diff --git a/CLAUDE.md b/CLAUDE.md
index 01885c7..bc67585 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -39,7 +39,7 @@ cargo test --features test-panic --release
 
 `ffi.load("qjson")` uses `dlopen`, which respects `LD_LIBRARY_PATH` — **not** LuaJIT's `package.cpath`. The Makefile sets `LD_LIBRARY_PATH=target/release` for `test`/`bench`; if you invoke `busted` or `luajit` directly, set it yourself.
 
-`make lint` runs clippy only (with `-D warnings`); `cargo fmt --check` is intentionally **not** part of the lint gate because the codebase uses manual column alignment in struct definitions and compact single-line literals that default rustfmt would reflow. See the README "Roadmap / Deferred" entry on fmt for context.
+`make lint` runs clippy only (with `-D warnings`); `cargo fmt --check` is intentionally **not** part of the lint gate because the codebase uses manual column alignment in struct definitions and compact single-line literals that default rustfmt would reflow.
 
 ## Architecture
 
@@ -97,4 +97,4 @@ If you add a scanner code path, run gate 2 locally; the cross-check test (`tests
 
 ## Conventions
 
-- Deferred / "we'll do this later" decisions go in `README.md` under **Roadmap / Deferred**, one bullet per item, so each can be picked up individually. Don't park them in code comments or scratch files.
+- Deferred / "we'll do this later" decisions should be tracked as GitHub Issues so each can be picked up individually. Don't park them in code comments or scratch files.
diff --git a/README.md b/README.md
index 0c410c8..702cdd3 100644
--- a/README.md
+++ b/README.md
@@ -114,21 +114,21 @@ AMD EPYC Rome, Zen 2, 4 vCPUs; 5 rounds, deterministic payload).
 
 ### Encode (unmodified) + modify-then-re-encode
 
-| Size | encode (unmodified) | modify top (cjson / qjson) | modify nested (cjson / qjson) | speedup vs. cjson |
-|---:|---:|---:|---:|---:|
-|   2 KB | 219,925 | 59,761 /  56,909 | 61,685 /  49,798 |  1.0× /  0.8× |
-|  60 KB | 143,843 |  4,590 / **44,370** |  4,616 / **196,386** |  9.7× / 42.5× |
-| 100 KB | 119,617 |  2,645 / **32,712** |  5,263 /  **59,809** | 12.4× / 11.4× |
-|   1 MB |  16,269 |    241 /  **3,108** |    516 /  **14,134** | 12.9× / 27.4× |
+Numbers from the same run as [`docs/benchmarks.md`](docs/benchmarks.md).
+
+| Size | encode (unmodified) | modify top + encode | modify nested + encode |
+|---:|---:|---:|---:|
+|    2 KB | 260,322 | 58,242 |  43,003 |
+|   60 KB | 141,563 | 37,498 | 134,590 |
+|  100 KB | 105,374 | 28,114 |  71,942 |
+|    1 MB |  16,269 |  3,125 |  13,649 |
 
 > **qjson.encode(unmodified)** re-emits the original byte range via `memcpy` —
 > no fields touched means zero serializer work.
 > **qjson modify+encode** materializes only the mutated subtree; unmodified
-> siblings stay on the fast path. cjson always does a full materialize +
-> re-serialize on every encode. At 60 KB+, qjson modify+encode is **10–43×**
-> faster than the cjson equivalent.
-> See [`docs/benchmarks.md`](docs/benchmarks.md) for the full size ladder,
-> memory numbers, and environment.
+> siblings stay on the fast path. See [`docs/benchmarks.md`](docs/benchmarks.md)
+> for the full size ladder, cjson comparisons, speedup ratios, memory
+> numbers, and environment.
 
 ```sh
 make bench       # qjson vs cjson and lua-resty-simdjson