Question on (hidden?) Alignment Layer

[SirRapid]

I am getting different answers from models ran in Odysseus vs running same model (via ollama) in terminal or ollama application (same as terminal obviously). Specifically I was using lamma2-uncensored. I was attracted to Odysseus because I wanted to run local models with no secondary alignment layers, but it seems that there is something built-in to Odysseus that is blocking me.
My question is what in Odysseus could be causing this change in alignment? And how can I edit or remove it?

[BarsatZulkarnine]

The behavior difference is likely caused by a few things Odysseus injects into every request:

1. Security policy prompt - Odysseus always prepends a prompt-injection defence message (UNTRUSTED_CONTEXT_POLICY in src/prompt_security.py). It tells the model to treat external content as data not instructions. This is not an alignment layer but it does influence output.

2. Active preset - if you have a preset selected, its system prompt is prepended. The default "Assistant" preset has its own instructions. Try setting the system prompt to empty in Settings.

3. Memory - if memory is enabled, retrieved memories are injected as system context.

To get closer to bare model behavior: clear/disable the preset system prompt and turn off memory. There is no UI toggle for the security policy, you would need to edit src/prompt_security.py directly to remove it.

[SirRapid]

Thanks for the reply! My guess is its not the memory (I am working off a fresh wipe), but I will test further. Could be the security policy...
I am having trouble finding active preset for the prepended prompt in settings?

[MauledRainbows]

I also had similar results. When ran through Konsole I got a differing response than when I was using Odysseus.
I was testing the boundaries of the AI and it seems it picked up a few extra when ran through Odysseus