diff --git a/docs/analysis/bench_baseline.md b/docs/analysis/bench_baseline.md
new file mode 100644
index 000000000..c1b3b7704
--- /dev/null
+++ b/docs/analysis/bench_baseline.md
@@ -0,0 +1,88 @@
+# Message-path benchmark — Phase 1 baseline
+
+Baseline capture on the **current branch before Phases 2/3**, produced by
+`apps/benchmarks/message_path_bench.nim` (see
+[plan_phase1_bench.md](plan_phase1_bench.md)). Re-run after each phase and diff
+against these numbers.
+
+## Environment
+
+| Item | Value |
+|---|---|
+| Commit (harness) | `cf8502eb4975ae6bc8a5667a5d179cd154e8e1d1` |
+| Branch | `experimental/chore-nocopy-e2e-perf-harness` |
+| Machine | Apple M4 |
+| OS | macOS 26.5 (arm64) |
+| Nim | 2.2.4, `--mm:refc` |
+| Build | `-d:msgPathCounters -d:chronicles_log_level=ERROR`, `--passL:librln_v2.0.2.a` |
+| Run | `nimble benchMessagePath` (or `./build/message_path_bench`) |
+| Workload | seed 42, payload mix 10/50/150 kB @ 25/50/25 %, N=1000 + 100 warmup |
+
+## Results (CSV)
+
+```
+scenario,msgs,payload_profile,wall_ms,msg_per_s,ns_per_msg_p50,ns_per_msg_p99,decodes_per_msg,hashes_per_msg,hashed_MB,decoded_MB,occupied_mem_delta_MB,gc_collections
+micro_run1,1000,10/50/150kB@25/50/25,828.591,1206.9,641375,1925042,2.000,2.000,130.00,130.11,1.54,7
+micro_run2,1000,10/50/150kB@25/50/25,795.166,1257.6,619167,1810625,2.000,2.000,130.00,130.11,1.51,8
+macro,1000,10/50/150kB@25/50/25,4367.314,229.0,3510750,10701375,6.000,5.000,325.00,390.33,141.58,10
+```
+
+| Scenario | msg/s | decodes/msg | hashes/msg | decoded_MB | hashed_MB | occ_mem_delta_MB |
+|---|---|---|---|---|---|---|
+| micro run1 | 1206.9 | 2.000 | 2.000 | 130.11 | 130.00 | 1.54 |
+| micro run2 | 1257.6 | 2.000 | 2.000 | 130.11 | 130.00 | 1.51 |
+| macro | 229.0 | 6.000 | 5.000 | 390.33 | 325.00 | 141.58 |
+
+Micro determinism: run1=1206.9 msg/s, run2=1257.6 msg/s → **variance 4.03 %**
+(< 5 % target; two earlier back-to-back runs were 1.90 % and 3.12 %). The
+decode/hash *counts* are byte-exact reproducible across every run.
+
+## Interpretation vs. the analysis
+
+**Decode / hash counts validate [async_copy_analysis.md](async_copy_analysis.md).**
+
+The counter is a single process-global (both in-process nodes share it), so the
+macro per-message numbers are the aggregate of publisher (node A) + receiver
+(node B). Breakdown of the macro **6 decodes / 5 hashes** per message:
+
+| Counter | Node A (publish) | Node B (receive) | Total |
+|---|---|---|---|
+| `WakuMessage.decode` | 1 (`onSend` observer) | 4 (ordered validator, `onRecv`, `onValidated`, subscribe topicHandler) | **6** |
+| `computeMessageHash` | 2 (`publish` :686, `onSend`→`logMessageInfo`) | 3 (`onValidated`→`logMessageInfo`, archive, store-sync) | **5** |
+
+- The **inbound-decode claim (≥4 decodes on a relayed message)** is confirmed:
+  node B alone decodes the same proto bytes **4×**
+  (`waku_relay/protocol.nim` validator :543, `onRecv` :271, `onValidated` :326,
+  topicHandler :603).
+- The micro scenario isolates the non-observer path and shows the **2 decodes /
+  2 hashes** it exercises (validator + topicHandler decode; archive + store-sync
+  hash) — the observers only fire under real libp2p dispatch, which the macro
+  scenario adds.
+
+### Caveats
+
+1. **Log level elides some hash sites.** Built at `chronicles_log_level=ERROR`,
+   `computeMessageHash` calls that appear *only* as disabled log arguments are
+   not evaluated (chronicles skips disabled-sink args) — notably the
+   `node.publish` `notice` at `waku_node/relay.nim:148`. Sites bound to a `let`
+   (archive :101, store-sync :78, `logMessageInfo` :217, `publish` :686) run
+   regardless of level and are the ones counted here.
+2. **Filter not mounted.** Per the plan, node B has archive + store-sync but not
+   filter. The analysis's "4–6 hashes" upper end assumes an active filter
+   (`waku_filter_v2` adds ~2 more receiver-side `computeMessageHash` passes,
+   :195 + :246). With filter mounted the receiver-side hash count would rise
+   from 3 toward the analysis's upper bound.
+3. **`occupied_mem_delta`** is a coarse `getOccupiedMem()` before/after delta on
+   the refc heap, not a per-message allocation trace; use it for order-of-
+   magnitude trend, not exact bytes. `decoded_MB` / `hashed_MB` (exact byte
+   totals fed to decode/hash) are the reliable volume metrics.
+
+## Acceptance gate (Phase 1)
+
+| # | Criterion | Verdict |
+|---|---|---|
+| 1 | Representative tests green (production untouched except no-op templates) | **PASS** — `tests/waku_core/test_message_digest` 5/5, `tests/waku_relay/test_protocol` 20/20 |
+| 2 | Production build without `-d:msgPathCounters` compiles (templates vanish) | **PASS** — `nim check apps/wakunode2` clean; relay test built without the define |
+| 3 | Two consecutive micro runs < 5 % msg/s variance | **PASS** — 4.03 % (and 1.90 % / 3.12 % on earlier runs) |
+| 4 | Macro `decodes_per_msg ≥ 4` and `hashes_per_msg` in 4–6 | **PASS** — decodes 6.0, hashes 5.0 (process aggregate; receiver-side decodes = 4) |
+| 5 | This baseline doc committed | **PASS** |