diff --git a/encoding/cbor/cbor.go b/encoding/cbor/cbor.go
index e7973ad9e..fe8449a82 100644
--- a/encoding/cbor/cbor.go
+++ b/encoding/cbor/cbor.go
@@ -1,34 +1,50 @@
 // Package cbor implements partial encoding/decoding of concise binary object
-// representation (CBOR) described in RFC 8949.
+// representation (CBOR) described in [RFC 8949].
 //
-// This package implements a subset of the specification required to support
-// the Smithy RPCv2-CBOR protocol and is NOT suitable for general application
-// use.
+// This package is intended for use only by the smithy client runtime. The
+// exported API therein is not considered stable and is subject to breaking
+// changes without notice. More specifically, this package implements a subset
+// of the RFC 8949 specification required to support the Smithy RPCv2-CBOR
+// protocol and is NOT suitable for general application use.
 //
-// As the encoding API operates strictly off of a constructed syntax tree, the
-// length of each data item in a Value will always be known and the encoder
-// will always generate definite-length encodings of container types (byte/text
-// string, list, map).
+// The following principal restrictions apply:
+//   - Map (major type 5) keys can only be strings.
+//   - Float16 (major type 7, 25) values can be read but not encoded. Any
+//     float16 encountered during decode is converted to float32.
+//   - Indefinite-length values can be read but not encoded. Since the encoding
+//     API operates strictly off of a constructed syntax tree, the length of each
+//     data item in a Value will always be known and the encoder will always
+//     generate definite-length variants.
 //
-// Conversely, the decoding API will handle both definite and indefinite
-// variations of encoded containers.
+// It is the responsibility of the caller to determine whether a decoded CBOR
+// integral or floating-point Value is suitable for its target (e.g. whether
+// the value of a CBOR Uint fits into a field modeled as a Smithy short).
+//
+// All CBOR tags (major type 6) are implicitly supported since the
+// encoder/decoder does not attempt to interpret a tag's contents. It is the
+// responsibility of the caller to both provide valid Tag values to encode and
+// to assert that a decoded Tag's contents are valid for its tag ID (e.g.
+// ensuring whether a Tag with ID 1, indicating an enclosed epoch timestamp,
+// actually contains a valid integral or floating-point CBOR Value).
+//
+// [RFC 8949]: https://www.rfc-editor.org/rfc/rfc8949.html
 package cbor
 
 // Value describes a CBOR data item.
 //
 // The following types implement Value:
-//   - Uint
-//   - NegInt
-//   - Slice
-//   - String
-//   - List
-//   - Map
-//   - Tag
-//   - Bool
-//   - Nil
-//   - Undefined
-//   - Float32
-//   - Float64
+//   - [Uint]
+//   - [NegInt]
+//   - [Slice]
+//   - [String]
+//   - [List]
+//   - [Map]
+//   - [Tag]
+//   - [Bool]
+//   - [Nil]
+//   - [Undefined]
+//   - [Float32]
+//   - [Float64]
 type Value interface {
 	len() int
 	encode(p []byte) int
@@ -49,10 +65,20 @@ var (
 	_ Value = Float64(0)
 )
 
-// Uint describes a CBOR uint (major type 0).
+// Uint describes a CBOR uint (major type 0) in the range [0, 2^64-1].
 type Uint uint64
 
-// NegInt describes a CBOR negative int (major type 1).
+// NegInt describes a CBOR negative int (major type 1) in the range [-2^64, -1].
+//
+// The "true negative" value of a type 1 is specified by RFC 8949 to be -1
+// minus the encoded value. The encoder/decoder applies this bias
+// automatically, e.g. the integral -100 is represented as NegInt(100), which
+// will which encode to/from hex 3863 (major 1, minor 24, argument 99).
+//
+// This implicitly means that the lower bound of this type -2^64 is represented
+// as the wraparound value NegInt(0). Deserializer implementations should take
+// care to guard against this case when deriving a value for a signed integral
+// type which was encoded as NegInt.
 type NegInt uint64
 
 // Slice describes a CBOR byte slice (major type 2).
diff --git a/encoding/cbor/float16.go b/encoding/cbor/float16.go
index 8c3ae2caf..eea42eed4 100644
--- a/encoding/cbor/float16.go
+++ b/encoding/cbor/float16.go
@@ -29,7 +29,7 @@ func splitf16(f uint16) (sign, exp, mantissa uint32) {
 // float32 where the hidden bit is 1, e.g.
 //
 // f16: 0    00000              0001010000 = 0.000101 * 2^(-14), which is equal to
-// f32: 0 00000000 00100000000000000000000 =     1.01 * 2^(-18)
+// f32: 0 01101101 01000000000000000000000 =     1.01 * 2^(-18)
 //
 // this is achieved by shifting the mantissa to the right until the leading bit
 // that == 1 reaches position 24, then the number of positions shifted over is
diff --git a/encoding/cbor/float16_test.go b/encoding/cbor/float16_test.go
new file mode 100644
index 000000000..ef97f3acc
--- /dev/null
+++ b/encoding/cbor/float16_test.go
@@ -0,0 +1,41 @@
+package cbor
+
+import (
+	"testing"
+)
+
+func TestFloat16To32(t *testing.T) {
+	for name, c := range map[string]struct {
+		In     uint16
+		Expect uint32
+	}{
+		"+infinity": {
+			0b0_11111_0000000000,
+			0b0_11111111_00000000000000000000000,
+		},
+		"-infinity": {
+			0b1_11111_0000000000,
+			0b1_11111111_00000000000000000000000,
+		},
+		"NaN": {
+			0b0_11111_0101010101,
+			0b0_11111111_01010101010000000000000,
+		},
+		"absolute zero": {0, 0},
+		"subnormal": {
+			0b0_00000_0001010000,
+			0b0_01101101_01000000000000000000000,
+		},
+		"normal": {
+			0b0_00001_0001010000,
+			0b0_0001110001_00010100000000000000000,
+		},
+	} {
+		t.Run(name, func(t *testing.T) {
+			if actual := float16to32(c.In); c.Expect != actual {
+				t.Errorf("%x != %x", c.Expect, actual)
+			}
+		})
+	}
+
+}