todo

# testing
- logit diff topk tests
- agent scores with ADL at different layers

## exp
why fda has an adl so low on fda and 1:1 has a datpoint whiele there is no modek?
- make sure that the exps overwrite instead of adding results to avoid illusion of smaller variance.






# amplification dashboard
- multi prompt add continue regenerate etc 
- results in multi prompt should have the same widgets as in the multi gen tab.
- cut the debugging or at least put in if debug mode is enabled.
- fix the fact that we need to reload the page once to load the multi gen tab cache

## refactoring?
- flatten `method_params` nesting in `kl` and `activation_analysis` configs/code to match `diff_mining`/`adl` style (params at top level, no `method_params` key)
- check for duplicated code between multi gen and multi prompt prompt building functions
- split the dashboard into subclasses (which store a reference to the main dashboard)

# clement
- none should not be a config (clement 1 day later: ???) (clement 1 week later: yeah the fact that is a config is a bit hacky, maybe there is some better way to do this)
- fix _MODEL_CACHE being dumb when using dashboard, we should probably use the streamlit cache instead?
- reduce code duplication in diffing methods (e.g. pca, sae_difference, activation_analysis all have the same code for computing activations in both models from the same input). See docs/DUPLICATION_REPORT.md for more details.
- support full finetunes in ActivationOracleMethod (currently only LoRA adapters supported)