Skip to content

LordOS — a free, fast and secure GNU/Linux distribution based on Debian, with its own desktop environment: display server and Wayland compositor in C, a React UI engine on QuickJS, and private native apps.

26.7 KB
Raw Blame

Validating the paper

The technical paper (https://lordos.org/paper/, docs/PAPER.md) makes claims about LordOS. This page was written against its 0.8.0 edition and lists each claim with how it is checked today and its state in 0.8.5, so anyone can see what is proven, what was measured once and what is not tested. The rows marked Outdated were corrected in the paper's 0.8.5 edition (2026-10-26; published with the 0.8.5-beta release on 2026-10-27), which also separates whole-system memory from a process's RSS (§7.1) and points to versioned results (§11.5, docs/RESULTS.md).

How a claim is checked

Mark Meaning
Test An automated test in the repository (make -C server test, test-full or node suites) fails if the claim stops being true.
Check System Information → Checks verifies it on the user's own computer: configured (the mechanism is there) or effective (it was tried and gave the right answer).
Measured A number measured by hand on a named machine; nothing repeats it yet. LordBench (0.8.6) is meant to.
Untested The paper itself says it wasn't tested (real hardware).
Outdated → corrected in 0.8.5 No longer true in 0.8.5: the paper must change.
Design A design rule, checked by review and AGENTS.md, not by a program.

§1–§2 Vision and architecture

§ Claim How State
1.2 JavaScript has no API to run programs, with three narrow exceptions Design (AGENTS.md); the exceptions grew since: Xwayland (0.8.3), the network connectors and the blocklist downloader (0.8.5) Outdated → corrected in 0.8.5 (§1.2 now lists them, the Store's download child included)
1.2 GPU composition optional, the CPU path always works Test (canvas-test, CPU path in every test); Check gpu says which path and why Holds
1.2 The compositor never waits on helpers or apps Design (non-blocking sockets) Design
2.2 lord-server runs as the user Check serverUser (configured) Holds
2.4 lord-netd only takes net-* actions from uid 0 peers Design; net-test covers the actions Partly tested
2.4 lord-fand writes only fans its scan found and restores them on stop or crash Test (fan-test, fake sysfs); Measured on the Pi 5 (§7.4) Holds
2.4 (0.8.6, in development) With no default chosen, a file opens with the system's default or the first installed app that declares its type, shown as "Automatic" Test (defaults-test, fake XDG tree); checked once in an isolated session on VM 116 (2026-10-29) Holds
2.4 Display changes revert after 15 s without confirmation Manual Not automated
3.6 Passwords are checked by the root process, not the login screen Design; remote-sec-test covers the parsers Design

§3 Compositor and remote desktop

§ Claim How State
3.1 Modesetting with our own DRM ioctls, the card with connectors Design (kms.c, no libdrm) Design
3.2 CPU path without a 3D card; software OpenGL falls back to it Check gpu names the reason (no-3d, software…) Holds (VM 116 shows it)
3.2 GPU path tested only with Mesa's software renderer Untested on AMD/Nvidia/Intel Untested
3.3 Touchpad gestures replayed in tests Test (tp-test, tests/touchpad-sim) Holds
3.4 Only libwayland-server plus generated protocol code Design Design
3.4 Xwayland not supported Xwayland launches on demand since 0.8.3 Outdated → corrected in 0.8.5
3.4 No screen-capture protocol offered to apps Design Design
3.5 Scale 100–200 % per display; two virtual heads tested, real monitors not Manual (VM 116) Untested on real monitors
3.6 Remote desktop off by default, port 5673 Check remoteClosed (effective: /proc/net) Holds
3.6 TLS 1.2+, self-signed P-256; cookie Secure, HttpOnly, SameSite=Strict Test (remote-sec-test, remote-fuzz) Holds
3.6 Failed logins wait 1, 2, 4… s up to 5 min Since 0.8.3: 3 failures block the source 15 min → 1 h → 24 h, answers at a fixed 1.25 s Outdated → corrected in 0.8.5
3.6 Screen updates are PNG still and JPEG moving WebRTC video (H.264, VP8) since 0.8.4 Outdated → corrected in 0.8.5
3.6 No audio Sound since 0.8.4 (Opus) Outdated → corrected in 0.8.5

§4–§5 UI engine and apps

§ Claim How State
4.2 React 19.3.1, QuickJS 2025-04-26; default theme 367 KB Build output Size changed → corrected in 0.8.5 (about 1.8 MB, 540 KB compressed)
4.2 Shaping identical to HarfBuzz on 409 of 409 sentences Test (shape-test, hb_compare.py in test-full) Holds
4.2, 7.4 NEON loops bit-identical to C Test (canvas-test on the Pi) Holds
4.4 English, Spanish, Portuguese; keys and brand names checked Test (tools/i18n-check.mjs) Holds
5.1 About thirty native apps Count 34 in apps/, System Information and its hidden probe included; the 0.8.5 edition says thirty-one (without the live installer and the hidden probe)
5.2 .lpk limits: 2,000 files, 64 MB Test (pkg-test, sign-test) Holds
5.3 InRelease signatures, plans equal to apt's in 12 cases Test (pkg-test, engine-test) Holds
5.3 Downloads over HTTP even for HTTPS sources lord-tls (0.8.5) fetches https:// sources over HTTPS; the Store's update and installs used it on VM 116 and the VMware machine Outdated → corrected in 0.8.5 (2026-10-27: §5.3 now says HTTPS since 0.8.5)
5.3 AI category: vendor repositories added with the pinned key; llama.cpp downloads models with -hf and runs on the CPU Measured: VMware machine, 2026-10-27 (llama.cpp from the Store, llama-server -hf ggml-org/Qwen3-0.6B-GGUF:Q8_0, about 33 tokens/s on the CPU); VM 116 (Cursor's and VS Code's sources added, apt update and the engine's update clean). GPU through Vulkan Holds; GPU Untested
5.4 Linux apps not sandboxed by default Still true; restrictions per app since 0.8.3 Holds, needs a note
5.5 WebAssembly 2.0 without SIMD, 26,290 of 26,290 assertions Test (wasm-test) Holds
5.5 WebAssembly 3–5× faster than JS on the example Measured once Measured
5.6 Install script checks the Ed25519 signature and SHA-256 Manual Not automated

§6 Security

§ Claim How State
6.4 Landlock ABI up to 6 for .lpk apps Check landlock (configured) + Sandbox Probe appSandbox (effective) Holds
6.4 Seccomp denies ptrace, mount, exec…; no new privileges; no capabilities Sandbox Probe (effective: each tried; NoNewPrivs, Seccomp 2, CapEff 0 read by lord-server); Test (sandbox-test) Holds
6.4 Per-app cgroup: memory.high 192 MB, memory.max 512 MB, one core Sandbox Probe appScope (memory.max) Holds in a session; memory.high and CPU not shown yet
6.4 No network permission: no internet sockets, no TCP Sandbox Probe (netInet refused) Holds
6.4 sandbox-test 14/14 and 14/14 with network Test (test-full) Holds
6.4 Malicious test app fails at everything Test (tests/lpk-evil) Holds
6.4, 7.1 Isolated app ~13.7 MB RSS, starts in 106 ms, sandbox in 0.89 ms (VM 122) Measured once Measured
6.5 The broker checks every request in C Sandbox Probe broker (effective: four requests refused); Test (broker-restrict-test, brokermail-test) Holds
6.6 Seven network eBPF programs plus device and LSM programs Check netperm (configured); Test (npverify-test, kernel, VM) Holds where installed (VM 116 lacks it)
6.6 480,000 random connections, 0 differences Test (npcompile-test) Holds
6.6 ~0.5 µs per UDP connect, no per-packet cost Measured once Measured
6.6 A denied app can't connect Not yet effective on the user's machine Pending (Checks)
6.7 Helpers: no sockets, no exec, seccomp on every thread, exit if it fails Sandbox Probe helperSandbox (effective); Test (sandbox-test helper, 17/17) Holds
6.8 Ed25519 and SHA-256 against NIST and RFC 8032: 55 of 55 Test (sign-test) Holds
6.8 Official .lpk packages signed with a development key server/src/sec/keys.h (checked 2026-10-26: only the development key) Holds; re-check before publishing
6.9 Release key fingerprint, SBOM in SPDX 2.3 Release scripts Manual
6.9 Every bundle verified against the signed manifest Check code (effective: this session's bundles) Holds
6.10 Vault: Argon2id, XChaCha20-Poly1305, 1,165 of 1,165 tamper tests Test (vault.mk suites) Holds
6.10 FIDO2 keys untested on real hardware — Untested
6.11 Protected windows black in screenshots, recordings, remote desktop Test (media-test covers capture paths) Partly tested
6.12 Aligned with ISO/IEC 29147, 30111, 5230, 18974; no audit Documents Design

§7 Performance

§ Claim How State
7.1 Live ISO idle 465 MB, installed idle 487 MB, CPU 99 % idle Measured once (VM 117, 120) Measured → LordBench
7.1 Boot 16.3 s with UEFI Measured once Measured → LordBench
7.1 lord-server without 3D ~32 MB RSS; session ~50 MB, login 41 MB Measured once; Check memory shows the current number (VM 116 today: ~100 MB in a headless session with apps open) Re-measure at idle
7.2 Window drag 1280×800 CPU path 6.7–6.9 ms, ~120 fps Measured once (VM 118) Measured → LordBench
7.2 Video: lord-server 42 % → 13 %, fullscreen 92 % → 15 % of a core Measured once Measured
7.1 — New in the 0.8.5-beta edition: the released package measured on a VMware VM updated from 0.8.4.1, whole system 592 MB used and the first frame 7.4 s after the kernel started (results/0.8.5-beta/vmware-fusion, RESULTS.md) Measured, versioned results (§7.1)
7.2 — New in 0.8.5: remote desktop CPU reference (REMOTE.md, "Measurements (0.8.5, CPU reference)") Added in the 0.8.5 edition (§7.2)
7.2 — New, 0.8.6 in development: dragged windows' rounded corners no longer turn black (the drag layer's corner restore mixes both pixels alpha included); a drag frame repaints only what changed, pixel-identical except an approved 1-level change at rounded corners below the window; GPU drag composes the window with the CPU's exact arithmetic after a self-test (WINDOWS.md, GPU.md) Test (engine-test: test_drag_corners against a full repaint on both paths, test_drag_exact's 2,417 golden frame hashes per path, the gpu-drag path; canvas-test: GPU drag's arithmetic over all 8,421,376 valid combinations, x86_64 only)
7.3 GPU benefits on real cards not measured — Untested
7.4 Pi 5: build 1 min 02 s, tests 23/0/0, NEON 2.0–2.3×, idle 36 MB, fan behaviour Measured once on the Pi Measured
7.4 — New, 0.8.6 in development: with a 1920×1080 60 Hz monitor, login screen 5.99 s after the kernel starts; desktop 59.4 fps alone (p99 13.0 ms, 1 % late) and 24 fps under a low-priority all-core load (96.5 % late); 3 minutes sustained at 5.9 GFLOP/s, 55.6 °C; software video decoding per codec. LordBench on the Pi (BENCHMARK.md, "Built so far") Measured with LordBench (one run each); the load result led to two fixes (compositor CPU weight: 31–33 % → 7–8 % late without the monitor; with the real-time priority, measured on the monitor: 24 → 28.9–30.5 fps, still 94–98 % late, the frame's build time under load still to explain)
7.4 — New, 0.8.6 in development: LordBench's square now turns; with it, the frame-cost work took the Pi 5 with its monitor from 53–58 to 59.8–60 fps alone and from 24–25 to about 31 fps under load; the frame pipeline (+2–5 fps under load, 17–29 ms more pointer latency) and GPU drag (no measurable gain) were measured and are off by default; the remaining cost under load is the GPU's time per frame (1.5 → 14–16 ms, no clock or thermal throttling; memory contention likely, not proven) (BENCHMARK.md, "The drag frame's cost") Measured with LordBench on the Pi 5 (2–3 runs each, 2026-10-28; final run 2026-10-29: both off 59.9 / 30.5–31.1 fps, pointer to screen p50 25.7–25.9 / 48.5–49.0 ms; GPU drag on 59.9 / 30.8–31.6; the pipeline 60.0 / 33.0–36.2 fps, pointer 42 / 78 ms; vcgencmd every 2 s)
7.5 No measurements on real PCs — Untested
7.5 — New, 0.8.6 in development: LordBench's interface workloads on VM 116 (layout and paint 37.0/39.9 fps, shaping 619/3,595 runs/s; compared only within the same LordOS version) and the helper-sandbox self-test (six tries, all refused on VM 116: seccomp EPERM, Landlock EACCES) (BENCHMARK.md, "Built so far") Measured with LordBench (one Full and four Quick runs); on the Pi 5 two Quick runs from a terminal (2026-10-29): 39.3–39.4 / 45.4 fps, 630–632 / 2,522–2,527 runs/s, not yet in its reference (a Full run)
7.5 — New, 0.8.6 in development: on a Raspberry Pi LordBench watches the firmware's under-voltage alarm (the only flag the LordOS kernel exposes), get_throttled on older kernels and 80 °C as the firmware's limit (a peak test stops as throttled); System Information's Raspberry Pi section reads the same through hw/rpifw.c and leaves out what the kernel doesn't show (BENCHMARK.md, "The Raspberry Pi firmware's throttling") Tested on fake sysfs trees (tests/lordbench_test.c, tests/sysreport_test.c) on VM 116; on the Pi 5 (2026-10-29) the rpi_volt alarm is present and reads 0 with the official supply (LordBench's piFirmware), get_throttled absent; an actual under-voltage not provoked

§10–§11 Limits and verification

§ Claim How State
10.2 Untested on real hardware: AMD/Nvidia GPUs, monitors on PCs (one 1920×1080 monitor tested on the Pi 5), laptops, Bluetooth, printers, scanners, FIDO2, mail providers System Information reports (0.8.7: community reports) Untested
10.2 Secure Boot must be off System Information shows Secure Boot Holds
11.2 make test 23 passed, 1 skipped Today (2026-10-27, VM 116) 23 passed, 0 failed, 12 skipped (the node suites; they pass on a machine with node) Holds
11.2 (0.8.6, in development) every frame compared with a full repaint; seven partial-repaint errors fixed; the 0.8.2 scene on a running desktop found no stale area build/engine-test (test_damage_pipeline, test_damage_fuzz); VM 116 isolated headless session, six runs of 54 checks (docs/MONITOR.md, "The 0.8.2 integration test, repeated") Holds (2026-10-29)
11.2 (0.8.6, in development) a blurred window hidden whole behind an opaque one is no longer repainted with every change near it (a translucent Terminal behind Monitor's moving map: 46 % of the screen per frame → 11–15 % in one box, 3–4 % in several rectangles) build/engine-test (test_hidden_blur_damage); VM 116 --profile and the repaint check in both modes, before and after rebasing onto the frame-cost work (docs/MONITOR.md, "A blurred window hidden behind Monitor") Holds (2026-10-29; rebased build: 11.2 % and 3.3 %)
11.2 30,953 parser checks with mutation fuzzing Tests (remote-fuzz, mail, doc, imager fuzzers) Holds; count to refresh
11.4 Port 5673 closed until enabled Check remoteClosed (effective) Holds

What the paper must change for 0.8.5

State in the 0.8.5 edition (web/content/paper/paper.en.md, checked 2026-10-26) after each item:

  1. §1.2: the approved exceptions are more than three (Xwayland, the network connectors, the blocklist downloader, the Store's download child, the Sandbox Probe's system package). Done; the Sandbox Probe is not listed as an exception there, since it starts through lord.apps.launchPackage(id) (§6.4 describes it).
  2. §3.4: Xwayland is supported on demand (0.8.3). Done.
  3. §3.6: remote login lockout (3 failures; 15 min → 1 h → 24 h; fixed 1.25 s answers); WebRTC video with H.264/VP8 and Opus sound instead of PNG/JPEG and no audio. Done.
  4. §7: re-measure lord-server's memory at idle and the theme's size; add the 0.8.5 remote desktop and CPU path measurements (REMOTE.md, GPU.md). Partly: the theme's size and the 0.8.5 CPU path are in; §7.1 now separates whole-system memory from a process's RSS, but lord-server's idle RSS is still the 0.8.0 number (not re-measured).
  5. §6: mention that anyone can now check the sandbox on their own computer (System Information → Checks). Done (§6.4).

What is still missing to validate everything

Gap Covers Plan
A denied connection, tried on the user's machine §6.6 effective An app with network access and a temporary rule (Checks)
Repeatable performance numbers §7 LordBench (0.8.6, in development: CPU, memory, GPU offscreen, storage, sustained load, software video decoding and encoders, the desktop's frame times and pointer-to-screen latency alone and under load (and the compositor's CPU weight and real-time priority they led to), the boot time per version, the N-boot test, lordos-sysreport bench, the interface layout/paint and shaping workloads and the helper-sandbox self-test are built, with VM 116 and the Pi 5 as references, BENCHMARK.md "Built so far"; idle memory and CPU, hardware video decoding and app start are not; hardware video decoding and the x86 PC reference moved to 0.8.7 on 2026-10-29)
Rendering, 3D and video on every GPU §3.2, §7.2–7.3 The visual test bench below
Every key, the mouse, touchpads, gamepads, and the Wayland protocols §3.3–§3.4 The visual test bench below
Real hardware §10.2 Reports from System Information (0.8.7)

The visual test bench (lordos-testbench)

Phase 1 built (2026-10-26): server/src/testbench/ (make -C server build/lordos-testbench), packaged as /usr/bin/lordos-testbench with packaging/desktop/lordos-testbench.desktop (Test Bench in the launcher). The test card is ported from go-link-console's pkg/testpattern (same author). It shows the card with a live clock and a marker that moves one step per frame, frames a second and frame times, wp_presentation feedback (shown, discarded, refresh, skipped vblanks, zero-copy), the pointer (exact crosshair, eight buttons, wheel in steps and 120ths, axis source), the keyboard (a full ISO/ANSI layout with keypad and navigation keys lit while down and marked once seen, codes of keys not drawn, keysym and text from the compositor's keymap, modifiers and layout, repeat rate and delay) and every global with its version. --auto [seconds] prints the same as JSON. On VM 116 (headless session, CPU path) it found that wp_presentation reports every frame as discarded and no refresh rate when the session has no monitor.

Test Bench on VM 116

Phase 2, the protocol checklist (2026-10-26, checks.c): runs at start (F5 runs it again; Pause freezes the picture for a still capture) and in --auto's JSON. Each protocol is bound at the advertised version and exercised, and a protocol error ends the connection with the failing check named (its first run caught the bench's own mistake, wl_shm.release on version 1, which LordOS's compositor rightly refused). First run on VM 116 (headless session); after the two compositor fixes below: 16 passed, 0 failed:

Check Result
wl_shm formats pass: ARGB8888 and XRGB8888
wl_output modes, zxdg_output layout pass: LORD-1 1280×800 at 60.00 Hz, scale 1, logical 1280×800 at 0,0
fractional scale pass: preferred 1.000
subsurface + viewport pass: shown (its frame callback came back), viewport 48 → 96
xdg popup pass: configured 200×100 and mapped
maximize, fullscreen pass: the configure carries the state, and loses it after unset
linux-dmabuf v4 feedback pass, with a finding: a headless session sends main device 0:0 (none) and 4 formats
clipboard, primary selection round trip failed, then fixed (same day): the compositor didn't send a new selection back to the focused client that set it (wl.c, offer_selection and prim_offer_to returned early for "its own"), which the protocol requires. Now the focused client gets every new selection and dev_set_selection offers it at once; both round trips come back intact
wp_presentation offered, but a headless session discards every frame and gives no refresh (the output's mode says 60 Hz)
cursor shape, relative pointer, gestures bound; they need the real pointer or a touchpad to go green
idle inhibit added (same day): zwp_idle_inhibit_manager_v1; while a shown surface holds an inhibitor, main.c keeps the idle timers (screen off, lock, sleep) from running. The bench creates and releases one
not offered text-input v3 (input methods for Linux apps), xdg-activation, single-pixel buffers, tearing control, content type, tablets, idle notify, drm-lease

Also seen: system shortcuts (F1 opens Help, F4 the launcher, ❖ combinations) never reach the app, so the keyboard test can't light them; the bench should list them as taken by the system.

Phase 3 (2026-10-26, gl3d.c and testbench.c):

  • 3D (G, or --gl): a rotating cube with EGL and OpenGL ES 2 on a subsurface through Mesa's Wayland platform, as a game draws: the renderer, vendor and version, whether it is a hardware GPU, a virtual GPU or software (llvmpipe and the like, i.e. the CPU), the time to draw a frame (glFinish), and its frames' presentation with the zero-copy flag (direct scanout). On VM 116: llvmpipe (LLVM 19.1.7), software, 0.5–8 ms a frame.
  • Input to screen: the time of each key or click (the event's timestamp) to the presentation time of the frame that shows it (wp_presentation, CLOCK_MONOTONIC). In the headless session it says "not measurable": no frame is presented there.
  • Generated video (V): content at 24, 30 or 60 fps with a square that steps once per content frame, and how many refreshes each content frame stayed (24 on 60 Hz should be 2 and 3 alternating).
  • Pointer lock (L): zwp_pointer_constraints_v1.lock_pointer, shown as locked or not.
  • Built without EGL the bench still runs and the 3D row says so.

3D on VM 116 (software renderer)

Phase 4 (2026-10-26): the report in System Information. The bench writes its last report to ~/.local/state/lordos/testbench.json (atomically, after its checklist and on exit); System Information → Checks shows it (Wayland protocols, the 3D renderer, whether frames reach the screen, input-to-screen latency; evaluateBench in apps/sysinfo/checks.js, tested) with Open Test Bench, which launches it like any Linux app.

The Test Bench in System Information

Keys LordOS's compositor keeps (wl.c, wl_key: Print Screen, the media keys, ❖ on its own) are outlined in purple on the bench's keyboard, so a key that never lights isn't mistaken for a broken one.

Shared code (2026-10-26). The measurement pieces moved to server/src/bench/, shared with LordBench (0.8.6): frame times now give p50, p99 and frames that missed the refresh budget (frameMsP50, frameMsP99, missedBudget in the report); build/bench-test (22 checks) found that vblanks skipped were never counted from presentation times when the first frame was at time 0, fixed.

Still to build: pointer confine, gamepads, and a run with the real monitor and GPU (presentation, latency, zero-copy).

The plan

LordOS writes its own compositor, so it needs its own tool to show, on any computer, that what it draws and what it receives are right. The idea is a test-pattern window like a TV test card (and like the test pattern of go-link-console), with live measurements around it:

  • What it is: a native Wayland client in C (tools/testbench/, installed as /usr/bin/lordos-testbench with a .desktop file). It speaks the protocol directly with libwayland-client, the way any app does, so it tests the compositor from the outside. It is launched like any Linux app (lord.apps.launch), so no new exception in lord-server, and it also runs on other compositors for comparison.
  • The pattern (rendering): colour bars, grey steps, a resolution grating (1-pixel lines that show scaling or blurring), a geometry grid and circle (wrong aspect or scaling shows at once), a live clock and a frame counter that changes every frame (a dropped or repeated frame is visible), scale and fractional-scale readout, colour format.
  • Timing: frames per second against the display's refresh, frame-time histogram, missed frames, using wl_surface.frame and wp_presentation feedback (when the frame really reached the screen).
  • 3D: a rotating textured scene with EGL/GLES2 on the GPU (linux-dmabuf), showing the renderer and whether the buffer was scanned out directly or composed; a CPU-only shm mode to compare.
  • Video: a generated moving pattern at 24/30/60 fps (no files, no decoders), to see judder and dropped frames; later the same through lord-media's hardware decoder.
  • Keyboard: a full keyboard drawn on screen; each key lights when its key event arrives, with the keycode, the keysym from the keymap the compositor sent, modifiers, repeat rate and delay; a list of keys never seen, to go through every key of a real keyboard.
  • Mouse and touchpad: pointer position and the hotspot, buttons, scroll (discrete and smooth), relative motion, pointer constraints and locking, gestures (swipe, pinch), the cursor shapes it asks for.
  • Gamepad: shown if the device is readable by the user (evdev); otherwise said so.
  • Protocol checklist (low level): every global the compositor advertises, its version, and a scripted run of each one (create, use, destroy) with pass or fail: wl_shm formats, xdg_wm_base (configure, resize, maximize, fullscreen, popups), xdg-decoration, linux-dmabuf formats and feedback, wp_viewporter, fractional-scale, presentation-time, pointer-constraints, relative-pointer, pointer-gestures, text-input, primary-selection, data device (copy and paste of text and images), idle-inhibit, xdg-activation, cursor-shape. Its output is a report (JSON and Markdown) that System Information can attach, with the GPU and the monitor, so the same run on many computers shows where LordOS's Wayland works and where it doesn't.
  • Automation: lordos-testbench --auto runs the protocol checklist and a timed render without a person, for CI and for the VMs (the input tests then use the remote desktop's injected keys and pointer, as the tests in this repository already do).