WHY IT'S BROKEN

Two monitors that miss it

Behavioral testing

you'd have to guess the trigger

The model is correct on virtually every input you try
To catch it you'd need the exact trigger in advance — you won't have it

Joint cross-model features

the popular interpretability fix

Crosscoders learn features over base + fine-tuned together
The backdoor competes with everything the model represents — and gets buried

So where IS the signal? In what the fine-tuning actually changed.