diff --git a/docs/SECURITY_crypto_handshake_RFC.md b/docs/SECURITY_crypto_handshake_RFC.md
new file mode 100644
index 00000000..26ea2c28
--- /dev/null
+++ b/docs/SECURITY_crypto_handshake_RFC.md
@@ -0,0 +1,80 @@
+# RFC: fix the mesh crypto handshake (N3-N5) — for review before implementation
+
+Status: PROPOSAL. Not implemented. Crypto stays hand-written Rust (t27 cannot
+express it — see #62); this design is for the human-written `src/crypto.rs` /
+`src/bin/trios_meshd.rs` and needs sign-off on the approach before coding.
+Anchor: phi^2 + phi^-2 = 3.
+
+## The three defects (from the wave audit, verified in code)
+
+- **N3 — handshake is not authenticated.** `NoiseXX::complete_initiator`
+  (`crypto.rs:132`) and `complete_responder` (`:151`) compute only `ee` and
+  `ss`, then call `combine_dh_shares(&ee, &ss, &ss)` — passing `ss` in BOTH the
+  `es` and `se` slots. The real `es`/`se` Diffie-Hellman operations are never
+  performed, and no transcript hash binds ephemeral<->static keys. The
+  `AllowList` (`:183`) exists but is never consulted inside `complete_*`. The
+  test `noise_xx_resistant_to_mitm` (`:738`) even asserts the MITM party
+  DECRYPTS the victim's traffic — encoding the break as expected behaviour.
+- **N4 — HELLO metric auth is dead + keyed by a constant.** `HELLO_MAC_KEY`
+  (`trios_meshd.rs:13`) is a source-embedded constant; `verify_mac`/`is_fresh`
+  are never called in the daemon RX path, so a neighbour can forge the `heard`
+  list and game ETX.
+- **N5 — static keys are public.** `seed_for(id) = Sha256("trios-mesh/demo/v1/
+  node/" || id)` (`trios_meshd.rs:37`) derives each node's static secret from
+  its public NodeId, so any reader of the open-source daemon recomputes every
+  node's private key.
+
+## Proposed design
+
+### 1. Correct the handshake (Noise-IK-style over the existing primitives)
+Keep X25519 + ChaCha20-Poly1305 + the existing `Session`/`ratchet` transport.
+Change only key agreement + peer authentication:
+
+- In `complete_initiator`, compute THREE distinct shares:
+  `ee = e_i . e_r`, `es = e_i . s_r`, `se = s_i . e_r`; mirror in
+  `complete_responder` (`es = s_i . e_r`, `se = e_i . s_r` by role). Feed the
+  real `(ee, es, se)` into `combine_dh_shares` (its signature already expects
+  three distinct inputs — only the call sites are wrong).
+- Mix a running transcript hash `h = H(h || each public key/DH output)` into the
+  KDF so the session key binds both parties' ephemeral AND static keys.
+- Make `complete_*` take `&AllowList` (or the expected peer NodeId->PublicKey)
+  and REJECT if `peer_static` is not the allow-listed key for the claimed
+  identity, returning `Result<Session, MeshError::Auth>` instead of an
+  always-succeeding `Session`.
+- Rewrite `noise_xx_resistant_to_mitm` to assert the MITM party FAILS to open
+  the victim's frames (the current assertion is inverted).
+
+### 2. Wire HELLO authentication (N4)
+- Derive the per-link HELLO MAC key from the established session key
+  (e.g. `HKDF(session_key, "hello-mac")`), not a global constant. Remove
+  `HELLO_MAC_KEY`.
+- In the daemon RX path, before `observe()` feeds a neighbour's `heard` list
+  into ETX, call `verify_mac` + `is_fresh` and track last-seen seq for replay;
+  drop on failure. Use a constant-time comparison (`subtle::ConstantTimeEq`).
+
+### 3. Provision real static keys (N5)
+- Load each node's static secret from an out-of-band keystore file (env-pointed,
+  not world-readable), never `from_seed(NodeId)`. Keep `from_seed` only behind a
+  `--demo` flag for local loopback.
+- Populate the `AllowList` (NodeId -> PublicKey) from the same provisioning, and
+  gate session establishment on it (step 1).
+
+## Test plan (must all pass)
+- MITM: an attacker substituting its own static key cannot open either party's
+  frames (rewritten `noise_xx_resistant_to_mitm`).
+- Allow-list: a peer whose static key is not allow-listed gets `MeshError::Auth`.
+- HELLO forge: a neighbour forging another node's `heard` list is rejected
+  (bad MAC) and does not move ETX.
+- Replay: a re-sent HELLO is rejected by `is_fresh`/seq tracking.
+- Regression: the existing 103 tests stay green; add the four above.
+
+## Formal check (optional, Lane B)
+Model the revised handshake in Tamarin (AutoTam, arXiv:2606.19937) to prove
+mutual authentication + forward secrecy, rather than asserting it.
+
+## Decision needed from the owner
+1. Noise-IK vs a full Noise-XX (three-message) — IK is one-round-trip and fits
+   a known-peer mesh; confirm.
+2. Keystore format / provisioning mechanism for static keys (file? per-node
+   config? an existing provisioning path?).
+3. Whether to gate this behind a feature flag during rollout.