diff --git a/crates/agent-lifecycle/src/k8s.rs b/crates/agent-lifecycle/src/k8s.rs
new file mode 100644
index 0000000..4514b0a
--- /dev/null
+++ b/crates/agent-lifecycle/src/k8s.rs
@@ -0,0 +1,194 @@
+//! K8s runtime driver — projects k8s Pod signals onto the canonical model.
+//!
+//! Mapping (ADR §6, the k8s column): Pod phase + readinessProbe ⇒ Starting /
+//! Running; `metadata.deletionTimestamp != null` ⇒ `DesiredStatus::Stopped`
+//! (Terminating: preStop + grace); `Unknown` phase (node lost) ⇒
+//! Unhealthy(fenced), same as a lost lease; CrashLoopBackOff ⇒ Unhealthy.
+//! Deliberately mirrors `ecs.rs`'s shape — this is a direct port of an
+//! already-approved ADR table, not a new design decision (studio#146, the
+//! observability follow-up `k8s_driver.rs`'s module doc flagged as deferred
+//! when the k8s `ProvisionDriver` write path landed).
+
+use crate::{DesiredStatus, Discriminator, Health, RuntimeDriver};
+
+/// The subset of k8s Pod `phase` relevant to the projection. `Unknown` is
+/// k8s's own "kubelet not reporting" signal — the direct analogue of ECS's
+/// node-loss case, not a leftover default.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum PodPhase {
+    Pending,
+    Running,
+    Succeeded,
+    Failed,
+    Unknown,
+}
+
+/// A single k8s Pod observation (the subset the projection needs).
+#[derive(Debug, Clone, Copy)]
+pub struct K8sPod {
+    pub phase: PodPhase,
+    /// `metadata.deletionTimestamp != null`.
+    pub deletion_timestamp_set: bool,
+    /// The Pod's `Ready` condition (readiness probe result when one is
+    /// declared; otherwise just reflects the container's running state).
+    pub ready: bool,
+    /// Whether a readiness probe is actually declared on the pod spec. When
+    /// it is not, k8s's `Ready` condition tracks plain container-running
+    /// state, not health — so `ready == false` there means "still starting",
+    /// not a fault. Mirrors `EcsTask::health_check_defined`'s "no probe ≠
+    /// fault" distinction; without it, a probe-less pod that just hasn't
+    /// finished starting would misread as permanently degraded.
+    pub ready_check_defined: bool,
+    /// A container in this pod is in CrashLoopBackOff (or an equivalent
+    /// terminal restart-loop waiting reason) — a fault signal distinct from
+    /// `ready`, since a crash-looping container can still report a phase
+    /// that looks superficially fine between restarts.
+    pub crash_loop_back_off: bool,
+    /// CP-issued lease still valid (heartbeat authorized).
+    pub lease_valid: bool,
+    /// CP/director cordon: `false` ⇒ not admitting new work (→ Paused).
+    pub accepting_work: bool,
+}
+
+/// Projects k8s Pod state onto the canonical lifecycle model.
+pub struct K8sDriver;
+
+impl RuntimeDriver for K8sDriver {
+    type Native = K8sPod;
+    /// Pod UID.
+    type InstanceId = String;
+
+    fn project(&self, pod: &K8sPod, verified_before: bool) -> Discriminator {
+        let desired_status = if pod.deletion_timestamp_set {
+            DesiredStatus::Stopped
+        } else {
+            DesiredStatus::Running
+        };
+
+        // `identity_verified` latches once the pod has ever reported phase
+        // Running with Ready true.
+        let identity_verified =
+            verified_before || (pod.phase == PodPhase::Running && pod.ready);
+
+        // Node-unreachable (Unknown phase), a crash-loop, or a lost lease all
+        // fault outright. A *declared* readiness probe failing faults too —
+        // but an undeclared one reporting `ready == false` just means "still
+        // starting" (see `ready_check_defined`'s doc), not a fault; that case
+        // only matters once `identity_verified` is already true, since
+        // `classify()` reads `Starting` ahead of `health` otherwise.
+        let health = if pod.phase == PodPhase::Unknown
+            || pod.crash_loop_back_off
+            || !pod.lease_valid
+            || (pod.ready_check_defined && !pod.ready)
+        {
+            Health::Faulted
+        } else {
+            Health::Ok
+        };
+
+        Discriminator {
+            desired_status,
+            accepting_work: pod.accepting_work,
+            health,
+            identity_verified,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::AgentState;
+
+    fn pod(
+        phase: PodPhase,
+        deleting: bool,
+        ready: bool,
+        lease: bool,
+        accepting: bool,
+    ) -> K8sPod {
+        // Default: no readiness probe declared (the common case for OAB agents).
+        pod_probe(phase, deleting, ready, lease, accepting, false, false)
+    }
+
+    fn pod_probe(
+        phase: PodPhase,
+        deleting: bool,
+        ready: bool,
+        lease: bool,
+        accepting: bool,
+        ready_check_defined: bool,
+        crash_loop_back_off: bool,
+    ) -> K8sPod {
+        K8sPod {
+            phase,
+            deletion_timestamp_set: deleting,
+            ready,
+            ready_check_defined,
+            crash_loop_back_off,
+            lease_valid: lease,
+            accepting_work: accepting,
+        }
+    }
+
+    #[test]
+    fn pending_pod_projects_to_starting() {
+        let d = K8sDriver.project(&pod(PodPhase::Pending, false, false, false, false), false);
+        assert_eq!(d.classify(), AgentState::Starting);
+    }
+
+    #[test]
+    fn running_ready_pod_projects_to_running() {
+        let d = K8sDriver.project(&pod(PodPhase::Running, false, true, true, true), false);
+        assert_eq!(d.classify(), AgentState::Running);
+    }
+
+    #[test]
+    fn cordoned_running_pod_is_paused() {
+        let d = K8sDriver.project(&pod(PodPhase::Running, false, true, true, false), true);
+        assert_eq!(d.classify(), AgentState::Paused);
+    }
+
+    #[test]
+    fn declared_probe_failing_after_verified_is_unhealthy() {
+        // Was Ready before, now the declared readiness probe fails ⇒
+        // Unhealthy (not Starting).
+        let d = K8sDriver.project(
+            &pod_probe(PodPhase::Running, false, false, true, false, true, false),
+            true,
+        );
+        assert_eq!(d.classify(), AgentState::Unhealthy);
+    }
+
+    #[test]
+    fn node_lost_unknown_is_unhealthy_not_stopped() {
+        // Unknown phase (kubelet not reporting) while verified ⇒
+        // Unhealthy(fenced), not Stopped.
+        let d = K8sDriver.project(&pod(PodPhase::Unknown, false, false, false, false), true);
+        assert_eq!(d.classify(), AgentState::Unhealthy);
+    }
+
+    #[test]
+    fn crash_loop_back_off_is_unhealthy_even_if_ready() {
+        let d = K8sDriver.project(
+            &pod_probe(PodPhase::Running, false, true, true, true, true, true),
+            true,
+        );
+        assert_eq!(d.classify(), AgentState::Unhealthy);
+    }
+
+    #[test]
+    fn running_without_declared_probe_and_not_ready_is_still_running() {
+        // No readiness probe declared ⇒ k8s's Ready condition isn't a health
+        // signal here — a running, leased instance stays Running (the ECS
+        // side's equivalent case: `running_unknown_without_health_check_is_running`).
+        let d = K8sDriver.project(&pod(PodPhase::Running, false, false, true, true), true);
+        assert_eq!(d.classify(), AgentState::Running);
+    }
+
+    #[test]
+    fn deletion_timestamp_is_stopping_while_observable() {
+        let d = K8sDriver.project(&pod(PodPhase::Running, true, true, true, false), true);
+        assert_eq!(d.classify(), AgentState::Stopping);
+    }
+}
diff --git a/crates/agent-lifecycle/src/lib.rs b/crates/agent-lifecycle/src/lib.rs
index 4b709b4..263f681 100644
--- a/crates/agent-lifecycle/src/lib.rs
+++ b/crates/agent-lifecycle/src/lib.rs
@@ -7,6 +7,7 @@
 //! never changes per runtime.
 
 pub mod ecs;
+pub mod k8s;
 
 /// Whether the control plane wants this instance running or stopped.
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]