diff --git a/components/component.go b/components/component.go
index d30771ce..b6573539 100644
--- a/components/component.go
+++ b/components/component.go
@@ -158,6 +158,9 @@ type Ctx struct {
 	MaxCachedIdx int
 	// FilterStats receives cmdfilter's per-filter ledger (which command families pay
 	// off, and which output shapes matched nothing). nil = not recording.
+	//
+	// Read it through Stats(), never directly: in observe mode nothing is forwarded, so
+	// recording into an enforced-namespace field would report savings that never happened.
 	FilterStats FilterStatsSink
 	// ExistingBreakpoints is how many prompt-cache breakpoints the RAW request already
 	// carries, counted across `system`, `tools` and `messages` — which is what the
@@ -207,6 +210,30 @@ func (c *Ctx) TailOnly(i int) bool {
 	return i > c.MaxCachedIdx
 }
 
+// Stats returns the per-filter ledger sink, or nil in observe mode.
+//
+// Observe computes what compaction WOULD have done and forwards the request untouched, so
+// every enforced-namespace metric must stay zero: a figure that cannot be told apart from
+// a real saving is worse than no figure, because it silently inflates the product's own
+// headline. The savings totals are already namespaced (potential_* / projected_*), but the
+// filter ledger is not — an observe-only run was reporting real-looking `cmdfilter_families`
+// and `cmdfilter_filters` entries with no mode label and no hypothetical counterpart.
+//
+// Gating here rather than at the call site is deliberate. A component author reaching for
+// c.FilterStats has no reason to think about modes, and the next sink added to Ctx would
+// reproduce the bug; an accessor makes the safe path the only convenient one.
+//
+// Two enforced fields are deliberately NOT suppressed in observe mode, because they are real
+// rather than hypothetical: cg_added_ms_avg (a true measurement of the enforced path, which
+// correctly reads ~0) and context-guru's own model spend (observe measures off-path, and that
+// costs real money). Those are labelled instead — see metrics.Snapshot's observe notices.
+func (c *Ctx) Stats() FilterStatsSink {
+	if c == nil || c.Mode == ModeObserve {
+		return nil
+	}
+	return c.FilterStats
+}
+
 // Report is the per-component result, modelled after lean-ctx's ToolOutput
 // token accounting and headroom's record_pipeline_run inputs. The pipeline
 // fills TokensBefore/After/DurationMs; the component fills CacheKey and may set
diff --git a/components/offload/cmdfilter.go b/components/offload/cmdfilter.go
index c328065b..95e88c21 100644
--- a/components/offload/cmdfilter.go
+++ b/components/offload/cmdfilter.go
@@ -96,13 +96,13 @@ func (f *Cmdfilter) Offload(req *schemas.BifrostChatRequest, rep *components.Rep
 		key := selectorKey(content)
 		filt := f.reg.Match(key)
 		if filt == nil {
-			if c.FilterStats != nil {
+			if fs := c.Stats(); fs != nil {
 				// The miss ledger: it turns "which filter to write next" into data
 				// instead of guesswork (after rtk's parse_failures table). Log only the
 				// FIRST line — the selector is multi-line, and keying the bounded ledger
 				// on whole multi-line blobs would make almost every entry unique and
 				// exhaust the cap on noise instead of ranking real shapes.
-				c.FilterStats.FilterMiss(firstLine(key))
+				fs.FilterMiss(firstLine(key))
 			}
 			continue
 		}
@@ -141,8 +141,8 @@ func (f *Cmdfilter) Offload(req *schemas.BifrostChatRequest, rep *components.Rep
 			rep.Irreversible = true
 		}
 		schema.SetMessageText(m, newText)
-		if c.FilterStats != nil {
-			c.FilterStats.FilterAct(filt.Family(), filt.Name, stashKey, before-after)
+		if fs := c.Stats(); fs != nil {
+			fs.FilterAct(filt.Family(), filt.Name, stashKey, before-after)
 		}
 		changed++
 	}
diff --git a/components/offload/cmdfilter_test.go b/components/offload/cmdfilter_test.go
index d32f0cc2..d859b161 100644
--- a/components/offload/cmdfilter_test.go
+++ b/components/offload/cmdfilter_test.go
@@ -250,3 +250,43 @@ func TestFamilyMetricsAndSelectorMisses(t *testing.T) {
 		t.Fatalf("expected the unmatched selector to be logged, got %v", fs.misses)
 	}
 }
+
+// The per-filter ledger is an ENFORCED-namespace field with no potential_* counterpart, so
+// an observe-mode run populating it reports savings that never happened — and a consumer
+// cannot tell them apart from real ones. #31 named that as its primary correctness risk: a
+// mislabelled hypothetical is worse than no number, because it inflates the product's own
+// headline claim.
+//
+// The gate lives on Ctx.Stats() rather than at this component's two call sites, so it also
+// covers the next sink added to Ctx — a component author reaching for c.FilterStats has no
+// reason to think about modes. Asserting that sync DOES record proves the gate rather than a
+// dead sink.
+func TestObserveModeDoesNotRecordFilterStats(t *testing.T) {
+	const aptOut = "Reading package lists...\nSetting up git (1:2.43.0-1ubuntu7.3) ...\n" +
+		"Processing triggers for libc-bin (2.39-0ubuntu8.6) ...\n"
+
+	for _, tc := range []struct {
+		mode       components.Mode
+		wantRecord bool
+	}{{components.ModeSync, true}, {components.ModeObserve, false}} {
+		sink := &recordingSink{}
+		f := newFilterComp(t, "min_size: 1\n")
+		req := &schemas.BifrostChatRequest{Provider: schemas.Anthropic,
+			Input: []schemas.ChatMessage{cmdToolMsg(aptOut)}}
+		c := &components.Ctx{Ctx: context.Background(), Session: "s",
+			Store: store.NewMemory(store.Options{}), MaxCachedIdx: -1,
+			Mode: tc.mode, FilterStats: sink}
+		if _, err := f.Offload(req, &components.Report{}, c); err != nil {
+			t.Fatalf("mode=%s: %v", tc.mode, err)
+		}
+		if got := sink.acts + sink.misses; (got > 0) != tc.wantRecord {
+			t.Errorf("mode=%s: %d ledger events (acts=%d misses=%d), wantRecord=%v",
+				tc.mode, got, sink.acts, sink.misses, tc.wantRecord)
+		}
+	}
+}
+
+type recordingSink struct{ acts, misses int }
+
+func (r *recordingSink) FilterAct(_, _, _ string, _ int) { r.acts++ }
+func (r *recordingSink) FilterMiss(string)               { r.misses++ }