
+
+### Model outputs
+
+TODO
+
+The list of in-scope genes and mutations will vary over time, with model releases (rather than with STAVE data releases). Every model release has a dependency on one STAVE data release.
+
+### STAVE data
+
+When a new STAVE data release is provided, it should be given a version name e.g. "2026.03.17", and committed in `scripts/input/stave/<version>/stave_data.rds`. Then, run the [process_stave.R](./scripts/process_stave.R) script:
+
+```sh
+Rscript ./scripts/process_stave.R 2026.03.17
+```
+
+This will create `./data/stave/<version>/survey_data.parquet`.
diff --git a/data/stave/.gitkeep b/data/stave/.gitkeep
new file mode 100644
index 0000000..e69de29
diff --git a/scripts/input/stave/2026.03.17/stave_data.rds b/scripts/input/stave/2026.03.17/stave_data.rds
new file mode 100644
index 0000000..96e43ed
Binary files /dev/null and b/scripts/input/stave/2026.03.17/stave_data.rds differ
diff --git a/scripts/process_stave.R b/scripts/process_stave.R
new file mode 100644
index 0000000..4d4b66f
--- /dev/null
+++ b/scripts/process_stave.R
@@ -0,0 +1,93 @@
+# Converts an .rds file containing a STAVE object of 3 tables (studies, surveys, counts)
+# into an expanded version of the studies table, having one row per survey per variant,
+# which is saved in parquet format.
+# This results in a table of about 253,000 rows at time of writing.
+# Most of the work and logic is done by the two packages STAVE and variantstring,
+# particularly STAVE's `$get_prevalence` function.
+
+library(arrow)
+library(cli)
+library(dplyr)
+library(tidyr)
+library(STAVE)
+library(variantstring)
+library(here)
+
+args <- commandArgs(trailingOnly = TRUE)
+if (length(args) == 0) {
+  cli_abort(c(
+    "Usage: {.code Rscript process_stave.R <stave_release>}",
+    "i" = "Example: {.code Rscript process_stave.R 2026.03.17}"
+  ))
+}
+
+current_stave_release <- args[[1]]
+output_filename = "survey_data.parquet"
+
+input_dir <- here("scripts", "input", "stave", current_stave_release)
+output_dir <- here("data", "stave", current_stave_release)
+
+dir.create(output_dir, recursive = TRUE, showWarnings = FALSE)
+
+input_file <- file.path(input_dir, "stave_data.rds")
+if (!file.exists(input_file)) {
+  cli_abort("Input file not found: {.file {input_file}}.")
+}
+
+stave_obj <- readRDS(input_file)
+
+variants <- stave_obj$get_variants()
+
+# For each variant, calculate imputed prevalence per survey
+# (only surveys with a non-zero denominator), and combine into one tibble.
+prevalence_tbl <- variants |>
+  lapply(function(v) {
+    stave_obj$get_prevalence(target_variant = v, return_full = FALSE) |>
+      mutate(variant = v)
+  }) |>
+  bind_rows()
+
+# Parse gene/locus/amino-acid from each variant string using variantstring's own parser.
+# We validate that all variant strings define a single variant.
+# If they do, they have only a single value between each colon.
+# If they don't, `variant_to_long` will unpack the strings into multiple variants, and
+# we'll catch this and abort.
+parsed_list <- variant_to_long(variants)  # one data.frame per variant
+n_rows <- vapply(parsed_list, nrow, integer(1))
+if (any(n_rows != 1)) {
+  bad <- variants[n_rows != 1]
+  cli_abort(c(
+    "Expected each variant to parse to exactly 1 row via {.fn variant_to_long}, but got unexpected row counts for: {.val {unique(bad)}}.",
+    "x" = "The variant string might not be single-locus.",
+    "i" = "Did you call {.code $get_variants(report_haplo = TRUE)}?"
+  ))
+}
+
+parsed <- bind_rows(Map(function(df, v) mutate(df, variant = v), parsed_list, variants))
+
+prevalence_tbl <- prevalence_tbl |>
+  left_join(
+    parsed |>
+      transmute(variant, gene, mutation = paste0(pos, aa)),
+    by = "variant"
+  )
+
+# Because of encoding errors in paper titles, we need to force conversion to UTF-8.
+# This function tries UTF-8 first, and if invalid, assumes Latin-1
+fix_utf8 <- function(x) {
+  bad <- !validEnc(x) | is.na(iconv(x, "UTF-8", "UTF-8")) # get vector of whether utf-8 encoding works for the string
+  x[bad] <- iconv(x[bad], from = "latin1", to = "UTF-8") # For just the flagged entries, reinterpret the raw bytes as Latin-1 and convert to utf-8
+  enc2utf8(x) # tag every string as declared-UTF-8
+}
+
+# Drop columns we don't need
+drop_cols <- c("description", "access_level", "PMID", "country_name",
+               "location_method", "location_notes", "time_method", "time_notes")
+
+prevalence_tbl <- prevalence_tbl |>
+  select(-all_of(drop_cols)) |>
+  mutate(across(where(is.character), fix_utf8))
+
+write_parquet(prevalence_tbl, file.path(output_dir, output_filename))
+
+cli_inform(c("v" = "Wrote {.file {output_filename}} with {nrow(prevalence_tbl)} rows to {.path {output_dir}}."))