Document experimental status and benchmark recommendation

Author: regenrek

README.md

@@ -1,5 +1,18 @@
 # Switchloom
 
+> **Experimental. I currently recommend against adopting this routing workflow.**
+>
+> I built Switchloom to see whether mixing models could get me the same quality
+> for less money. My benchmarks did not deliver that result. In the second
+> Pokédex run, the team with Jev cost almost as much as Astra alone and hit the
+> time limit before final acceptance. Astra finished and passed all 14 independent
+> functional checks.
+>
+> I am keeping the code and results public, but I would not add this layer to my
+> everyday workflow based on these results. This is evidence from two pilots on
+> one application, not proof that all model routing fails.
+> [Read the benchmark and its limitations](https://switchloom.ai/benchmarks/pokedex-round-2).
+
 **One workflow prompt for persistent Codex tasks, with optional TypeSafe/Jev routing.**
 
 Assign capabilities to models, then copy the prompt into Codex Desktop with
@@ -28,9 +41,13 @@ Jev-versus-vanilla measurements remain open in [TASKS.md](TASKS.md).
 
 ## Setup
 
-Both modes use the same handoff. Vanilla sets an explicit capability and does
-not call TypeSafe, so it needs no API key. Jev mode, enabled by default, needs
-the CLI, the skill and a local key:
+The copied desktop workflow currently requires the Switchloom CLI and skill in
+both modes. The CLI prepares routing decisions and handoff arguments. Codex
+creates the tasks, sends the messages and runs the work. The public Playground
+does not require a local CLI installation.
+
+Vanilla sets an explicit capability and does not call TypeSafe, so it needs no
+API key. Jev mode, enabled by default, additionally needs a local TypeSafe key:
 
 ```sh
 cargo install --path . --locked