diff --git a/desktop/daemon/src/main.rs b/desktop/daemon/src/main.rs
index e946c4e..89deb32 100644
--- a/desktop/daemon/src/main.rs
+++ b/desktop/daemon/src/main.rs
@@ -33,6 +33,8 @@
 //! installed read-only in `/usr/bin`, while the database must be writable and
 //! must survive an upgrade.
 
+mod runner;
+
 use std::path::PathBuf;
 use std::process::ExitCode;
 
@@ -76,6 +78,9 @@ fn main() -> ExitCode {
         );
     }
 
+    let usage = blocker_core::engine::UsageSnapshot::default();
+    let context = blocker_core::engine::PlatformContext::default();
+
     let schedules = match store.load_schedules() {
         Ok(s) => s,
         Err(e) => {
@@ -98,7 +103,75 @@ fn main() -> ExitCode {
         println!("backend: none — blocking cannot work; see docs/LIMITATIONS.md section 3");
     }
 
-    println!("enforcement lands in M2-Linux; see docs/DECISIONS.md for the ADR log");
+    // The event loop, running for real.
+    //
+    // X11 is wired here rather than KWin because the KWin path needs the D-Bus
+    // watcher server — the daemon owning `vu.valiq.blocker` and exporting the
+    // interface `callDBus` targets — which is not built yet. Rather than stub it,
+    // the loop runs on the backend that works today and says which one it picked.
+    // The honest consequence is in the startup line below: on a Wayland session
+    // this sees XWayland clients only.
+    let (tx, rx) = std::sync::mpsc::channel();
+    match blocker_platform_linux::x11::X11Backend::connect() {
+        Ok(backend) => {
+            println!("backend: x11 — native Wayland apps are NOT visible through this");
+            if let Err(e) = backend.subscribe_events() {
+                eprintln!("blocker-daemon: cannot subscribe to X events: {e}");
+                return ExitCode::FAILURE;
+            }
+            let _ = backend.request_active_window();
 
+            // The X11 subscription runs on its own thread and parks in
+            // poll_for_event. The main thread parks in recv_timeout. Neither
+            // ticks, so an idle system costs two blocked threads and no CPU.
+            let quit = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false));
+            let worker_quit = quit.clone();
+            let worker = std::thread::spawn(move || {
+                backend.run(tx, worker_quit.as_ref());
+            });
+
+            let ctx = runner::LoopContext {
+                schedules: &schedules,
+                usage: &usage,
+                platform_context: &context,
+                boot_id: blocker_core::engine::BootId(0),
+                enforcement: blocker_platform_linux::Enforcement::Freeze,
+            };
+            let clock = blocker_platform_linux::SystemMonotonic;
+
+            println!("enforcement: SIGSTOP (configurable); entering the event loop");
+            'events: loop {
+                // The wait IS the next wakeup. An idle system parks here rather
+                // than ticking, which is how the section 5 idle budget is met.
+                let (action, blocked) =
+                    runner::step(&rx, std::time::Duration::from_secs(3600), &ctx, &clock, now());
+                if blocked > 0 {
+                    println!("blocked an app ({action:?})");
+                }
+                if action == runner::Action::Idle {
+                    // The channel disconnected: the X11 watcher thread is gone,
+                    // and nothing can arrive to wake us again. Exiting is
+                    // correct, and better than spinning on a dead channel.
+                    break 'events;
+                }
+            }
+            let _ = worker.join();
+        }
+        Err(e) => {
+            println!("backend: none — {e}");
+            println!("blocking cannot work on this session; see docs/LIMITATIONS.md section 3");
+            return ExitCode::FAILURE;
+        }
+    };
     ExitCode::SUCCESS
 }
+
+/// Civil time, from the platform layer.
+///
+/// A placeholder reading rather than a clock read: the loop's correctness
+/// depends on `next_wakeup` arithmetic, and wiring a real timezone-aware source
+/// is part of the next PR. Deliberately returns a fixed instant so the loop is
+/// obviously not yet time-driven, instead of appearing to work by accident.
+fn now() -> blocker_core::engine::CivilTime {
+    blocker_core::engine::CivilTime { epoch_minutes: 0, utc_offset_minutes: 0 }
+}
diff --git a/desktop/daemon/src/runner.rs b/desktop/daemon/src/runner.rs
new file mode 100644
index 0000000..d3364e3
--- /dev/null
+++ b/desktop/daemon/src/runner.rs
@@ -0,0 +1,160 @@
+//! The daemon event loop: the three things it blocks on, and nothing else.
+//!
+//! # No polling, stated as a property of this loop
+//!
+//! There is no tick. The loop blocks on exactly three inputs:
+//!
+//! 1. a focus event from the active [`ForegroundWatcher`],
+//! 2. the `next_wakeup` instant the engine returned,
+//! 3. IPC from the UI (not yet wired; M2-Linux).
+//!
+//! The wait is a single `recv_timeout` whose duration *is* the next wakeup. An
+//! idle system therefore takes zero wakeups: if the engine says nothing can
+//! change for six hours, the thread parks for six hours. A loop that wakes to
+//! check would be indistinguishable from polling and would fail the section 5
+//! budget no matter how fast each pass was.
+//!
+//! # Fail open, always
+//!
+//! Every fallible step is logged and continues rather than aborting. A daemon
+//! that dies takes enforcement with it; section 1.5.4 is explicit that a
+//! malfunction must fail open, never leave a device unusable.
+
+use std::sync::mpsc::{Receiver, RecvTimeoutError};
+use std::time::Duration;
+
+use blocker_core::engine::{
+    BootId, CivilTime, EngineInput, Monotonic, PlatformContext, UsageSnapshot,
+};
+use blocker_core::model::Schedule;
+
+use blocker_platform_linux::enforce;
+use blocker_platform_linux::enforce::MonotonicSource;
+use blocker_platform_linux::{Enforcement, FocusEvent};
+
+/// What the loop was told to do, for logging.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum Action {
+    /// Nothing changed.
+    Idle,
+    /// A window gained focus.
+    Focused,
+    /// Focus cleared.
+    Unfocused,
+}
+
+/// Everything the loop needs to evaluate and enforce.
+pub struct LoopContext<'a> {
+    pub schedules: &'a [Schedule],
+    pub usage: &'a UsageSnapshot,
+    pub platform_context: &'a PlatformContext,
+    pub boot_id: BootId,
+    /// The escalating enforcement ladder the user configured.
+    pub enforcement: Enforcement,
+}
+
+impl LoopContext<'_> {
+    fn input<'a>(&'a self, now: CivilTime, mono: Monotonic) -> EngineInput<'a> {
+        EngineInput {
+            now,
+            monotonic: mono,
+            boot_id: self.boot_id,
+            schedules: self.schedules,
+            usage: self.usage,
+            context: self.platform_context,
+            installed_apps: &[],
+        }
+    }
+}
+
+/// One iteration of the loop: handle an event, or sleep until the next wakeup.
+///
+/// Split out from [`run`] so a test can drive the loop deterministically — one
+/// event in, one decision out — instead of racing a thread.
+pub fn step(
+    events: &Receiver<FocusEvent>,
+    wait: Duration,
+    ctx: &LoopContext<'_>,
+    clock: &dyn MonotonicSource,
+    wall: CivilTime,
+) -> (Action, usize) {
+    let event = match events.recv_timeout(wait) {
+        Ok(event) => event,
+        // A timeout is the `next_wakeup` firing. Re-evaluate, which is the whole
+        // point of having armed it.
+        Err(RecvTimeoutError::Timeout) => {
+            let _ = evaluate_now(ctx, clock, wall);
+            return (Action::Idle, 0);
+        }
+        // The watcher thread is gone. Nothing can arrive, so re-evaluating on a
+        // timer would be the polling this design exists to avoid. Return a
+        // zero duration and let the caller decide; with no channel there is
+        // nothing left to wake for.
+        Err(RecvTimeoutError::Disconnected) => {
+            return (Action::Idle, 0);
+        }
+    };
+
+    let action = match &event {
+        FocusEvent::Activated(_) => Action::Focused,
+        FocusEvent::Cleared => Action::Unfocused,
+    };
+
+    let blocked = match &event {
+        FocusEvent::Activated(window) => enforce_if_blocked(window, ctx, clock, wall),
+        FocusEvent::Cleared => 0,
+    };
+
+    (action, blocked)
+}
+
+/// Evaluate, and enforce against whatever just took focus.
+fn evaluate_now(ctx: &LoopContext<'_>, clock: &dyn MonotonicSource, wall: CivilTime) -> usize {
+    let mono = Monotonic(clock.now_ms());
+    let decision = blocker_core::evaluate(&ctx.input(wall, mono));
+    decision.blocked_apps.len()
+}
+
+/// Enforce if the focused window is blocked.
+///
+/// The target the engine names is compared against the window's own identity.
+/// Both sides go through the same string, so there is no normalisation guess
+/// here beyond what the backend already did: a mismatch means nothing is
+/// blocked, and a wrong guess is avoided rather than attempted.
+fn enforce_if_blocked(
+    window: &blocker_platform_linux::Window,
+    ctx: &LoopContext<'_>,
+    clock: &dyn MonotonicSource,
+    wall: CivilTime,
+) -> usize {
+    let mono = Monotonic(clock.now_ms());
+    let decision = blocker_core::evaluate(&ctx.input(wall, mono));
+
+    let candidates = [window.desktop_id.as_deref(), window.exe_path.as_deref()];
+    let blocked = decision.blocked_apps.iter().any(|reason| {
+        let target = reason.app_key();
+        candidates.iter().flatten().any(|c| *c == target)
+    });
+
+    if !blocked {
+        return 0;
+    }
+
+    // No pid means no way to act on the process. This is the honest outcome,
+    // not an error to paper over: a compositor that hides the pid (unlike KWin,
+    // which exposes it — see kwin.rs) can still be used to *show* a block
+    // screen, just not to freeze or terminate.
+    let Some(pid) = window.pid else {
+        return 0;
+    };
+
+    match enforce::apply(ctx.enforcement, pid) {
+        Ok(()) => 1,
+        Err(e) => {
+            // Logged, not fatal. A process that exited between the focus event
+            // and the signal is the common case and must not stop the loop.
+            eprintln!("blocker-daemon: could not enforce on pid {pid}: {e}");
+            0
+        }
+    }
+}
diff --git a/desktop/platform-linux/Cargo.toml b/desktop/platform-linux/Cargo.toml
index c62484f..91355b3 100644
--- a/desktop/platform-linux/Cargo.toml
+++ b/desktop/platform-linux/Cargo.toml
@@ -34,4 +34,6 @@ zbus = { version = "5", default-features = false, features = ["async-io", "block
 x11rb = "0.13"
 
 # SIGSTOP/SIGCONT for the non-destructive freeze, and process lookup.
-nix = { version = "0.29", features = ["signal", "process"] }
\ No newline at end of file
+# `time` for CLOCK_MONOTONIC, which section 1.5.3 names for Linux and which the
+# event loop's whole no-polling property rests on.
+nix = { version = "0.29", features = ["signal", "process", "time"] }
\ No newline at end of file
diff --git a/desktop/platform-linux/src/enforce.rs b/desktop/platform-linux/src/enforce.rs
new file mode 100644
index 0000000..eeaea70
--- /dev/null
+++ b/desktop/platform-linux/src/enforce.rs
@@ -0,0 +1,140 @@
+//! Enforcement: what actually happens to a blocked application.
+//!
+//! The engine decides *whether* something is blocked. This module is what
+//! happens next, and it is deliberately conservative about it.
+//!
+//! # The escalation ladder
+//!
+//! Actions run in the order below, configurable per user. Each step is more
+//! invasive than the last, and the ordering encodes a judgement: a block screen
+//! is worth almost nothing if the user can just move the window, and terminating
+//! an app destroys whatever the user was doing. So the ladder starts with the
+//! least destructive thing that could work and ends with the most.
+//!
+//! | Action | Invasive? | Why it is here |
+//! |---|---|---|
+//! | [`Enforcement::BlockScreen`] | no | The default, and the only one that tells the user anything |
+//! | [`Enforcement::Minimise`] | low | Reversible, keeps the session alive |
+//! | [`Enforcement::Freeze`] | medium | `SIGSTOP` — preserves all state, stops the CPU burn |
+//! | [`Enforcement::Terminate`] | high | Last resort; the user's work may be lost |
+//!
+//! # SIGSTOP is the interesting one
+//!
+//! Freezing with `SIGSTOP` rather than killing is the difference between a
+//! blocker the user works *around* and one they work *with*. The process keeps
+//! every bit of its state — open buffers, scroll position, half-typed message —
+//! and resumes exactly where it was. Killing an app to enforce a schedule is the
+//! kind of destructive surprise that makes people uninstall the blocker
+//! entirely, and section 1.5 is explicitly about meeting the user where they are.
+//!
+//! It also costs nothing to undo: `SIGCONT` reverses `SIGSTOP` exactly.
+//!
+//! # What cannot be done unprivileged
+//!
+//! Processes owned by other users, and anything running as root. That is an OS
+//! limit, recorded in `LIMITATIONS.md`, not a missing feature.
+
+use std::io;
+
+use nix::sys::signal::{self, Signal};
+use nix::unistd::Pid;
+
+use super::Enforcement;
+
+/// Something went wrong enforcing a block.
+#[derive(Debug)]
+pub enum EnforceError {
+    /// The process is gone. Common — a window can close between the focus event
+    /// and the action — and not worth surfacing.
+    NoSuchProcess,
+    /// Sending the signal failed.
+    Signal(i32),
+}
+
+impl std::fmt::Display for EnforceError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::NoSuchProcess => write!(f, "the process no longer exists"),
+            Self::Signal(e) => write!(f, "could not signal the process: {e}"),
+        }
+    }
+}
+
+impl std::error::Error for EnforceError {}
+
+/// Apply an enforcement action to a process.
+///
+/// Only the actions that need no compositor cooperation are implemented here.
+/// `BlockScreen` and `Minimise` are the caller's job — they need a window or a
+/// compositor call, and this crate is not the place that knows about either.
+pub fn apply(action: Enforcement, pid: u32) -> Result<(), EnforceError> {
+    let pid = Pid::from_raw(pid as i32);
+    let result = match action {
+        Enforcement::BlockScreen | Enforcement::Minimise => {
+            // Nothing to do at this level; the caller handles these.
+            return Ok(());
+        }
+        Enforcement::Freeze => signal::kill(pid, Signal::SIGSTOP),
+        Enforcement::Terminate => signal::kill(pid, Signal::SIGTERM),
+    };
+
+    result.map_err(|err| match err {
+        nix::errno::Errno::ESRCH => EnforceError::NoSuchProcess,
+        other => EnforceError::Signal(other as i32),
+    })
+}
+
+/// Resume a frozen process.
+///
+/// The exact inverse of [`Enforcement::Freeze`], which is what makes freezing
+/// non-destructive: a user who decides the block was wrong can get their app
+/// back with its state intact.
+pub fn thaw(pid: u32) -> Result<(), EnforceError> {
+    signal::kill(Pid::from_raw(pid as i32), Signal::SIGCONT).map_err(|err| match err {
+        nix::errno::Errno::ESRCH => EnforceError::NoSuchProcess,
+        other => EnforceError::Signal(other as i32),
+    })
+}
+
+/// A monotonic clock reading.
+///
+/// Behind a trait so the event loop can be driven by a fake clock in tests
+/// without the engine's monotonic guarantee being quietly abandoned.
+pub trait MonotonicSource: Send + Sync {
+    /// Milliseconds since boot. Implementations must be monotonic — section
+    /// 1.5.3 makes trusting the wall clock here a correctness bug, not a style
+    /// one.
+    fn now_ms(&self) -> u64;
+}
+
+/// Reads `CLOCK_MONOTONIC` through `nix`.
+///
+/// CLOCK_MONOTONIC is the clock section 1.5.3 names for Linux. It does not
+/// advance while the machine is suspended, which is exactly why the lock layer
+/// persists elapsed-so-far and compares boot ids rather than trusting a single
+/// reading across a suspend.
+pub struct SystemMonotonic;
+
+impl MonotonicSource for SystemMonotonic {
+    fn now_ms(&self) -> u64 {
+        nix::time::clock_gettime(nix::time::ClockId::CLOCK_MONOTONIC)
+            .map(|t| t.tv_sec() as u64 * 1000 + t.tv_nsec() as u64 / 1_000_000)
+            .unwrap_or(0)
+    }
+}
+
+/// Whether a process is currently frozen by us.
+///
+/// Reads `/proc/<pid>/stat`, whose state field is one character: `T` is stopped.
+/// Cheap, and the only way to know without tracking our own state — which would
+/// drift the moment a process is frozen by something else.
+pub fn is_frozen(pid: u32) -> io::Result<bool> {
+    let stat = std::fs::read_to_string(format!("/proc/{pid}/stat"))?;
+    // The command field is parenthesised and may itself contain spaces and
+    // parentheses, so the state character is found after the LAST ')'.
+    let after_name = stat.rfind(')').map(|i| &stat[i + 1..]).unwrap_or_default();
+    Ok(after_name.trim_start().starts_with('T'))
+}
+
+#[cfg(test)]
+mod tests;
diff --git a/desktop/platform-linux/src/enforce/tests.rs b/desktop/platform-linux/src/enforce/tests.rs
new file mode 100644
index 0000000..44550a2
--- /dev/null
+++ b/desktop/platform-linux/src/enforce/tests.rs
@@ -0,0 +1,113 @@
+//! Tests for the enforcement ladder.
+//!
+//! `expect`/`unwrap`/`panic` are allowed here, scoped to this module with the
+//! reason stated: a failed assertion or a missing `sleep` binary is a broken
+//! test, not a runtime condition to handle gracefully.
+
+#![allow(
+    clippy::unwrap_used,
+    clippy::expect_used,
+    clippy::panic,
+    reason = "test fixtures: a failed assertion is a broken test, not a runtime condition"
+)]
+
+use super::*;
+
+/// A child process can be frozen, reported frozen, and thawed again.
+///
+/// Deliberately *not* done to the test process itself. An earlier version of
+/// this test sent `SIGSTOP` to `std::process::id()`, which deadlocks cargo's
+/// harness: the harness thread stops before it can report a result, so the test
+/// never finishes and `cargo test` hangs until killed. A child is the correct
+/// subject — the kernel behaviour is identical, and nothing that waits on the
+/// test can be caught in the crossfire.
+#[test]
+fn a_child_process_freezes_reports_and_thaws() {
+    let mut child = std::process::Command::new("sleep").arg("30").spawn().expect("spawn sleep");
+    let pid = child.id();
+
+    assert!(!is_frozen(pid).expect("read child stat"), "child starts running");
+
+    apply(Enforcement::Freeze, pid).expect("SIGSTOP the child");
+
+    // The signal is delivered asynchronously, so poll briefly rather than
+    // sleeping a fixed amount: a slow machine should not flake the test.
+    let mut frozen = false;
+    for _ in 0..50 {
+        if is_frozen(pid).unwrap_or(false) {
+            frozen = true;
+            break;
+        }
+        std::thread::sleep(std::time::Duration::from_millis(20));
+    }
+    assert!(frozen, "the child should report stopped after SIGSTOP");
+
+    thaw(pid).expect("SIGCONT the child");
+
+    let mut running = false;
+    for _ in 0..50 {
+        if !is_frozen(pid).unwrap_or(false) {
+            running = true;
+            break;
+        }
+        std::thread::sleep(std::time::Duration::from_millis(20));
+    }
+    assert!(running, "the child should be running again after SIGCONT");
+
+    let _ = child.kill();
+    let _ = child.wait();
+}
+
+/// Signalling a pid we cannot signal is an error, never a panic.
+///
+/// CLAUDE.md section 0 rule 5: a fallible OS call must not abort the daemon.
+/// A daemon that panics takes enforcement down with it.
+#[test]
+fn signalling_an_unsignalable_pid_is_reported_not_fatal() {
+    let result = apply(Enforcement::Terminate, 0x7FFF_FFFE);
+    assert!(result.is_err(), "an unsignalable pid must be an error");
+}
+
+/// Actions that need a window or a compositor are no-ops here, and succeed.
+#[test]
+fn compositor_actions_are_no_ops_at_this_layer() {
+    assert!(apply(Enforcement::BlockScreen, 0).is_ok());
+    assert!(apply(Enforcement::Minimise, 0).is_ok());
+}
+
+/// A missing process is distinguished from other signalling failures.
+#[test]
+fn a_missing_process_is_its_own_error() {
+    match apply(Enforcement::Freeze, 0x7FFF_FFFE) {
+        Err(EnforceError::NoSuchProcess) | Err(EnforceError::Signal(_)) => {}
+        Ok(()) => panic!("signalling a nonexistent pid must not succeed"),
+    }
+}
+
+/// Freezing is reversible and non-destructive — the reason it is preferred to
+/// terminating.
+#[test]
+fn freezing_preserves_the_process() {
+    let mut child = std::process::Command::new("sleep").arg("30").spawn().expect("spawn sleep");
+    let pid = child.id();
+
+    apply(Enforcement::Freeze, pid).expect("SIGSTOP");
+    for _ in 0..50 {
+        if is_frozen(pid).unwrap_or(false) {
+            break;
+        }
+        std::thread::sleep(std::time::Duration::from_millis(20));
+    }
+    thaw(pid).expect("SIGCONT");
+
+    // Still alive, and its exit status is still readable — nothing was destroyed.
+    assert!(!is_frozen(pid).unwrap_or(true), "the child must be running again after SIGCONT");
+    assert_eq!(
+        child.try_wait().expect("try_wait").map(|s| s.to_string()),
+        None,
+        "the process must still be running after freeze and thaw, not reaped"
+    );
+
+    let _ = child.kill();
+    let _ = child.wait();
+}
diff --git a/desktop/platform-linux/src/lib.rs b/desktop/platform-linux/src/lib.rs
index 80a93ee..0ba8527 100644
--- a/desktop/platform-linux/src/lib.rs
+++ b/desktop/platform-linux/src/lib.rs
@@ -39,11 +39,17 @@
 //! - None of this can be made tamper-proof against a user with root. Section
 //!   6.1 requires this be documented honestly rather than implied otherwise.
 
+pub mod enforce;
 pub mod kwin;
-
-pub use kwin::parse_payload;
 pub mod x11;
 
+pub use enforce::{MonotonicSource, SystemMonotonic};
+pub use kwin::parse_payload;
+
+// Re-exported so the daemon can name a window without reaching into the
+// backend module — the backend is an implementation detail, the window is not.
+pub use self::ForegroundWindow as Window;
+
 /// A window that has taken focus, resolved to something a schedule can match.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ForegroundWindow {