diff --git a/docs/design/native-sol-client/05-raw-body-parity.md b/docs/design/native-sol-client/05-raw-body-parity.md
new file mode 100644
index 000000000..dc3d1ff89
--- /dev/null
+++ b/docs/design/native-sol-client/05-raw-body-parity.md
@@ -0,0 +1,68 @@
+# Native Sol Raw-Body Parity Vector Design
+
+This note extends the parity corpus response shape documented in
+`02-design.md`.
+
+## Why
+
+`decode_response` parsed response bodies with serde_json's default best-effort
+float precision. Server-emitted entity scores `22.184388937230562` and
+`9.785157957944147` therefore rendered as `22.18438893723056` and
+`9.785157957944149`, while the Python oracle emitted the original bytes.
+
+The fix is to enable serde_json's `float_roundtrip` feature. That feature is an
+empty feature list: it adds no dependency edge, leaves `core/Cargo.lock`
+unchanged, and affects parsing only. Serialization behavior is unchanged.
+
+## `response.raw_body`
+
+The `response` shape inside each transport request now supports exactly one body
+source:
+
+| Field | Meaning |
+|---|---|
+| `json` | Parsed JSON value. The harness serializes it back to response body bytes. |
+| `raw_body` | String carrying the exact UTF-8 response body bytes. |
+
+A vector response must set exactly one of `json` or `raw_body`. Both-present,
+neither-present, and non-string `raw_body` cases panic or raise with a message
+naming the vector id.
+
+Use `raw_body` for any vector asserting float byte-parity. It exists so the
+numeric literal is never round-tripped through serde_json at vector-load time.
+
+## Harness Symmetry
+
+The rule is enforced in both
+`core/crates/solstone-core-sol-client-cli/tests/parity.rs` and
+`tests/native_sol/run_python_parity.py`. Both harnesses read the same committed
+corpus, so any future schema extension must land in both places.
+
+## Coverage Classification
+
+`scripts/check_native_sol_coverage.py` classifies success and failure from
+`response.status`, not from `response.json`. Vectors using `raw_body` therefore
+classify the same way as existing success responses.
+
+## Reference Vector
+
+`entities-overview-raw-body-numeric-roundtrip` in
+`core/fixtures/native-sol/parity/entities.jsonl` carries both observed score
+literals as raw response bytes. Its expected stdout is the exact
+`json.dumps(body, indent=2, ensure_ascii=False)` output for that overview body.
+
+## Shared-Client Regression
+
+`core/crates/solstone-core-sol-client/tests/json_numeric_roundtrip.rs` starts
+from raw `Vec<u8>` bodies and crosses `decode_response` plus all five
+`json_format` modes.
+
+| Assertion class | Values |
+|---|---|
+| Python-parity | `22.184388937230562`, `9.785157957944147`, `-9.785157957944147` |
+| Renderer policy | `5e-324`, `1.7976931348623157e+308`, `1e+300`, `2.2250738585072014e-308` |
+
+The Python-parity assertions pin byte-exact output against the oracle. The
+exponent and subnormal cases assert accurate finite-`f64` decoding plus existing
+renderer policy; they do not assert preservation of the server's original token
+spelling.