diff --git a/Cargo.lock b/Cargo.lock
index a4bd3044..6a5dd471 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -169,6 +169,7 @@ dependencies = [
  "data-encoding",
  "serde",
  "serde_json",
+ "serde_json_canonicalizer",
  "thiserror 1.0.69",
 ]
 
@@ -2508,6 +2509,12 @@ version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"
 
+[[package]]
+name = "ryu-js"
+version = "1.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "04d056b875a9d2e6cb9a61d127afee9ac5999b9f87bcb32079d1318e505be714"
+
 [[package]]
 name = "same-file"
 version = "1.0.6"
@@ -2629,6 +2636,17 @@ dependencies = [
  "zmij",
 ]
 
+[[package]]
+name = "serde_json_canonicalizer"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fe52319a927259afbfa5180c5157cd8167edfd3e8c254f9558c7fef44c5649f2"
+dependencies = [
+ "ryu-js",
+ "serde",
+ "serde_json",
+]
+
 [[package]]
 name = "serde_spanned"
 version = "0.6.9"
diff --git a/atomic-canonical/Cargo.toml b/atomic-canonical/Cargo.toml
index 58ceac78..c6c48ce5 100644
--- a/atomic-canonical/Cargo.toml
+++ b/atomic-canonical/Cargo.toml
@@ -16,4 +16,5 @@ bs58 = { workspace = true }
 chrono = { workspace = true }
 thiserror = { workspace = true }
 data-encoding = { workspace = true }
+serde_json_canonicalizer = "0.3"
 atomic-identity = { workspace = true }
diff --git a/atomic-canonical/src/jcs.rs b/atomic-canonical/src/jcs.rs
index b1ec4d09..d0682a2b 100644
--- a/atomic-canonical/src/jcs.rs
+++ b/atomic-canonical/src/jcs.rs
@@ -9,63 +9,42 @@
 //! hash (`hash.rs`) and the Data Integrity proof (`proof.rs`) go through
 //! `canonicalize`, so the two can never drift.
 //!
-//! Object keys are sorted by UTF-16 code units as RFC 8785 §3.2.3 specifies
-//! (not by UTF-8 bytes — the two differ once keys leave the BMP, e.g. an
-//! emoji key sorts after `\u{ff61}` in UTF-8 but before it in UTF-16).
+//! # The algorithm is delegated rather than written here
 //!
-//! Scope note: numbers are emitted via `serde_json`'s formatter, which
-//! matches RFC 8785 for the integer values our vocabulary admits; the full
-//! ECMAScript number-to-string algorithm is only needed if floating-point
-//! payloads are ever admitted.
+//! The bytes come from `serde_json_canonicalizer`. Two of the three parts of
+//! RFC 8785 are straightforward to write by hand and the third is not:
+//!
+//! * Object keys sort by UTF-16 code units as §3.2.3 specifies, not by UTF-8
+//!   bytes — the two differ once a key leaves the BMP, e.g. an emoji key sorts
+//!   after `\u{ff61}` in UTF-8 but before it in UTF-16.
+//! * Strings take the short form for the seven named escapes and lowercase
+//!   `\u00xx` for the rest of C0 (§3.2.2.2).
+//! * **Numbers take the ECMAScript `Number::toString` algorithm** (§3.2.2.3),
+//!   which `serde_json`'s formatter is not. `serde_json` writes `-0.0` where the
+//!   algorithm writes `0`, keeps a `.0` on an integer-valued double where the
+//!   algorithm drops it, and crosses between decimal and exponent notation at
+//!   different magnitudes: `1e-6` for `0.000001`, `2.9514790517935283e+20` for
+//!   `295147905179352830000`. The delegate formats through `ryu_js`, which is
+//!   the ECMAScript variant the section names, and it serializes through an
+//!   explicit heap stack rather than the call stack. The cases that used to
+//!   diverge are pinned in `tests/vectors/`.
+//!
+//! Scope note: the entry point takes an already-parsed [`Value`], so faults that
+//! only exist in the wire bytes cannot be decided here — a repeated object
+//! member is gone before this function is called, and RFC 8785 admits an integer
+//! past 2^53 by rounding it to its double. Those belong to a strict decoder at
+//! the boundary where the bytes arrive; `tests/vectors/INGEST-BOUNDARY.md` names
+//! the cases and what closes them.
 
 use serde_json::Value;
 
 /// Canonicalize a JSON value into its RFC-8785 string form.
 pub fn canonicalize(value: &Value) -> String {
-    let mut out = String::new();
-    write_value(&mut out, value);
-    out
-}
-
-fn write_value(out: &mut String, value: &Value) {
-    match value {
-        Value::Null => out.push_str("null"),
-        Value::Bool(true) => out.push_str("true"),
-        Value::Bool(false) => out.push_str("false"),
-        Value::Number(n) => out.push_str(&n.to_string()),
-        Value::String(s) => write_json_string(out, s),
-        Value::Array(items) => {
-            out.push('[');
-            for (i, item) in items.iter().enumerate() {
-                if i > 0 {
-                    out.push(',');
-                }
-                write_value(out, item);
-            }
-            out.push(']');
-        }
-        Value::Object(map) => {
-            let mut keys: Vec<&String> = map.keys().collect();
-            keys.sort_by(|a, b| a.encode_utf16().cmp(b.encode_utf16()));
-            out.push('{');
-            for (i, key) in keys.iter().enumerate() {
-                if i > 0 {
-                    out.push(',');
-                }
-                write_json_string(out, key);
-                out.push(':');
-                write_value(out, &map[*key]);
-            }
-            out.push('}');
-        }
-    }
-}
-
-/// Emit a JSON string with standard escaping. `serde_json` produces a valid,
-/// minimally-escaped JSON string literal (quotes included), which matches JCS
-/// for the ASCII content in our records.
-fn write_json_string(out: &mut String, s: &str) {
-    out.push_str(&serde_json::to_string(s).expect("string serialization is infallible"));
+    // Infallible for a `Value`: there is no writer to fail against, every
+    // member name is already a Rust `String`, and the delegate's only other
+    // error path is a non-finite float, which `Value` cannot hold. This mirrors
+    // the expectation the hand-written string helper carried before it.
+    serde_json_canonicalizer::to_string(value).expect("canonicalizing a Value is infallible")
 }
 
 #[cfg(test)]
diff --git a/atomic-canonical/tests/jcs_vectors.rs b/atomic-canonical/tests/jcs_vectors.rs
new file mode 100644
index 00000000..3b668362
--- /dev/null
+++ b/atomic-canonical/tests/jcs_vectors.rs
@@ -0,0 +1,118 @@
+//! Canonicalization vectors, one fixture per divergence.
+//!
+//! Each file in `tests/vectors/` carries an input document, the divergence it
+//! pins, and the canonical form RFC 8785 requires. The number cases are Appendix
+//! B rows: the canonical text of a double is one string and no other, so a
+//! canonicalizer that writes `1e-6` where the algorithm writes `0.000001`
+//! produces a different hash for the same value, and a second implementation
+//! then rejects a proof this one accepts.
+//!
+//! `tests/vectors/INGEST-BOUNDARY.md` lists the conformance cases this entry
+//! point cannot decide, because it receives an already-parsed value rather than
+//! the bytes.
+
+use std::fs;
+use std::path::{Path, PathBuf};
+
+use atomic_canonical::jcs;
+use serde_json::Value;
+
+fn vector_dir() -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR")).join("tests/vectors")
+}
+
+fn fixture(name: &str) -> Value {
+    let path = vector_dir().join(format!("{name}.json"));
+    let raw = fs::read_to_string(&path).unwrap_or_else(|e| panic!("read {path:?}: {e}"));
+    serde_json::from_str(&raw).unwrap_or_else(|e| panic!("parse {path:?}: {e}"))
+}
+
+/// Parse the fixture's input the way any caller receiving JSON does, canonicalize
+/// it, and require the exact bytes.
+fn assert_canonical(name: &str) {
+    let f = fixture(name);
+    let text = f["input"]
+        .as_str()
+        .expect("fixture carries an input string");
+    let value: Value = serde_json::from_str(text).expect("fixture input is valid JSON");
+    let expected = f["expect"]["canonical"]
+        .as_str()
+        .expect("fixture expects a canonical form");
+    assert_eq!(
+        jcs::canonicalize(&value),
+        expected,
+        "{name} ({}): {}",
+        f["vector"].as_str().unwrap_or("?"),
+        f["divergence"].as_str().unwrap_or("")
+    );
+}
+
+#[test]
+fn number_decimal_below_exponent_threshold() {
+    assert_canonical("number-decimal-below-exponent-threshold");
+}
+
+#[test]
+fn number_decimal_2pow68() {
+    assert_canonical("number-decimal-2pow68");
+}
+
+#[test]
+fn number_decimal_999999999999999700000() {
+    assert_canonical("number-decimal-999999999999999700000");
+}
+
+#[test]
+fn number_decimal_999999999999999900000() {
+    assert_canonical("number-decimal-999999999999999900000");
+}
+
+#[test]
+fn number_negative_small_decimal() {
+    assert_canonical("number-negative-small-decimal");
+}
+
+#[test]
+fn number_negative_zero() {
+    assert_canonical("number-negative-zero");
+}
+
+#[test]
+fn number_exponent_9_999999999999997e22() {
+    assert_canonical("number-exponent-9.999999999999997e22");
+}
+
+#[test]
+fn number_exponent_1_0000000000000001e23() {
+    assert_canonical("number-exponent-1.0000000000000001e23");
+}
+
+#[test]
+fn number_exponent_9_999999999999997e_minus_7() {
+    assert_canonical("number-exponent-9.999999999999997e-7");
+}
+
+#[test]
+fn number_rounded_to_its_double() {
+    assert_canonical("number-rounded-to-its-double");
+}
+
+/// The accept half of the depth condition: a document at 128 containers
+/// canonicalizes, and the delegate walks it on the heap rather than the call
+/// stack. The refusal half needs a fallible boundary -- see
+/// `tests/vectors/INGEST-BOUNDARY.md`.
+#[test]
+fn depth_at_the_cap_is_canonicalized() {
+    let f = fixture("depth-at-the-cap-is-canonicalized");
+    assert!(f["input_generated"].as_str().is_some());
+
+    let mut value = Value::Null;
+    for _ in 0..128 {
+        value = Value::Array(vec![value]);
+    }
+    let canonical = jcs::canonicalize(&value);
+    assert_eq!(
+        canonical,
+        format!("{}null{}", "[".repeat(128), "]".repeat(128))
+    );
+}
diff --git a/atomic-canonical/tests/vectors/INGEST-BOUNDARY.md b/atomic-canonical/tests/vectors/INGEST-BOUNDARY.md
new file mode 100644
index 00000000..20f86019
--- /dev/null
+++ b/atomic-canonical/tests/vectors/INGEST-BOUNDARY.md
@@ -0,0 +1,30 @@
+# Vectors this entry point cannot decide
+
+`jcs::canonicalize` takes an already-parsed `serde_json::Value`. Three conformance
+vectors ask for a refusal that no function of that shape can give, because the
+fault either vanished during the parse or is not a fault under RFC 8785 at all.
+They are listed here rather than dropped, so the follow-up has its brief in-tree.
+
+| vector | condition | what it asks for | why not here |
+| --- | --- | --- | --- |
+| `v0f4f2093061d303f` | duplicate member | reject `{"a":1,"a":2}` | `serde_json` keeps one of the two members while parsing, so the repeat is gone before `canonicalize` is called. Measured: the document canonicalizes to `{"a":2}`, and `{"a":2,"a":1}` to `{"a":1}` -- two wire documents, two canonical forms, and a signature over either verifies. |
+| `v679f56481420e45a` | unsafe integer | reject `9007199254740993` | RFC 8785 defers number formatting to ECMAScript, which has one numeric type, so the specification *admits* the token and writes the double it rounds to. Refusing it is the RFC 7493 I-JSON profile, which is a stricter profile rather than RFC 8785 itself. Pinned as an accept in `number-rounded-to-its-double.json`. |
+| `v97f5d8777e514257` | non-integer in a signed field | reject `0.7` | Same shape: RFC 8785 admits a fractional number. Refusing one is a field-level profile decision, not a canonicalization rule. |
+
+A fourth case is half-covered. `vd94ac70c9f0d84bf` asks a canonicalizer to refuse
+a document nested one container past a stated cap with a catchable error. The
+accept half is pinned in `depth-at-the-cap-is-canonicalized.json`; the refusal
+half needs a fallible boundary, and an infallible `canonicalize(&Value) -> String`
+has nowhere to put it.
+
+## What closes all four
+
+A strict decoder on the raw bytes, ahead of this function: it refuses a repeated
+member, caps nesting at 128 with an error rather than a stack walk, refuses a
+string that is not a sequence of Unicode scalar values, and optionally applies the
+I-JSON safe-integer profile. `jcs-admit` on crates.io does exactly that and then
+hands canonical output to the same `serde_json_canonicalizer` this file already
+uses, so adopting it adds refusals without changing a single byte of output.
+
+The place it belongs is wherever a document arrives as bytes rather than as a
+value built in-process.
diff --git a/atomic-canonical/tests/vectors/depth-at-the-cap-is-canonicalized.json b/atomic-canonical/tests/vectors/depth-at-the-cap-is-canonicalized.json
new file mode 100644
index 00000000..c0a3a890
--- /dev/null
+++ b/atomic-canonical/tests/vectors/depth-at-the-cap-is-canonicalized.json
@@ -0,0 +1,8 @@
+{
+  "vector": "vd94ac70c9f0d84bf (accept half)",
+  "divergence": "a document nested 128 containers deep canonicalizes, which is the accept side of the depth condition",
+  "input_generated": "128 nested arrays around null",
+  "expect": {
+    "canonical_shape": "128 open brackets, null, 128 close brackets"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-decimal-2pow68.json b/atomic-canonical/tests/vectors/number-decimal-2pow68.json
new file mode 100644
index 00000000..57d98b14
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-decimal-2pow68.json
@@ -0,0 +1,8 @@
+{
+  "vector": "v362600d69975d14c",
+  "divergence": "decimal-to-exponent threshold: a magnitude below 1e21 stays decimal",
+  "input": "{\"n\":295147905179352830000}",
+  "expect": {
+    "canonical": "{\"n\":295147905179352830000}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-decimal-999999999999999700000.json b/atomic-canonical/tests/vectors/number-decimal-999999999999999700000.json
new file mode 100644
index 00000000..11a1c1b0
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-decimal-999999999999999700000.json
@@ -0,0 +1,8 @@
+{
+  "vector": "vaa33c34f7bd1058b",
+  "divergence": "decimal notation retained just below the 1e21 threshold",
+  "input": "{\"n\":999999999999999700000}",
+  "expect": {
+    "canonical": "{\"n\":999999999999999700000}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-decimal-999999999999999900000.json b/atomic-canonical/tests/vectors/number-decimal-999999999999999900000.json
new file mode 100644
index 00000000..6c188bf1
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-decimal-999999999999999900000.json
@@ -0,0 +1,8 @@
+{
+  "vector": "vb177ef3b3a945a72",
+  "divergence": "decimal notation retained for the last double below 1e21",
+  "input": "{\"n\":999999999999999900000}",
+  "expect": {
+    "canonical": "{\"n\":999999999999999900000}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-decimal-below-exponent-threshold.json b/atomic-canonical/tests/vectors/number-decimal-below-exponent-threshold.json
new file mode 100644
index 00000000..96129517
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-decimal-below-exponent-threshold.json
@@ -0,0 +1,8 @@
+{
+  "vector": "vaf1b67f038ffff84",
+  "divergence": "decimal-to-exponent threshold: 1e-6 is written in decimal notation",
+  "input": "{\"n\":0.000001}",
+  "expect": {
+    "canonical": "{\"n\":0.000001}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-exponent-1.0000000000000001e23.json b/atomic-canonical/tests/vectors/number-exponent-1.0000000000000001e23.json
new file mode 100644
index 00000000..21bb9d5e
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-exponent-1.0000000000000001e23.json
@@ -0,0 +1,8 @@
+{
+  "vector": "vb7dd5fb8dcb6e345",
+  "divergence": "seventeen significant digits kept where sixteen would not round-trip",
+  "input": "{\"n\":1.0000000000000001e+23}",
+  "expect": {
+    "canonical": "{\"n\":1.0000000000000001e+23}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-exponent-9.999999999999997e-7.json b/atomic-canonical/tests/vectors/number-exponent-9.999999999999997e-7.json
new file mode 100644
index 00000000..4f2e4350
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-exponent-9.999999999999997e-7.json
@@ -0,0 +1,8 @@
+{
+  "vector": "vf2f6991c55096b1a",
+  "divergence": "shortest round-tripping digits below the 1e-6 threshold",
+  "input": "{\"n\":9.999999999999997e-7}",
+  "expect": {
+    "canonical": "{\"n\":9.999999999999997e-7}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-exponent-9.999999999999997e22.json b/atomic-canonical/tests/vectors/number-exponent-9.999999999999997e22.json
new file mode 100644
index 00000000..50f2ebfa
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-exponent-9.999999999999997e22.json
@@ -0,0 +1,8 @@
+{
+  "vector": "v843566cf72cd33c3",
+  "divergence": "shortest round-tripping digits in exponent notation",
+  "input": "{\"n\":9.999999999999997e+22}",
+  "expect": {
+    "canonical": "{\"n\":9.999999999999997e+22}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-negative-small-decimal.json b/atomic-canonical/tests/vectors/number-negative-small-decimal.json
new file mode 100644
index 00000000..afc61d9d
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-negative-small-decimal.json
@@ -0,0 +1,8 @@
+{
+  "vector": "vd2659c36ece7eb47",
+  "divergence": "a negative magnitude above 1e-6 stays decimal",
+  "input": "{\"n\":-0.0000033333333333333333}",
+  "expect": {
+    "canonical": "{\"n\":-0.0000033333333333333333}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-negative-zero.json b/atomic-canonical/tests/vectors/number-negative-zero.json
new file mode 100644
index 00000000..ad01f931
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-negative-zero.json
@@ -0,0 +1,8 @@
+{
+  "vector": "v9a8b364b8bc121de",
+  "divergence": "negative zero is written as 0, and -0.0 is not valid ECMAScript output at all",
+  "input": "{\"n\":-0.0}",
+  "expect": {
+    "canonical": "{\"n\":0}"
+  }
+}
diff --git a/atomic-canonical/tests/vectors/number-rounded-to-its-double.json b/atomic-canonical/tests/vectors/number-rounded-to-its-double.json
new file mode 100644
index 00000000..c23749de
--- /dev/null
+++ b/atomic-canonical/tests/vectors/number-rounded-to-its-double.json
@@ -0,0 +1,8 @@
+{
+  "vector": "v679f56481420e45a",
+  "divergence": "RFC 8785 treats every number as a double, so 2^53+1 canonicalizes to the double it rounds to",
+  "input": "{\"n\":9007199254740993}",
+  "expect": {
+    "canonical": "{\"n\":9007199254740992}"
+  }
+}