diff --git a/readme.md b/readme.md
index f9cd2b6..d16acd9 100644
--- a/readme.md
+++ b/readme.md
@@ -21,6 +21,17 @@ For a few years RPNAB had a daughter company called Recamp Nordic AB that everyt
 
 Code will eventually be kept on a private gitea instance (for issue handling), but everything is to be considered open source. No expectations of assistance with writing code, but if the project and the company is successful a dream of the maintainer is for other companies and developers to make use of it. However, due to the bespoke nature of the functionality provided this seems unlikely.
 
+## Toolchain
+
+- Rust via rustup, stable channel (installed with `brew install rustup-init;
+  rustup default stable` on macOS; rustup's shims live in
+  `/opt/homebrew/opt/rustup/bin` — keep it on PATH).
+- Build and test from the `Application/` directory: `cargo build`,
+  `cargo test`, `cargo fmt --check`, `cargo clippy`.
+- Python 3 (stdlib + openpyxl + pypdf) is used only by
+  `Application/scripts/` for regenerating sanitized test fixtures; it is
+  not part of the product toolchain.
+
 ## Technology
 
 The whole stack is Rust: one language, one toolchain, one test runner,
@@ -90,6 +101,19 @@ committed, and their values must never be embedded in committed test files.
 Tests may *reference* sample files by path (the files stay local), or use
 synthetic/sanitized values in the repo.
 
+Sanitized copies of the samples — committed, safe for any repo consumer —
+live under `Application/data/fixtures/` (epsilon batch files 405-412 with
+synthetic card/customer ids, batches renumbered 9405+ and dates +2y;
+a cumulative slice, batches 5001+, dates -6y; tsdrms xlsx files with
+synthetic R/A / DBR / location ids, dates +2y; subfranchise statements as
+extracted text with all identity and amounts replaced). Amounts, GL codes
+and descriptions are kept verbatim where they carry no personal data;
+cross-value arithmetic in the subfranchise fixtures (EUR × rate = SEK) is
+deliberately NOT preserved. Regenerate with
+`python3 Application/scripts/sanitize_samples.py` (reads the local raw
+samples) and verify sanitization with `check_fixture_leaks.py` (exit 0 =
+clean). Feature branches should build parser tests against the fixtures.
+
 The raw formats are canonical: the application parses what the source systems
 deliver. Filenames carry metadata (batch / month / invoice number) and are
 the natural keys for idempotent re-ingestion — ingesting the same file twice