diff --git a/apps/machine/electron/hal-service.ts b/apps/machine/electron/hal-service.ts
index 2ec863b..5453fea 100644
--- a/apps/machine/electron/hal-service.ts
+++ b/apps/machine/electron/hal-service.ts
@@ -87,19 +87,40 @@ export async function initializeHal(config: HalConfig): Promise<HalInstance> {
 
   const { validator: valConfig, dispenser: dispConfig } = config
 
-  // Create hardware instances
-  const dispenser: BillDispenser = hal.createDispenser(dispConfig.type, {
-    device: dispConfig.device,
-  })
+  // Start dispenser (optional — mirrors the validator handling below).
+  //
+  // A cash-in-only machine is a legitimate configuration: the Raspberry Pi
+  // reference build has a bill acceptor and no dispenser at all. This used to
+  // create and init the dispenser unconditionally, so a missing device threw
+  // and aborted the WHOLE of initializeHal — taking the validator with it,
+  // even though the validator was present and working. The Pi bring-up hit
+  // exactly that: "cannot open /dev/ttyDispenser-not-fitted", then an endless
+  // renderer-reload loop, with a perfectly good acceptor on ttyValidator0.
+  //
+  // The validator has been optional since it was written; the asymmetry was
+  // the bug.
+  let dispenser: BillDispenser | null = null
 
-  // Initialize dispenser. `dispenserInitData` is `let` because
-  // `setCassettes` swaps it in to re-init with a new layout (also used by
-  // the on-error re-init path at dispenseCash).
+  // `dispenserInitData` is `let` because `setCassettes` swaps it in to re-init
+  // with a new layout (also used by the on-error re-init path at dispenseCash).
   let dispenserInitData = {
     fiatCode: valConfig.fiatCode,
     cassettes: dispConfig.cassettes,
   }
-  await dispenser.init(dispenserInitData)
+
+  try {
+    const fs = await import('node:fs')
+    if (dispConfig.device && fs.existsSync(dispConfig.device)) {
+      dispenser = hal.createDispenser(dispConfig.type, { device: dispConfig.device })
+      await dispenser.init(dispenserInitData)
+      console.log('[HAL] Dispenser started')
+    } else {
+      console.log('[HAL] Dispenser device not found, running cash-in only')
+    }
+  } catch (err) {
+    console.warn('[HAL] Dispenser failed to start, running cash-in only:', err)
+    dispenser = null
+  }
   console.log('[HAL] Dispenser initialized')
 
   // Start validator (optional — proceed without if device is missing or fails)
@@ -251,6 +272,17 @@ export async function initializeHal(config: HalConfig): Promise<HalInstance> {
     dispenseCash: async (amounts): Promise<DispenseResult> => {
       console.log('[HAL] Dispensing:', amounts)
 
+      // Cash-in-only machine: refuse the ask rather than throwing a null
+      // dereference into the renderer's dispense path.
+      if (!dispenser) {
+        return {
+          bills: [],
+          cassettes: [],
+          dispensed: false,
+          error: 'No dispenser fitted on this machine — cash-out unavailable',
+        }
+      }
+
       // Re-initialize dispenser if it was closed after a previous error
       if (!dispenser.initialized) {
         console.log('[HAL] Dispenser not initialized, re-initializing...')
@@ -391,6 +423,12 @@ export async function initializeHal(config: HalConfig): Promise<HalInstance> {
           count: c.count ?? 0,
         }))
       dispenserInitData = { fiatCode: valConfig.fiatCode, cassettes }
+      // Without a dispenser the layout is still worth recording (the operator
+      // config consumer keeps calling this), but there is nothing to re-init.
+      if (!dispenser) {
+        console.log('[HAL] Cassettes recorded; no dispenser fitted, nothing to re-init')
+        return
+      }
       // Close + re-init the dispenser so its internal per-bay state matches
       // the new layout. Errors here surface to the caller (operator-config
       // consumer) — the renderer can decide whether to retry.
@@ -407,7 +445,7 @@ export async function initializeHal(config: HalConfig): Promise<HalInstance> {
       return new Promise<void>((resolve) => {
         validator?.disable()
         validator?.lightOff()
-        dispenser.close()
+        dispenser?.close()
         if (validator) {
           validator.close((err?: Error) => {
             if (err) console.error('[HAL] Validator close error:', err)
diff --git a/apps/machine/src/config/device.ts b/apps/machine/src/config/device.ts
index cffdaa3..9a06c9a 100644
--- a/apps/machine/src/config/device.ts
+++ b/apps/machine/src/config/device.ts
@@ -15,7 +15,15 @@ import type { HalConfig, CassetteConfig } from '@/services/hal'
 /**
  * Supported machine models
  */
-export type MachineModel = 'sintra' | 'tejo' | 'douro' | 'gaia' | 'batm3' | 'custom'
+export type MachineModel =
+  | 'sintra'
+  | 'tejo'
+  | 'douro'
+  | 'gaia'
+  | 'batm3'
+  | 'rpi4'
+  | 'rpi5'
+  | 'custom'
 
 /**
  * Full device configuration
@@ -28,7 +36,10 @@ export interface DeviceConfig {
   /** Bill validator configuration */
   validator: {
     /** Validator protocol type */
-    type: 'id003' | 'ebds'
+    // 'apex' = Pyramid Apex RS-232. The driver landed with the Pi 5 work
+    // (packages/hal ValidatorType) but these app-side unions were never
+    // widened, so no machine could actually be configured to use it.
+    type: 'id003' | 'ebds' | 'apex'
     /** Serial device path(s) */
     device: string | string[]
   }
@@ -117,6 +128,52 @@ export const MACHINE_PRESETS: Record<MachineModel, Omit<DeviceConfig, 'fiatCode'
    * - Dispenser: Fujitsu F56
    * Note: Device paths may vary - verify on hardware
    */
+  /**
+   * Raspberry Pi 4 / CM4 reference build (aarch64).
+   * - Validator: Pyramid Apex 7600 over RS-232, via a USB-serial adapter
+   * - Dispenser: NONE WIRED YET — cash-in only
+   *
+   * The device path is the udev symlink raspberry-pi-4.nix creates for an
+   * FTDI bridge. Swap to ttyValidator1 (CP210x) or ttyValidator2 (CH340) to
+   * match the adapter actually fitted; `ls -l /dev/ttyValidator*` after
+   * plugging it in will say which one appeared.
+   *
+   * The dispenser block is a placeholder, not a claim. DispenserType has no
+   * 'none' variant and DeviceConfig requires the field, so it points at a
+   * path that does not exist and carries no cassettes. Cash-out is not
+   * available on this board until real hardware and a real path land here.
+   */
+  rpi4: {
+    model: 'rpi4',
+    validator: {
+      type: 'apex',
+      device: '/dev/ttyValidator0',
+    },
+    dispenser: {
+      type: 'f56',
+      device: '/dev/ttyDispenser-not-fitted',
+      cassettes: [],
+    },
+  },
+
+  /**
+   * Raspberry Pi 5 reference build (aarch64). Same shape as rpi4 — same
+   * validator, same absent dispenser, same udev symlinks from
+   * raspberry-pi-5.nix.
+   */
+  rpi5: {
+    model: 'rpi5',
+    validator: {
+      type: 'apex',
+      device: '/dev/ttyValidator0',
+    },
+    dispenser: {
+      type: 'f56',
+      device: '/dev/ttyDispenser-not-fitted',
+      cassettes: [],
+    },
+  },
+
   gaia: {
     model: 'gaia',
     validator: {
diff --git a/apps/machine/src/services/hal.ts b/apps/machine/src/services/hal.ts
index f48fda8..4ab83ee 100644
--- a/apps/machine/src/services/hal.ts
+++ b/apps/machine/src/services/hal.ts
@@ -23,7 +23,7 @@ export interface CassetteConfig {
 
 export interface HalConfig {
   validator: {
-    type: 'id003' | 'ebds'
+    type: 'id003' | 'ebds' | 'apex'
     device: string | string[]
     fiatCode: string
   }
diff --git a/deploy/nixos/README.md b/deploy/nixos/README.md
index 2bfc333..7dff65a 100644
--- a/deploy/nixos/README.md
+++ b/deploy/nixos/README.md
@@ -14,7 +14,9 @@ deploy/nixos/
 ├── hardware/
 │   ├── douro.nix           # Dell OptiPlex 9030 AIO (stock Douro motherboard; SATA SSD, eGalax touch)
 │   ├── batm3.nix           # GeneralBytes BATM3 chassis with a Dell OptiPlex 9030 AIO grafted in (custom mod; WireGuard wired in)
-│   └── upboard.nix         # Aaeon UP Board (Sintra + tejo; eMMC root via sdhci-acpi + mmc_block)
+│   ├── upboard.nix         # Aaeon UP Board (Sintra + tejo; eMMC root via sdhci-acpi + mmc_block)
+│   ├── raspberry-pi-5.nix  # Raspberry Pi 5 (aarch64) — DIY build; board glue on top of nixos-hardware
+│   └── raspberry-pi-4.nix  # Raspberry Pi 4 (aarch64) — same, previous-generation board
 ├── udev/
 │   └── 99-bitspire-hardware.rules   # additional udev rules (loaded via configuration.nix)
 ├── provision-atm.sh        # Push LNbits credentials to a deployed ATM via SSH
@@ -36,6 +38,11 @@ Each ATM model has two flake outputs:
 
 Models: `douro`, `tejo`, `sintra`, `batm3`.
 
+The Raspberry Pi boards (`rpi5`, `rpi4`) are aarch64 and follow a different
+pipeline — SD-card image instead of GPT disk image, U-Boot instead of
+systemd-boot, and they need an aarch64 builder. See
+[`docs/raspberry-pi-setup.md`](../../docs/raspberry-pi-setup.md).
+
 ```bash
 # Build a Sintra disk image
 nix build .#disk-image-sintra
diff --git a/deploy/nixos/hardware/raspberry-pi-4.nix b/deploy/nixos/hardware/raspberry-pi-4.nix
new file mode 100644
index 0000000..5b136dc
--- /dev/null
+++ b/deploy/nixos/hardware/raspberry-pi-4.nix
@@ -0,0 +1,181 @@
+# Raspberry Pi 4 hardware module (aarch64).
+#
+# The Pi 4 twin of raspberry-pi-5.nix. Kernel, firmware, bootloader and device
+# tree come from the nixos-hardware `raspberry-pi-4` module (paired with this
+# file in flake.nix's piBoards); here we set only the bitSpire-specific
+# hardware glue: serial for the bill validators, the kiosk display driver, and
+# no-suspend. The wiring notes in raspberry-pi-5.nix apply unchanged — same
+# validators over USB-serial, same QR scanner, same touchscreen options.
+#
+# What differs from the Pi 5:
+#   - GPU/KMS: the Pi 5 module enables vc4/v3d modesetting by default; on the
+#     Pi 4 it is an opt-in (`fkms-3d`) that also injects the CMA + vc4 device
+#     tree overlays. Without it X falls back to the plain framebuffer and
+#     Electron renders in software.
+#   - Memory: 4 GB is the floor for Electron + the kiosk; 8 GB is comfortable.
+#     The shared Pi runtime's MemoryMax=2G leaves headroom on either.
+#   - No PCIe (the Pi 5's NVMe path); boot/root is SD or USB-SATA only.
+{ config, lib, pkgs, ... }:
+
+{
+  # aarch64 target. (The flake instantiates this config with aarch64 pkgs; this
+  # line documents/asserts it.)
+  nixpkgs.hostPlatform = lib.mkDefault "aarch64-linux";
+
+  # Mainline kernel, NOT the Raspberry Pi vendor one.
+  #
+  # nixos-hardware's raspberry-pi/4 module mkDefaults boot.kernelPackages to
+  # the vendor kernel (linux-rpi, via common/kernel.nix). That kernel is in no
+  # binary cache — Hydra does not build nixos-hardware's overlays, and it is
+  # not in aiolabs.cachix.org either — so every Pi compiles a kernel from
+  # source, on an SD card, and recompiles on every bump. The first bring-up
+  # attempt spent hours on `CC [M] fs/overlayfs/inode.o` before anyone noticed
+  # what it was doing.
+  #
+  # Mainline aarch64 kernels are cached, and mainline demonstrably boots a
+  # Pi 4: it is what the stock NixOS aarch64 SD image runs. The vendor kernel's
+  # Pi-specific patches buy nothing this kiosk needs — display, USB serial and
+  # WiFi are all mainline, and vc4/v3d KMS has been mainline for years.
+  #
+  # Watch the graphics path when changing this. fkms-3d below is a
+  # nixos-hardware overlay built around the vendor kernel's firmware-KMS route;
+  # the mainline equivalent is full KMS (vc4-kms-v3d). If X ends up on the
+  # framebuffer with Electron rendering in software, that overlay is where to
+  # look — not the kernel choice, which is worth keeping either way.
+  boot.kernelPackages = pkgs.linuxPackages;
+
+  # Bootloader: the aarch64 sd-image uses the extlinux-compatible generator;
+  # nixos-hardware's rpi4 module wires the firmware/u-boot. No systemd-boot.
+  boot.loader.grub.enable = lib.mkDefault false;
+  boot.loader.generic-extlinux-compatible.enable = lib.mkDefault true;
+
+  # Primary UART (GPIO 14/15) available for a GPIO-wired validator. Keep the
+  # serial console OFF it so the validator owns the line — mirrors upboard.nix
+  # keeping ttyS4 free for the dispenser. USB-serial adapters are unaffected.
+  #
+  # Not mkDefault: kernelParams is list-merged, and only definitions at the
+  # highest priority survive. nixpkgs sets loglevel/lsm at normal priority, so
+  # a mkDefault list here is dropped entirely — and with no console= at all
+  # the kernel falls back to the device tree's stdout-path, i.e. this UART.
+  # cma=256M: the vc4 display pipeline allocates its framebuffer from the
+  # contiguous memory area, and the default reservation on this board is 32MiB
+  # with about 11MiB free. A 3840x1080 framebuffer is ~16.6MB before double
+  # buffering, so X got as far as picking the mode and then died:
+  #
+  #   Output HDMI-1 using initial mode 3840x1080 +0+0
+  #   (EE) AddScreen/ScreenInit failed for driver 0
+  #
+  # nixos-hardware's fkms-3d injected a CMA overlay alongside the display one;
+  # this replaces that half of it. 256M is generous for any panel an ATM will
+  # carry and trivial against 4-8GB of RAM.
+  boot.kernelParams = [ "console=tty0" "cma=256M" ];
+
+  # ── REQUIRES A MANUAL STEP ON THE FIRMWARE PARTITION ────────────────
+  # This board boots the FIRMWARE's vendor DTB, not the DTBs NixOS builds.
+  # Confirmed on the CM4: the live device tree carries __symbols__ and the
+  # mainline DTBs in dtbs-filtered do not, and U-Boot found no FDTDIR match for
+  # compatible "raspberrypi,4-compute-module" so it passed the firmware's DTB
+  # through. That means hardware.deviceTree.overlays cannot reach the running
+  # device tree, and the display has to be enabled by the firmware instead.
+  #
+  # In the vendor DTB every display node (hvs, gpu, all pixelvalves, both hdmi)
+  # ships `disabled`. So /boot/firmware/config.txt needs:
+  #
+  #     dtoverlay=vc4-kms-v3d,noaudio
+  #
+  # and /boot/firmware/overlays/ needs to be populated from raspberrypifw --
+  # the NixOS sd-image writes the DTBs there but NOT the overlays, so the
+  # directory ships empty and the dtoverlay line fails silently. Copy the whole
+  # directory (2MB, 356 files); copying only vc4-kms-v3d.dtbo is not enough
+  # because the firmware remaps that to vc4-kms-v3d-pi4.dtbo on this board.
+  #
+  # `noaudio` is required, not cosmetic. With HDMI audio enabled vc4_hdmi cannot
+  # register its PCM component, returns -517 (EPROBE_DEFER) forever, and the DRM
+  # device never registers -- so X finds no card at all. We removed the audio
+  # stack anyway, so there is nothing to lose.
+  #
+  # This is a reflash-losing manual step and it should be folded into the image
+  # builder. Tracked as a follow-up; noted here so the next person does not
+  # rediscover it from a blank screen.
+
+  # Kiosk display: mainline full KMS, NOT nixos-hardware's fkms-3d.
+  #
+  # fkms-3d applies the rpi4-cma-overlay and rpi4-vc4-fkms-v3d-overlay device
+  # tree overlays. Those target nodes that exist in the Raspberry Pi VENDOR
+  # kernel's DTBs and not in mainline's, so with the mainline kernel above the
+  # overlay step fails outright:
+  #
+  #   Applying overlay rpi4-vc4-fkms-v3d-overlay
+  #   libfdt.FdtException: pylibfdt error -1: FDT_ERR_NOTFOUND
+  #
+  # It is also unnecessary. "fkms" is FIRMWARE KMS, the older route where the
+  # VideoCore firmware owns the display and Linux drives it at arm's length.
+  # Mainline does full KMS instead, and mainline's own bcm2711-rpi-4-b.dtb
+  # already describes the hardware — it carries brcm,bcm2711-vc5 and
+  # brcm,2711-v3d nodes, checked with dtc. The vc4 and v3d drivers bind to
+  # those directly with no overlay involved.
+  #
+  # fkms-3d used to set services.xserver.videoDrivers as a side effect. Nothing
+  # needs to replace it: the shared configuration.nix already declares
+  # modesetting, which is the correct driver for full KMS and what the x86
+  # machines use. Setting it again here only produced a duplicate entry.
+
+  hardware.enableRedistributableFirmware = true;
+
+  # pcscd MUST be enabled, and not because this board has a card reader.
+  #
+  # The app constructs @pokusew/pcsclite at startup. That calls
+  # SCardEstablishContext(), which calls SCardCheckDaemonAvailability(), which
+  # — when there is no pcscd to find — BUSY-LOOPS in fstatat64 at ~92% CPU
+  # instead of returning an error. It runs on Electron's main thread, before
+  # the BrowserWindow is created, so the window never appears and the panel
+  # stays white forever. Nothing is logged, nothing throws, and V8's own
+  # inspector cannot be serviced because the thread never yields: CDP
+  # Debugger.pause and Profiler.stop both hang. It took a native gdb backtrace
+  # to see it at all:
+  #
+  #   #0 fstatat64                     libc
+  #   #1 SCardCheckDaemonAvailability  libpcsclite
+  #   #2 SCardEstablishContext         libpcsclite
+  #   #3 PCSCLite::PCSCLite()          pcsclite.node
+  #
+  # The x86 machines never hit this because upboard.nix and batm3.nix both
+  # enable pcscd for their actual readers. This module did not, which is the
+  # entire difference. A running pcscd with no reader attached just idles, so
+  # this is cheap insurance rather than a claim about the hardware.
+  services.pcscd.enable = true;
+
+  # pcscd gates client access via polkit; without a rule the `bitspire` service
+  # user is "Rejected unauthorized PC/SC client". Same wiring as upboard.nix.
+  security.polkit.extraConfig = ''
+    polkit.addRule(function(action, subject) {
+      if ((action.id == "org.debian.pcsc-lite.access_pcsc" ||
+           action.id == "org.debian.pcsc-lite.access_card") &&
+          subject.user == "bitspire") {
+        return polkit.Result.YES;
+      }
+    });
+  '';
+
+  # Kiosk: never sleep.
+  systemd.targets = {
+    sleep.enable = false;
+    suspend.enable = false;
+    hibernate.enable = false;
+    hybrid-sleep.enable = false;
+  };
+
+  # Stable device symlinks for USB-serial bill-validator adapters, so the ATM
+  # config can point at /dev/ttyValidator0 regardless of enumeration order.
+  # Identical to the Pi 5 module — same adapters, same bridges. If two adapters
+  # of the SAME chip are used, disambiguate by KERNELS/serial instead — tune
+  # during bring-up.
+  services.udev.extraRules = lib.mkAfter ''
+    # FTDI (e.g. FT232R) → ttyValidator0
+    SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", ATTRS{idProduct}=="6001", SYMLINK+="ttyValidator0"
+    # Silicon Labs CP210x → ttyValidator1
+    SUBSYSTEM=="tty", ATTRS{idVendor}=="10c4", ATTRS{idProduct}=="ea60", SYMLINK+="ttyValidator1"
+    # WCH CH340 → ttyValidator2
+    SUBSYSTEM=="tty", ATTRS{idVendor}=="1a86", ATTRS{idProduct}=="7523", SYMLINK+="ttyValidator2"
+  '';
+}
diff --git a/docs/raspberry-pi-setup.md b/docs/raspberry-pi-setup.md
new file mode 100644
index 0000000..d6c525f
--- /dev/null
+++ b/docs/raspberry-pi-setup.md
@@ -0,0 +1,145 @@
+# Deploying bitSpire to a Raspberry Pi (4 or 5)
+
+The DIY reference build: a Raspberry Pi, a bill validator on USB-serial, a
+touchscreen and a QR scanner. Both boards share one runtime in `flake.nix`
+(`piBaseModules`) and differ only in their hardware glue module:
+
+| Board | `nixosConfigurations` | Flashable image | Hardware glue |
+|---|---|---|---|
+| Raspberry Pi 5 | `rpi5-installed`, `rpi5-image` | `packages.aarch64-linux.sd-image-rpi5` | `deploy/nixos/hardware/raspberry-pi-5.nix` |
+| Raspberry Pi 4 | `rpi4-installed`, `rpi4-image` | `packages.aarch64-linux.sd-image-rpi4` | `deploy/nixos/hardware/raspberry-pi-4.nix` |
+
+`<board>-image` is what you flash; `<board>-installed` is what a running Pi
+rebuilds itself against afterwards. They share every module except the root
+filesystem declaration and the SD-image builder — see the comments above
+`mkPiInstalled` / `mkPiImage` in `flake.nix` for why they're split.
+
+**Status:** the Pi 4 target is evaluation-verified (it instantiates, and its
+configuration differs from the Pi 5's only in the expected board-specific
+places) but has not yet been booted on hardware. Treat the first bring-up as
+exactly that.
+
+## What's different from the x86 fleet
+
+- **aarch64.** Any Pi build needs an aarch64 builder — a Pi itself, an ARM
+  box, or `boot.binfmt.emulatedSystems = [ "aarch64-linux" ]` on an x86 host.
+  The Electron app closure will not build on plain x86.
+- **Boot.** Raspberry Pi firmware + U-Boot + extlinux (from `nixos-hardware`),
+  not systemd-boot. The image is an SD-card image, not a GPT disk image.
+- **No `determinate`, no `atm-tui`** in the Pi runtime yet (neither ships an
+  aarch64 package). Add them when they do.
+- **Cachix is pre-wired** (`aiolabs.cachix.org` in `nix.settings`), so an
+  in-place rebuild substitutes the heavy closure rather than compiling on the
+  Pi — provided the closure was pushed there first.
+
+## Board differences that matter
+
+| | Pi 5 | Pi 4 |
+|---|---|---|
+| SoC / device tree | BCM2712 (`bcm2712*-rpi-*.dtb`) | BCM2711 (`bcm2711-rpi-4*.dtb`) |
+| Kiosk GPU / KMS | vc4/v3d on by default | opt-in via `hardware.raspberry-pi."4".fkms-3d` (the glue module enables it; it also injects the CMA + vc4 overlays) |
+| RAM | 4–16 GB | 4 GB is the floor for Electron; 8 GB is comfortable |
+| Root storage | SD, USB, or NVMe over PCIe | SD or USB only |
+| Power | 5 V / 5 A USB-C PD | 5 V / 3 A; Electron startup is the peak |
+
+Everything else — validator serial symlinks, `console=tty0` keeping the GPIO
+UART free, no-suspend, the bitspire service — is identical between the two
+glue modules by design. Keep it that way: a change to one almost certainly
+belongs in the other.
+
+## 1. Build the image
+
+On (or via) an aarch64 builder:
+
+```bash
+nix build .#packages.aarch64-linux.sd-image-rpi4     # or sd-image-rpi5
+ls result/sd-image/
+# → nixos-image-sd-card-<version>-aarch64-linux.img.zst
+```
+
+## 2. Flash
+
+```bash
+lsblk -f                      # identify the SD card — NOT your main disk
+zstd -dc result/sd-image/*.img.zst | sudo dd of=/dev/sdX bs=4M status=progress conv=fsync
+```
+
+The image carries two labelled partitions the installed config expects:
+`FIRMWARE` (vfat, Pi firmware + U-Boot) and `NIXOS_SD` (ext4 root). The root
+partition grows to fill the card on first boot.
+
+## 3. First boot
+
+Insert the card, connect Ethernet and power. The Pi boots into the kiosk with
+no pairing, so the screen shows the pairing wizard — with a camera it waits
+for a QR; without one it says so and tells you to provision `VITE_SPIRE_SEED`
+instead. It also comes up with sshd and password auth enabled, same as the
+x86 installed configs — this is the provisioning window.
+
+Provision exactly as for a Sintra — from the dev box, with the spire seed
+minted by spirekeeper:
+
+```bash
+SPIRE_SEED='spire-seed:v1:…' bash deploy/nixos/provision-atm.sh <pi-ip> 22
+```
+
+That writes `/var/lib/bitspire/.env` and restarts the service; the seed
+carries the relay and the LNbits transport pubkey, so nothing else is needed.
+Alternatively, show the seed's QR to the machine's camera and let the
+on-screen wizard do the same thing. Verify with:
+
+```bash
+ssh bitspire@<pi-ip> 'journalctl -u bitspire -n 50 --no-pager | grep "\["'
+# expect [Signer] Pairing to bunker … then [Lightning] LNbits client initialized
+```
+
+The dev-only `VITE_ATM_PRIVATE_KEY` fallback works here too for a bench
+setup without a bunker — see `provision-atm.sh`'s header for the variables.
+
+## 4. Updating in place
+
+Once flashed, never re-flash for a software update. The `-installed` target
+is the aarch64 twin of the fleet's rebuild ritual:
+
+```bash
+sudo nixos-rebuild switch --flake \
+  "git+ssh://forgejo@git.atitlan.io/aiolabs/bitspire.git?ref=dev#rpi4-installed"
+```
+
+(Pi 5: `#rpi5-installed`.) With the closure on cachix this is a download, not
+a build. If it starts compiling Electron on the Pi, the closure wasn't pushed
+— stop, build on the aarch64 builder, `cachix push aiolabs`, retry.
+
+## 5. Wiring the peripherals
+
+- **Bill validator** — over a USB-serial adapter. The glue modules give
+  stable symlinks so `.env` never depends on enumeration order:
+  FTDI → `/dev/ttyValidator0`, CP210x → `/dev/ttyValidator1`,
+  CH340 → `/dev/ttyValidator2`. Two adapters of the *same* chip need
+  disambiguating by `KERNELS`/serial in the udev rule — do that at bring-up.
+  A validator wired straight to the GPIO UART (pins 14/15) also works: the
+  kernel console is pinned to `tty0` precisely so that UART stays free.
+- **QR scanner** — USB HID keyboard-emulation, no configuration.
+- **Touchscreen** — DSI or HDMI. X runs on the Pi's KMS driver.
+- **Serial console for debugging** — there isn't one by default (see above).
+  Use SSH, or temporarily add `console=ttyAMA0,115200` to
+  `boot.kernelParams` in the glue module.
+
+## Hardware notes for 24/7 operation
+
+- **Storage.** SD cards have finite write endurance; for anything beyond a
+  bench build put root on a USB-SATA SSD (both boards) or NVMe (Pi 5).
+  `state.db` and the journal are the writers.
+- **Thermal.** Both boards throttle under sustained load without cooling.
+  Heatsinks at minimum; the official active cooler for the Pi 5.
+- **Power.** Brown-outs on Electron startup look like random reboots. Use the
+  official supply or one rated above the board's peak, never a hub.
+- **Swap.** The runtime provisions a 2 GB swapfile so memory pressure degrades
+  instead of hard-freezing. On a 4 GB Pi 4 that is not optional.
+
+## Related
+
+- `deploy/nixos/README.md` — the x86 fleet pipeline this mirrors
+- `docs/machine-installation.md` — why images rather than `nixos-install`
+- `deploy/nixos/hardware/raspberry-pi-{4,5}.nix` — the per-board glue
+- `flake.nix` — `piBoards`, `piBaseModules`, `mkPiInstalled`, `mkPiImage`
diff --git a/flake.nix b/flake.nix
index 705fd4c..029a879 100644
--- a/flake.nix
+++ b/flake.nix
@@ -308,14 +308,30 @@
         fiatCode = fiatCodeForModel.${machineModel} or "USD";
       };
 
+      # Supported Pi boards, keyed by machine model. Each pairs the
+      # nixos-hardware board module (kernel, firmware, bootloader, device tree)
+      # with our own hardware glue (validator serial, kiosk display, no-suspend).
+      # Everything else in the Pi runtime is board-agnostic.
+      piBoards = {
+        rpi5 = {
+          hardware = nixos-hardware.nixosModules.raspberry-pi-5;
+          glue = ./deploy/nixos/hardware/raspberry-pi-5.nix;
+        };
+        rpi4 = {
+          hardware = nixos-hardware.nixosModules.raspberry-pi-4;
+          glue = ./deploy/nixos/hardware/raspberry-pi-4.nix;
+        };
+      };
+
       # Shared module list (everything EXCEPT the root fs and the sd-image
       # builder). atm-app/fiatCode are threaded in so both products share one
       # evaluated app closure.
-      piBaseModules = { machineModel, atm-app, fiatCode }: [
-        nixos-hardware.nixosModules.raspberry-pi-5
+      piBaseModules = { machineModel, atm-app, fiatCode }:
+        let board = piBoards.${machineModel}; in [
+        board.hardware
         ./deploy/nixos/configuration.nix
         ./deploy/nixos/bitspire-atm.nix
-        ./deploy/nixos/hardware/raspberry-pi-5.nix
+        board.glue
         ({ config, lib, pkgs, pkgs-unstable, ... }: {
           services.bitspire = {
             enable = true;
@@ -471,6 +487,12 @@
         rpi5-installed = mkPiInstalled "rpi5";
         rpi5-image = mkPiImage "rpi5";
 
+        # Raspberry Pi 4 (aarch64) — same runtime and products as rpi5, on the
+        # previous-generation board (see deploy/nixos/hardware/raspberry-pi-4.nix
+        # for what differs). 4 GB minimum for Electron; 8 GB comfortable.
+        rpi4-installed = mkPiInstalled "rpi4";
+        rpi4-image = mkPiImage "rpi4";
+
         # USB-bootable variant of batm3-installed. This is the config the
         # flashed USB stick actually runs — distinct fs labels so stage-1 can't
         # latch the internal drive, nofail /boot, no growPartition, autoUpgrade
@@ -685,13 +707,16 @@
         iso = self.nixosConfigurations.douro.config.system.build.isoImage;
       };
 
-      # ── Packages (aarch64-linux — Raspberry Pi 5 build) ───────────
-      # Flashable SD image for the Pi 5. Build on an aarch64 builder (native Pi
-      # / arm box / `boot.binfmt` emulation on this x86 host):
+      # ── Packages (aarch64-linux — Raspberry Pi builds) ────────────
+      # Flashable SD images for the Pi boards. Build on an aarch64 builder
+      # (native Pi / arm box / `boot.binfmt` emulation on this x86 host):
       #   nix build .#packages.aarch64-linux.sd-image-rpi5
+      #   nix build .#packages.aarch64-linux.sd-image-rpi4
       packages.aarch64-linux = {
         sd-image-rpi5 = self.nixosConfigurations.rpi5-image.config.system.build.sdImage;
         atm-app-rpi5 = mkAtmAppAarch64 { model = "rpi5"; fiatCode = "USD"; };
+        sd-image-rpi4 = self.nixosConfigurations.rpi4-image.config.system.build.sdImage;
+        atm-app-rpi4 = mkAtmAppAarch64 { model = "rpi4"; fiatCode = "USD"; };
       };
     }
     //
diff --git a/nix/mkAtmApp.nix b/nix/mkAtmApp.nix
index 68a4c37..272bd33 100644
--- a/nix/mkAtmApp.nix
+++ b/nix/mkAtmApp.nix
@@ -172,6 +172,27 @@ pkgs.stdenv.mkDerivation (finalAttrs: {
     cp -rL "$bcpp_store/node_modules/@serialport/bindings-cpp/prebuilds" $out/node_modules/@serialport/bindings-cpp/prebuilds
     cp "$bcpp_store/node_modules/@serialport/bindings-cpp/package.json" $out/node_modules/@serialport/bindings-cpp/package.json
 
+    # bindings-cpp ships prebuilds for every platform it supports: android,
+    # win32, darwin, and linux for several arches in both glibc and musl. Keep
+    # only the one this system can actually load.
+    #
+    # This is load-bearing on aarch64, not just tidiness. On x86_64 autoPatchelf
+    # skipped the foreign prebuilds because their ELF architecture did not match
+    # the host. On aarch64 the android-arm64 prebuild IS the host architecture,
+    # so autoPatchelf tries to patch it and fails hunting for Android's
+    # liblog.so and libc++_shared.so, which do not exist on NixOS. First Pi
+    # build died exactly there.
+    #
+    # Pruning rather than extending autoPatchelfIgnoreMissingDeps: teaching
+    # autoPatchelf to tolerate a binary we never load, for a platform we do not
+    # target, is the wrong shape of fix. The musl entry in that list below is
+    # the same problem solved the other way, and is now redundant.
+    keep_prebuild=${if pkgs.stdenv.hostPlatform.isAarch64 then "linux-arm64" else "linux-x64"}
+    find $out/node_modules/@serialport/bindings-cpp/prebuilds -mindepth 1 -maxdepth 1 \
+      ! -name "$keep_prebuild" -exec rm -rf {} +
+    rm -f $out/node_modules/@serialport/bindings-cpp/prebuilds/*/*.musl.node
+    echo "serialport prebuilds kept: $(ls $out/node_modules/@serialport/bindings-cpp/prebuilds)/$(ls $out/node_modules/@serialport/bindings-cpp/prebuilds/"$keep_prebuild")"
+
     copy_pnpm_pkg @serialport/bindings-interface $out/node_modules/@serialport/bindings-interface
     copy_pnpm_pkg @serialport/binding-mock $out/node_modules/@serialport/binding-mock
 
diff --git a/packages/hal/src/validators/apex/__tests__/apex-rs232.test.ts b/packages/hal/src/validators/apex/__tests__/apex-rs232.test.ts
index 95c1607..bafb50d 100644
--- a/packages/hal/src/validators/apex/__tests__/apex-rs232.test.ts
+++ b/packages/hal/src/validators/apex/__tests__/apex-rs232.test.ts
@@ -6,6 +6,7 @@ import {
   parseStatus,
   parseResponse,
 } from '../apex-rs232.js'
+import { denomForChannel } from '../denominations.js'
 
 // Cassette-present bit; OR it into event bytes so parseStatus doesn't short to
 // 'stackerOpen'.
@@ -20,6 +21,27 @@ describe('Apex RS-232 protocol', () => {
     })
   })
 
+  describe('computeChecksum spans the frame, not a fixed range', () => {
+    it('matches the two reset frames the spec spells out literally', () => {
+      // Rev G gives these verbatim, checksum included, so they are the only
+      // ground truth available for the XOR range without hardware.
+      const a = Buffer.from([0x02, 0x08, 0x61, 0x7f, 0x7f, 0x7f, 0x03, 0x16])
+      const b = Buffer.from([0x02, 0x08, 0x60, 0x7f, 0x7f, 0x7f, 0x03, 0x17])
+      expect(computeChecksum(a)).toBe(0x16)
+      expect(computeChecksum(b)).toBe(0x17)
+    })
+
+    it('covers the six data bytes of an 11-byte reply', () => {
+      // The reply is longer than the host frame. A checksum hardcoded to the
+      // host range silently mis-validates every reply the acceptor sends.
+      const reply = [0x02, 0x0b, 0x20, 0x01, 0x10, 0x00, 0x00, 0x12, 0x34, 0x03, 0x00]
+      let want = 0
+      for (let i = 1; i <= 8; i++) want ^= reply[i] as number
+      reply[10] = want
+      expect(computeChecksum(Buffer.from(reply))).toBe(want)
+    })
+  })
+
   describe('buildFrame', () => {
     it('lays out the 8-byte poll frame with the ACK bit and checksum', () => {
       const f = buildFrame(0, 0x7f, 0x00)
@@ -32,6 +54,14 @@ describe('Apex RS-232 protocol', () => {
       expect(f[7]).toBe(0x66)
     })
 
+    it('carries escrow, stack and return in the command byte', () => {
+      // Rev G BYTE 1: bit 4 escrow enable, bit 5 stack, bit 6 return. Escrow
+      // is an enable held across polls, so it rides alongside the action bit.
+      expect(buildFrame(0, 0x7f, 0x10)[4]).toBe(0x10)
+      expect(buildFrame(0, 0x7f, 0x30)[4]).toBe(0x30)
+      expect(buildFrame(0, 0x7f, 0x50)[4]).toBe(0x50)
+    })
+
     it('sets the stack command bit (0x20) in the command byte', () => {
       const f = buildFrame(0, 0x7f, 0x20)
       expect(f[4]).toBe(0x20)
@@ -82,11 +112,14 @@ describe('Apex RS-232 protocol', () => {
   })
 
   describe('parseResponse', () => {
-    const usd = [1, 5, 10, 20, 50, 100]
-    const resolve = (ch: number) => usd[ch - 1] ?? null
+    // Resolve through the real module, not a hand-copied array. The previous
+    // local copy duplicated the shipping table's off-by-one and so asserted
+    // the bug instead of catching it.
+    const resolve = (ch: number) => denomForChannel('USD', ch)
 
     it('resolves the escrowed note denomination from the credit channel', () => {
-      // state=escrowed, event=present, credit=channel 4 ($20)
+      // state=escrowed, event=present, credit=channel 4. Per spec Rev G the
+      // USD channel order is $1 $2 $5 $10 $20 $50 $100, so channel 4 is $10.
       const frame = Buffer.from([
         0x02,
         0x0b,
@@ -102,7 +135,7 @@ describe('Apex RS-232 protocol', () => {
       ])
       const r = parseResponse(frame, resolve)
       expect(r.status).toBe('billsRead')
-      expect(r.bill?.denomination).toBe(20)
+      expect(r.bill?.denomination).toBe(10)
     })
 
     it('returns no bill when no channel is credited', () => {
@@ -124,8 +157,9 @@ describe('Apex RS-232 protocol', () => {
       expect(r.bill).toBeUndefined()
     })
 
-    it('yields a null denomination for an unmapped channel', () => {
-      // channel 7 not present in the 6-entry USD table
+    it('maps the top channel to the largest note', () => {
+      // Channel 7 is $100. It read as unmapped while the table omitted $2,
+      // which is exactly the shift this test now pins down.
       const frame = Buffer.from([
         0x02,
         0x0b,
@@ -140,7 +174,25 @@ describe('Apex RS-232 protocol', () => {
         0x00,
       ])
       const r = parseResponse(frame, resolve)
-      expect(r.bill?.denomination).toBeNull()
+      expect(r.bill?.denomination).toBe(100)
+    })
+
+    it('pins the whole USD channel order from the spec', () => {
+      // Rev G, BYTE 2 bits 3-5: 001=$1 010=$2 011=$5 100=$10 101=$20
+      // 110=$50 111=$100. A note credited at the wrong value is silent and
+      // costs real money, so the full mapping is asserted rather than sampled.
+      expect([1, 2, 3, 4, 5, 6, 7].map((ch) => denomForChannel('USD', ch))).toEqual([
+        1, 2, 5, 10, 20, 50, 100,
+      ])
+    })
+
+    it('has no denomination for channel 0, which means no note', () => {
+      expect(denomForChannel('USD', 0)).toBeNull()
+    })
+
+    it('has no denomination when the currency is unknown', () => {
+      expect(denomForChannel(null, 3)).toBeNull()
+      expect(denomForChannel('ZZZ', 3)).toBeNull()
     })
   })
 })
diff --git a/packages/hal/src/validators/apex/apex-rs232.ts b/packages/hal/src/validators/apex/apex-rs232.ts
index d8bcb7e..e8fc12f 100644
--- a/packages/hal/src/validators/apex/apex-rs232.ts
+++ b/packages/hal/src/validators/apex/apex-rs232.ts
@@ -6,23 +6,27 @@
  *
  * Implemented from Pyramid's PUBLIC protocol facts only — the wire format,
  * bit masks and serial parameters documented in Pyramid's "RS-232 Serial
- * Interface Specification" (https://pyramidacceptors.com/pdf/RS_232.pdf) and
- * mirrored by their published integrator samples. No third-party (or
- * lamassu-machine) source is copied; the byte layout below is a functional
+ * Interface Specification", document RS_232, Rev G 12/03/14. No third-party
+ * (or lamassu-machine) source is copied; the byte layout below is a functional
  * spec, re-expressed for bitSpire under AGPL.
  *
+ * The interface is Mars/MEI GL5-compatible, which is why it looks so much like
+ * the EBDS driver next door. The acceptor is a pure slave: it answers polls and
+ * never speaks first. Polls must not fall more than 5s apart or the acceptor
+ * may dump an escrowed note and stop accepting until the host resumes.
+ *
  * Frame (host → acceptor), fixed 8 bytes:
  *   [0] STX   0x02
  *   [1] LEN   0x08
- *   [2] CTRL  0x10 | ack        (ack toggles 0↔1 every message)
- *   [3] ENA   denomination enable bitmask (0x7F = all, 0x00 = none)
- *   [4] CMD   0x00 base; | 0x20 stacks the escrowed note
- *   [5] RSVD  0x00
+ *   [2] CTRL  msg type 1 (master) in bits 4-6, ack in bit 0 (toggles every message)
+ *   [3] ENA   BYTE 0 — per-note enable bits: bit 0 = note 1 … bit 6 = note 7
+ *   [4] CMD   BYTE 1 — bit 4 escrow enable, bit 5 stack, bit 6 return
+ *   [5] RSVD  BYTE 2 — reserved, 0x00
  *   [6] ETX   0x03
- *   [7] CHK   XOR of bytes [1..5]
+ *   [7] CHK   XOR of all bytes except STX, ETX and itself
  *
- * Frame (acceptor → host), length-prefixed like the host frame; the fields
- * this driver consumes:
+ * Frame (acceptor → host), 11 bytes — STX, LEN, CTRL, six data bytes, ETX,
+ * CHK. The fields this driver consumes:
  *   [3] STATE bits  1=idling 2=accepting 4=escrowed 8=stacking
  *                   16=stacked 32=returning 64=returned
  *   [4] EVENT bits  0x01=cheated 0x02=rejected 0x04=jammed
@@ -39,8 +43,11 @@ const STX = 0x02
 const ETX = 0x03
 const HOST_FRAME_LEN = 0x08
 
-// Host command byte (frame[4])
-const CMD_STACK = 0x20
+// Host command byte (frame[4]) — spec Rev G, "Data Fields for Messages Sent
+// By the Master", BYTE 1.
+const CMD_ESCROW = 0x10 // bit 4: set to 1 to ENABLE escrow mode
+const CMD_STACK = 0x20 // bit 5: stack the escrowed note
+const CMD_RETURN = 0x40 // bit 6: return the escrowed note
 
 // Response STATE byte (frame[3]) bit masks
 const STATE_IDLING = 0x01
@@ -90,10 +97,18 @@ export interface ApexRs232Config {
 // Pure functions — checksum, frame building, response parsing
 // ---------------------------------------------------------------------------
 
-/** XOR checksum over bytes [1..5] (LEN through RSVD), matching the host frame. */
-export function computeChecksum(frame: number[] | Buffer): number {
+/**
+ * XOR checksum over every byte except STX, ETX and the checksum itself — spec
+ * Rev G: "calculated on all bytes (except: STX, ETX and the checksum byte
+ * itself)". For the 8-byte host frame that is bytes 1..5; for the 11-byte
+ * reply it is bytes 1..8, which is why this is derived from the length rather
+ * than hardcoded. Confirmed against the two reset frames the spec spells out
+ * literally (02 08 61 7f 7f 7f 03 16 and 02 08 60 7f 7f 7f 03 17).
+ */
+export function computeChecksum(frame: number[] | Buffer, length?: number): number {
+  const n = length ?? frame.length
   let cs = 0x00
-  for (let i = 1; i <= 5; i++) cs ^= frame[i] ?? 0
+  for (let i = 1; i <= n - 3; i++) cs ^= frame[i] ?? 0
   return cs
 }
 
@@ -278,18 +293,25 @@ export class ApexRs232 extends EventEmitter {
 
   /** Send one poll, carrying the current mask + latched escrow action. */
   poll(): void {
-    // 'return' is expressed by disabling all channels while a note is escrowed,
-    // which makes the acceptor hand the note back (Apex has no distinct return
-    // opcode). 'stack' asserts CMD_STACK. Both are re-asserted until the device
-    // leaves escrow. NOTE: verify the return-by-disable behaviour on the 7600
-    // during bench bring-up; some firmware returns only on escrow timeout.
-    let enableByte = this.enabledMask
-    let cmdByte = 0x00
-    if (this.pendingAction === 'stack') cmdByte = CMD_STACK
-    else if (this.pendingAction === 'return') enableByte = 0x00
+    // Escrow is asserted on EVERY poll. It is an enable bit, not a one-shot:
+    // with it clear the acceptor never stops at escrow, so the host is never
+    // offered the stack/return decision and notes are banked before anything
+    // has validated them. This driver's whole FSM is built around that
+    // decision point.
+    //
+    // Stack and return are the spec's own bits. An earlier version expressed
+    // return by zeroing the enable mask, on the assumption that the Apex had
+    // no return opcode; it has one, and disabling channels mid-escrow is not
+    // what it means.
+    //
+    // Both are re-asserted until the device leaves escrow, so a single dropped
+    // frame cannot strand a note.
+    let cmdByte = CMD_ESCROW
+    if (this.pendingAction === 'stack') cmdByte |= CMD_STACK
+    else if (this.pendingAction === 'return') cmdByte |= CMD_RETURN
 
     this.ack ^= 0x01
-    this.serial?.write(buildFrame(this.ack, enableByte, cmdByte))
+    this.serial?.write(buildFrame(this.ack, this.enabledMask, cmdByte))
   }
 
   stack(): void {
@@ -331,13 +353,20 @@ export class ApexRs232 extends EventEmitter {
       this.poll()
       return
     }
-    // Checksum is validated leniently: a mismatch is logged once but the frame
-    // is still parsed. Pyramid's published host samples don't verify the reply
-    // checksum, and the exact XOR range for the *reply* isn't confirmable from
-    // the (scanned) spec — so we don't want a wrong assumption to blackhole
-    // every otherwise-valid frame. Tighten to a hard drop once verified on hw.
-    if (frame[len - 1] !== computeChecksum(frame)) {
-      console.warn('[APEX] reply checksum mismatch (parsing anyway pending hw verification)')
+    // The XOR range is now confirmed from the spec, so a mismatch is a hard
+    // drop rather than the previous parse-anyway. A corrupted frame carries a
+    // denomination field, and crediting a note from a frame we know is damaged
+    // is the one outcome worth refusing outright. The raw bytes are logged so
+    // a systematic framing error is still diagnosable rather than silent.
+    const want = computeChecksum(frame, len)
+    if (frame[len - 1] !== want) {
+      console.warn(
+        `[APEX] reply checksum mismatch: got ${frame[len - 1]?.toString(16)} ` +
+          `want ${want.toString(16)} — frame dropped: ${frame.toString('hex')}`
+      )
+      this.emit('badFrame')
+      this.poll()
+      return
     }
 
     const result = parseResponse(frame, this.denomForChannel)
diff --git a/packages/hal/src/validators/apex/denominations.ts b/packages/hal/src/validators/apex/denominations.ts
index 9874da4..95410ed 100644
--- a/packages/hal/src/validators/apex/denominations.ts
+++ b/packages/hal/src/validators/apex/denominations.ts
@@ -1,19 +1,29 @@
 /**
  * Pyramid Apex Denomination Tables
  *
- * The Apex RS-232 reply reports a 1-based CREDIT CHANNEL (1..7), not a value —
- * the channel→value mapping is fixed by the bill dataset programmed into the
- * acceptor's firmware for its country. The arrays below are ascending value
- * lists indexed by (channel - 1); they MUST match the dataset flashed on your
- * specific Apex 7600, or a credited note will be booked at the wrong value.
- * Verify against the unit's configuration card during bring-up.
+ * Indexed by (credit channel - 1). The channel is NOT an index into "the notes
+ * this unit happens to have enabled" — it is a fixed protocol constant.
+ * Pyramid's RS-232 spec (Rev G, "Data Fields for Messages sent by the Slave",
+ * BYTE 2 bits 3-5) fixes the USD mapping:
  *
- * Defaults follow Pyramid's standard datasets (US = $1/$5/$10/$20/$50/$100;
- * $2 channel omitted as it's rarely enabled).
+ *   001 = $1    010 = $2    011 = $5    100 = $10
+ *   101 = $20   110 = $50   111 = $100
+ *
+ * $2 occupies channel 2 whether or not the unit accepts $2 notes. An earlier
+ * version of this table omitted it as "rarely enabled", which shifted every
+ * larger note down one slot: a $5 credited as $10, a $20 as $50, a $50 as
+ * $100, and a $100 as nothing at all. Never drop an unused channel from these
+ * arrays — pad it instead.
+ *
+ * For non-USD the spec says only "Foreign currencies are in sequential order
+ * as note 1-7", so channel N is the Nth note type of whatever dataset is
+ * flashed on the unit. The orderings below are the conventional ascending sets
+ * and are UNVERIFIED against a real configuration card. Check the card before
+ * a machine takes money in any of them.
  */
 
 export const denominations: Record<string, number[]> = {
-  USD: [1, 5, 10, 20, 50, 100],
+  USD: [1, 2, 5, 10, 20, 50, 100],
   EUR: [5, 10, 20, 50, 100, 200, 500],
   GBP: [5, 10, 20, 50],
   CAD: [5, 10, 20, 50, 100],
diff --git a/packages/hal/src/validators/apex/index.ts b/packages/hal/src/validators/apex/index.ts
index 463ee06..5d98c03 100644
--- a/packages/hal/src/validators/apex/index.ts
+++ b/packages/hal/src/validators/apex/index.ts
@@ -48,6 +48,11 @@ export class ApexValidator extends EventEmitter implements BillValidator {
   constructor(config: ValidatorConfig) {
     super()
     this.config = config
+    // Seed from config, as id003 effectively does by threading config.fiatCode
+    // into its rs232 config. Nothing in the app calls setFiatCode(), so a
+    // driver that relies on it alone resolves every credit channel to null and
+    // rejects every note. setFiatCode() stays available as a later override.
+    this.fiatCode = config.fiatCode ?? config.rs232.fiatCode ?? null
     this._throttledError = throttle((err: Error) => this.emit('error', err), 2000)
   }
 
@@ -76,7 +81,20 @@ export class ApexValidator extends EventEmitter implements BillValidator {
     this.fsm.on('billsAccepted', () => this.emit('billsAccepted'))
     this.fsm.on('billsRead', (bill: { denomination: number | null; code: string }) => {
       if (!bill.denomination) {
-        console.log('[APEX] Bill rejected: unsupported/unmapped channel')
+        // Say WHICH of the two causes this is. "unmapped channel" alone reads
+        // as a hardware/dataset mismatch and sent us looking at DIP switches
+        // when the real cause was a null fiat code rejecting every note.
+        if (!this.fiatCode) {
+          console.error(
+            '[APEX] Bill rejected: no fiat code set on the driver, so NO channel ' +
+              'can resolve to a value. Every note will be returned until this is fixed.'
+          )
+        } else {
+          console.log(
+            `[APEX] Bill rejected: channel ${bill.code || '?'} is not mapped in the ` +
+              `${this.fiatCode} dataset — check the acceptor's configuration card`
+          )
+        }
         this.rs232?.reject()
         return
       }
@@ -92,6 +110,13 @@ export class ApexValidator extends EventEmitter implements BillValidator {
     this.fsm.on('standby', () => this.emit('standby'))
     this.fsm.on('error', (err: Error) => this.emit('error', err))
 
+    if (!this.fiatCode) {
+      console.error(
+        '[APEX] starting with NO fiat code — denomination lookup will return null ' +
+          'for every credit channel and the acceptor will reject every note.'
+      )
+    }
+
     this.rs232.open((err) => {
       if (err) return cb(err)
       this.rs232!.reset() // start disabled
diff --git a/packages/hal/src/validators/ebds/index.ts b/packages/hal/src/validators/ebds/index.ts
index 44dd540..52a749b 100644
--- a/packages/hal/src/validators/ebds/index.ts
+++ b/packages/hal/src/validators/ebds/index.ts
@@ -53,6 +53,11 @@ export class EbdsValidator extends EventEmitter implements BillValidator {
   constructor(config: ValidatorConfig) {
     super()
     this.config = config
+    // Seed from config, as id003 effectively does by threading config.fiatCode
+    // into its rs232 config. Nothing in the app calls setFiatCode(), so a
+    // driver that relies on it alone resolves every credit channel to null and
+    // rejects every note. setFiatCode() stays available as a later override.
+    this.fiatCode = config.fiatCode ?? config.rs232.fiatCode ?? null
     this._throttledError = throttle((err: Error) => this.emit('error', err), 2000)
   }