diff --git a/cmd/gc/build_desired_state.go b/cmd/gc/build_desired_state.go
index fe46ec017c..5b906b029a 100644
--- a/cmd/gc/build_desired_state.go
+++ b/cmd/gc/build_desired_state.go
@@ -357,7 +357,7 @@ func buildDesiredStateWithSessionBeads(
 		}
 		poolWorkBeads := filterAssignedWorkBeadsForPoolDemand(cfg, cityPath, sessionBeads.Open(), assignedWorkBeads, assignedWorkStoreRefs)
 		bp.assignedWorkBeads = poolWorkBeads
-		poolDesiredStates := ComputePoolDesiredStatesTraced(cfg, poolWorkBeads, sessionBeads.Open(), scaleCheckCounts, trace)
+		poolDesiredStates := ComputePoolDesiredStatesDebouncedTraced(cfg, poolWorkBeads, sessionBeads.Open(), scaleCheckCounts, poolScaleCheckPartialTemplates, trace)
 		for _, poolState := range poolDesiredStates {
 			cfgAgent := findAgentByTemplate(cfg, poolState.Template)
 			if cfgAgent == nil {
diff --git a/cmd/gc/city_runtime.go b/cmd/gc/city_runtime.go
index 34adaf0d8e..616fceece9 100644
--- a/cmd/gc/city_runtime.go
+++ b/cmd/gc/city_runtime.go
@@ -1601,8 +1601,8 @@ func (cr *CityRuntime) beadReconcileTick(ctx context.Context, result DesiredStat
 	if poolDesired == nil {
 		poolWorkBeads := filterAssignedWorkBeadsForPoolDemand(cr.cfg, cr.cityPath, sessionBeads.Open(), assignedWorkBeads, assignedWorkStoreRefs)
 		poolDesired = retainScaleCheckPartialPoolDesired(
-			PoolDesiredCounts(ComputePoolDesiredStatesTraced(
-				cr.cfg, poolWorkBeads, sessionBeads.Open(), result.ScaleCheckCounts, trace)),
+			PoolDesiredCounts(ComputePoolDesiredStatesDebouncedTraced(
+				cr.cfg, poolWorkBeads, sessionBeads.Open(), result.ScaleCheckCounts, result.PoolScaleCheckPartialTemplates, trace)),
 			sessionBeads,
 			result.PoolScaleCheckPartialTemplates,
 		)
@@ -2133,8 +2133,8 @@ func (cr *CityRuntime) controlDispatcherTick(ctx context.Context) {
 	open := filterSessionBeadsByName(updated, cfgNames)
 	poolWorkBeads := filterAssignedWorkBeadsForPoolDemand(filteredCfg, cr.cityPath, open, wfcResult.AssignedWorkBeads, wfcResult.AssignedWorkStoreRefs)
 	poolDesired := retainScaleCheckPartialPoolDesired(
-		PoolDesiredCounts(ComputePoolDesiredStates(
-			filteredCfg, poolWorkBeads, open, wfcResult.ScaleCheckCounts)),
+		PoolDesiredCounts(ComputePoolDesiredStatesDebounced(
+			filteredCfg, poolWorkBeads, open, wfcResult.ScaleCheckCounts, wfcResult.PoolScaleCheckPartialTemplates)),
 		newSessionBeadSnapshot(open),
 		wfcResult.PoolScaleCheckPartialTemplates,
 	)
@@ -2298,8 +2298,8 @@ func (cr *CityRuntime) loadDemandSnapshot(
 		}
 		poolWorkBeads := filterAssignedWorkBeadsForPoolDemand(cr.cfg, cr.cityPath, openSessionBeads, result.AssignedWorkBeads, result.AssignedWorkStoreRefs)
 		result.PoolDesiredCounts = retainScaleCheckPartialPoolDesired(
-			PoolDesiredCounts(ComputePoolDesiredStatesTraced(
-				cr.cfg, poolWorkBeads, openSessionBeads, result.ScaleCheckCounts, trace)),
+			PoolDesiredCounts(ComputePoolDesiredStatesDebouncedTraced(
+				cr.cfg, poolWorkBeads, openSessionBeads, result.ScaleCheckCounts, result.PoolScaleCheckPartialTemplates, trace)),
 			sessionBeads,
 			result.PoolScaleCheckPartialTemplates,
 		)
diff --git a/cmd/gc/cmd_start.go b/cmd/gc/cmd_start.go
index b5e3041553..5ad4974798 100644
--- a/cmd/gc/cmd_start.go
+++ b/cmd/gc/cmd_start.go
@@ -879,8 +879,8 @@ func doStartStandalone(args []string, controllerMode bool, stdout, stderr io.Wri
 	dt := newDrainTracker()
 	poolWorkBeads := filterAssignedWorkBeadsForPoolDemand(cfg, cityPath, open, dsResult.AssignedWorkBeads, dsResult.AssignedWorkStoreRefs)
 	poolDesired := retainScaleCheckPartialPoolDesired(
-		PoolDesiredCounts(ComputePoolDesiredStates(
-			cfg, poolWorkBeads, open, dsResult.ScaleCheckCounts)),
+		PoolDesiredCounts(ComputePoolDesiredStatesDebounced(
+			cfg, poolWorkBeads, open, dsResult.ScaleCheckCounts, dsResult.PoolScaleCheckPartialTemplates)),
 		sessionBeads,
 		dsResult.PoolScaleCheckPartialTemplates,
 	)
diff --git a/cmd/gc/main_test.go b/cmd/gc/main_test.go
index 29655e3be1..9217fe7d33 100644
--- a/cmd/gc/main_test.go
+++ b/cmd/gc/main_test.go
@@ -216,6 +216,11 @@ func TestMain(m *testing.M) {
 	}
 	configureFSPressureForTests()
 	configureSupervisorHooksForTests()
+	// Disable the pool scale_check confirm window for tests so existing
+	// single-call assertions about spawn behavior continue to hold without
+	// having to seed the per-template history. Debouncer-specific tests opt
+	// back in via setPoolScaleCheckDebouncerWindowForTesting.
+	setPoolScaleCheckDebouncerWindowForTesting(1)
 	testscript.Main(newDoltLeakGuardedTestingM(m, testTempRoot, testTempRoot, gcHome, runtimeDir, providerStubDir, sharedTestFormulaDir, sharedTestCityDir), map[string]func(){
 		"gc": func() {
 			configureTestscriptEnvDefaults()
diff --git a/cmd/gc/pool_desired_state.go b/cmd/gc/pool_desired_state.go
index d41a66d9d1..80da674f11 100644
--- a/cmd/gc/pool_desired_state.go
+++ b/cmd/gc/pool_desired_state.go
@@ -75,6 +75,38 @@ func ComputePoolDesiredStatesTraced(
 	return computePoolDesiredStates(cfg, assignedWorkBeads, sessionBeads, scaleCheckCounts, trace)
 }
 
+// ComputePoolDesiredStatesDebounced wraps ComputePoolDesiredStates with the
+// process-scoped scale_check confirm window. Reconciler call sites should
+// prefer this over the raw entry point so a single-tick flap on
+// scaleCheckCounts (e.g. a transient assignee release from an upstream race)
+// can't trigger a fresh pool spawn that has nothing to do. partialTemplates
+// is the set of templates whose probe failed this tick — those pass through
+// the debouncer unchanged so transient probe errors don't poison the window.
+func ComputePoolDesiredStatesDebounced(
+	cfg *config.City,
+	assignedWorkBeads []beads.Bead,
+	sessionBeads []beads.Bead,
+	scaleCheckCounts map[string]int,
+	partialTemplates map[string]bool,
+) []PoolDesiredState {
+	debounced := debouncePoolScaleCheckCounts(scaleCheckCounts, partialTemplates)
+	return computePoolDesiredStates(cfg, assignedWorkBeads, sessionBeads, debounced, nil)
+}
+
+// ComputePoolDesiredStatesDebouncedTraced is the traced variant of
+// ComputePoolDesiredStatesDebounced. See that function for the semantics.
+func ComputePoolDesiredStatesDebouncedTraced(
+	cfg *config.City,
+	assignedWorkBeads []beads.Bead,
+	sessionBeads []beads.Bead,
+	scaleCheckCounts map[string]int,
+	partialTemplates map[string]bool,
+	trace *sessionReconcilerTraceCycle,
+) []PoolDesiredState {
+	debounced := debouncePoolScaleCheckCounts(scaleCheckCounts, partialTemplates)
+	return computePoolDesiredStates(cfg, assignedWorkBeads, sessionBeads, debounced, trace)
+}
+
 func computePoolDesiredStates(
 	cfg *config.City,
 	assignedWorkBeads []beads.Bead,
diff --git a/cmd/gc/pool_scale_check_debouncer.go b/cmd/gc/pool_scale_check_debouncer.go
new file mode 100644
index 0000000000..cede45ca4a
--- /dev/null
+++ b/cmd/gc/pool_scale_check_debouncer.go
@@ -0,0 +1,128 @@
+package main
+
+import "sync"
+
+// defaultPoolScaleCheckConfirmWindow is the number of consecutive scale_check
+// samples a pool template's count must clear before the reconciler acts on a
+// non-zero value as new demand. The pathology this prevents: a single tick
+// where the scale_check shell returns 1 (because the bead briefly appeared
+// unassigned via some upstream race) is enough to spawn a fresh pool session,
+// which then runs work_query, finds nothing actually claimable, and drains —
+// repeating every ~tick interval. Confirming over a window collapses that
+// loop without changing steady-state demand semantics.
+//
+// 3 is the smallest window that suppresses the dominant single-tick flap
+// while still tolerating a single dropped sample inside an otherwise
+// sustained run. The window applies per template, so a slow reconciler tick
+// cadence (multi-second) yields a confirm latency on the order of N ticks
+// before legitimate new demand is acted on — acceptable for pool spawn
+// because in-flight session beads from a prior tick continue to count as
+// "spent" new demand via poolInFlightNewRequests during the wait.
+const defaultPoolScaleCheckConfirmWindow = 3
+
+// poolScaleCheckDebouncer holds the per-template history of recent
+// scale_check samples used to confirm new pool demand. Methods are safe
+// for concurrent use.
+type poolScaleCheckDebouncer struct {
+	mu      sync.Mutex
+	windowN int
+	history map[string][]int
+}
+
+// defaultPoolScaleCheckDebouncerSingleton is the process-scoped debouncer
+// used by the reconciler. One gascity process supervises one city, so a
+// package-level singleton keyed only by agent template name is sufficient
+// to isolate history within the process.
+//
+// Tests should reset this via resetPoolScaleCheckDebouncerForTesting (and
+// optionally setPoolScaleCheckDebouncerWindowForTesting to shrink the
+// window) to avoid cross-test bleed from accumulated history.
+var defaultPoolScaleCheckDebouncerSingleton = newPoolScaleCheckDebouncer(defaultPoolScaleCheckConfirmWindow)
+
+func newPoolScaleCheckDebouncer(windowN int) *poolScaleCheckDebouncer {
+	if windowN < 1 {
+		windowN = 1
+	}
+	return &poolScaleCheckDebouncer{
+		windowN: windowN,
+		history: make(map[string][]int),
+	}
+}
+
+// debounce records this tick's raw scale_check counts and returns the
+// confirmed counts the reconciler should act on. For each template:
+//   - If the template is in partialTemplates, the raw count passes through
+//     and history is left untouched: a probe failure isn't a real
+//     observation of demand, and admitting it would poison the window.
+//   - Otherwise the sample is appended to the per-template ring. While the
+//     ring has fewer than windowN samples, the raw count passes through —
+//     so a cold-start supervisor still acts on demand without first
+//     accumulating windowN ticks of history. Once the ring is full, the
+//     returned count is the minimum across the window: a single tick at 0
+//     in an otherwise positive window forces 0, killing the 1-tick flap.
+func (d *poolScaleCheckDebouncer) debounce(raw map[string]int, partialTemplates map[string]bool) map[string]int {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	out := make(map[string]int, len(raw))
+	for template, count := range raw {
+		if partialTemplates[template] {
+			out[template] = count
+			continue
+		}
+		h := d.history[template]
+		h = append(h, count)
+		if len(h) > d.windowN {
+			h = h[len(h)-d.windowN:]
+		}
+		d.history[template] = h
+
+		if len(h) < d.windowN {
+			out[template] = count
+			continue
+		}
+
+		minCount := h[0]
+		for _, v := range h[1:] {
+			if v < minCount {
+				minCount = v
+			}
+		}
+		out[template] = minCount
+	}
+	return out
+}
+
+// debouncePoolScaleCheckCounts is the package-level entry point the
+// reconciler uses to apply the singleton's confirm window to a fresh
+// scale_check sample.
+func debouncePoolScaleCheckCounts(raw map[string]int, partialTemplates map[string]bool) map[string]int {
+	return defaultPoolScaleCheckDebouncerSingleton.debounce(raw, partialTemplates)
+}
+
+// resetPoolScaleCheckDebouncerForTesting clears the singleton's per-template
+// history. Tests that exercise the reconciler across multiple buildDesiredState
+// or ComputePoolDesiredStatesDebounced calls should reset to keep the window
+// deterministic.
+func resetPoolScaleCheckDebouncerForTesting() {
+	defaultPoolScaleCheckDebouncerSingleton.mu.Lock()
+	defer defaultPoolScaleCheckDebouncerSingleton.mu.Unlock()
+	defaultPoolScaleCheckDebouncerSingleton.history = make(map[string][]int)
+}
+
+// setPoolScaleCheckDebouncerWindowForTesting overrides the singleton's
+// confirm window. windowN<=1 makes the debouncer a no-op (it returns the
+// current sample unchanged), which is the default for cmd/gc tests so
+// existing assertions about single-call spawn behavior remain valid.
+// Returns the previous window for restoration.
+func setPoolScaleCheckDebouncerWindowForTesting(windowN int) int {
+	if windowN < 1 {
+		windowN = 1
+	}
+	defaultPoolScaleCheckDebouncerSingleton.mu.Lock()
+	defer defaultPoolScaleCheckDebouncerSingleton.mu.Unlock()
+	prev := defaultPoolScaleCheckDebouncerSingleton.windowN
+	defaultPoolScaleCheckDebouncerSingleton.windowN = windowN
+	defaultPoolScaleCheckDebouncerSingleton.history = make(map[string][]int)
+	return prev
+}
diff --git a/cmd/gc/pool_scale_check_debouncer_test.go b/cmd/gc/pool_scale_check_debouncer_test.go
new file mode 100644
index 0000000000..715f851668
--- /dev/null
+++ b/cmd/gc/pool_scale_check_debouncer_test.go
@@ -0,0 +1,156 @@
+package main
+
+import (
+	"reflect"
+	"testing"
+)
+
+// withPoolScaleCheckDebouncerWindow installs a fresh debouncer at the given
+// window for one test and restores the prior window on cleanup. Use this
+// inside debouncer-specific tests because main_test.go's TestMain pins the
+// singleton at window=1 so unrelated tests don't get debouncing applied.
+func withPoolScaleCheckDebouncerWindow(t *testing.T, windowN int) {
+	t.Helper()
+	prev := setPoolScaleCheckDebouncerWindowForTesting(windowN)
+	t.Cleanup(func() {
+		setPoolScaleCheckDebouncerWindowForTesting(prev)
+	})
+}
+
+func TestPoolScaleCheckDebouncer_ColdStartPassesThroughBelowWindow(t *testing.T) {
+	withPoolScaleCheckDebouncerWindow(t, 3)
+
+	got := debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": 1}, nil)
+	if want := map[string]int{"local-core.builder": 1}; !reflect.DeepEqual(got, want) {
+		t.Fatalf("first sample: got %v, want %v (history smaller than window must pass raw count)", got, want)
+	}
+	got = debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": 1}, nil)
+	if want := map[string]int{"local-core.builder": 1}; !reflect.DeepEqual(got, want) {
+		t.Fatalf("second sample: got %v, want %v (still below window)", got, want)
+	}
+}
+
+func TestPoolScaleCheckDebouncer_SteadyDemandConfirmsAfterWindow(t *testing.T) {
+	withPoolScaleCheckDebouncerWindow(t, 3)
+
+	for i := 0; i < 3; i++ {
+		got := debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": 1}, nil)
+		if got["local-core.builder"] != 1 {
+			t.Fatalf("tick %d: got %d, want 1 (every sample is 1, min over window must be 1)", i, got["local-core.builder"])
+		}
+	}
+	// Tenth tick still 1 — window slides forward, but min stays 1.
+	for i := 0; i < 10; i++ {
+		got := debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": 1}, nil)
+		if got["local-core.builder"] != 1 {
+			t.Fatalf("sustained tick %d: got %d, want 1", i, got["local-core.builder"])
+		}
+	}
+}
+
+func TestPoolScaleCheckDebouncer_SuppressesSingleTickFlap(t *testing.T) {
+	withPoolScaleCheckDebouncerWindow(t, 3)
+
+	// Mirror the fm-0ubml4 pathology: scale_check flaps 1, 0, 1, 0, 1, 0...
+	// Once the window is full (after 3 samples), the minimum over any window
+	// containing at least one 0 must be 0, which kills the spawn loop.
+	samples := []int{1, 0, 1, 0, 1, 0}
+	results := make([]int, 0, len(samples))
+	for _, s := range samples {
+		got := debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": s}, nil)
+		results = append(results, got["local-core.builder"])
+	}
+	// First two samples pass through (window not yet full).
+	// Samples 3+ are min over [1,0,1]=0, [0,1,0]=0, [1,0,1]=0, [0,1,0]=0.
+	want := []int{1, 0, 0, 0, 0, 0}
+	if !reflect.DeepEqual(results, want) {
+		t.Fatalf("flap sequence: got %v, want %v", results, want)
+	}
+}
+
+func TestPoolScaleCheckDebouncer_PartialTemplatesPassThroughAndDoNotPoisonWindow(t *testing.T) {
+	withPoolScaleCheckDebouncerWindow(t, 3)
+
+	// A partial sample must not be recorded into the per-template history —
+	// otherwise a probe failure (count defaults to 0) would force min=0 for
+	// the next two ticks even after the probe recovers and reports steady 1.
+	for i := 0; i < 3; i++ {
+		got := debouncePoolScaleCheckCounts(
+			map[string]int{"local-core.builder": 0},
+			map[string]bool{"local-core.builder": true},
+		)
+		if got["local-core.builder"] != 0 {
+			t.Fatalf("partial tick %d: got %d, want 0 (partial passes raw through)", i, got["local-core.builder"])
+		}
+	}
+	// Three real samples of 1 should now confirm — partial ticks weren't
+	// recorded, so the window fills cleanly from real samples.
+	for i := 0; i < 3; i++ {
+		got := debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": 1}, nil)
+		if got["local-core.builder"] != 1 {
+			t.Fatalf("post-partial real tick %d: got %d, want 1", i, got["local-core.builder"])
+		}
+	}
+}
+
+func TestPoolScaleCheckDebouncer_WindowOfOneIsNoOp(t *testing.T) {
+	withPoolScaleCheckDebouncerWindow(t, 1)
+
+	// Window=1 is what TestMain installs so existing tests behave as before.
+	// Whatever the raw count is, it must be returned unchanged on every tick.
+	for _, s := range []int{1, 0, 5, 0, 1} {
+		got := debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": s}, nil)
+		if got["local-core.builder"] != s {
+			t.Fatalf("window=1 sample %d: got %d, want %d (debouncer must be a no-op)", s, got["local-core.builder"], s)
+		}
+	}
+}
+
+func TestPoolScaleCheckDebouncer_PerTemplateIsolation(t *testing.T) {
+	withPoolScaleCheckDebouncerWindow(t, 3)
+
+	// Builder flaps; planner is steady. Each template's history must be
+	// independent — the planner's confirmed count must not collapse to 0
+	// just because the builder's window saw a 0 sample.
+	type tick struct {
+		builder int
+		planner int
+	}
+	ticks := []tick{
+		{builder: 1, planner: 1},
+		{builder: 0, planner: 1},
+		{builder: 1, planner: 1},
+		{builder: 0, planner: 1},
+	}
+	for i, tk := range ticks {
+		got := debouncePoolScaleCheckCounts(
+			map[string]int{"local-core.builder": tk.builder, "local-core.planner": tk.planner},
+			nil,
+		)
+		if got["local-core.planner"] != 1 {
+			t.Fatalf("tick %d planner: got %d, want 1 (steady demand must not be debounced by sibling template's flap)", i, got["local-core.planner"])
+		}
+	}
+}
+
+func TestPoolScaleCheckDebouncer_ResetClearsHistory(t *testing.T) {
+	withPoolScaleCheckDebouncerWindow(t, 3)
+
+	// Fill the window with 0s — without reset, the next real sample of 1
+	// would be suppressed to 0 by the min over [0,0,1].
+	for i := 0; i < 3; i++ {
+		debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": 0}, nil)
+	}
+	got := debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": 1}, nil)
+	if got["local-core.builder"] != 0 {
+		t.Fatalf("pre-reset: got %d, want 0 (history [0,0,1] → min=0 confirms suppression)", got["local-core.builder"])
+	}
+
+	resetPoolScaleCheckDebouncerForTesting()
+
+	// After reset, the next sample is a cold start again — passes through.
+	got = debouncePoolScaleCheckCounts(map[string]int{"local-core.builder": 1}, nil)
+	if got["local-core.builder"] != 1 {
+		t.Fatalf("post-reset: got %d, want 1 (history is empty, raw count must pass through)", got["local-core.builder"])
+	}
+}