diff --git a/cmd/cost_explain.go b/cmd/cost_explain.go
index 6defcd5..95a8af7 100644
--- a/cmd/cost_explain.go
+++ b/cmd/cost_explain.go
@@ -104,7 +104,7 @@ func Explain(ctx context.Context, report *variancev1.Result) (*ExplainReport, er
 		visible := false
 		for _, evidence := range report.MissingEvidence {
 			detail := strings.TrimLeft(evidence, " \t\n\r")
-			if strings.HasPrefix(detail, "actual cost:") && strings.Contains(detail[len("actual cost:"):], report.UnavailableReason) {
+			if strings.HasPrefix(detail, "actual cost:") && strings.TrimSpace(detail[len("actual cost:"):]) == strings.TrimSpace(report.UnavailableReason) {
 				visible = true
 				break
 			}
@@ -152,6 +152,9 @@ func Explain(ctx context.Context, report *variancev1.Result) (*ExplainReport, er
 				}
 			}
 			sum += c.Amount.Amount
+			if math.IsNaN(sum) || math.IsInf(sum, 0) {
+				return nil, fmt.Errorf("contributions[%d] and residual total must be finite", i)
+			}
 			out.ConfidenceScore = math.Min(out.ConfidenceScore, c.ConfidenceScore)
 			action := "Review workload measurements and cost evidence before adjusting this driver"
 			switch c.Kind {
diff --git a/cmd/cost_explain_acceptance_test.go b/cmd/cost_explain_acceptance_test.go
new file mode 100644
index 0000000..cb045ea
--- /dev/null
+++ b/cmd/cost_explain_acceptance_test.go
@@ -0,0 +1,56 @@
+package cmd
+
+import (
+	"context"
+	"math"
+	"testing"
+
+	costv1 "github.com/IntelIP/ProfitCtl/pkg/contracts/cost/v1"
+	variancev1 "github.com/IntelIP/ProfitCtl/pkg/variance/v1"
+	"github.com/stretchr/testify/require"
+)
+
+// Operator-owned regressions from rejected agent candidates. Workers may read,
+// but cannot edit this file or the controller's acceptance commands.
+func TestCostExplainAcceptanceActualReason(t *testing.T) {
+	f, _ := varianceFixture(t, "forecast", 10, 2, 20, costv1.DriverFixed)
+	for _, tc := range []struct {
+		name, evidence string
+		valid          bool
+	}{
+		{"canonical actual reason", "actual cost: invoice not delivered", true},
+		{"contradictory substring", "actual cost: not invoice not delivered", false},
+		{"reason only in forecast clause", "actual cost: ledger pending; forecast invoice not delivered", false},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			result, err := Diff(context.Background(), &f, nil)
+			require.NoError(t, err)
+			result.UnavailableReason = "invoice not delivered"
+			result.MissingEvidence = []string{tc.evidence}
+			report, err := Explain(context.Background(), result)
+			if !tc.valid {
+				require.ErrorContains(t, err, "missing_evidence")
+				return
+			}
+			require.NoError(t, err)
+			require.Equal(t, result.UnavailableReason, report.UnavailableReason)
+			require.Nil(t, report.Actual)
+		})
+	}
+}
+
+func TestCostExplainAcceptanceAccumulationOverflow(t *testing.T) {
+	f, fd := varianceFixture(t, "forecast", 10, 2, 20, costv1.DriverFixed)
+	a, ad := varianceFixture(t, "actual", 20, 2, 40, costv1.DriverFixed)
+	result, err := variancev1.Compare(variancev1.Input{Forecast: f, Actual: &a, Drivers: []variancev1.DriverPair{{Forecast: fd, Actual: ad}}})
+	require.NoError(t, err)
+	require.NotEmpty(t, result.Contributions)
+	contribution := result.Contributions[0]
+	result.Contributions = result.Contributions[:0]
+	for _, amount := range []float64{math.MaxFloat64, math.MaxFloat64, -math.MaxFloat64, -math.MaxFloat64} {
+		contribution.Amount.Amount = amount
+		result.Contributions = append(result.Contributions, contribution)
+	}
+	_, err = Explain(context.Background(), &result)
+	require.Error(t, err, "Finite inputs must not allow an infinite accumulated contribution total")
+}
diff --git a/cmd/cost_variance_test.go b/cmd/cost_variance_test.go
index 8567b82..1f210f9 100644
--- a/cmd/cost_variance_test.go
+++ b/cmd/cost_variance_test.go
@@ -400,7 +400,11 @@ func TestCostExplainUnavailableReasonEvidence(t *testing.T) {
 		{"missing evidence", "invoice not delivered", nil, "missing_evidence"},
 		{"unrelated evidence", "invoice not delivered", []string{"actual cost: ledger pending"}, "missing_evidence"},
 		{"blank evidence", "invoice not delivered", []string{"  "}, "missing_evidence"},
+		{"negated actual reason", "invoice not delivered", []string{"actual cost: not invoice not delivered"}, "missing_evidence"},
+		{"forecast-only reason", "invoice not delivered", []string{"actual cost: ledger pending; forecast invoice not delivered"}, "missing_evidence"},
+		{"unrelated entry with reason", "invoice not delivered", []string{"actual cost: ledger pending", "forecast: invoice not delivered"}, "missing_evidence"},
 		{"valid supplied reason", "invoice not delivered", []string{"actual cost: invoice not delivered"}, ""},
+		{"valid reason among other evidence", "invoice not delivered", []string{"forecast: pending", " actual cost: invoice not delivered "}, ""},
 	} {
 		t.Run(tc.name, func(t *testing.T) {
 			result := *base
@@ -413,9 +417,15 @@ func TestCostExplainUnavailableReasonEvidence(t *testing.T) {
 			}
 			require.NoError(t, err)
 			require.Equal(t, tc.reason, report.UnavailableReason)
-			require.Contains(t, report.MissingEvidence, "actual cost: "+tc.reason)
-			require.Nil(t, report.Actual)
-			require.Nil(t, report.VariancePercent)
+			require.Equal(t, tc.evidence, report.MissingEvidence)
+			payload, err := json.Marshal(report)
+			require.NoError(t, err)
+			var fields map[string]json.RawMessage
+			require.NoError(t, json.Unmarshal(payload, &fields))
+			require.JSONEq(t, `"invoice not delivered"`, string(fields["unavailable_reason"]))
+			for _, key := range []string{"actual", "variance_absolute", "variance_percent", "residual"} {
+				require.Equal(t, "null", string(fields[key]), key)
+			}
 		})
 	}
 }
@@ -436,6 +446,38 @@ func TestCostExplainMissingActualEvidenceIdentity(t *testing.T) {
 	require.Contains(t, report.MissingEvidence, "actual cost: invoice not delivered")
 }
 
+func TestCostExplainRejectsRunningTotalOverflow(t *testing.T) {
+	f, fd := varianceFixture(t, "forecast", 10, 2, 20, costv1.DriverFixed)
+	a, ad := varianceFixture(t, "actual", 20, 2, 40, costv1.DriverFixed)
+	base, err := variancev1.Compare(variancev1.Input{Forecast: f, Actual: &a, Drivers: []variancev1.DriverPair{{Forecast: fd, Actual: ad}}})
+	require.NoError(t, err)
+	require.NotEmpty(t, base.Contributions)
+	for _, tc := range []struct {
+		name     string
+		residual float64
+		amounts  []float64
+	}{
+		{"contributions cancel after overflow", 0, []float64{math.MaxFloat64, math.MaxFloat64, -math.MaxFloat64, -math.MaxFloat64, 20}},
+		{"residual plus contribution overflows", math.MaxFloat64, []float64{math.MaxFloat64}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			result := base
+			residual := *base.Residual
+			residual.Amount = tc.residual
+			result.Residual = &residual
+			result.Contributions = nil
+			for _, amount := range tc.amounts {
+				contribution := base.Contributions[0]
+				contribution.Amount.Amount = amount
+				result.Contributions = append(result.Contributions, contribution)
+			}
+			_, err := Explain(context.Background(), &result)
+			require.ErrorContains(t, err, "total must be finite")
+			require.NotContains(t, err.Error(), "reconcile")
+		})
+	}
+}
+
 func TestCostExplainInvalidReconciliation(t *testing.T) {
 	result := variancev1.Result{SchemaVersion: variancev1.SchemaVersion, Status: "available", Actual: &costv1.Money{Amount: 2, Currency: "USD"}, AbsoluteVariance: &costv1.Money{Amount: 2, Currency: "USD"}, Residual: &costv1.Money{Amount: math.NaN(), Currency: "USD"}}
 	_, err := Explain(context.Background(), &result)
diff --git a/docs/cost-variance.md b/docs/cost-variance.md
index 4a315a9..fa440b0 100644
--- a/docs/cost-variance.md
+++ b/docs/cost-variance.md
@@ -15,6 +15,6 @@ Input is a `variance/v1.Input` JSON object:
 
 The abbreviated evidence placeholders above must be replaced by complete `cost/v1` evidence objects with source identity, capture date, confidence and rationale. For an available comparison, set `actual` to a validated observation and omit `unavailable_reason`. To explain drivers supply `drivers` as pairs of full `{ "forecast": CostDriver, "actual": CostDriver }` records referenced by both observations. Optional `exchange_rates` provide snapshot currency conversions and `receipts` provide exact observation-paired delivery evidence. Missing exchange rates fail; no live rates are fetched. The forecast and actual windows and workload must match.
 
-The diff JSON matches [`result.schema.json`](../schemas/cost-variance/v1/result.schema.json): `profitctl.cost-variance/v1`, status, window, observation IDs, currency-denominated forecast/actual/absolute variance and residual, percentage variance, contributions, missing evidence, rounding tolerance, and timestamped source references. A zero forecast gives `percentage_variance: null`; unavailable actual gives null actual, variance and residual, not a fabricated zero. Contributions are ranked by absolute amount; the residual is never hidden. `explain` returns the same amounts under `variance_absolute` and `variance_percent`, plus ranked `driver_contributions` containing their evidence, confidence score, receipts, and review actions. Review actions are suggestions, never automatic remediation or asserted root causes. Without an exact receipt, delivery attribution is `unknown` and missing evidence is named. A source label alone cannot establish causation.
+The diff JSON matches [`result.schema.json`](../schemas/cost-variance/v1/result.schema.json): `profitctl.cost-variance/v1`, status, window, observation IDs, currency-denominated forecast/actual/absolute variance and residual, percentage variance, contributions, missing evidence, rounding tolerance, and timestamped source references. A zero forecast gives `percentage_variance: null`; unavailable actual gives null actual, variance and residual, not a fabricated zero. To explain an unavailable actual, `missing_evidence` must include a direct `actual cost: <unavailable_reason>` entry (for example, `actual cost: invoice not delivered`). Negation such as `actual cost: not invoice not delivered`, or a reason appearing only in another clause such as `actual cost: ledger pending; forecast invoice not delivered`, does not count. Contributions are ranked by absolute amount; the residual is never hidden. `explain` returns the same amounts under `variance_absolute` and `variance_percent`, plus ranked `driver_contributions` containing their evidence, confidence score, receipts, and review actions. Review actions are suggestions, never automatic remediation or asserted root causes. Without an exact receipt, delivery attribution is `unknown` and missing evidence is named. A source label alone cannot establish causation.
 
 Malformed/invalid input and aliased output paths return exit code 2; output failures return code 3. No partial JSON is printed for invalid input. The exported `cmd.Diff(ctx, forecast, actual)` offers a totals-only comparison returning `*variancev1.Result`; `cmd.Explain(ctx, result)` returns `*cmd.ExplainReport`. For full attribution call `variancev1.Compare(variancev1.Input)` first. The peer engine's concrete exported type is `Result` (not `DiffReport`) and its comparison entry point is `Compare`.