diff --git a/internal/object/float_tos_test.go b/internal/object/float_tos_test.go
new file mode 100644
index 00000000..b63d1ab7
--- /dev/null
+++ b/internal/object/float_tos_test.go
@@ -0,0 +1,61 @@
+// Copyright (c) the go-embedded-ruby/ruby authors
+//
+// SPDX-License-Identifier: BSD-3-Clause
+
+package object
+
+import (
+	"math"
+	"testing"
+)
+
+// TestFloatToSMatchesMRI pins Float#to_s against MRI's flo_to_s (numeric.c):
+// the shortest round-tripping digits are laid out in fixed notation while the
+// decimal point falls within -3..DBL_DIG (15), and in exponential notation
+// outside that, where the mantissa always carries a fractional digit.
+//
+// Every expected value is what `ruby` 4.0.5 prints for the same literal.
+func TestFloatToSMatchesMRI(t *testing.T) {
+	cases := []struct {
+		in   float64
+		want string
+	}{
+		// Fixed notation, including the "1.0 not 1" rule.
+		{0, "0.0"},
+		{math.Copysign(0, -1), "-0.0"},
+		{1, "1.0"},
+		{3.14, "3.14"},
+		{-2.5, "-2.5"},
+		{0.1, "0.1"},
+		{100, "100.0"},
+		{1e6, "1000000.0"}, // Go's 'g' verb gives "1e+06" here
+		// The upper boundary: decpt 15 stays fixed, 16 goes exponential.
+		{1e14, "100000000000000.0"},
+		{123456789012345, "123456789012345.0"},
+		{1e15, "1.0e+15"},
+		{1234567890123456, "1.234567890123456e+15"},
+		// The lower boundary: decpt -3 stays fixed, -4 goes exponential.
+		{1e-4, "0.0001"},
+		{1e-5, "1.0e-05"},
+		// Exponential, with and without a fraction of its own.
+		{1e-6, "1.0e-06"},
+		{1.5e-6, "1.5e-06"},
+		{-1e-6, "-1.0e-06"},
+		{1e20, "1.0e+20"},
+		{1e100, "1.0e+100"},
+		{5e-324, "5.0e-324"},
+		{math.MaxFloat64, "1.7976931348623157e+308"},
+		// The non-finite values.
+		{math.Inf(1), "Infinity"},
+		{math.Inf(-1), "-Infinity"},
+		{math.NaN(), "NaN"},
+	}
+	for _, c := range cases {
+		if got := Float(c.in).ToS(); got != c.want {
+			t.Errorf("Float(%v).ToS() = %q, want %q", c.in, got, c.want)
+		}
+		if got := Float(c.in).Inspect(); got != c.want {
+			t.Errorf("Float(%v).Inspect() = %q, want %q", c.in, got, c.want)
+		}
+	}
+}
diff --git a/internal/object/object.go b/internal/object/object.go
index e5097323..9692dcbe 100644
--- a/internal/object/object.go
+++ b/internal/object/object.go
@@ -101,18 +101,39 @@ func BigOf(v Value) (*big.Int, bool) {
 type Float float64
 
 func (f Float) ToS() string {
-	if math.IsInf(float64(f), 1) {
+	v := float64(f)
+	if math.IsInf(v, 1) {
 		return "Infinity"
 	}
-	if math.IsInf(float64(f), -1) {
+	if math.IsInf(v, -1) {
 		return "-Infinity"
 	}
-	if math.IsNaN(float64(f)) {
+	if math.IsNaN(v) {
 		return "NaN"
 	}
-	// Ruby always shows a decimal point for floats (1.0 not 1).
-	s := strconv.FormatFloat(float64(f), 'g', -1, 64)
-	if !strings.ContainsAny(s, ".eE") {
+	// MRI's flo_to_s (numeric.c) lays the shortest round-tripping digits out in
+	// fixed notation when the decimal point falls within -3..DBL_DIG (15), and in
+	// exponential notation otherwise. Go's 'g' verb decides differently — it
+	// compares the exponent against the number of significant digits — so 1e6
+	// rendered as "1e+06" where Ruby prints "1000000.0", and 1e15 as
+	// "1e+15" where Ruby prints "1.0e+15".
+	//
+	// 'e' with precision -1 gives the shortest digits plus the exponent, which is
+	// what decides the layout. Atoi cannot fail on an exponent Go just produced.
+	e := strconv.FormatFloat(v, 'e', -1, 64)
+	at := strings.IndexByte(e, 'e')
+	exp, _ := strconv.Atoi(e[at+1:])
+	if decpt := exp + 1; decpt < -3 || decpt > 15 {
+		// Ruby always shows a fractional digit in the mantissa: 1.0e-06, not 1e-06.
+		mant := e[:at]
+		if !strings.ContainsRune(mant, '.') {
+			mant += ".0"
+		}
+		return mant + e[at:]
+	}
+	s := strconv.FormatFloat(v, 'f', -1, 64)
+	// And a decimal point in fixed notation: 1.0, not 1.
+	if !strings.ContainsRune(s, '.') {
 		s += ".0"
 	}
 	return s
diff --git a/internal/vm/process_limits_internal_test.go b/internal/vm/process_limits_internal_test.go
index 4cbb2c97..8c80ab4e 100644
--- a/internal/vm/process_limits_internal_test.go
+++ b/internal/vm/process_limits_internal_test.go
@@ -384,11 +384,10 @@ p Process.times.utime > u`, "true\n"},
 		{`p Process.clock_getres(:TIME_BASED_CLOCK_REALTIME, :nanosecond)`, "1000000000\n"},
 		{`p Process.clock_getres(:GETRUSAGE_BASED_CLOCK_PROCESS_CPUTIME_ID, :nanosecond)`, "1000\n"},
 		{`p Process.clock_getres(Process::CLOCK_REALTIME, :nanosecond)`, "1000\n"},
-		// 1e-06 second, the same value MRI reports; the spelling differs only
-		// because this VM's Float#inspect drops the ".0" before the exponent
-		// (`p 0.000001` prints 1e-06 here and 1.0e-06 in MRI) — a core
-		// Float-formatting divergence, not a clock one.
-		{`p Process.clock_getres(Process::CLOCK_MONOTONIC)`, "1e-06\n"},
+		// 1e-06 second, the same value and now the same spelling as MRI: the
+		// Float#inspect divergence this case used to document (1e-06 here,
+		// 1.0e-06 in MRI) is fixed in internal/object.
+		{`p Process.clock_getres(Process::CLOCK_MONOTONIC)`, "1.0e-06\n"},
 		// argv0 is frozen and is the same object every time.
 		{`p [Process.argv0.is_a?(String), Process.argv0.frozen?, Process.argv0.equal?(Process.argv0)]`,
 			"[true, true, true]\n"},