diff --git a/docs/05_architecture_overview.md b/docs/05_architecture_overview.md
index 081e470..2d329d5 100644
--- a/docs/05_architecture_overview.md
+++ b/docs/05_architecture_overview.md
@@ -181,31 +181,64 @@ have no path from a frame name to the actual byte layout for the currently
 loaded LDF — and worse, the test bench would happily ship bytes encoded
 against a *stale* LDF baked into the generator's last run.
 
-Concretely, a single `fio.send` call traverses both layers:
+### Three independent entry points, one wire
 
-```
-test code
-   |
-   |  AlmReqA.send(fio, AmbLightColourRed=0, ...)
-   v
-tests/hardware/_generated/lin_api.py   <-- typed names, compile-time check
-   |
-   |  fio.send("ALM_Req_A", AmbLightColourRed=0, ...)
-   v
-tests/hardware/frame_io.py             <-- per-instance frame cache
-   |
-   |  ldf.frame("ALM_Req_A").pack(AmbLightColourRed=0, ...)
-   v
-ecu_framework/lin/ldf.py               <-- runtime pack/unpack
-   |
-   |  raw_frame.encode_raw({...})
-   v
-ldfparser                              <-- bit-level layout from LDF on disk
+A tester has three legitimate ways to drive the bus, all converging at
+`LinInterface`. They are **parallel paths**, not a single nested stack —
+`FrameIO` deliberately has no static dependency on `ecu_framework/lin/ldf.py`
+(its only `ecu_framework` import is `LinInterface` + `LinFrame` from
+`lin/base.py`), so the `ldf` it receives can be any object with a
+`.frame(name)` method.
+
+```mermaid
+flowchart TB
+  T[test code]
+
+  subgraph Paths[three independent ways to address a frame]
+    GEN["gen_lin_api typed wrapper<br/>AlmReqA.send&#40;fio, ...&#41;<br/>compile-time name check"]
+    FIO["FrameIO stringly-typed<br/>fio.send&#40;'ALM_Req_A', ...&#41;<br/>per-instance frame cache"]
+    LDFDIRECT["LdfDatabase directly<br/>ldf.frame&#40;'ALM_Req_A'&#41;.pack&#40;...&#41;<br/>returns bytes, no I/O"]
+  end
+
+  T --> GEN
+  T --> FIO
+  T --> LDFDIRECT
+
+  GEN -.delegates.-> FIO
+  FIO -.duck-typed lookup.-> LDFOBJ[ldf-like object<br/>currently LdfDatabase]
+  LDFDIRECT --> LDFOBJ
+  LDFOBJ --> LDFPARSER[ldfparser - bit layout]
+
+  FIO --> LIN[LinInterface.send / receive]
+  LDFDIRECT -->|caller invokes lin.send<br/>with the packed bytes| LIN
+  LIN --> WIRE[wire]
 ```
 
-Each layer's responsibility is unique to that layer; removing either
-collapses a distinct kind of check (compile-time name validation, or
-runtime LDF-driven byte layout) that the other layer cannot provide.
+What each path buys you:
+
+- **`gen_lin_api`** — compile-time name validation. Typo a frame or signal
+  name and the IDE / mypy / pytest collection rejects it before any LDF
+  is read. Delegates the actual packing to `fio.send`.
+- **`FrameIO`** — stringly-typed I/O over the wire. Caches frame
+  lookups, supports raw escape hatches (`send_raw` / `receive_raw`) that
+  bypass the LDF object entirely.
+- **`LdfDatabase` directly** — schema-only access. Useful when a test
+  wants to inspect frame layout, pack a buffer without sending, or hand
+  the bytes to a non-FrameIO transport.
+
+The LDF object (currently `LdfDatabase`) is consumed by both `FrameIO`
+and any direct-use code path. `FrameIO`'s use is via injection — it
+never imports `LdfDatabase` and can be tested against a stub.
+
+Removing any of the three entry points collapses a distinct affordance:
+
+- Drop `gen_lin_api` → tests keep stringly-typed `fio.send("ALM_Req_A", …)`
+  and hand-copied state constants, both of which silently drift when the
+  LDF changes.
+- Drop `FrameIO` → every test that wants high-level I/O has to wire
+  `LinInterface` + LDF lookup + pack/unpack itself.
+- Drop direct `LdfDatabase` usage → tests can no longer pack a frame
+  without sending it, or inspect frame metadata without an I/O attempt.
 
 ## Extending the architecture