diff --git a/cli/cmd/apply.go b/cli/cmd/apply.go
index 44c6f5a3..19f0b4a5 100644
--- a/cli/cmd/apply.go
+++ b/cli/cmd/apply.go
@@ -510,10 +510,23 @@ func toBoolPtr(v interface{}) *bool {
 // scenarioBoolKeys are the spec keys scenarioToRequest sends as true or false.
 var scenarioBoolKeys = []string{"enableIdempotence", "enableTransactions", "enableCrc"}
 
+// scenarioIntKeys are the spec keys scenarioToRequest reads as numbers. A
+// quoted number, a unit such as 10m, or a bare key was left out of the
+// request, and the run used its type's default without a word.
+var scenarioIntKeys = []string{
+	"records", "parallelProducers", "recordSizeBytes", "durationSeconds", "batchSize", "lingerMs", "numConsumers",
+	"replicationFactor", "partitions", "minInsyncReplicas", "targetThroughput", "fetchMinBytes", "fetchMaxWaitMs",
+}
+
 // scenarioSpecProblems names the spec keys of a scenario that
 // scenarioToRequest cannot send as written.
 func scenarioSpecProblems(s TestScenario) []string {
 	var problems []string
+	for _, key := range scenarioIntKeys {
+		if v, ok := s.Spec[key]; ok && toIntPtr(v) == nil {
+			problems = append(problems, fmt.Sprintf("spec.%s is %s; write a number", key, mcpScnShow(v)))
+		}
+	}
 	for _, key := range scenarioBoolKeys {
 		if v, ok := s.Spec[key]; ok && toBoolPtr(v) == nil {
 			problems = append(problems, fmt.Sprintf("spec.%s is %v; write true or false", key, v))
diff --git a/cli/cmd/apply_test.go b/cli/cmd/apply_test.go
index 1ffd1f32..9c85ddef 100644
--- a/cli/cmd/apply_test.go
+++ b/cli/cmd/apply_test.go
@@ -455,3 +455,48 @@ func TestApply_RefusesAFlagThatIsNotABoolean(t *testing.T) {
 		t.Errorf("%d requests reached the backend", requests)
 	}
 }
+
+// A number key that holds no number is refused too. A quoted number, a unit
+// or a bare key was left out of the request, so the run used its type's
+// default records and duration, and --wait graded that shorter run.
+func TestScenarioSpecProblems_ANumberKeyThatIsNotANumber(t *testing.T) {
+	var sf ScenarioFile
+	if err := yaml.Unmarshal([]byte("scenarios:\n  - name: x\n    type: LOAD\n    spec:\n      records: \"5000000\"\n      durationSeconds: 10m\n      lingerMs:\n"), &sf); err != nil {
+		t.Fatal(err)
+	}
+	problems := scenarioSpecProblems(sf.Scenarios[0])
+	if len(problems) != 3 || !slices.Contains(problems, `spec.records is "5000000"; write a number`) ||
+		!slices.Contains(problems, "spec.lingerMs is empty; write a number") {
+		t.Errorf("problems = %q", problems)
+	}
+	sf.Scenarios[0].Spec = map[string]any{"records": 5000000, "lingerMs": 0, "durationSeconds": 1.5}
+	if problems := scenarioSpecProblems(sf.Scenarios[0]); len(problems) != 0 {
+		t.Errorf("numbers, 0 among them, are fine: %q", problems)
+	}
+}
+
+// kates test apply refuses a file with such a number before it starts any of
+// its tests.
+func TestApply_RefusesANumberThatIsNotANumber(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "s.yaml")
+	data := "scenarios:\n  - name: first\n    type: LOAD\n  - name: second\n    type: LOAD\n    spec:\n      records: \"5000000\"\n"
+	if err := os.WriteFile(path, []byte(data), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	var requests int
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { requests++ }))
+	defer ts.Close()
+	apiClient = client.New(ts.URL)
+	output.ResetForTesting()
+	applyFile, applyWait = path, false
+	defer func() { applyFile = "" }()
+
+	err := testApplyCmd.RunE(testApplyCmd, nil)
+
+	if err == nil || !strings.Contains(err.Error(), `scenario 2 (second): spec.records is "5000000"; write a number`) {
+		t.Errorf("err = %v", err)
+	}
+	if requests != 0 {
+		t.Errorf("%d requests reached the backend", requests)
+	}
+}
diff --git a/cli/cmd/mcp_tools_scenario.go b/cli/cmd/mcp_tools_scenario.go
index 1376612d..aa51b280 100644
--- a/cli/cmd/mcp_tools_scenario.go
+++ b/cli/cmd/mcp_tools_scenario.go
@@ -898,9 +898,9 @@ func mcpScnCheckSpecKey(i int, key string, v any, testType string, fs *mcpScnFin
 	}
 }
 
-// mcpScnCheckSpecNumber flags a value scenarioToRequest leaves out (one that
-// is not a number, or a 0 where the field leaves one out), a number it cuts,
-// and one it sends outside the range the backend accepts.
+// mcpScnCheckSpecNumber flags a value that is not a number, which kates test
+// apply refuses, a 0 where the field leaves one out, a number scenarioToRequest
+// cuts, and one it sends outside the range the backend accepts.
 func mcpScnCheckSpecNumber(i int, field string, k mcpScnSpecKey, v any, fs *mcpScnFindings) {
 	var n int64
 	switch x := v.(type) {
@@ -916,12 +916,7 @@ func mcpScnCheckSpecNumber(i int, field string, k mcpScnSpecKey, v any, fs *mcpS
 			fs.add(i, mcpScnWarning, field, fmt.Sprintf("it is %v, which scenarioToRequest cuts to %d", x, n))
 		}
 	default:
-		read := "which scenarioToRequest reads as 0"
-		if k.sendsZero {
-			read = "which is not a number"
-		}
-		fs.add(i, mcpScnWarning, field, "it is "+mcpScnShow(v)+", "+read+", so it is left out "+
-			"of the request and the type default applies")
+		fs.add(i, mcpScnInvalid, field, "it is "+mcpScnShow(v)+"; kates test apply refuses it: write a number")
 		return
 	}
 	scale := max(k.scale, 1)
@@ -1276,7 +1271,7 @@ var mcpCaveatsScenario = []mcpCaveat{
 			"produce, and RPO is measured only when a resilience run marks a fault, so a scenario run never has one. " +
 			"Keys ValidationSpec does not name, such as maxDuplicatePercent, are dropped when the file is read.",
 		Refs: []string{
-			mcpAnchoredRef("cli/cmd/apply.go:29-39,217-250,339-351,525-580,597-622",
+			mcpAnchoredRef("cli/cmd/apply.go:29-39,217-250,339-351,538-593,610-635",
 				"type ValidationSpec struct {", `yaml:"maxCrcFailures,omitempty"`,
 				"if !applyWait {", "nonNil(validateSLAs(finalResult, scenario.Validate))",
 				"} else if r.SLA != nil {", `"✓ SLA Pass"`,
diff --git a/cli/cmd/mcp_tools_scenario_test.go b/cli/cmd/mcp_tools_scenario_test.go
index aa6f9e67..3d095908 100644
--- a/cli/cmd/mcp_tools_scenario_test.go
+++ b/cli/cmd/mcp_tools_scenario_test.go
@@ -455,7 +455,7 @@ validate:
 	if s := sc.Request.Spec; sc.Request.TestType != "STRESS" || s.Records != 0 || s.LingerMs == nil || *s.LingerMs != 0 {
 		t.Errorf("request = %+v", sc.Request.Spec)
 	}
-	mcpScnHasFinding(t, out, 0, "spec.records", mcpScnWarning, "reads as 0")
+	mcpScnHasFinding(t, out, 0, "spec.records", mcpScnInvalid, "kates test apply refuses it")
 	mcpScnHasFinding(t, out, 0, "spec.throughput", mcpScnWarning, "the scenario key for the producer rate is targetThroughput")
 	mcpScnHasFinding(t, out, 0, "spec.numRecords", mcpScnWarning, "the scenario key is records")
 	if f := mcpScnFindingsOn(out, 0, "spec.lingerMs"); len(f) != 0 {
@@ -768,6 +768,19 @@ func TestMCPDraftScenarioSpecKeysMatchScenarioToRequest(t *testing.T) {
 	if !slices.Equal(read, ours) {
 		t.Errorf("scenarioToRequest reads %v; draft_scenario describes %v", read, ours)
 	}
+	// scenarioSpecProblems checks every number and true/false key, so a key
+	// added to scenarioToRequest cannot skip apply's refusal.
+	kinds := map[byte][]string{}
+	for key, k := range mcpScnSpecKeys {
+		kinds[k.kind] = append(kinds[k.kind], key)
+	}
+	for kind, checked := range map[byte][]string{'i': scenarioIntKeys, 'b': scenarioBoolKeys} {
+		checked = slices.Sorted(slices.Values(checked))
+		slices.Sort(kinds[kind])
+		if !slices.Equal(checked, kinds[kind]) {
+			t.Errorf("scenarioSpecProblems checks %v; the %c keys are %v", checked, kind, kinds[kind])
+		}
+	}
 	// Each key's wire name is the JSON name of the field it sets.
 	for key, k := range mcpScnSpecKeys {
 		sc := TestScenario{Type: "LOAD", Spec: map[string]any{key: 7}}
diff --git a/cli/cmd/scenario_diff_test.go b/cli/cmd/scenario_diff_test.go
index d1431909..bd802ed7 100644
--- a/cli/cmd/scenario_diff_test.go
+++ b/cli/cmd/scenario_diff_test.go
@@ -273,14 +273,12 @@ func TestScenarioDiff_KeysApplyDoesNotSendAreNotCompared(t *testing.T) {
     spec:
       numRecords: 500000
       throughput: 5000
-      records: "50000"
       topic: ""
       partitions: 0
 `
 	d := diffScenarioJSON(t, "s.yaml", file,
 		diffRun("STRESS", `{"numRecords": 1000000, "throughput": -1, "partitions": 3, "topic": "stress-test"}`), "")
 	want := []scenarioDiffSkip{
-		{"spec.records", "it is not a number, so kates test apply leaves it out"},
 		{"spec.topic", "it is empty, so kates test apply leaves it out"},
 		{"spec.partitions", "kates test apply reads it as 0 and leaves it out"},
 		{"spec.numRecords", "kates test apply does not read it; the scenario key is records"},
@@ -384,6 +382,16 @@ func TestScenarioDiff_RefusesAFlagApplyRefuses(t *testing.T) {
 	}
 }
 
+// kates test apply refuses a number key that holds no number, so no run came
+// from such a scenario either.
+func TestScenarioDiff_RefusesANumberApplyRefuses(t *testing.T) {
+	const file = "scenarios:\n  - name: quoted\n    type: LOAD\n    spec:\n      records: \"50000\"\n"
+	_, err := runScenarioDiff(t, "s.yaml", file, diffRun("LOAD", `{"numRecords": 50000}`), "", "table")
+	if err == nil || !strings.Contains(err.Error(), `scenario 1 (quoted): spec.records is "50000"; write a number`) {
+		t.Errorf("err = %v", err)
+	}
+}
+
 // scenarioDiffFields compares every spec key scenarioToRequest reads (the keys
 // of mcpScnSpecKeys, which TestMCPDraftScenarioSpecKeysMatchScenarioToRequest
 // holds to apply.go), and reads each from the field its key sets in the
diff --git a/docs/book/13-scenario-files.md b/docs/book/13-scenario-files.md
index b21337e6..2c85599e 100644
--- a/docs/book/13-scenario-files.md
+++ b/docs/book/13-scenario-files.md
@@ -91,7 +91,7 @@ The `spec` object controls all test parameters. Every field is optional. The [Ka
 ::: {.callout-important}
 The Kates API merges `spec` with the test type's defaults, and every key below reaches the run. `targetThroughput` sets the producer rate in place of the type's default. `parallelProducers` counts only for STRESS and CAPACITY, and no test type reads `numConsumers`, so a LOAD scenario runs one producer and one consumer whatever they say.
 
-The three `enable` keys take `true` or `false`. `kates test apply` refuses a file where one holds anything else, such as `yes`, `on` or nothing, naming the scenario and the key, before it starts any of the file's tests.
+The three `enable` keys take `true` or `false`. `kates test apply` refuses a file where one holds anything else, such as `yes`, `on` or nothing, naming the scenario and the key, before it starts any of the file's tests. The number keys take a number: `kates test apply` and `kates scenario-diff` refuse a file where one holds a quoted number (`"5000000"`), a unit (`10m`, `5M`) or nothing, naming the scenario and the key, before any test starts.
 
 A key the scenario's type or benchmark backend cannot apply is refused: `kates test apply` gets a `400` naming it, and the scenario does not start. That is a rate other than -1 for SPIKE or CAPACITY, which run unthrottled; `consumerGroup` or a fetch setting for a type that takes no consumer settings (every type but LOAD, ENDURANCE and INTEGRITY); `enableCrc: true` for any type but INTEGRITY; [`enableIdempotence: true`](appendix-a-glossary.md#gl-idempotent-producer) or `enableTransactions: true` when [`acks`](appendix-a-glossary.md#gl-acks) is not `all` (SPIKE's default is `1`); `enableTransactions: true` with `enableIdempotence: false`, or on the `trogdor` benchmark backend. The [API Reference](11-api-reference.md#post-apitests) lists the rules; [Test Types Deep Dive](05-test-types.md) and [Data Integrity Verification](08-data-integrity.md) cover what the options do.
 :::